Exotic Option Pricing and Advanced L´evy Models
Exotic Option Pricing and Advanced L´evy Models
Edited by
Andreas E...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Exotic Option Pricing and Advanced L´evy Models

Exotic Option Pricing and Advanced L´evy Models

Edited by

Andreas E. Kyprianou, Wim Schoutens and Paul Wilmott

Copyright 2005

John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England Telephone (+44) 1243 779777

Email (for orders and customer service enquiries): [email protected] Visit our Home Page on www.wiley.com All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1T 4LP, UK, without the permission in writing of the Publisher. Requests to the Publisher should be addressed to the Permissions Department, John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England, or emailed to [email protected], or faxed to (+44) 1243 770620. This publication is designed to provide accurate and authoritative information in regard to the subject matter covered. It is sold on the understanding that the Publisher is not engaged in rendering professional services. If professional advice or other expert assistance is required, the services of a competent professional should be sought.

Other Wiley Editorial Ofﬁces John Wiley & Sons Inc., 111 River Street, Hoboken, NJ 07030, USA Jossey-Bass, 989 Market Street, San Francisco, CA 94103-1741, USA Wiley-VCH Verlag GmbH, Boschstr. 12, D-69469 Weinheim, Germany John Wiley & Sons Australia Ltd, 42 McDougall Street, Milton, Queensland 4064, Australia John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809 John Wiley & Sons Canada Ltd, 22 Worcester Road, Etobicoke, Ontario, Canada M9W 1L1 Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic books.

British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN-13 978-0-470-01684-8 ISBN-10 0-470-01684-1 Typeset in 10/12pt Times by Laserwords Private Limited, Chennai, India Printed and bound in Great Britain by Antony Rowe Ltd, Chippenham, Wiltshire This book is printed on acid-free paper responsibly manufactured from sustainable forestry in which at least two trees are planted for each one used for paper production.

Contents

Contributors

xi

Preface

xiii

About the Editors

xvii

About the Contributors

xix

1 L´evy Processes in Finance Distinguished by their Coarse and Fine Path Properties Andreas E. Kyprianou and R. Loeffen 1.1 1.2 1.3

Introduction L´evy processes Examples of L´evy processes in ﬁnance 1.3.1 Compound Poisson processes and jump-diffusions 1.3.2 Spectrally one-sided processes 1.3.3 Meixner processes 1.3.4 Generalized tempered stable processes and subclasses 1.3.5 Generalized hyperbolic processes and subclasses 1.4 Path properties 1.4.1 Path variation 1.4.2 Hitting points 1.4.3 Creeping 1.4.4 Regularity of the half line 1.5 Examples revisited 1.5.1 Compound Poisson processes and jump-diffusions 1.5.2 Spectrally negative processes 1.5.3 Meixner process 1.5.4 Generalized tempered stable process 1.5.5 Generalized hyperbolic process 1.6 Conclusions References

1

1 2 4 5 6 6 7 9 10 10 12 14 16 17 17 17 17 19 23 24 26

vi

Contents

2 Simulation Methods with L´evy Processes Nick Webber 2.1 2.2

Introduction Modelling price and rate movements 2.2.1 Modelling with L´evy processes 2.2.2 Lattice methods 2.2.3 Simulation methods 2.3 A basis for a numerical approach 2.3.1 The subordinator approach to simulation 2.3.2 Applying the subordinator approach 2.4 Constructing bridges for L´evy processes 2.4.1 Stratiﬁed sampling and bridge methods 2.4.2 Bridge sampling and the subordinator representation 2.5 Valuing discretely reset path-dependent options 2.6 Valuing continuously reset path-dependent options 2.6.1 Options on extreme values and simulation bias 2.6.2 Bias correction for L´evy processes 2.6.3 Variation: exceedence probabilities 2.6.4 Application of the bias correction algorithm 2.7 Conclusions References 3 Risks in Returns: A Pure Jump Perspective H´elyette Geman and Dilip B. Madan 3.1 3.2 3.3

Introduction CGMY model details Estimation details 3.3.1 Statistical estimation 3.3.2 Risk neutral estimation 3.3.3 Gap risk expectation and price 3.4 Estimation results 3.4.1 Statistical estimation results 3.4.2 Risk neutral estimation results 3.4.3 Results on gap risk expectation and price 3.5 Conclusions References 4 Model Risk for Exotic and Moment Derivatives Wim Schoutens, Erwin Simons and Jurgen Tistaert 4.1 4.2

Introduction The models 4.2.1 The Heston stochastic volatility model 4.2.2 The Heston stochastic volatility model with jumps

29

29 30 30 31 32 33 34 35 36 36 37 39 40 42 43 44 45 48 48 51

51 54 57 58 59 60 60 61 61 61 63 65 67

67 68 69 69

Contents

4.2.3 The Barndorff-Nielsen–Shephard model 4.2.4 L´evy models with stochastic time 4.3 Calibration 4.4 Simulation 4.4.1 NIG L´evy process 4.4.2 VG L´evy process 4.4.3 CIR stochastic clock 4.4.4 Gamma-OU stochastic clock 4.4.5 Path generation for time-changed L´evy process 4.5 Pricing of exotic options 4.5.1 Exotic options 4.5.2 Exotic option prices 4.6 Pricing of moment derivatives 4.6.1 Moment swaps 4.6.2 Moment options 4.6.3 Hedging moment swaps 4.6.4 Pricing of moments swaps 4.6.5 Pricing of moments options 4.7 Conclusions References 5 Symmetries and Pricing of Exotic Options in L´evy Models Ernst Eberlein and Antonis Papapantoleon 5.1 5.2 5.3 5.4

Introduction Model and assumptions General description of the method Vanilla options 5.4.1 Symmetry 5.4.2 Valuation of European options 5.4.3 Valuation of American options 5.5 Exotic options 5.5.1 Symmetry 5.5.2 Valuation of barrier and lookback options 5.5.3 Valuation of Asian and basket options 5.6 Margrabe-type options References 6 Static Hedging of Asian Options under Stochastic Volatility Models using Fast Fourier Transform Hansj¨org Albrecher and Wim Schoutens 6.1 6.2

Introduction Stochastic volatility models 6.2.1 The Heston stochastic volatility model 6.2.2 The Barndorff-Nielsen–Shephard model 6.2.3 L´evy models with stochastic time

vii

70 71 74 78 78 79 79 79 79 80 80 82 86 89 89 90 91 93 93 95 99

99 100 105 106 106 111 113 114 114 115 117 119 124

129

129 131 131 132 133

viii

Contents

6.3 6.4

Static hedging of Asian options Numerical implementation 6.4.1 Characteristic function inversion using FFT 6.4.2 Static hedging algorithm 6.5 Numerical illustration 6.5.1 Calibration of the model parameters 6.5.2 Performance of the hedging strategy 6.6 A model-independent static super-hedge 6.7 Conclusions References 7 Impact of Market Crises on Real Options Pauline Barrieu and Nadine Bellamy 7.1 7.2

136 138 138 140 140 140 140 145 145 145 149

Introduction The model 7.2.1 Notation 7.2.2 Consequence of the modelling choice 7.3 The real option characteristics 7.4 Optimal discount rate and average waiting time 7.4.1 Optimal discount rate 7.4.2 Average waiting time 7.5 Robustness of the investment decision characteristics 7.5.1 Robustness of the optimal time to invest 7.5.2 Random jump size 7.6 Continuous model versus discontinuous model 7.6.1 Error in the optimal proﬁt–cost ratio 7.6.2 Error in the investment opportunity value 7.7 Conclusions Appendix References

149 151 151 153 155 156 156 157 158 159 160 161 161 163 165 165 167

8 Moment Derivatives and L´evy-type Market Completion Jos´e Manuel Corcuera, David Nualart and Wim Schoutens

169

8.1 8.2

8.3

Introduction Market completion in the discrete-time setting 8.2.1 One-step trinomial market 8.2.2 One-step ﬁnite markets 8.2.3 Multi-step ﬁnite markets 8.2.4 Multi-step markets with general returns 8.2.5 Power-return assets The L´evy market 8.3.1 L´evy processes 8.3.2 The geometric L´evy model 8.3.3 Power-jump processes

169 170 170 172 173 174 174 177 177 178 178

Contents

ix

Enlarging the L´evy market model 8.4.1 Martingale representation property 8.5 Arbitrage 8.5.1 Equivalent martingale measures 8.5.2 Example: a Brownian motion plus a ﬁnite number of Poisson processes 8.6 Optimal portfolios 8.6.1 Optimal wealth 8.6.2 Examples References

179 180 183 183

8.4

9 Pricing Perpetual American Options Driven by Spectrally One-sided L´evy Processes Terence Chan 9.1 9.2

Introduction First-passage distributions and other results for spectrally positive L´evy processes 9.3 Description of the model, basic deﬁnitions and notations 9.4 A renewal equation approach to pricing 9.5 Explicit pricing formulae for American puts 9.6 Some speciﬁc examples Appendix: use of fast Fourier transform References Epilogue Further references

10 On Asian Options of American Type Goran Peskir and Nadia Uys 10.1 Introduction 10.2 Formulation of the problem 10.3 The result and proof 10.4 Remarks on numerics Appendix References 11 Why be Backward? Forward Equations for American Options Peter Carr and Ali Hirsa 11.1 11.2 11.3 11.4 11.5 11.6

Introduction Review of the backward free boundary problem Stationarity and domain extension in the maturity direction Additivity and domain extension in the strike direction The forward free boundary problem Summary and future research

185 186 187 188 192

195

195 198 202 204 207 209 213 214 215 215 217

217 218 220 231 233 234 237

237 239 242 245 247 250

x

Contents

Appendix: Discretization of forward equation for American options References 12 Numerical Valuation of American Options Under the CGMY Process Ariel Almendral

251 257 259

12.1 Introduction 12.2 The CGMY process as a L´evy process 12.2.1 Options in a L´evy market 12.3 Numerical valuation of the American CGMY price 12.3.1 Discretization and solution algorithm 12.4 Numerical experiments Appendix: Analytic formula for European option prices References

259 260 261 263 263 270 271 275

13 Convertible Bonds: Financial Derivatives of Game Type Jan Kallsen and Christoph K¨uhn

277

13.1 Introduction 13.2 No-arbitrage pricing for game contingent claims 13.2.1 Static no-arbitrage prices 13.2.2 No-arbitrage price processes 13.3 Convertible bonds 13.4 Conclusions References

277 279 279 282 286 289 289

14 The Spread Option Optimal Stopping Game Pavel V. Gapeev

293

14.1 Introduction 14.2 Formulation of the problem 14.3 Solution of the free-boundary problem 14.4 Main result and proof 14.5 Conclusions References

293 294 296 299 302 304

Index

307

Contributors Hansj¨org Albrecher Department of Mathematics, Graz University of Technology, Steyrergasse 30, A-8010 Graz, Austria Ariel Almendral Norwegian Computing Center, Gaustadalleen 23, Postbox 114, Blindern, N-0314 Oslo, Norway Pauline Barrieu Statistics Department, London School of Economics, Houghton Street, London, WC2A 2AE, UK Nadine Bellamy Equipe d’Analyse et Probabilit´es, Universit´e d’Evry Val d’Essonne, Rue du P`ere Jarlan, 91025 Evry Cedex, France Peter Carr Bloomberg LP, 731 Lexington Avenue, New York, NY 10022, USA Terence Chan School of Mathematical and Computer Sciences, Heriot-Watt University, Edinburgh, EH14 4AS, Scotland, UK Jos´e Manuel Corcuera Facultat de Matematiques, Universitat de Barcelona, Gran Via de les Corts Catalanes 585, E-08007 Barcelona, Spain Ernst Eberlein Department of Mathematical Stochastics, University of Freiburg, Eckerstraße 1, D-79104, Freiburg, Germany Pavel V. Gapeev Institute of Control Sciences, Russian Academy of Sciences, Profsoyuznaya Str. 65, 117997 Moscow, Russia H´elyette Geman University of Paris Dauphine, Paris, France and ESSEC-Finance Department, 95021 Cergy-Pontoise, France

xii

Contributors

Ali Hirsa Caspian Capital Management, LLC, 745 Fifth Avenue, 28th Floor, New York, NY 10151, USA Jan Kallsen HVB-Institute for Mathematical Finance, Munich University of Technology, D-85747 Garching, Germany Christoph Kuehn Johann Wolfgang Goethe-Universit¨at, Fachbereich Mathematik (Fach 187), D-60054 Frankfurt am Main, Germany Andreas E. Kyprianou School of Mathematical and Computer Sciences, Heriot Watt University, Edinburgh, EH14 4AS, Scotland, UK R. Loeffen Department of Mathematics, University of Utrecht, PO Box 80.010, 3508 TA Utrecht, The Netherlands Dilip B. Madan Department of Finance, Robert H. School of Business, Van Munching Hall, University of Maryland, College Park, MD 20742, USA David Nualart Facultat de Matematiques, Universitat de Barcelona, Gran Via de les Corts Catalanes 585, E-08007 Barcelona, Spain Antonis Papapantoleon Department of Mathematical Stochastics, University of Freiburg, Eckerstraße 1, D-79104, Freiburg, Germany Goran Peskir Department of Mathematical Sciences, University of Aarhus, Ny Munkegade, DK-8000, Aarhus, Denmark Wim Schoutens Katholieke Universiteit Leuven - U.C.S, W. de Croylaan 54, B-3001 Leuven, Belgium Irwin Simons ING SWE, Financial Modelling, Marnixlaan 24, B-1000 Brussels, Belgium Jurgen Tistaert ING SWE, Financial Modelling, Marnixlaan 24, B-1000 Brussels, Belgium Nadia Uys Programme for Advanced Mathematics of Finance, School of Computational and Applied Mathematics, University of the Witwatersrand, Private Bag 3, Witwatersrand 2050, South Africa Nick Webber Warwick Business School, University of Warwick, Coventry CV4 7AL, UK

Preface Since around the turn of the millennium there has been a general acceptance that one of the more practical improvements one may make in the light of the shortfalls of the classical Black–Scholes model is to replace the underlying source of randomness, a Brownian motion, by a L´evy process. Working with L´evy processes allows one to capture distributional characteristics in the stock returns such as semi-heavy tails and asymmetry, as well as allowing for jumps in the price process with the interpretation as market shocks and effects due to trading taking place in business time rather than real time. In addition, L´evy processes in general, as well as having the same properties as Brownian motion in the form of stationary independent increments, have many well understood probabilistic and analytical properties which make them attractive as mathematical tools. At the same time, exotic derivatives are gaining increasing importance as ﬁnancial instruments and are traded nowadays in large quantities in over the counter markets. The consequence of working with markets driven by L´evy processes forces a number of new mathematical challenges with respect to exotic derivatives. Many exotic options are based on the evolving historical path of the underlying. In terms of pricing and hedging, this requires an understanding of ﬂuctuation theory, stochastic calculus and distributional decompositions associated with L´evy processes. This current volume is a compendium of articles, each of which consists of a discursive review and recent research on the topic of Exotic Option Pricing and Advanced L´evy Models written by leading scientists in this ﬁeld. This text is organized as follows. The ﬁrst two chapters can be seen as an introduction to L´evy processes and their applications. The ﬁrst chapter, by A. E. Kyprianou and R. Loeffen, gives a brief introduction to L´evy processes, providing several examples which are commonly used in ﬁnance, as well as examining in more detail some of their ﬁne and coarse path properties. To apply L´evy processes in practice one needs good numerics. In Chapter 2, N. Webber discusses recent progress in the development of simulation methods suitable for most of the widely used L´evy processes. Speed-up methods, bridge algorithms and stratiﬁed sampling are some of the many ingredients. These techniques are applied in the context of the valuation of different kinds of exotic options. In the second part, one can see L´evy-driven equity models at work. In Chapter 3, H. Geman and D. Madan use pure jump models, in particular from the CGMY class, for the evolution of stock prices and investigate in this setting the relationship between the statistical and risk-neutral densities. Statistical estimation is conducted on different world indexes. Their conclusions depart from the standard applications of utility theory to asset pricing which assume a representative agent who is long the market. They argue that one

xiv

Preface

must have at a minimum a two-agent model in which some weight is given to an agent who is short the market. In Chapter 4, W. Schoutens, E. Simons and J. Tistaert calibrate different L´evy-based stochastic volatility models to a real market option surface and price by Monte Carlo techniques a range of exotics options. Although the different models discussed can all be nicely calibrated to the option surface – leading to almost identical vanilla prices – exotic option prices under the different models discussed can differ considerably. This investigation is pushed further by looking at the prices of moment derivatives, a new kind of derivative paying out realized higher moments. Even more pronounced differences are reported in this case. The study reveals that there is a clear issue of model risk and warns of blind use of fancy models in the realm of exotic options. The third part is devoted to pricing, hedging and general theory of different exotics options of a European nature. In Chapter 5, E. Eberlein and A. Papapantoleon consider time-inhomogeneous L´evy processes (or additive processes) to give a better explanation of the so-called ‘volatility smile’, as well as the ‘term structure of smiles’. They derive different kinds of symmetry relations for various exotic options. Their contribution also contains an extensive review of current literature on exotics driven in L´evy markets. In Chapter 6, H. Albrecher and W. Schoutens present a simple static super-hedging strategy for the Asian option, based on stop-loss transforms and comonotonic theory. A numerical implementation is given in detail and the hedging performance is illustrated for several stochastic volatility models. Real options form the main theme of Chapter 7, authored by P. Barrieu and N. Bellamy. There, the impact of market crises on investment decisions is analysed through real options under a jump-diffusion model, where the jumps characterize the crisis effects. In Chapter 8, J.M. Corcuera, D. Nualart and W. Schoutens show how moment derivatives can complete L´evy-type markets in the sense that, by allowing trade in these derivatives, any contingent claim can be perfectly hedged by a dynamic portfolio in terms of bonds, stocks and moment-derivative related products. In the fourth part, exotics of an American nature are considered. Optimal stopping problems are central here. Chapter 9 is a contribution at the special request of the editors. This consists of T. Chan’s original unpublished manuscript dating back to early 2000, in which many important features of the perpetual American put pricing problem are observed for the case of a L´evy-driven stock which has no positive jumps. G. Peskir and N. Uys work in Chapter 10 under the traditional Black–Scholes market but consider a new type of Asian option where the holder may exercise at any time up to the expiry of the option. Using recent techniques developed by Peskir concerning local time–space calculus, they are able to give an integral equation characterizing uniquely the optimal exercise boundary. Solving this integral equation numerically brings forward stability issues connected with the Hartman– Watson distribution. In Chapter 11, P. Carr and A. Hirsa give forward equations for the value of an American put in a L´evy market. A numerical scheme for the VG case for very fast pricing of an American put is given in its Appendix. In the same spirit, A. Almendral discusses the numerical valuation of American options under the CGMY model. A numerical solution scheme for the Partial-Integro-Differential Equation is provided; computations are accelerated by the Fast-Fourier Transform. Pricing American options and their early exercise boundaries can be carried out within seconds. The ﬁnal part considers game options. In Chapter 13, C. K¨uhn and J. Kallsen give a review of the very recent literature concerning game-type options, that is, options in which both holder and writer have the right to exercise. Game-type options are very closely related to convertible bonds and K¨uhn and Kallsen also bring this point forward in their contribution.

Preface

xv

Last, but by far not least, P. Gapeev gives a concrete example of a new game-type option within the Black–Scholes market for which an explicit representation can be obtained. We should like to thank all contributors for working hard to keep to the tempo that has allowed us to compile this text within a reasonable period of time. We would also like to heartily thank the referees, all of whom responded gracefully to the ﬁrm request to produce their reports within a shorter than normal period of time and without compromising their integrity. This book grew out of the 2004 Workshop, Exotic Option Pricing under Advanced L´evy Models, hosted at EURANDOM in The Netherlands. In addition to the excellent managerial and organizational support offered by EURANDOM, it was generously supported by grants from Nederlands Organizatie voor Wetenschappelijk Onderzoek (The Dutch Organization for Scientiﬁc Research), Koninklijke Nederlandse Akademie van Wetenschappen (The Royal Dutch Academy of Science) and The Journal of Applied Econometrics. Special thanks goes to Jef Teugels and Lucienne Coolen. Thanks also to wilmott.com and mathfinance.de for publicizing the event. A. E. Kyprianou, Edinburgh, UK W. Schoutens, Leuven, Belgium P. Wilmott, London, UK

About the Editors Andreas E. Kyprianou Address: School of Mathematical and Computer Sciences, Heriot Watt University, Edinburgh, EH14 4AS, Scotland, UK E-mail: [email protected] Afﬁliation: Heriot Watt University, Scotland, UK Andreas Kyprianou has a degree in mathematics from Oxford University and a PhD in probability theory from Shefﬁeld University. He has held academic positions in the Mathematics and/or Statistics Departments at The London School of Economics, Edinburgh University, Utrecht University and, currently, Heriot Watt University. He has also worked for nearly two years as a research mathematician with Shell International Exploration and Production. His research interests are focused on pure and applied probability with recent focus on L´evy processes. He has taught a range of courses on probability theory, stochastic analysis, ﬁnancial stochastics and L´evy processes on the Amsterdam–Utrecht Masters programme in Stochastics and Financial Mathematics and the MSc programme in Financial Mathematics at Edinburgh University. Wim Schoutens Address: Katholieke Universiteit Leuven – UCS, W. De Croylaan 54, B-3001 Leuven, Belgium E-mail: [email protected] Afﬁliation: Katholieke Universiteit Leuven, Belgium Wim Schoutens has a degree in Computer Science and a PhD in Science (Mathematics). He is a research professor at the Department of Mathematics at the Catholic University of Leuven (Katholieke Universiteit Leuven), Belgium. He has been a consultant to the banking industry and is the author of the Wiley book L´evy Processes in Finance – Pricing Financial Derivatives. His research interests are focused on ﬁnancial mathematics and stochastic processes. He currently teaches several courses related to ﬁnancial engineering in different Master programmes.

xviii

About the Editors

Paul Wilmott Address: ‘Wherever I lay my hat’ E-mail: [email protected] Afﬁliation: Various Paul Wilmott has undergraduate and DPhil degrees in mathematics. He has written over 100 articles on mathematical modeling and ﬁnance, as well as internationally acclaimed books including Paul Wilmott on Quantitative Finance, published by Wiley. Paul has extensive consulting experience in quantitative ﬁnance with leading US and European ﬁnancial institutions. He has founded a university degree course and the popular Certiﬁcate in Quantitative Finance. Paul also manages wilmott.com.

About the Contributors Hansjoerg Albrecher is Associate Professor of Applied Mathematics at the Graz University of Technology. He studied Mathematics in Graz, Limerick and Baltimore, receiving his doctorate in 2001. He held visiting appointments at the Katholieke Universiteit Leuven and the University of Aarhus. Research interests include ruin theory, stochastic simulation and quantitative ﬁnance. Ariel Almendral will take up a research position at the Norwegian Computing Center, starting in August 2005. In 2004, he obtained his PhD from the University of Oslo, Norway. In his thesis he focused on numerical methods for ﬁnancial derivatives in the presence of jump processes, from a differential equation perspective. Parts of his PhD research were carried out at Delft University of Technology, The Netherlands, where he held a postdoctoral position for a year. Pauline Barrieu has been a lecturer in the Department of Statistics at the London School of Economics since 2002, after obtaining a PhD in ﬁnance (doctorate HEC, France) and a PhD in Mathematics (University of Paris 6, France). Her research interests are mainly problems at the interface of insurance and ﬁnance, in particular, optimal design of new types of derivatives and securitization. She also works on quantitative methods for assessing ﬁnancial and non-ﬁnancial risks, on stochastic optimization and environmental economics. Nadine Bellamy is Associate Professor in Mathematics at the University of Evry, France. Her PhD thesis (University of Evry, 1999) deals with hedging and pricing in markets driven by discontinuous processes and her current research interests are related to optimization and real options problems. Dr Peter Carr heads Quantitative Research at Bloomberg LP. He also directs the Masters in Mathematical Finance program at NYU’s Courant Institute. Formerly, Dr Carr was a ﬁnance professor for eight years at Cornell University. Since receiving his PhD in Finance from UCLA in 1989, he has published extensively in both academic and industry-oriented journals. He has recently won awards from Wilmott Magazine for Cutting Edge Research and from Risk Magazine for Quant of the Year. Terence Chan completed his PhD at Cambridge University UK after which he obtained his current position at Heriot-Watt University, Edinburgh. Among his research interests are L´evy processes but he only occasionally dabbles in ﬁnancial mathematics to maintain the illusion that he is doing something of practical use!

xx

About the Contributors

Jos´e Manuel Corcuera is an associate professor since 1997 at the Faculty of Mathematics of the University of Barcelona. His main research interest is in the theoretical aspects of statistics and quantitative ﬁnance. Ernst Eberlein is Professor of Stochastics and Mathematical Finance at the University of Freiburg. He is a co-founder of the Freiburg Center for Data Analysis and Modeling (FDM), an elected member of the International Statistical Institute and at present Executive Secretary of the Bachelier Finance Society. His current research focuses on statistical analysis and realistic modeling of ﬁnancial markets, risk management, as well as pricing of derivatives. Pavel Gapeev was born in Moscow in 1976. He studied and obtained his PhD in Stochastics at Moscow State University in 2001. He is now working as a Senior Researcher at the Institute of Control Sciences, Russian Academy of Sciences in Moscow. He has held a visiting appointment at Humboldt University, Berlin (2001/2002) in addition to some short term research visits to Aarhus, Bochum, Copenhagen, Frankfurt, Helsinki, and Zurich. His main ﬁeld of research is stochastic analysis and its applications into ﬁnancial mathematics, optimal control, optimal stopping, and quickest detection. Apart from mathematics he is interested in arts, sports and travelling, and enjoys playing the violin. H´elyette Geman is a Professor of Finance at the University Paris Dauphine and ESSEC Graduate Business School. She is a graduate of Ecole Normale Superieure in Mathematics, holds a Masters degree in theoretical physics and a PhD in mathematics from the University Pierre et Marie Curie and a PhD in Finance from the University Pantheon Sorbonne. Professor Geman has published more than 60 papers in major ﬁnance journals including the Journal of Finance, Mathematical Finance, Journal of Financial Economics, Journal of Banking and Finance and Journal of Business. Professor Geman’s research includes asset price modelling using jump-diffusions and L´evy processes, commodity forward curve modelling and exotic option pricing for which she won the ﬁrst prize of the Merrill Lynch Awards. She has written a book entitled Commodities and Commodity Derivatives (John Wiley & Sons Ltd, 2005). Ali Hirsa joined Caspian Capital Management as the Head of Analytical Trading Strategy in April 2004. At CCM his responsibilities include design and testing of new trading strategies. Prior to his current position, Ali worked at Morgan Stanley for four years. Ali is also an adjunct professor at Columbia University and New York University where he teaches in the mathematics of ﬁnance program. Ali received his PhD in applied mathematics from University of Maryland at College Park under the supervision of Dilip B. Madan. Jan Kallsen is a Professor of Mathematical Finance at Munich University of Technology. His research interests include pricing and hedging in incomplete markets and the general theory of stochastic processes. Christoph Kuhn is Junior Professor at the Frankfurt MathFinance Institute. He holds a ¨ diploma in Mathematical Economics from the University of Marburg and a PhD in mathematics from Munich University of Technology. His main research interests are pricing and hedging of derivatives in incomplete markets and the microstruture of ﬁnancial markets. Ronnie Loeffen was born in 1981 in the Netherlands and has recently received a Master’s degree in Mathematics at the University Utrecht. The subject of his Master’s thesis was American options on a jump-diffusion model.

About the Contributors

xxi

Dilip B. Madan is Professor of Finance at the Robert H. Smith School of Business, University of Maryland. He is co-editor of Mathematical Finance and served as President of the Bachelier Finance Society 2002–2003. He has been a consultant to Morgan Stanley since 1996. He now also consults for Bloomberg and Caspian Capital. His primary research focus is on stochastic processes as they are applied to the management and valuation of ﬁnancial risks. David Nualart is Professor at the Faculty of Mathematics of the University of Barcelona. His research interests include a variety of topics in stochastic analysis, with emphasis on stochastic partial differential equations, Malliavin calculus and fractional Brownian motion. He is the author of the monograph Malliavin Calculus and Related Topics. Antonis Papapantoleon is a research assistant at the Department for Mathematical Stochastics, University of Freiburg. He received a Diploma in Mathematics from the University of Patras (2000) and an MSc in Financial Mathematics from the University of Warwick (2001). From January to August 2002 he worked at the FX Quantitative Research group of Commerzbank in Frankfurt. Goran Peskir is the Chair in Probability at the School of Mathematics, University of Manchester. In the period 1996–2005 he was an Associate Professor at the Department of Mathematical Sciences, University of Aarhus in Denmark. He is an internationally leading expert in the ﬁeld of Optimal Stopping and author to over sixty papers dealing with various problems in the ﬁeld of probability and its applications (optimal stopping, stochastic calculus, option pricing). Together with Albert Shiryaev he has co-authored the book Optimal Stopping and Free-Boundary Problems. Erwin Simons works in Quantitative Modeling at ING Brussels. After 3 years of frontofﬁce experience in Equity derivatives pricing, over the last year he switched to Interest-Rate derivatives modeling. He holds a PhD in Applied Mathematics from the Catholic University Leuven, von Karman Institute for Fluid Dynamics on the subject of large-scale computing of incompressible turbulent ﬂows. Jurgen Tistaert joined the Credit Risk Management Department of ING Brussels at the end of 1996 where he developed several rating, exposure and risk/performance models. He moved to Financial Markets in 2001, where the team is focusing on the R&D of pricing models for a broad range of derivative products. Before joining ING, he was a research assistant at the Quantitative Methods Group of K.U. Leuven Applied Economics Faculty, where he currently is appointed as a Fellow. Nadia Uys completed her Bachelors in Economic Science, majoring in Mathematical Statistics and Actuarial Science, at the University of the Witwatersrand in 2000, followed with Honours in Advanced Mathematics of Finance in 2001. Her MSc dissertation entitled ‘Optimal Stopping Problems and American Options’ was completed under the supervision of Professor G. Peskir (University of Aarhus) and Mr H. Hulley (Sydney Polytechnic) and received a distinction in 2005. She is currently teaching in the Programme in Advanced Mathematic of Finance at the University of the Witwatersrand and engaging in research toward a PhD under the supervision of Professor F. Lombard (University of Johannesburg). Nick Webber is Director of the Financial Options Research Centre, University of Warwick. Formerly Professor of Computational Finance at Cass Business School, he is interested not

xxii

About the Contributors

only in theoretical ﬁnancial mathematics, but also in methods for the fast evaluation of options prices under a variety of assumptions for returns distributions. As well as work with L´evy processes and numerical methods he has also worked on copulas, credit models and interest rates.

1 L´evy Processes in Finance Distinguished by their Coarse and Fine Path Properties Andreas E. Kyprianou Heriot Watt University, Edinburgh, UK and

R. Loeffen University of Utrecht, The Netherlands Abstract ´ We give a brief introduction to Levy processes and indicate the diversity of this class of stochastic processes by quoting a number of complete characterizations of coarse and ﬁne path properties. The theory is exempliﬁed by distinguishing such properties for ´ Levy processes which are currently used extensively in ﬁnancial models. Speciﬁcally, we treat jump-diffusion models (including Merton and Kou models), spectrally one-sided processes, truncated stable processes (including CGMY and Variance Gamma models), Meixner processes and generalized hyperbolic processes (including hyperbolic and normal inverse Gaussian processes).

1.1

INTRODUCTION

The main purpose of this text is to provide an entr´ee to the compilation Exotic Options and Advanced L´evy Models. Since path ﬂuctuations of L´evy processes play an inevitable role in the computations which lead to the pricing of exotic options, we have chosen to give a review of what subtleties may be encountered there. In addition to giving a brief introduction to the general structure of L´evy processes, path variation and its manifestation in the L´evy–Khintchine formula, we shall introduce classiﬁcations of drifting and oscillation, regularity of the half line, the ability to visit ﬁxed points and creeping. The theory is exempliﬁed by distinguishing such properties for L´evy processes which are currently used extensively in ﬁnancial models. Speciﬁcally, we treat jump-diffusion models (including Merton and Kou models), spectrally one-sided processes, truncated stable processes (including CGMY and variance gamma models), Meixner processes and generalized hyperbolic processes (including hyperbolic and normal inverse Gaussian processes). To support the presentation of more advanced path properties and for the sake of completeness, a number of known facts and properties concerning these processes are reproduced from the literature. We have relied heavily upon the texts by Schoutens (2003) and Cont and Tankov (2004) for inspiration. Another useful text in this respect is that of Boyarchenko and Levendorskii (2002). Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

2

Exotic Option Pricing and Advanced L´evy Models

The job of exhibiting the more theoretical facts concerning path properties have been greatly eased by the existence of the two indispensable monographs on L´evy processes, namely Bertoin (1996) and Sato (1999); see, in addition, the more recent monograph of Applebaum (2004) which also contains a section on mathematical ﬁnance. In the course of this text, we shall also brieﬂy indicate the relevance of the path properties considered to a number of exotic options. In some cases, the links to exotics is rather vague due to the fact that the understanding of pricing exotics and advanced L´evy models is still a ‘developing market’, so to speak. Nonetheless, we believe that these issues will in due course become of signiﬁcance as research progresses.

1.2

´ LEVY PROCESSES

We start with the deﬁnition of a real valued L´evy process followed by the L´evy–Khintchine characterization. Deﬁnition 1 A L´evy process X = {Xt : t ≥ 0} is a stochastic process deﬁned on a probability space (, F, P) which satisﬁes the following properties: (i) (ii) (iii) (iv)

The paths of X are right continuous with left limits almost surely. X0 = 0 almost surely. X has independent increments; for 0 ≤ s ≤ t, Xt − Xs is independent of σ (Xu : u ≤ s). X has stationary increments; for 0 ≤ s ≤ t, Xt − Xs is equal in distribution to Xt−s .

It turns out that there is an intimate relationship between L´evy processes and a class of distributions known as inﬁnitely divisible distributions which gives a precise impression of how varied the class of L´evy processes really is. To this end, let us devote a little time to discussing inﬁnitely divisible distributions. Deﬁnition 2 We say that a real valued random variable has an inﬁnitely divisible distribution if for each n = 1, 2, . . . . there exists a sequence of iid random variables 1 , . . . , n such that d

= 1,n + · · · + n,n d

where = is equality in distribution. Alternatively, we could have expressed this relation in terms of probability laws. That is to say, the law µ of a real valued random variable is inﬁnitely divisible if for each n = 1, 2, . . . there exists another law µn of a real valued random variable such that µ = µ∗n n , the n-fold convolution of µn . The full extent to which we may characterize inﬁnitely divisible distributions is carried out via their characteristic function (or Fourier transform of their law) and an expression known as the L´evy–Khintchine formula. Theorem 3 (L´evy–Khintchine formula) A probability law µ of a real valued random variable is inﬁnitely divisible with characteristic exponent , eiux µ (dx) = e−(u) for u ∈ R, R

L´evy Processes in Finance–Coarse and Fine Path Properties

3

if and only if there exists a triple (γ , σ, ), where γ ∈ R, σ ≥ 0 and is a measure supported on R\{0} satisfying R 1 ∧ x 2 (dx) < ∞, such that 1 (u) = iγ u + σ 2 u2 + (1 − eiux + iux1(|x| 0, Xt is a random variable whose law belongs to the class of inﬁnitely divisible distributions. This follows from the fact that for any n = 1, 2, . . . Xt = Xt/n + (X2t/n − Xt/n ) + · · · + (Xt − X(n−1)t/n )

(1.1)

together with the fact that X has stationary independent increments. Suppose now that we deﬁne for all u ∈ R, t ≥ 0 t (u) = − log E eiuXt then by using equation (1.1) twice we have for any two positive integers m, n that m1 (u) = m (u) = nm/n (u) and hence for any rational t > 0

t (u) = t1 (u) .

(1.2)

If t is an irrational number, then we can choose a decreasing sequence of rationals {tn : n ≥ 1} such that tn ↓ t as n tends to inﬁnity. Almost sure right continuity of X implies right continuity of exp{−t (u)} (by dominated convergence) and hence equation (1.2) holds for all t ≥ 0. In conclusion, any L´evy process has the property that E eiuXt = e−t(u) where (u) := 1 (u) is the characteristic exponent of X1 which has an inﬁnitely divisible distribution. Deﬁnition 5 In the sequel we shall also refer to (u) as the characteristic exponent of the L´evy process. Note that the law of a L´evy process is uniquely determined by its characteristic exponent. This is because the latter characterizes uniquely all one-dimensional distributions of X. From the property of stationary independent increments, it thus follows that the characteristic exponent characterizes uniquely all ﬁnite dimensional distributions which themselves uniquely characterize the law of X.

4

Exotic Option Pricing and Advanced L´evy Models

It is now clear that each L´evy process can be associated with an inﬁnitely divisible distribution. What is not clear is whether given an inﬁnitely divisible distribution, one may construct a L´evy process such that X1 has that distribution. This latter issue is resolved by the following theorem which gives the L´evy–Khintchine formula for L´evy processes. Theorem 6 Suppose that γ ∈ R, σ ≥ 0 and is a measure on R\{0} such that R (1 ∧ |x|2 )(dx) < ∞. From this triple deﬁne for each u ∈ R 1 2 2 (1.3) (u) = iγ u + σ u + [1 − eiux + iux1(|x| 0 is the initial value of the asset and X is a L´evy process. There are essentially four main classes of L´evy processes which feature heavily in current mainstream literature on market modeling with pure L´evy processes (we exclude from the discussion stochastic volatility models such as those of Barndorff–Nielsen and Shephard (2001)). These are the jump-diffusion processes (consisting of a Brownian motion with drift plus an independent compound Poisson process), the generalized tempered stable processes (which include more speciﬁc examples such as Variance Gamma processes and CGMY),

L´evy Processes in Finance–Coarse and Fine Path Properties

5

Generalized Hyperbolic processes and Meixner processes. There is also a small minority of papers which have proposed to work with the arguably less realistic case of spectrally onesided L´evy processes. Below, we shall give more details on all of the above key processes and their insertion into the literature. 1.3.1 Compound Poisson processes and jump-diffusions Compound Poisson processes form the simplest class of L´evy processes in the sense of understanding their paths. Suppose that ξ is a random variable with honest distribution F supported on R but with no atom at 0. Let Xt :=

Nt

t ≥0

ξi ,

i=1

where {ξi : i ≥ 1} are independent copies of ξ and N := {Nt : t ≥ 0} is an independent Poisson process with rate λ > 0. Then, X = {Xt : t ≥ 0} is a compound Poisson process. The fact that X is a L´evy process can easily be veriﬁed by computing the joint characteristic of the variables Xt − Xs and Xv − Xu for 0 ≤ v ≤ u ≤ s ≤ t < ∞ and showing that it factorizes. Indeed, standard facts concerning the characteristic function of the Poisson distribution leads to the following expression for the characteristic exponent of X, (u)) = (1 − eiux )λF (dx) (u) = λ(1 − F R

(u) = E(eiuξ ). Consequently, we can easily identify the L´evy triple via σ = 0 and where F γ = − R xλF (dx) and (dx) = λF (dx). Note that has ﬁnite total mass. It is not difﬁcult to reason that any L´evy process whose L´evy triple has this property must necessarily be a compound Poisson process. Since the jumps of the process X are spaced out by independent exponential distributions, the same is true of X and hence X is pathwise piecewise constant. Up to adding a linear drift, compound Poisson processes are the only L´evy processes which are piecewise linear. The ﬁrst model for risky assets in ﬁnance which had jumps was proposed by Merton (1976) and consisted of the log-price following an independent sum of a compound Poisson process, together with a Brownian motion with drift. That is, Xt = −γ t + σ Bt +

Nt

ξi ,

t ≥0

i=1

where γ ∈ R, {Bt : t ≥ 0} is a Brownian motion and {ξi : i ≥ 0} are normally distributed. Kou (2002) assumed the above structure, the so called jump-diffusion model, but chose the jump distribution to be that of a two-sided exponential distribution. Kou’s choice of jump distribution was heavily inﬂuenced by the fact that analysis of ﬁrst passage problems become analytically tractable which itself is important for the valuation of American put options (see Chapter 11 below). Building on this idea, Asmussen et al. (2004) introduce a jump-diffusion model with two-sided phasetype distributed jumps. The latter form a class of distributions which generalize the two-sided exponential distribution and like Kou’s model, have the desired property that ﬁrst passage problems are analytically tractable.

6

Exotic Option Pricing and Advanced L´evy Models

1.3.2 Spectrally one-sided processes Quite simply, spectrally one-sided processes are characterized by the property that the support of the L´evy measure is restricted to the upper or the lower half line. In the latter case, that is (0, ∞) = 0, one talks of spectrally negative L´evy processes. Without loss of generality we can and shall restrict our discussion to this case unless otherwise stated in the sequel. Spectrally negative L´evy processes have not yet proved to be a convincing tool for modeling the evolution of a risky asset. The fact that the support of the L´evy measure is restricted to the lower half line does not necessarily imply that the distribution of the L´evy process itself is also restricted to the lower half line. Indeed, there are many examples of spectrally negative processes whose ﬁnite time distributions are supported on R. One example, which has had its case argued for in a ﬁnancial context by Carr and Wu (2003) and Cartea and Howison (2005), is a spectrally negative stable process of index α ∈ (1, 2). To be more precise, this is a process whose L´evy measure takes the form (dx) = 1(x 0 and whose parameter σ is identically zero. A lengthy calculation reveals that this process has the L´evy–Khintchine exponent πα (u) = c|u|α 1 + i tan signu . 2 Chan (2000, 2004), Mordecki (1999, 2002) and Avram et al. (2002, 2004), have also worked with a general spectrally negative L´evy process for the purpose of pricing American put and Russian options. In their case, the choice of model was based purely on a degree of analytical tractability centred around the fact that when the path of a spectrally negative process passes from one point to another above it, it visits all other points between them. 1.3.3 Meixner processes The Meixner process is deﬁned through the Meixner distribution which has a density function given by

β(x − µ) (2 cos(β/2))2δ i(x − µ) 2 exp fMeixner (x; α, β, δ, µ) = δ + 2απ (2δ) α α where α > 0, −π < β < π, δ > 0, m ∈ R. The Meixner distribution is inﬁnitely divisible with a characteristic exponent

2δ cos(β/2) Meixner (u) = − log − iµu, cosh(αu − iβ)/2 and therefore there exists a L´evy process with the above characteristic exponent. The L´evy triplet (γ , σ, ) is given by ∞ sinh(βx/α) γ = −αδ tan(β/2) + 2δ dx − µ, sinh(π x/α) 1

L´evy Processes in Finance–Coarse and Fine Path Properties

7

σ = 0 and (dx) = δ

exp(βx/α) dx. x sinh(π x/α)

(1.4)

The Meixner process appeared as an example of a L´evy process having a particular martingale relation with respect to orthogonal polynomials (see Schoutens and Teugels (1998) and Schoutens (2000)). Grigelionis (1999) and Schoutens (2001, 2002) established the use of the Meixner process in mathematical ﬁnance. Relationships between Mexiner distributions and other inﬁnitely divisible laws also appear in the paper of Pitman and Yor (2003). 1.3.4 Generalized tempered stable processes and subclasses The generalized tempered stable process has L´evy density ν := d/dx given by ν(x) =

cp −λp x e 1{x>0} 1+α x p

+

cn eλn x 1{x 0, λn > 0, cp > 0 and cn > 0. These processes take their name from stable processes which have L´evy measures of the form

cp cn (dx) = 1{x>0} + 1{x 0. Stable processes with index α ∈ (0, 1] have no moments and when α ∈ (1, 2) only a ﬁrst moment exists. Generalized tempered stable processes differ in that they have an exponential weighting in the L´evy measure. This guarantees the existence of all moments, thus making them suitable for ﬁnancial modelling where a moment-generating function is necessary. Since the shape of the L´evy measure in the neighbourhood of the origin determines the occurrence of small jumps and hence the small time path behaviour, the exponential weighting also means that on small time scales stable processes and generalized tempered stable processes behave in a very similar manner. Generalized tempered stable processes come under a number of different names. They are sometimes called KoBoL processes, named after the authors Koponen (1995) and Boyarchenko and Levendorskii (2002). Carr et al. (2002, 2003) have also studied this sixparameter family of processes and as a consequence of their work they are also referred to as generalized CGMY processes or, for reasons which will shortly become clear, CCGMYY processes. There seems to be no uniform terminology used for this class of processes at the moment and hence we have simply elected to follow the choice of Cont and Tankov (2004). Since |x|ν(x)dx < ∞ R\(−1,1)

it turns out to be more convenient to express the L´evy–Khintchine formula in the form

(u) = iuγ +

∞

−∞

(1 − eiux + iux)ν(x)dx

(1.5)

8

Exotic Option Pricing and Advanced L´evy Models

where γ = γ − by

R\(−1,1) xν(x)dx

< ∞. In this case, the characteristic exponent is given

(u) = iuγ − Ap − An , where

iu iuc + c (λ − iu) log 1 − p p p λp

iu iu + log 1 − −cp Ap = λ λ p p

iuαp iu αp αp 1− −1+ (−αp )λp cp λp λp

iu −iucn + cn (λn + iu) log 1 + λn

iu iu −cn − + log 1 + An = λ λ n n

αn

iu iuαn 1+ −1− (−αn )λαnn cn λn λn

if αp = 1 if αp = 0 otherwise if αn = 1 if αn = 0 otherwise

(see Cont and Tankov (2004), p. 122). When αp = αn = Y , cp = cn = C, λp = M and λn = G, the generalized tempered stable process becomes the so called CGMY process, named after the authors who ﬁrst introduced it, i.e. Carr et al. (2002). The characteristic exponent of the CGMY process for Y = 0 and Y = 1 is often written as CGMY (u) = −C (−Y )[(M − iu)Y − M Y + (G + iu)Y − GY )] − iuµ,

(1.6)

which is the case for an appropriate choice of γ , namely

Y GY Y MY + iµ. − γ = C (−Y ) M G The properties of the CGMY process can thus be inferred from the properties of the generalized tempered stable process. Note that in this light, generalized tempered stable processes are also referred to as CCGMYY. As a limiting case of the CGMY process, but still within the class of generalized tempered stable processes, we have the variance gamma process. The latter was introduced as a predecessor to the CGMY process by Madan and Seneta (1987) and treated in a number of further papers by Madan and co-authors. The variance gamma process can be obtained by starting with the parameter choices for the CGMY but then taking the limit as Y tends to zero. This corresponds to a generalized tempered stable process with αp = αn = 0. Working with γ = −C/M + C/G + µ, we obtain the variance gamma process with the characteristic exponent

iu iu + log 1 + − iuµ. (1.7) VG (u) = C log 1 − M G The characteristic exponent is usually written as

1 2 2 1 VG (u) = log 1 − iθ κu + σ κu − iuµ, κ 2

L´evy Processes in Finance–Coarse and Fine Path Properties

where

C = 1/κ,

M=

2

θ 2 + 2 σκ − θ σ2

and G =

9

2 θ 2 + 2 σκ + θ σ2

for θ ∈ R and κ > 0. Again, the properties of the variance gamma process can be derived from the properties of the generalized tempered stable process. 1.3.5 Generalized hyperbolic processes and subclasses The density of a generalized hyperbolic distribution is given by λ 1 fGH (x; α, β, λ, δ, µ) = C(δ 2 + (x − µ)2 ) 2 − 4 Kλ− 1 α δ 2 + (x − µ)2 eβ(x−µ) , 2

where

− β 2 )λ/2 C= √ 2π α λ−1/2 δ λ Kλ δ α 2 − β 2 (α 2

and with α > 0, 0 ≤ |β| < α, λ ∈ R, δ > 0 and µ ∈ R. The function Kλ stands for the modiﬁed Bessel function of the third kind with index λ. This distribution turns out to be inﬁnitely divisible with a characteristic exponent GH (u) = − log

α2 − β 2 α 2 − (β + iu)2

λ/2

Kλ (δ α 2 − (β + iu)2 ) − iµu. Kλ (δ α 2 − β 2 )

These facts are non-trivial to prove–see Halgreen (1979) who gives the proofs. The corresponding L´evy measure is rather complicated, being expressed as integrals of special functions. We refrain from offering the L´evy density here on account of its complexity and since we shall not use it in the sequel. Generalized hyperbolic processes were introduced within the context of mathematical ﬁnance by Barndorff-Nielsen (1995, 1998) and Erbelein and Prause (1998). When λ = 1, we obtain the special case of a hyperbolic process and when λ = − 12 , the normal inverse Gaussian process is obtained. Because the modiﬁed Bessel function has a simple form when λ = − 12 , namely K− 1 (z) = 2

π − 1 −z z 2e , 2

the characteristic exponent can be simpliﬁed to NI G (u) = δ

α 2 − (β + iu)2 −

α2 − β 2 .

Eberlein and Hammerstein (2002) investigated some limiting cases of generalized hyperbolic distributions and processes. Because for λ > 0 Kλ ∼

z −λ 1

(λ) 2 2

when z → 0,

10

Exotic Option Pricing and Advanced L´evy Models

we have that

λ 2δ α 2 − β 2 2δ α 2 − (β + iu)2

2 u2 α − (β + iu)2 2βiu = λ log 1 + = λ log − α2 − β 2 α2 − β 2 α2 − β 2 when δ → 0 and for µ = 0. Here we write f ∼ g when u → ∞ to mean that limu→∞ f (u)/ g(u) = 1. So, we see that when δ → 0 and for µ = 0, λ = 1/κ, β = θ/σ 2 and α =

α2 − β 2 GH (u) ∼ − log 2 α − (β + iu)2

λ/2

(2/κ)+(θ 2 /σ 2 ) , σ2

the characteristic exponent of the generalized hyperbolic process converges to the characteristic exponent of the variance gamma process. Because the variance gamma process is obtained by a limiting procedure, its path properties cannot be deduced directly from those of the generalized hyperbolic process. Indeed, we shall see they are fundamentally different processes.

1.4

PATH PROPERTIES

In the following sections, we shall discuss a number of coarse and ﬁne path properties of general L´evy processes. These include path variation, hitting of points, creeping and regularity of the half line. With the exception of the last property, none of the above have played a prominent role in mainstream literature on the modeling of ﬁnancial markets. Initial concerns of L´evydriven models were focused around the pricing of vanilla-type options, that is, options whose value depends on the distribution of the underlying L´evy process at a ﬁxed point in time. Recently, more and more attention has been paid to exotic options which are typically path dependent. Fluctuation theory and path properties of Brownian motion being well understood has meant that many examples of exotic options under the assumptions of the classical Black–Scholes models can and have been worked out in the literature. We refer to objects such as American options, Russian options, Asian options, Bermudan options, lookback options, Parisian options, Israeli or game options, Mongolian options, and so on. However, dealing with exotic options in L´evy-driven markets has proved to be considerably more difﬁcult as a consequence of the more complicated, and to some extent, incomplete nature of the theory of ﬂuctuations of L´evy processes. Nonetheless, it is clear that an understanding of course and ﬁne path properties plays a role in the evaluation of exotics. In the analysis below, we shall indicate classes of exotics which are related to the described path property. 1.4.1 Path variation Understanding the path variation for a L´evy process boils down to a better understanding of the L´evy–Khintchine formula. We therefore give a sketch proof of Theorem 6 which shows that for any given L´evy triple (γ , σ, ) there exists a L´evy process whose characteristic exponent is given by the L´evy–Khintchine formula. Reconsidering the formula for , note that we may write it in the form 1 2 2 iux (1 − e )(dx) (u) = iuγ + σ u + 2 R\(−1,1) iux + (1 − e + ixu)(dx) 0 0, consider the L´evy processes X(3,) deﬁned by (3,) () = Yt − t x(dx), t ≥ 0 (1.8) Xt 0} be the set of points that a L´evy process can hit. We say a L´evy process can hit points if C = ∅. Kesten (1969) and Bretagnolle (1971) give the following classiﬁcation. Theorem 7 Suppose that X is not a compound Poisson process. Then X can hit points if and only if

1 du < ∞. (1.11) 1 + (u) R Moreover, (i) If σ > 0, then X can hit points and C = R. (ii) If σ = 0, but X is of unbounded variation and X can hit points, then C = R. (iii) If X is of bounded variation, then X can hit points, if and only if, d = 0 where d is the drift in the representation (equation (1.10)) of its L´evy–Khintchine exponent . In this case, C = R unless X or −X is a subordinator and then C = (0, ∞) or C = (−∞, 0), respectively.

L´evy Processes in Finance–Coarse and Fine Path Properties

13

The case of a compound Poisson process will be discussed in Section 1.5.1. Excluding the latter case, from the L´evy–Khintchine formula we have that 1 ((u)) = σ 2 u2 + (1 − cos(ux))(dx) 2 R\{0} and

((u)) = γ u +

R\{0}

(− sin(ux) + ux1{|x| 0 we have

p 1 du < ∞ 1 + (u) −p and

−p

−∞

∞ 1 1 du = du 1 + (u) 1 + (u) p

and the question as to whether the integral (equation is ﬁnite or inﬁnite depends (1.11)) 1 on what happens when u → ∞. If, for example, 1+(u) g(u) when u → ∞, then we can use g to deduce whether the integral (equation (1.11)) is ﬁnite or inﬁnite. Note, we use the notation f g to mean that there exists a p > 0, a > 0 and b > 0 such that ag(u) ≤ f (u) ≤ bg(u) for all u ≥ p, This technique will be used quite a lot in the examples we consider later on in the text. An example of an exotic option which in principle makes use of the ability of a L´evy process to hit points is the so-called callable put option. This option belongs to a more general class of exotics called Game or Israeli options, described in Kifer (2000) (see also the review by K¨uhn and Kallsen (2005) in this volume). Roughly speaking, these options have the same structure as American-type options but for one signiﬁcant difference. The writer also has the option to cancel the contract at any time before its expiry. The consequence of the writer cancelling the contract is that the holder is paid what they would have received had they exercised at that moment, plus an additional amount (considered as a penalty for the writer). When the claim of the holder is the same as that of the American put and the penalty of the writer is a constant, δ, then this option has been named a callable put in K¨uhn and Kyprianou (2005) (also an Israeli δ-penalty put option in Kyprianou (2004)). In the latter two papers, the value and optimal strategies of writer and holder of this exotic option have been calculated explicitly for the Black–Scholes market. It turns out there that the optimal strategy of the writer is to cancel the option when the value of the underlying asset hits precisely the strike price, providing that this happens early on enough in the

14

Exotic Option Pricing and Advanced L´evy Models

contract. Clearly, this strategy takes advantage of case (i) of the above theorem. Suppose now for the same exotic option that instead of an exponential Brownian motion we work with an exponential L´evy process which cannot hit points. What would be the optimal strategies of the writer (and hence the holder)? 1.4.3 Creeping Deﬁne for each x ≥ 0 the ﬁrst passage time τx+ = inf{t > 0 : Xt > x}. Here, we work with the deﬁnitions inf ∅ = ∞ and if τx+ = ∞, then Xτx+ = ∞. We say that a L´evy process X creeps upwards if for all x ≥ 0 P (Xτx+ = x) > 0 and that X creeps downwards if −X creeps upwards. Creeping simply means that with positive probability, a path of a L´evy process continuously passes a ﬁxed level instead of jumping over it. A deep and yet enchanting aspect of L´evy processes, excursion theory, allows for the following non-trivial deduction concerning the range of {Xτx+ : x ≥ 0}. With probability one, the random set {Xτx+ : x ≥ 0} ∩ [0, ∞) corresponds precisely to the range of a certain subordinator, killed at an independent exponential time with parameter q ≥ 0. The case that q = 0 should be understood to mean that there is no killing and hence that τx+ < ∞ almost surely for all x ≥ 0. In the obvious way, by considering −X, we may draw the same conclusions for the range of {−Xτx− : x ≥ 0} ∩ [0, ∞) where τx− := inf{t > 0 : Xt < x}. Suppose that κ(u) and κ (u) are the characteristic exponents of the aforementioned subordinators for the ranges of the upward and downward ﬁrst passage processes, respectively. Note, for example, that for u ∈ R κ(u) = q − iau + (1 − eiux )π(dx)

(0,∞)

for some π satisfying 0∞ (1 ∧ x)π(dx) < ∞ and a ≥ 0 (recall that q is the killing rate). It is now clear from Theorem 7 that X creeps upwards, if and only if, a > 0. The so-called Wiener–Hopf factorization tells us where these two exponents κ and κ are to be found: (u) = κ(u) κ (−u).

(1.12)

Unfortunately, there are very few examples of L´evy processes for which the factors κ and κ are known. Nonetheless, the following complete characterization of upward creeping has been established. Theorem 8 The L´evy process X creeps upwards, if and only if, one of the following three situations occurs: (i) X has bounded variation and d > 0 where d is the drift in the representation (equation (1.10)) of its L´evy–Khintchine exponent .

L´evy Processes in Finance–Coarse and Fine Path Properties

(ii) X has a Gaussian component, (σ > 0). (iii) X has unbounded variation, no Gaussian component and 1 x([x, ∞)) dx < ∞. 0 y 0 −x −1 ((−∞, u])dudy

15

(1.13)

This theorem is the collective work of Miller (1973) and Rogers (1984), with the crowning conclusion in case (iii) being given recently by Vigon (2002). As far as collective statements about creeping upwards and downwards are concerned, the situation is fairly straightforward to resolve with the help of the following easily proved lemma. (See Bertoin (1996), p. 16). Lemma 9 Let X be a L´evy process with characteristic exponent (u). (i) If X has ﬁnite variation then lim

u↑∞

(u) = −id u

where d is the drift appearing in the representation (equation (1.10)) of . (ii) For a Gaussian coefﬁcient σ ≥ 0, lim

u↑∞

(u) 1 = σ 2. u2 2

From the above lemma we see, for example, that lim

u↑∞

κ(u) = 0, u

if and only if, X creeps upwards. Consequently, from the Wiener–Hopf factorization (equation (1.12)) the following well-established result holds (see Bertoin (1996), p. 175). Lemma 10 A L´evy process creeps both upwards and downwards, if and only if it has a Gaussian component. There is also a relation between hitting points and creeping. Clearly, a process which creeps can hit points. In the case of bounded variation we see that hitting points is equivalent to creeping upwards or downwards. However, in the case of unbounded variation, it can be that a process does not creep upwards or downwards, but still can hit points. We will see an example of this later on–see Remark 17. A process which hits a point but does not creep over it must therefore do so by jumping above and below that point an inﬁnite number of times before hitting it. When considering the relevance of creeping to exotic option pricing, one need only consider any kind of option involving ﬁrst passage. This would include, for example, barrier options as well as Russian and American put options. Taking the latter case with inﬁnite horizon, the optimal strategy is given by ﬁrst passage below a ﬁxed value of the underlying L´evy process. The value of this option may thus be split into two parts, namely, the premium for exercise by jumping clear of the boundary and the premium for creeping over the boundary. For the ﬁnite expiry case, it is known that the optimal strategy of the holder is

16

Exotic Option Pricing and Advanced L´evy Models

to exercise when the underlying L´evy process crosses a time-varying barrier. In this case, a more general concept of creeping over moving boundaries may be introduced and it would be interesting to know whether the ability to creep over the optimal exercise boundary has any inﬂuence on the continuity or smoothness properties of the boundary as a function of time. 1.4.4 Regularity of the half line For a L´evy process X (which starts at zero) we say that 0 is regular for (0, ∞) (equiv. the upper half line) if X enters (0, ∞) immediately. That is to say, if P(τ (0,∞) = 0) = 1,

where

τ (0,∞) = inf{t > 0 : Xt ∈ (0, ∞)}.

Because of the Blumenthal 0–1 law, the probability P(τ (0,∞) = 0) is necessarily zero or one. When this probability is zero, we say that 0 is irregular for (0, ∞). We also say that 0 is regular for (−∞, 0) (equiv. the lower half line) if −X is regular for the upper half line. The following theorem is the conclusion of a number of works and gives a complete characterization of regularity for the upper half line (see Shtatland (1965), Rogozin (1968) and Bertoin (1997)). Theorem 11 For a L´evy process X, the point 0 is regular for (0, ∞), if and only if, one of the following three situations occurs: (i) X is a process of unbounded variation. (ii) X is a process of bounded variation and d > 0 where d is the drift in the representation (equation (1.10)) of its L´evy–Khintchine exponent . (iii) X is a process of bounded variation, d = 0 (with d as in (ii)) and

1 0

x 0

x(dx) = ∞. (−∞, −y)dy

(1.14)

Regularity of the lower half line has already proved to be of special interest to the pricing of American put options. In Alili and Kyprianou (2004). a perpetual American put is considered where the underlying market is driven by a general L´evy process. Building on the work of Mordecki (1999, 2002), Boyarchenko and Levendorskii (2002a) and Chan (2000, 2004), it is shown that the traditional condition of smooth pasting at the optimal exercise boundary may no longer be taken for granted. Indeed necessary and sufﬁcient conditions are given for no smooth pasting. This condition is quite simply the regularity of (−∞, 0) for 0 (in other words the regularity of the upper half line for −X). It was conjectured in Alili and Kyprianou (2005) that the very same condition would also characterize the appearance of smooth ﬁt for the ﬁnite expiry American put where the boundary is time varying. Indeed, numerical simulations in Matache et al. (2003) and Almendral (2004) support this conjecture. A ﬁnancial interpretation of a non-smooth ﬁt condition has yet to be clariﬁed.

L´evy Processes in Finance–Coarse and Fine Path Properties

1.5

17

EXAMPLES REVISITED

1.5.1 Compound Poisson processes and jump-diffusions Suppose that X is a compound Poisson process. Clearly, X has paths of bounded variation, cannot creep upwards or downwards and is irregular for the upper and lower half lines. Since the L´evy measure is bounded, it is easy to reason that the real and imaginary part 1 of its characteristic (u), and hence 1+(u) , is bounded away from zero and that the integral (equation (1.11)) is inﬁnite. Nonetheless, certain compound Poisson processes can hit points. Take the simple example of a Poisson process. This is a process which hits {0, 1, 2, . . . .}. Other similar examples where the jump distribution is supported on a lattice are possible. This is the reason for the exclusion of compound Poisson processes from Theorem 7. However, it can be said that so long as the jump distribution F is diffuse, a compound Poisson process can hit no point other than 0, its initial holding point. If X is a jump-diffusion then the above properties change drastically. In particular, if the Gaussian component is non-zero then this will dominate the paths of the process. This is because, until the ﬁrst jump, which occurs at arbitrarily large times with positive probability, the process behaves as a Brownian motion with drift. It is clear that paths will be of unbounded variation, there will be regularity for the upper and lower half lines, the process may creep both upwards and downwards and any point can be hit with positive probability. Note the latter fact is a well-known property of Brownian motion and does not require Theorem 7. 1.5.2 Spectrally negative processes By deﬁnition, spectrally negative processes creep upwards and hence can hit points. Therefore, unless there is a Gaussian component present, they cannot creep downwards. It is possible to have such processes of both bounded and unbounded variation according to the 0 ﬁniteness of the integral −1 |x|(dx). Clearly, if it is a process of unbounded variation, then there is regularity for the upper and lower half lines. If it is a process of bounded variation and not the negative of the subordinator, then by reconsidering equation (1.10) we see that necessarily the process must take the form of a strictly positive drift minus a subordinator. Consequently, from Theorem 11 in this case, there is regularity for the upper half line but not for the lower half line. This, in turn, implies that for spectrally negative L´evy processes, regularity of the lower half line coincides with having paths of inﬁnite variation. 1.5.3 Meixner process We begin with a known fact concerning path variation. Proposition 12 The Meixner process is of unbounded variation and hence is regular for the upper and lower half lines. Proof. Denote ν(x) as the density of the L´evy measure (equation (1.4)). We have to prove that (−1,1) |x|ν(x)dx is inﬁnite. For x ∈ (0, 1) we have |x|ν(x) = δ

e(β+π)x/α 1 exp(βx/α) = 2δ 2πx/α ≥ 2δ 2πx/α sinh(π x/α) e −1 e −1

18

Exotic Option Pricing and Advanced L´evy Models

and so

1

1

|x|ν(x)dx ≥

0

2δ 0

showing in particular that

1 e2πx/α

−1

dx

1 (2π/α)x 1 = 2δ log e − 1 − x = ∞ 2π/α 0

(−1,1) |x|ν(x)dx

= ∞.

Proposition 13 A Meixner process cannot hit points and therefore cannot creep. Proof. To see whether the Meixner process can hit points we have to employ the integral given in equation (1.11). In order to use this, we ﬁrst split the characteristic exponent into its real and imaginary part. First note that

1 1 1 1 αu − iβ = cos β cosh αu − sin β sinh αu i. cosh 2 2 2 2 2 Then with

2

2 1 1 1 1 β cosh αu β sinh αu + sin 2 2 2 2 1 1 − sin 2 β sinh 2 αu q = arctan cos 12 β cosh 12 αu r=

cos

and

we have (u) = −2δ log(cos(β/2)) + 2δ log(r) + (2δq − mu)i and hence

r 1 + 2δ log cos(β/2) − (2δq − mu)i 1 = and 2 1 + (u) r 1 + 2δ log cos(β/2) + (2δq − mu)2 r

1 + 2δ log cos(β/2) 1 . = 2 1 + (u) r 1 + 2δ log cos(β/2) + (2δq − mu)2

When u → ∞, then cosh(αu/2) sinh(αu/2) eαu/2 , and so r eαu/2 and log(r/ cos(β/2)) 12 αu when u → ∞. Further, (2δq − mu)2 m2 u2 when u → ∞, because arctan(z) ∈ (− 12 π, 12 π ) for all z ∈ R. So

1 + δαu 1 u−1 when u → ∞. 1 + (u) (1 + δαu)2 + m2 u2 ∞ Because for all p > 0, p u−1 du = ∞, we ﬁnd that the integral (equation (1.11)) is inﬁnite and therefore the Meixner process cannot hit points.

L´evy Processes in Finance–Coarse and Fine Path Properties

19

1.5.4 Generalized tempered stable process We again begin with a known statement concerning path variation. Proposition 14 The generalized tempered stable process has bounded variation, if and only if, αp < 1 and αn < 1. Proof. We have to determine whether the integral given in equation (1.9) is ﬁnite or inﬁnite, where this integral is given by 1 0 1 cp −λp x cn |x|ν(x)dx = e dx + eλn x dx. α αn p −1 0 x −1 (−x) It is clear, however, that this boils down to whether 0 1 −αp x dx + (−x)−αn dx −1

0

is ﬁnite or inﬁnite and the above expression is only ﬁnite when αp < 1 and αn < 1. Proposition 15 In the case of unbounded variation, a generalized tempered stable process creeps upwards, if and only if, αp < αn . Proof. Because the integral given in equation (1.13) is independent of cp and cn , we can assume without loss of generality that cp = cn = 1. In the following calculations, b1 , b2 , . . . are constants in R. Recall from the previous proposition that we have unbounded variation if αp ∈ [1, 2) or αn ∈ [1, 2). We shall therefore prove the result under the additional assumption that there is unbounded variation because αn ∈ [1, 2). Similar arguments then deliver the same conclusions when we assume that there is unbounded variation because αp ∈ [1, 2). For u ∈ (0, 1] we have ([u, ∞)) ∞ e−λp x x −(1+αp ) dx = u

∞ 1 −λp −(1+αp ) e x dx + e−λp x dx = b1 + b2 e−λp u 1 u 1 ∞ ≤ e−λp x dx + x −1 dx = b3 − log(u) 1 u 1 ∞ e−λp x dx + x −(1+αp ) dx = b4 + b5 u−αp

if αp ≤ −1 if αp = 0 if αp ∈ (−1, 2)\{0}

u

1

and for u ∈ [−1, 0) ((−∞, u]) = ≥

u

−∞ u −1

eλn x (−x)−(1+αn ) dx

e−λn (−x)−(1+αn ) dx = b6 (−u)−αn + b7

20

Exotic Option Pricing and Advanced L´evy Models

if αn ∈ [1, 2). Hence y b8 (−y)−αn +1 + b9 + b10 (y + 1) if αn ∈ (1, 2) ((−∞, u]) du ≥ if αn = 1 b11 log(−y) + b12 (y + 1) −1 and then 0 −x

y −1

((−∞, u]) du dy ≥

b13 x −αn +2 + b14 x 2 + b15 x b16 (x log(x) − x) + b17 x 2 + b18 x

if αn ∈ (1, 2) if αn = 1.

We then have for αp ∈ (−1, 2)\{0} and αn ∈ (1, 2) 0 y

x([x, ∞))

−x −1 ((−∞, u]) du dy

≤

b4 x + b5 x −αp +1 −α b13 x n +2 + b14 x 2 + b15 x

=

b4 + b5 x −αp , b13 x −αn +1 + b14 x + b15

for x ∈ (0, 1]. Deﬁne the right side of the above inequality f (x) and note that it is continuous for all x ∈ (0, 1]. When x → 0, then f x αn −1 for αp < 0 and f x αn −1−αp

for αp > 0.

1 Hence, 0 f (x)dx < ∞ for αp < αn and therefore the integral (equation (1.13)) is ﬁnite for these parameter values. When αp ≤ −1 or αp = 0 or αn = 1, there is a similar upper bound for which the same conclusions can be drawn. We have thus so far shown that there is creeping upwards if αp < αn . To prove the ‘only-if’ part, note that the L´evy density of −X is ν(−x) and this density is the same as ν(x) except that the p-parameters and the n-parameters have switched places. So, we can immediately conclude from the previous analysis that X creeps downwards if αp > αn and then from Lemma 10 we see that since there is no Gaussian component, X cannot creep upwards if αp > αn . Now only the case remains when αp = αn ∈ [1, 2). In this case, we can use for u ∈ (0, 1] the lower bound for ((−∞, −u]) as a lower bound for ([u, ∞)) and the upper bound for ([u, ∞)) as an upper bound for ((−∞, −u]) in order to create a lower bound for the integral (equation (1.13)) which turns out to be inﬁnite. Proposition 16 In the case of unbounded variation, a generalized tempered stable process can hit points, unless αp = αn = 1 and cp = cn . Proof. Because this process creeps upwards or downwards when αp = αn , we only have to prove that the process can hit points when αp = αn = α ∈ [1, 2). Let rp = 2 qp = arctan − λup , rn = 1 + λu2 and qn = arctan λun . Then n

Ap (u) = βp (rpα cos(αqp ) − 1) + iβp rpα sin(αqp ) + αu λp αu α α An (u) = βn (rn cos(αqn ) − 1) + iβn rn sin(αqn ) − λn ,

1+

u2 , λ2p

L´evy Processes in Finance–Coarse and Fine Path Properties

21

with βp = (−α)λαp cp and βn = (−α)λαn cn . So, for αp = αn ∈ (1, 2)

1 1 + (u)

=

1 − (Ap (u)) − (An (u)) [1 − (Ap (u)) − (An (u))]2 + [uγ − (Ap (u)) − (An (u))]2

≤

1 − (Ap (u)) − (An (u)) 1 . = [1 − (Ap (u)) − (An (u))]2 1 − (Ap (u)) − (An (u))

1 We see that the above upper bound of 1+(u) as a function of u is continuous and symmetric. So, the question whether the integral of this function from minus inﬁnity to inﬁnity is ﬁnite or inﬁnite depends on how the function behaves when u → ∞. When u → ∞, then rp rn u, qp → − 12 π and qn → 12 π . So, cos(αqp ) = cos(αqn ) → a for u → ∞, where a is a constant smaller than zero. Because (−α) > 0 for α ∈ (1, 2), we have ∞ that 1 − (Ap (u)) − (An (u)) uα when u → ∞. Because for all t > 0 t u−α du < ∞, the integral of the upper bound is ﬁnite and hence this process can hit points when αp = αn ∈ (1, 2). Now, let αp = αn = 1. Then by using the same expressions for rp , rn , qp and qn as above, Ap (u) = cp (λp log(rp ) + uqp ) + icp (u + λp qp − u log(rp )) An (u) = cn (λn log(rn ) − uqn ) + icn (−u + λn qn + u log(rn )) and then when u → ∞, 1 − (Ap (u)) − (An (u)) u if cp = cn uγ − (Ap (u)) − (An (u)) u log(u) uγ − (Ap (u)) − (An (u)) u or uγ − (Ap (u)) − (An (u)) 1 if cp = cn . So, when cp = cn , then

1 1+(u)

∞ t

1 u log2 (u)

and because for large t,

1 1 ∞ =− < ∞, log(u) t u log2 (u)

the process can hit points. For the case where cp = cn , then integral given in equation (1.11) is inﬁnite.

1 1+(u)

u−1 and the

Remark 17 The last two propositions give us an example of a L´evy process which can hit points but cannot creep. Take the example of a CGMY process where the parameter Y ∈ (1, 2). To some extent this is not surprising. As noted earlier, the small time behaviour of generalized tempered stable processes should in principle be similar to the behaviour of stable processes due to the similarities in their L´evy measures in the neighbourhood of the origin. In this sense, the class of CGMY processes mentioned are closely related to a symmetric stable processes of unbounded variation and for this class it is well known that they can hit points but cannot creep. To see the latter fact, note from Lemma 10 that it is clear that a symmetric stable process (or indeed any L´evy process which is symmetric without a Gaussian component) cannot creep upwards nor downwards on account of symmetry. On the

22

Exotic Option Pricing and Advanced L´evy Models

other hand, it is well known (cf. Chapter VIII in Bertoin (1996)) that for a symmetric stable process of index α, (u) = c|u|α for some constant c > 0, and there is unbounded variation when α ∈ (1, 2). It is easily veriﬁed that with this choice of , the integral (equation (1.11)) is ﬁnite. In the case that a generalized tempered stable process has bounded variation, that is, when αp < 1 and αn < 1, we can use the drift d to determine whether the process can hit points, creeps or whether 0 is regular for (0, ∞). However, then we have to know ∞ what the drift looks like. Comparing equations (1.2) and (1.6), we see that d = −γ − −∞ xν(x)dx. The latter integral can be computed explicitly; however, it is easier to use Lemma 9 (i) to identify the drift term since the L´evy–Khintchine formula is polynomial when there is bounded variation. Indeed, it is easy to see by inspection that d = −γ − dp − dn , where

α −1 −cp αp (−αp )λpp cp dp = λp cn αn (−αn )λnαn −1 if cn dn = if − λn

if αp ∈ (−∞, 1)\{0} and

if αp = 0 αn ∈ (−∞, 1)\{0} αn = 0.

Remark 18 For the special cases of the CGMY and variance gamma processes, we note that the representation of the L´evy–Khintchine formula given in equations (1.6) and (1.7) yields in both cases that the drift term d = µ. Proposition 19 When a generalized tempered stable process has bounded variation and drift equal to zero, then 0 is regular for (0, ∞), if and only if, αp ≥ αn and at least one of these two parameters is not smaller than zero. Proof. We can use the integral (equation (1.14)) to determine whether 0 is regular for (0, ∞). For y ∈ (0, 1) we have −y −y c n e λn x c n e λn x dx ≥ dx ((−∞, −y)) = 1+αn 1+αn −∞ (−x) −1 (−x) −y c n e λn b1 y −αn + b2 if αn = 0 dx = ≥ 1+α n b3 log(y) if αn = 0 −1 (−x) and ((−∞, −y)) = ≤

−1 −∞ −1 −∞

=

c n e λn x dx + (−x)1+αn λn x

cn e dx + (−x)1+αn

b4 + b5 y −αn b6 + b7 log(y)

−y −1 −y −1

c n e λn x dx (−x)1+αn cn dx (−x)1+αn

if αn = 0 if αn = 0,

L´evy Processes in Finance–Coarse and Fine Path Properties

where b1 , b2 , . . . are constants in R. So, x b2 x + b8 x −αn +1 ≤ 0 ((−∞, −y))dy ≤ b4 x + b9 x −αn +1

b3 (x log(x) − x) ≤

x 0

23

if αn = 0

((−∞, −y))dy ≤ b6 x + b7 (x log(x) − x)

if αn = 0.

Note that the constants b1 , b2 , . . . have values such that the above upper and lower boundaries are strictly positive for x > 0. Now, when αp < 1, αn < 1 and αn = 0, then x 0

xcp x −1−αp e−λp x cp x −1−αp e−λp x xν(x) ≤ = . −α +1 n b2 x + b8 x b2 + b8 x −αn ((−∞, −y))dy

Let f (x) be the value of the right-hand side of above inequality for x ∈ (0, 1). Then, as x → 0, then f (x) x −1−αp

if αn < 0 and f (x) x αn −αp −1

if αn > 0.

We conclude that the upper bound is ﬁnite if αp < 0 and αn < 0 and if αp < αn and hence therefore 0 is irregular for (0, ∞) in these cases. Because the lower bound on x 0 ((−∞, −y))dy has the same form as the upper bound, we can immediately conclude that in the other cases when αn = 0, 0 is regular for (0, ∞). Now only the case remains when αn = 0. Here we have x 0

cp x −1−αp e−λp x xcp x −1−αp e−λp x xν(x) = . ≤ b3 (x log(x) − x) b3 (log(x) − 1) ((−∞, −y))dy

Let g(x) be the value of the right-hand side of above inequality for x ∈ (0, 1). Then, when x → 0, g(x)

−1 . x 1+αp log(x)

Because for all t < 1, 0

t

−1 dx < ∞, log(x)

x 1+αp

if and only if, αp < 0, the upper bound is ﬁnite in this case and hence the integral (equation (1.14)) is ﬁnite. The lower bound has again the same form as the upper bound and so we conclude that this integral (equation (1.14)) is inﬁnite when αp ≥ 0 and αn = 0. 1.5.5 Generalized hyperbolic process Because the L´evy measure of this process is very complicated, it is very difﬁcult to use this measure to determine whether the process is of ﬁnite of inﬁnite variation. However, this can also be determined by using the characteristic exponent with the help of Lemma 9. We follow the ideas in given Cont and Tankov (2004). Proposition 20 A generalized hyperbolic process is of unbounded variation and has no Gaussian component and hence 0 is regular for the upper and lower half lines.

24

Exotic Option Pricing and Advanced L´evy Models

Proof. Using the properties of the logarithm we have λ λ (u) = − log(α 2 − β 2 ) + log(α 2 − (β + iu)2 ) 2 2 − log Kλ δ α 2 − (β + iu)2 + log Kλ δ α 2 − β 2 − iµu. Let r =

−2βu (α 2 − β 2 + u2 )2 + 4β 2 u2 and q = arctan α 2 −β 2 +u2 . Then √ 1 λ λ (u) = − log(α 2 − β 2 ) + (log(r) + iq) − log Kλ δ re 2 qi 2 2 √ 1 + log Kλ δ α 2 − β 2 − iµu. − log Kλ δ re 2 qi

When u → ∞, r ∼ u2 and q → 0. The modiﬁed Bessel function, Kλ , has the following property: if a → ∞ then Kλ (a + bi) ∼ e−(a+bi) √ 1 √ 1 qi So, Kλ δ re 2 qi ∼ e−δ re 2

π √ 1 qi 2δ re 2

π . 2(a + bi)

and therefore

√ 1 √ √ 1 1 1 qi log Kλ δ re 2 q + log(π ) − log 2δ r ∼ −δ r cos 2 2 2 and

√ 1 √ 1 1 ∼ −δ r sin log Kλ δ re 2 qi q − q 2 4 when u → ∞. So, ((u)) ∼ δu and we conclude from Lemma 9 that the process is of inﬁnite variation and has no Gaussian component. Proposition 21 A generalized hyperbolic process cannot hit points and hence cannot creep. Proof. We have seen from the proof above that ((u)) ∼ δu and that ((u)) ∼ 1 √ 1 δ r sin 2 q + µu, when u → ∞. This implies that 1+(u) u−1 when u → ∞ and therefore the process cannot hit points.

1.6 CONCLUSIONS Let us conclude with some tables with our ﬁndings for some of the more popular models we have mentioned. It will be useful to recall the notation C = {x ∈ R : P (Xt = x for at least one t > 0) > 0}.

L´evy Processes in Finance–Coarse and Fine Path Properties

25

Meixner processes

2δ cos(β/2) Meixner (u) = − log − iµu. cosh((αu − iβ)/2 Path variation: Hitting points: Creeping: Regularity:

Unbounded variation. C = ∅. No creeping upwards or downwards. Always regular for (0, ∞) and (−∞, 0). CGMY processes

CGMY (u) = −C (−Y )[(M − iu)Y − M Y + (G + iu)Y − GY )] − iuµ. Path variation: Hitting points: Creeping: Regularity:

Unbounded variation ⇔ Y ∈ [1, 2). C = ∅ ⇔ Y = 1 or Y ∈ (0, 1) and µ = 0, otherwise C = R. Upwards creeping ⇔ Y ∈ (0, 1) and µ > 0. Downwards creeping ⇔ Y ∈ (0, 1) and µ < 0. Irregular for (0, ∞) ⇔ Y ∈ (0, 1) and µ < 0. Irregular for (−∞, 0) ⇔ Y ∈ (0, 1) and µ > 0. Variance gamma processes

iu iu + log 1 + − iuµ. VG (u) = C log 1 − M G

Path variation: Hitting points: Creeping: Regularity:

Bounded variation C = ∅ ⇔ µ = 0, otherwise C = R. Upwards creeping ⇔ µ > 0, Downwards creeping ⇔ µ < 0 Regular for (0, ∞) ⇔ µ ≥ 0. Regular for (−∞, 0) ⇔ µ ≤ 0. Generalized hyperbolic processes

GH (u) = − log Path variation: Hitting points: Creeping: Regularity:

α2 − β 2 α 2 − (β + iu)2

λ/2

Kλ (δ α 2 − (β + iu)2 ) − iµu. Kλ (δ α 2 − β 2 )

Unbounded variation. C=∅ No creeping upwards or downwards. Always regular for (0, ∞) and (−∞, 0).

ACKNOWLEDGEMENTS We would like to express our thanks to Antonis Papapantoleon and a referee for careful reading of this manuscript.

26

Exotic Option Pricing and Advanced L´evy Models

REFERENCES [1] Alili, L. and Kyprianou, A. (2005), “Some remarks on ﬁrst passage of L´evy process, the American put and pasting principles”, Annals of Applied Probability, to be published. [2] Almendral, A. (2005), “Numerical valuation of American options under the CGMY process”, in A. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 259–276. [3] Applebaum, D. (2004), L´evy Processes and Stochastic Calculus, Cambridge Studies in Advanced Mathematics, Vol. 93, Cambridge University Press, Cambridge, UK. [4] Asmussen, S., Avram, F. and Pistorius, M.R. (2004), “Russian and American put options under exponential phase-type L´evy models”, Stochastic Processes and their Applications, 109, 79–111. [5] Avram F., Chan T. and Usabel M. (2002), “On the valuation of constant barrier options under spectrally one-sided exponential L´evy models and Carr’s approximation for American puts”, Stochastic Processes and their Applications, 100, 75–107. [6] Avram, F., Kyprianou, A.E. and Pistorius, M.R. (2004), “Exit problems for spectrally negative L´evy processes and applications to Russian, American and Canadized options”, Annals of Applied Probability, 14, 215–238. [7] Barndorff-Nielsen, O.E. (1995), “Processes of normal inverse Gaussian type”, Finance and Stochastics, 2, 41–68. [8] Barndorff-Nielsen, O.E. (1998), “Normal inverse Gaussian distributions and the modeling of stock returns”, Research Report No. 300, Department of Theoretical Statistics, Aarhus University, Denmark. [9] Barndorff-Nielsen, O.E. and Shephard, N. (2001), “Non-Gaussian Ornstein–Ulenbeck based models and some of their uses in ﬁnancial econometrics”, Journal of the Royal Statistical Society, B, 63, 167–241. [10] Bertoin, J. (1996), L´evy Processes, Cambridge University Press, Cambridge Tracts in Mathematics, Vol. 121, Cambridge, UK. [11] Bertoin, J. (1997), “Regularity of the half-line for L´evy processes”, Bulletin des Sciences Math´ematiques, 121, 345–354. [12] Boyarchenko, S.I. and Levendorskii, S.Z. (2002), Non-Gaussian Merton–Black–Scholes theory, Advanced Series on Statistical Science and Applied Probability, Vol. 9, World Scientiﬁc, River Edge, NJ, USA. [13] Boyarchenko, S.I and Levendorskii, S.Z. (2002a), “Perpetual American options under L´evy processes”, SIAM Journal on Control and Optimization, 40, 1663–1696. [14] Bretagnolle, J. (1971), “R´esultats de Kesten sur les processus a` accroissements ind´ependants”, in S´eminaire de Probabilit´es V, (Universitaire Strasbourg, Ann´ee Universitaire 1969–1970 ), Lecture Notes in Mathematics, Vol. 191, Springer-Verlag, Berlin, Germany, pp. 21–36 (in French). [15] Carr, P. and Wu, L. (2003), “The ﬁnite moment log stable process and option pricing”, Journal of Finance, 58, 753–778. [16] Carr, P., Geman, H., Madan, D.B. and Yor, M. (2002), “The ﬁne structure of asset returns: an empirical investigation”, Journal of Business, 75, 305–332. [17] Carr, P., Geman, H., Madan, D.B. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [18] Cartea, A. and Howison, A. (2005), “Option pricing with L´evy-stable process”, to be published. [19] Chan, T. (2000), “American options driven spectrally by one sided L´evy processes”, Unpublished manuscript (see also Chapter 9 in this volume). [20] Chan, T. (2004), “Some applications of L´evy processes in insurance and ﬁnance”, Finance, 25, 71–94.

L´evy Processes in Finance–Coarse and Fine Path Properties

27

[21] Cont, R. (2001), “Empirical properties of asset returns: stylized facts and statistical issues”, Quantitive Finance, 1, 223–236. [22] Cont, R. and Tankov, P. (2004), Financial Modeling with Jump Processes, Chapman and Hall/CRC Financial Mathematics Series, Boca Raton, FL, USA. [23] Eberlein, E. and van Hammerstein, E.A. (2004), “Generalized hyperbolic and inverse Gaussian distributions: limiting cases and approximation of processes”, Seminar on Stochastic Analysis, Random Fields and Applications IV, Progress in Probability, Vol. 58, Birkh¨auser Verlag, Basel, Switzerland, pp. 221–264. [24] Eberlein, E. and Prause, K. (2002), “The generalized hyperbolic model: ﬁnancial derivatives and risk measures”, in H. Geman, D. Madan, S. Pliska and T. Vorst (Eds), Mathematical FinanceBachelier Congress 2000, Springer-Verlag, Berlin, Germany, pp. 245–267. [25] Grigelionis, B. (1999), “Processes of Meixner type”, Lithuanian Mathematics Journal, 39, 33–41. [26] Halgreen, C. (1979), “Self-decomposability of the generalized inverse Gaussian and hyperbolic distributions”, Zeitschrift f¨ur Wahrscheinlichkeitstheorie und Verwandte Gebiete, 47, 13–17. [27] Kesten, H. (1969), “Hitting probabilities of single points for processes with stationary independent increments”, in Memoirs of the American Mathematical Society, Vol. 93, American Mathematical Society, Providence, RI, USA, pp. 1–129. [28] Kifer, Y. (2000), “Game options”, Finance and Stochastics, 4, 443–463. [29] Koponen, I. (1995), “Analytic approach to the problem of convergence of truncated L´evy ﬂights towards the Gaussian stochastic process”, Physical Review, E, 52, 1197–1199. [30] Kou, S.G. (2002), “A jump-diffusion model for option pricing”, Management Science, 48, 1086–1101. [31] K¨uhn, C. and Kyprianou, A.E. (2003), “Callable puts as composite exotic options”, to be published. [32] K¨uhn, C. and Kallsen, J. (2005), “Convertible bonds: ﬁnancial derivatives of game type”, in A. Kyprianou, W. Schoutens, and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK. [33] Kyprianou, A.E. (2004), “Some calculations for Israeli options”, Finance and Stochastics, 8, 73–86. [34] Madan, D.P. and Seneta, E. (1987), “The VG for share market returns”, Journal of Business, 63, 511–524. [35] Matache, A.M. Nitsche, P.A. and Schwab, C. (2003), “Wavelet Galerkin pricing of American options on L´evy driven assets”, to be published. [36] Merton, R. (1976), “Option pricing when underlying stock returns are discontinuous”, Journal of Financial Economics, 3, 125–144. [37] Miller, P.W. (1973), “Exit properties of stochastic processes with stationary independent increments”, Transactions of the American Mathematical Society, 178, 459–479. [38] Mordecki, E. (1999), “Optimal stopping for a diffusion with jumps”, Finance and Stochastics, 3, 227–236. [39] Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493. [40] Pitman, J. and Yor, M. (2003), “Inﬁnitely divisible laws associated with hyperbolic functions”, Canadian Journal of Mathematics, 55, 292–330. [41] Rogers, L.C.G. (1984), “A new identity for real L´evy processes”, Annals of the Institute of Henri Poincar´e, 20, 21–34. [42] Rogozin, B.A. (1968), “The local behavior of processes with independent increments”, Teoriga Veroyatnoste¨ui i ee Primeneniya (Theory of Probability and its Applications), 13, 507–512 (in Russian).

28

Exotic Option Pricing and Advanced L´evy Models

[43] Sato, K.-I. (1999), L´evy Processes and Inﬁnitely Divisible Distributions (translated from the 1990 Japanese original, revised by the author), Cambridge Studies in Advanced Mathematics, Vol. 68, Cambridge University Press, Cambridge, UK. [44] Schoutens, W. (2000), Stochastic Processes and Orthogonal Polynomials, Lecture Notes in Statistics, Vol. 146, Springer-Verlag, New York, NY, USA. [45] Schoutens, W. (2001), “The Meixner process in ﬁnance”, EURANDOM Report 2001–2002, EURANDOM, Eindhoven, The Netherlands. [46] Schoutens, W. (2002), “Meixner processes: theory and applications in ﬁnance”, EURANDOM report 2002–2004, EURANDOM, Eindhoven, The Netherlands. [47] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [48] Schoutens, W. and Teugels, J.L. (1998), “L´evy processes, polynomials and martingales”, Communications in Statistics: Stochastic Models, 14, 335–349. [49] Shtatland, E.S. (1965), “On local properties of processes with independent increments”, Teoriga Veroyatnoste¨ui i ee Primeneniya (Theory of Probability and its Applications), 10, 317–322 (in Russian). [50] Vigon, V. (2002), “Votre L´evy rampe-t-il? (Does your Le´vy process creep?)”, Journal of the London Mathematical Society, 2, 65(1), (2), 243–256 (in French).

2 Simulation Methods with L´evy Processes Nick Webber University of Warwick, Coventry, UK Abstract ´ processes are increasingly important tools for modelling asset return’s processes and Levy ´ interest rates. Although when Levy processes are used, direct integration methods are sometimes available to price standard European options, other numerical techniques must generally be employed to price instruments whose pay-offs are either path-dependent or American. This article discusses recent progress made in developing simulation methods suit´ able for some of the most widely used Levy processes. Bridge algorithms are given for the VG and NIG processes and these algorithms are applied to valuing average rate options. We consider the valuation of barrier options. It is shown how simulation bias can be reduced in this case. Once bias is absent, speed-up methods can be applied. Results are presented illustrating the bias reduction achieved for up-and-out and up-and-in barrier options.

2.1

INTRODUCTION

An important objective in ﬁnancial mathematics is to ﬁnd models of asset returns, interest rates and other ﬁnancial processes in order to value and hedge derivative securities, and for VaR and risk management purposes. Once a model has been found it is hoped that the relevant values, hedge ratios, reserves and risk factors can be computed. The standard Black–Scholes assumption, used in many applications and not just for simple options pricing, is that asset returns are normally distributed, and that joint returns distributions are normal. Alas, this assumption is false. Marginal distributions are not normal and joint distributions are not jointly normal (so that, in particular, the joint distribution does not have a Gaussian copula). The side-effects of assuming joint normality are unfortunately not ignorable. Market option implied volatilities have distinct smiles and historical returns distributions are fat tailed and skewed – unlike those predicted by a joint normality assumption. To overcome this problem, a number of different approaches are possible, some of which are discussed below. This article focuses on modelling univariate returns as L´evy processes and the application of simulation methods for option valuation. We are particularly concerned with ﬁnding effective numerical methods that can be used in practice to facilitate the calculation of option values when asset returns processes are L´evy.

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

30

Exotic Option Pricing and Advanced L´evy Models

The next section brieﬂy discusses modelling returns as L´evy processes. In Section 2.3, we discuss an approach towards ﬁnding numerical methods based on the subordinator decomposition of a L´evy process. Sections 2.4 and 2.5 present theory and results related to fast simulation methods, while Section 2.6 applies the methods, together with bias reduction methods, to continuously reset barrier options. The ﬁnal section provides a summary of our conclusions.

2.2

MODELLING PRICE AND RATE MOVEMENTS

Write St for the value at time t ≥ 0 of an asset value, conditional upon S0 . Different approaches to modelling asset price movements are possible, a number of which have been explored in the literature. 1. A standard modelling technique in mathematical ﬁnance is to model S = (St )t≥0 by specifying the SDE it satisﬁes. For example, by specifying SDEs for the asset process and for a stochastic volatility (for instance, Heston (1993) [15]). 2. By specifying the conditional distributions of S. For instance, by (a) giving the conditional distributions F (St | S0 ) themselves, or by (b) giving the densities f (St | S0 ), if they exist (for instance option pricing by log-normal mixtures (see Brigo and Mercurio (2001) [6])). (c) Giving the inverse distribution functions, F −1 (St | S0 ) (for instance, Corrado (2001) [9]). 3. As a time-changed Brownian motion (for instance, Geman et al. (2001) [13]). 4. As a L´evy process determined by its L´evy triple, (a, σ, υ). The process might have a L´evy density k (x), where υ (dx) = k (x) dx. (In practical applications one might approximate the L´evy process as a compound Poisson process.) Some of the many contributions here are cited in the next section. 5. By its time copula (see Bouy´e et al. (2000) [4]). Which ever way one models S, how might one calibrate to prices? It may be possible to use an implied pricing method to ﬁt exactly to an implied volatility surface. More usually, a functional form is speciﬁed, either explicitly or implicitly, and parameter values in the functional form chosen to ﬁt as closely as possible, by some criterion, to prices. We choose here to assume that a L´evy Process is given, whose triple (a, σ, υ) is parameterized, and whose parameter values can be chosen to match to observed prices. 2.2.1 Modelling with L´evy processes Consider an asset price process S = (St )t≥0 . Under the pricing measure with respect to the accumulator account numeraire suppose that St = S0 exp (rt + Lt − ωt)

(2.1)

short rate and ω = ln E exp (L1 ) where L = (Lt )t≥0 is a L´evy process, r is the constant compensates for the drift in L, so that St e−rt t≥0 is a martingale. Equation (2.1) is a

Simulation Methods with L´evy Processes

31

standard assumption in the literature (Madan et al. (1998) [19], Barndorff-Nielsen and Shephard (2000) [2], Eberlein and Keller (1995) [11], etc). Much work has been carried out, for various choices of L, to price options and to calibrate to empirical distributions. Important contributions include Carr et al. (2001) [7], Eberlein (2000) [10], Barndorff-Nielsen (2000) [2] and Rydberg (1999) [25]. There are also a number of interest rate models powered by L´evy processes in the literature (for instance, Eberlein and Raible (1999) [12]). One could consider a model for the short rate r = (rt )t≥0 in which under the pricing measure drt = α (µ (t) − rt ) dt + σ dLt ,

(2.2)

but we do not pursue interest rate models any further here. The problem with equations (2.1) and (2.2) is that, in general, it is hard to price nonvanilla derivatives in models incorporating these processes. For instance, how would one value path-dependent options or Bermudan or American puts? Even for vanilla European options, ﬁnding a price may require a difﬁcult numerical integration of a density function approaching the pathological. In a ‘Black–Scholes world’, where L is a multiple of a Wiener process, even if analytical solutions are unavailable, PDE, Monte Carlo integration and lattice methods can generally be found that provide adequate numerical solutions to many pricing problems. Lattice methods are cheap, ﬂexible and accurate, and can price American and Bermudan options. Monte Carlo methods can often be a long step and very good speed-up methods are available, so that path-dependent options are often quick and easy to price. In a L´evy world, the behaviour of the L´evy measure over short time horizons can cause immense practical problems. Analytic solutions may involve difﬁcult numerical integration (Madan et al. (1998) [19]). Fourier transform methods can work well for European options as long as the time horizon is not too short (Carr and Madan (1999) [8]). Monte Carlo methods can be used, either directly on the asset price process (equation (2.2) or indirectly through a representation of L as a subordinated Brownian motion (Rydberg (1997) [24]), a mean-variance mixture, or though an approximation as a compound Poisson process. However, sampling over a small time step is very hard if a L´evy density becomes unbounded near zero. PDE methods, such as the method of lines, can work for certain processes, but for a general process can prove very hard to use. Lattice methods need very high order branching and again a time step that is not too small. The problem is that on the one hand there are too many small jumps, and on the other there are too many big jumps. 2.2.2 Lattice methods Given the caveats noted above, can a lattice method work at all? The answer turns out to be yes, and hinges on one of the deﬁning properties of L´evy processes: convergence in probability. A process X = (Xt )t≥0 converges in probability if ∀ε > 0, Pr [|Xt − Xs | > ε] → 0 as s → t.

(2.3)

32

Exotic Option Pricing and Advanced L´evy Models

For a process which does not converge in probability, a lattice method might indeed be hard to construct. Consider the following example. Let t ∈ R+ . Deﬁne (Xt )t≥0 as if t ∈ Q then if t ∈ R\Q then

Pr [Xt = 0] = 12 , Pr [Xt = 2] = 12 , Pr [Xt = 0] = 12 ,

(2.4)

Pr [Xt = 1] = 12 .

This process clearly does not converge in probability. A computer only sees (essentially) rational values of t, a set of measure zero and a numerical method would not see the process taking values 1, even though this possibility occurs on a set of measure 1. If a process does converge in probability then, by deﬁnition, as t → 0, the probability of branching further away from a node at Xt than a ﬁxed distance X goes to zero, so that lattice methods cannot immediately be ruled out. Kellezi and Webber (2003) [17] devised a lattice method and applied it to VG and NIG processes. They obtained discrete branching probabilities from several alternative representations of the L´evy processes L. 1. Directly from the density function of Lt . Essentially, this is equivalent to ﬁtting to the characteristic function of the L´evy process. 2. From a representation as a subordinated Brownian motion, when the subordinator representation is known. 3. From the representation as a L´evy triple (a, σ, υ), via an approximation as a compound Poisson process. 4. From the moments of the process. The last possibility can be rejected very rapidly. Even if the moments of the process are known, it is still possible to match only a ﬁnite number of them. In any case, moment matching is equivalent to ﬁtting the characteristic function only at zero. Kellezi and Webber (2004) [17] found it preferable to construct a lattice directly from the density function (known for the examples they give). Nevertheless, their lattice has very high order branching, is relatively slow, still runs into problems when attempting to price American options, and in any case cannot value path-dependent options. Instead of investigating lattice methods any further, this article now turns its attention to simulation methods. Although these may not be usable with Bermudan or American options (although perhaps primal-dual methods could still work), they can value path-dependent options. 2.2.3 Simulation methods It is often not possible to directly and accurately simulate a L´evy process; it may not be possible to sample directly from the distribution of Lt . When it is not possible to simulate directly from some distribution, either a terminal distribution in a long-step Monte Carlo method, or a distribution, exact or otherwise, representing a small time step t, several alternatives are possible. For instance: 1. Represent the distribution as a mean-variance mixture.

Simulation Methods with L´evy Processes

33

2. Express an underlying process as a time-changed Brownian motion.1 3. In the worst case, approximate the L´evy measure as compound Poisson. Given two densities, f (x | α) and g(α), where α ∈ B and g is a density on B, their mixture distribution has the density f (x | y) g (y) dy.

fg (x) =

(2.5)

B

When f depends upon a single parameter α via its mean and variance, fg (x) is a meanvariance mixture. Given such a representation, it may be possible to (i) draw from g to get a y, and then (ii) draw from f (x | y) for x. If a time-change representation exists, so that Xt = wh(t) for a Brownian motion w and a time change h (see below), then it may be possible to (i) sample from h (t) to get a random time τ , and then (ii) sample from wτ . The time-change representation is used in the rest of this article.

2.3

A BASIS FOR A NUMERICAL APPROACH

We exploit the time-change representation of a L´evy process. Let X = (Xt )t≥0 be a onedimensional semi-martingale, and then (Monroe (1978) [20]) X is representable as a timechanged Brownian motion, Xt = wh(t) ,

(2.6)

where w is a Brownian motion, with drift µ and volatility σ , say, and h = (ht )t≥0 is a stochastic time change. A L´evy process L is a semi-martingale and so this representation can be used. In general, h and w need not be independent. However, we assume that they are independent; this assumption is valid for all the processes we consider below. We consider only the case when the time-change h is an increasing L´evy process. Then, X will also be a L´evy process. Since wt = µt + σ zt , for a Wiener process (zt )t≥0 , we can write Lt = µht + σ zh(t) . Both the variance gamma (VG) and the normal inverse Gaussian (NIG) processes have time changes h, whose conditional distributions are taken from a set of generalized inverse Gaussian (GIG) distributions. Write δt = δt. If ht ∼ GIG(δt , λ, γ ) has a GIG distribution, then the density ftGIG of ht is ftGIG (h; δt , λ, γ ) =

λ

1 γ 1 δt2 hλ−1 exp − + γ 2h δt 2Kλ (δt γ ) 2 h

The set of GIG distributions is not closed under convolution. 1

A time-change representation may yield a mean-variance representation.

(2.7)

34

Exotic Option Pricing and Advanced L´evy Models

The VG process. The gamma distribution is the limit of the GIG distribution as δ → 0 and λ → t/v. If ht ∼ (t, tv) is a gamma variate, its density ft conditional on h0 = 0 is ft

t x ν −1 exp − xv (x) = . t ν ν νt

(2.8)

If L is VG we have Lt ≡ LVG (t | σ, v, µ) = µht + σ zh(t)

(2.9)

where ht has the density given by equation (2.8). For the VG process, the compensator ω in equation (2.1) is

1 1 ω = − ln 1 − µν − σ 2 ν . (2.10) ν 2 The NIG process. The inverse Gaussian process is GIG with λ = − 12 . If ht ∼ IG (δt , γ ) is an inverse Gaussian process, then the density ftIG (x) of ht is

2 2 γ δ 3 1 δ t t x− ftIG (x) = √ x − 2 exp − . (2.11) 2 x γ 2π An NIG process L can be written Lt ≡ LNIG (t | µ, θ, δ, γ ) = µt + βht + zh(t)

(2.12)

where ht has the density given by equation (2.11). For the NIG case, the compensator ω is

(2.13) ω = µ + δ γ − α 2 − (β + 1)2 where α 2 = γ 2 + β 2 . 2.3.1 The subordinator approach to simulation Suppose we have a European derivative, with payoff HT at time T and value ct at times t ≤ T , so that cT = HT . In the martingale valuation framework, we have cT ] , ct = Et [

(2.14)

where cT = cT ppTt for a numeraire pt , with expectations Et [•] ≡ E [• | Ft ] taken with respect to the martingale measure associated with pt . Suppose that HT and pt depend on a single-state variable St , which in turn depends upon a L´evy process Lt . In particular, suppose that, as in equation (2.1), under the pricing measure St = S0 exp (rt + Lt − ωt) where r is a constant short rate.

(2.15)

Simulation Methods with L´evy Processes

35

Now, suppose that L has a subordinator representation Lt = wh(t) . Since wt and h (t) are independent, the ﬁltration Ft decomposes as Ft = Ftw × Fth and we can iterate the expectation: cT | Ft ] ct = E [ cT | Ftw × Fth =E =E E cT | Ftw × FTh | Ftw × Fth .

(2.16)

ct = Et [E [ cT | h]]

(2.19)

(2.17) (2.18)

Informally, we can write

where E [ cT | h] represents the expected value of cT , conditional upon knowing the path of h up to time T . The inner and outer expectations can be simulated separately and so a possible procedure to value ct by simulating Lt is to: 1. 2. 3. 4.

Simulate a path {hti }i=1,... ,N of h up to time T . Given {hti }, generate a path for w at times {hti } and set Lti = wh(ti ) . Compute cT from the path of L. Repeat 10 000 times (say) and average.

This procedure can be used to value not just vanilla European options but also options that payoff at a directly determined stopping time, such as barrier options.2 2.3.2 Applying the subordinator approach There are two expectations from equation (2.19). Each could, in principle, be computed by using either a lattice valuation method or a Monte Carlo valuation method.3 Each possible pairing of valuation methods can be assessed for its appropriateness or inappropriateness for valuing path-dependent options (P or NP), for valuing options such as American or Bermudan type which may be exercised early (A or NA), and for its ease of calibration (C or NC). We obtain Table 2.1. Using a Monte Carlo method for the inner expectation and a lattice method for the outer expectation results in the random lattice method (Kuan and Webber (2003) [18]). Table 2.1

Valuation methods

Method

Inner

2 3

MC Lattice

Outer MC

Lattice

P, NA, NC NP, ∼A, C

NP, A, NC NP, A, C

However, not options whose stopping times are determined by optimality conditions, such as American options. It is not immediately clear how a PDE method might be used.

36

Exotic Option Pricing and Advanced L´evy Models

Rydberg (1997) [24] effectively used a Monte Carlo method for both the inner and outer expectation to simulate the NIG process. The procedure was later developed by Ribeiro and Webber (2002, 2003a, b) [21]–[23] who showed how to apply effective speed-up methods and bias-reduction methods. We expound their approach in the remainder of this article.

2.4

´ CONSTRUCTING BRIDGES FOR LEVY PROCESSES

Plain Monte Carlo methods are very slow to achieve acceptable accuracy.4 Various speedup methods need to be employed to produce reasonable computation times. One efﬁcient method is path regularization by stratiﬁed sampling. This ensures that sample points form a less clustered draw from the sample space than a plain Monte Carlo method would produce. If an entire sample path is needed, rather than just a draw from the terminal time, then stratiﬁed sampling has to be used in conjunction with a bridge method. 2.4.1 Stratiﬁed sampling and bridge methods We discuss here the background to stratiﬁed sampling and bridge methods. Good reviews can be found in J¨ackel (2002) [16] and Glasserman (2003) [14]. Stratiﬁed sampling. It is easy to generate a stratiﬁed sample from the unit interval [0, 1]. Let vi ∼ U [0, 1], i = 1, . . . , Q, be a sample of size Q from the uniform distribution U [0, 1], and then ui = i+vQi −1 is a stratiﬁed sample of size Q from U [0, 1]. The set {ui } is guaranteed to have minimal clustering above the scale 1/Q. Stratiﬁed samples can be produced from other distributions by inverse transform. Suppose that X ∼ FX is a random variate with distribution function FX and that FX−1 is computable. Let u ∼ U [0, 1] be a uniform variate, and then FX−1 (u) ∼ FX has distribution FX . Given a stratiﬁed sample ui , i = 1, . . . , Q from U [0, 1], the set FX−1 (ui ), i = 1, . . . , Q, is a stratiﬁed sample from FX . Bridge sampling. Given a L´evy process L, where Lt has distribution Ft at time t (conditional on L0 ), suppose that we have found a sample, Li,N , i = 1, . . . , Q, of LtN from FtN , possibly stratiﬁed. Given a value for L0 at time 0 = t0 , we would like to construct an entire sample path L0 = Li,0 , . . . , Li,N with the correct conditional distributions. This means being able to sample Li,j , at time 0 < tj < tN , conditional upon the values of Li,0 and Li,N . In general, suppose that X ∼ FX , Y ∼ FY and Z ∼ FZ , with densities fX , fY and fZ , respectively, are random variates such that Z = Y + X. For instance, X, Y and Z could be increments in a L´evy process L. Given a draw z of Z, we want to sample from the conditional distribution X | Z. Write fX,Y (x, y) for the joint density of X and Y . Then fX|Z (x) = = 4

fX,Y (x, z − x) , fZ (z)

(2.20)

fX (x) fY (z − x) when X, Y are independent. fZ (z)

(2.21)

Measured by the square root of the second moment of the Monte Carlo estimate. An internally generated estimate of this is the Monte Carlo standard error.

Simulation Methods with L´evy Processes

37

For a L´evy process L, take X ∼ Fti , Y ∼ Ftj , Z ∼ Fti +tj , say, and then fti |ti +tj (x) ≡ fX|Z (x) is the bridge density of L. If the densities ft are known, then the bridge density of L can be computed and perhaps a sampling method can be found. 2.4.2 Bridge sampling and the subordinator representation When Lt = wh(t) has a subordinator representation we could construct a stratiﬁed bridge sample of L by ﬁrst obtaining a stratiﬁed bridge sample of h, and then constructing a stratiﬁed bridge sample of w at the times given by our path for h. The bridge distribution of a Brownian motion is well known and it is easy to sample from it. We need only to know the bridge distribution of the subordinator h, and to be able to sample from it. Ribeiro and Webber (2002, 2003b) [21], [23] ﬁnd the bridge density and a sampling method for the bridge density, for both the gamma and the inverse Gaussian processes. We summarize their results here. Let 0 = h0 < · · · < hN be a series of values at increasing times for the subordinator process h. Given hi and hk at times ti < tk , we want to sample hj at an intermediate time ti < tj < tk . Write z = hk − hi , x = hj − hi , and y = hk − hj , and set τz = tk − ti , τx = tj − ti , and τy = tk − tj . The gamma process. Given hi and hk , the bridge density of x from a gamma process with parameter ν is τ

τy τy τx −1 x ν x ν −1 1 νx + ν 1− fX|Z (x) = . z τνx τνy z z

(2.22)

An algorithm to sample from this density is 1. Generate bj ∼ B(τx /ν, τy /ν), where B is the beta distribution. 2. Set hj = hi + bj (hk − hi ). Existing algorithms to sample (by inverse transform) from the beta distribution are relatively slow. Even so, the numerical results in Section 2.5 below demonstrate the very great speed-ups possible with this sampling method. The inverse Gaussian process. The bridge density of x from an inverse Gaussian process with parameter δ is δ τx τy fX|Z (x) = √ 2π τz

xy z

− 3 2

1 exp − δ 2 2

where y = z − x. An algorithm to sample from this density is 1. Generate q ∼ χ12 .

τy2 τ2 τx2 + − z x y z

,

(2.23)

38

Exotic Option Pricing and Advanced L´evy Models

2. Set λ =

δ 2 τx2 x

and µ = τy /τx , and compute roots s1 and s2 , s1 = µ + s2 =

µ µ2 q − 4µλq + µ2 q 2 , 2λ 2λ

µ2 . s1

(2.24) (2.25)

3. Set s = s1 with probability p, p= or else set s = s2 . 4. Finally, set hj = hi +

1 1+s

µ (1 + s1 ) , (1 + µ) (µ + s1 )

(2.26)

(hk − hi ).

Since sampling from a χ12 distribution, by inverse transform or otherwise, is very fast, sampling from the inverse Gaussian distribution is also very quick. Using the bridge. Given a method for sampling from the subordinator process ht , we adopt the following algorithm. 1. 2. 3. 4.

Given h0 = 0, construct a stratiﬁed sample hi,N of hN . Using the bridge distribution, construct hi,N/2 conditional on h0 and hi,N . Using binary chop, continue to generate a path hi,j for times tj , j = 0, . . . , N . Generate whi,j , and hence Si,j conditional on hi,j .

At each intermediate time it is possible to continue to stratify the sample path. Generating a sample path requires a sequence of uniform variates from which samples with the desired distributions are obtained (by inverse transform). If n uniform variates are needed, this is equivalent to a draw from a unit hypercube of dimension n. To stratify m ≤ n of these draws, one performs a stratiﬁed sample on an m-dimensional hypercube, selecting the remaining n − m draws without stratiﬁcation. In practice, it is only possible to make a fully stratiﬁed draw from a hypercube of dimension 3 or so. To make a stratiﬁed draw from a hypercube of dimension m > 3, low discrepancy sampling is often used. For the VG process it takes one uniform variate to generate (by inverse transform) a draw from the gamma bridge distribution for hi and one more for each normal variate whi ; for the NIG process, it takes two uniform variates to draw for each hi and one more for whi . So, the VG process requires two uniform variates at each time step, and the NIG process requires three. Even if sampling with low discrepancy sequences it may only be possible in practice to sample reliably from a unit hypercube of dimension at most a few dozen. A freely available downloadable code for generating Sobol sequence numbers5 goes up to dimension 39. This means that (with binary chop) for the VG process it is possible to stratify at up to 16 times and for the NIG process one can stratify at up to 8 times. Draws for other times have to be made with ordinary non-stratiﬁed sampling. Even with this restriction, very good speed-ups are possible. 5

See Bratley and Bennett (1988) [5]. Code is downloadable from www.netlib.org/toms/659.

Simulation Methods with L´evy Processes

2.5

39

VALUING DISCRETELY RESET PATH-DEPENDENT OPTIONS

A discretely reset path-dependent option is one in which the option payoff is computed from observations of the underlying asset value at certain discrete times (the reset times). In this section, we value discretely reset path-dependent options when returns to the underlying asset are either VG or NIG processes. We present convergence results (taken from Ribeiro and Webber (2002, 2003b) [21] [23]) and show how true standard errors are improved relative to plain Monte Carlo. When a Monte Carlo method is used with stratiﬁed sampling, successive Monte Carlo estimates are correlated with one another. This means that the internally generated standard error measure is not a true reﬂection of the standard deviation of the Monte Carlo estimate. In the tables below, the true standard deviation is estimated by taking the sample standard deviation of the Monte Carlo estimates obtained from 100 replications of the Monte Carlo procedure. To compare two Monte Carlo methods, we use the efﬁciency gain of one method over the other. Suppose we have two Monte Carlo procedures. Monte Carlo method i, i = 1, 2, gives an estimate with standard deviation σi ina time ti . If the time taken t is proportional to the number of sample paths Q, and if σ is O − 12 in Q, then the efﬁciency gain E1,2 , deﬁned as E1,2 =

t2 σ22 , t1 σ12

(2.27)

is how many times faster method 1 is to achieve the same standard deviation as method 2. In the following tables, K is the number of stratiﬁcation times. K = 0 is plain Monte Carlo with Q = 106 paths. Bridge Monte Carlo uses Q = 104 paths. The benchmark is full low discrepancy with Q = 106 . The initial asset value is S0 = 100, the short rate is r = 0.1 and the option strike is X = 100 with maturity time T = 1. The VG case uses parameter values σ = 0.12136, v = 0.3 and µ = −0.1436 (based upon Madan et al. (1998) [19]). The NIG case uses parameters α = 75.49, β = −4.089, δ = 3 and µ = 0 (based upon Rydberg (1997) [24]). Programmes were written in VBA 6.0 run on a 900 Mhz PC. Tables 2.2 and 2.3 give values, standard deviations and computation times (in seconds) for a discretely reset average rate option when the underlying asset has either VG or NIG returns process. More illuminating are Tables 2.4 and 2.5, which give the efﬁciency gains in each case. In the VG case, speed-ups of up to a factor of about 400 are possible and up to about 200 for the NIG case. Speed-ups tend to improve as both the number of reset dates and the number of stratiﬁcation times increase, although this is not so evident in this example for the NIG case. Here, we have given only a few of the results of Ribeiro and Webber. They investigate many more cases, including discrete barrier and lookback options, demonstrating in each case that worthwhile speed-ups are attainable. Further speed-ups would seem to be possible. In both the VG and NIG cases, the number of stratiﬁcation dates was constrained by the dimension of the low discrepancy sequence generated by the available software. There have also been criticisms of the quality of this generator (J¨ackel (2002) [16]). With a better quality generator, capable of producing low discrepancy sequences of higher dimension, increased speed-ups would be possible.

40

Exotic Option Pricing and Advanced L´evy Models

Table 2.2 Values, standard deviations and computation times for average rate call options: comparison of plain and bridge Monte Carlo methods for the VG case (Ribeiro and Webber (2003b) [23]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

0

6.7720 (0.0064) [85.6]

6.0666 (0.0058) [175.0]

5.7274 (0.0055) [335.4]

5.5497 (0.0053) [647.5]

5.4625 (0.0052) [1277]

5.4075 (0.0052) [5034]

1

6.7993 (0.029) [2.4]

6.0290 (0.025) [3.5]

5.7234 (0.023) [5.5]

5.5242 (0.025) [9.3]

5.4830 (0.029) [17.0]

5.3874 (0.024) [52.2]

2

6.7635 (0.011) [4.0]

6.0741 (0.012) [5.0]

5.7187 (0.014) [6.9]

5.5208 (0.014) [10.8]

5.4761 (0.013) [18.5]

5.376 (0.013) [54.9]

4

6.7594 (—) [13.0]

6.0656 (0.0065) [14.0]

5.7149 (0.0067) [16.0]

5.5510 (0.0072) [19.9]

5.4765 (0.0063) [27.6]

5.3934 (0.0064) [62.9]

8

—

6.0711 (—) [33.8]

5.7283 (0.0029) [35.8]

5.5465 (0.0033) [39.6]

5.4667 (0.0035) [47.4]

5.3997 (0.0035) [83.0]

16

—

—

5.5527 (0.0014) [80.7]

5.4627 (0.0017) [88.1]

5.4008 (0.0017) [123.1]

Benchmark

6.7626 (—) [1297]

6.0702 (—) [3370]

5.7245 (—) [77.3] 5.7250 (—) [7638]

—

—

—

Reproduced by permission of L.D. Donaldson.

In the VG case, the stratiﬁcation algorithm for the gamma bridge distribution is constrained by the method used to generate stratiﬁed beta variates by inverse transform.6 An improved algorithm for the inverse transform of the beta distribution would result in even greater speed-ups in this case.

2.6 VALUING CONTINUOUSLY RESET PATH-DEPENDENT OPTIONS We would like to obtain for continuously reset barrier and lookback options the speed-ups that were found for discretely reset barrier and lookback options by Ribeiro and Webber (2002, 2003b) [21][23]. These same authors (2003a) [22] ﬁnd an approximate method to achieve this result. Particular problems arise when applying numerical methods to continuous barrier options. It turns out that the values of discretely reset barrier options converge only very slowly to values of corresponding continuously reset barrier options as the number of reset dates increases. Since numerical methods are set in discrete time, the values they ﬁnd for continuously reset barrier options may converge only very slowly to the true value. 6

The algorithm uses a root searching method.

Simulation Methods with L´evy Processes

41

Table 2.3 Values, standard deviations and computation times for average rate call options: comparison of plain and bridge Monte Carlo methods for the NIG case (Ribeiro and Webber (2002) [21]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

0, Q = 106

8.5856 (0.0103) [71.6]

7.7892 (0.0094) [141.4]

7.4059 (0.0089) [278.6]

7.2312 (0.0087) [551.8]

7.1286 (0.0086) [1101.4]

7.0698 (0.0086) [4390.1]

1

8.6326 (0.044) [1.9]

7.8205 (0.042) [2.4]

7.3874 (0.048) [4.9]

7.2347 (0.041) [7.7]

7.0763 (0.047) [15.1]

7.0656 (0.044) [58.8]

2

8.5530 (0.021) [1.8]

7.7695 (0.022) [2.1]

7.4282 (0.022) [4.8]

7.2457 (0.026) [7.7]

7.0721 (0.026) [15.0]

7.0497 (0.021) [58.8]

4

8.5695 (—) [1.8]

7.7963 (0.010) [2.0]

7.4181 (0.011) [4.7]

7.2105 (0.011) [7.6]

7.1249 (0.011) [14.9]

7.0430 (0.012) [58.6]

8

—

7.7959 (—) [1.7]

7.4045 (0.0048) [4.3]

7.2121 (0.0059) [7.3]

7.1296 (0.0051) [14.6]

7.0519 (0.0059) [58.4]

Benchmark

8.5807 (—) [169]

7.8072 (—) [169]

—

—

—

—

Table 2.4 Efﬁciency gains for average rate call options: comparison of plain and bridge Monte Carlo methods for the VG case (Ribeiro and Webber (2003b) [23]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

1 2 4 8 16

1.7 7.2 — — —

2.7 8.2 10 — —

3.5 7.5 14 34 —

3.1 8.6 18 42 115

2.4 11 32 60 136

4.5 15 53 134 383

Reproduced by permission of L.D. Donaldson.

Table 2.5 Efﬁciency gains for average rate call options: comparison of plain and bridge Monte Carlo methods for the NIG case (Ribeiro and Webber (2002) [21]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

1 2 4 8

2.1 9.8 — —

3.0 12.4 58.8 —

2.0 9.9 42.3 219.4

3.2 8.2 42.8 167.1

2.4 8.2 45.7 216.2

2.9 12.2 40.1 157.0

42

Exotic Option Pricing and Advanced L´evy Models

For a Monte Carlo method, this feature is called ‘simulation bias’. A discrete sample path for an underlying asset may not exceed a barrier level, but on the continuous path sampled by the discrete path, the barrier may have been hit in between times observed by the discrete sample path. For lookback options, the maximum (minimum) found along a discrete path will always be less than (more than) the true maximum (minimum) achieved along the continuous path. When the underlying asset has a geometric Brownian motion, simulation bias correction methods are available (Beaglehole et al. (1997) [3] and El Babsiri and Noel (1998) [1]). In this section, following Ribeiro and Webber, we show how these ideas can be extended to asset processes driven by L´evy processes. Only when bias has been removed, or at least signiﬁcantly reduced, does it makes sense to apply speed-up methods such as the bridge methods discussed in previous sections. 2.6.1 Options on extreme values and simulation bias Consider a continuously reset option maturing at time T , and let B be a barrier level. Given a (continuous) sample path {St }t∈[0,T ] for at asset value, set M0,T = max {St }

(2.28)

m0,T = min {St } ,

(2.29)

t∈[0,T ] t∈[0,T ]

An up-and-in barrier option with barrier level B has payoff HT (ω) I{M0,T ≥B } , where HT (ω) is the payoff at time T of the knocked-in option. Similarly, an up-and-out barrier option has payoff HT (ω) I{M0,T t. Denote by pt,T (Ut , L) the statistical density of this event; the corresponding risk neutral density is qt,T (Ut , L). These two densities interact in the determination, for example, of the value at risk in a contingent claim with cash ﬂow cT (L) at time T . The change in value over an interval of length h, of this cash ﬂow on a marked to market basis is the random variable ∞ ∞ −r(T −t+h) −r(T −t) cT (L)qt+h,T (Ut+h , L) dL − e cT (L)qt,T (Ut , L) dL e 0

0

= (Ut , Ut+h ), where it is supposed that interest rates are constant at the continuously compounded rate of r. The risk in this position is assessed by the statistical density pt ,t+h (Ut , Ut+h ) and the value at risk at the 0.95 conﬁdence interval deﬁned as the corresponding quantile of the statistical distribution. We note importantly that the contingent claim example is in fact quite general. It is now recognized explicitly that even bonds are claims contingent on the absence of counterparty default while equity is itself an option in the presence of outstanding bonds, or even otherwise when we take particular note of limited liability. An equally important entity is the ratio of the two densities yt,T (Ut , L) =

qt,T (Ut , L) pt,T (Ut , L)

which is called the change of measure density (or the Radon–Nikodym derivative of the measure q with respect to the measure p). By incorporating the measure change into the valuations above, one may perform all expectations with respect to the statistical measure and write the change in value as ∞ e−r(T −t+h) cT (L)yt+h,T (Ut+h , L)pt+h,T (Ut+h , L) dL −e−r(T −t)

0 ∞

cT (L)yt,T (Ut , L)pt,T (Ut , L) dL 0

= (Ut , Ut+h ) It is clear from these expressions that an understanding of the measure change yt,T (Ut , L) makes important contributions to risk management and investment decisions. The difﬁculty

Risks in Returns: A Pure Jump Perspective

53

however, lies in making observations on the measure change. This is because, although one may extract qt,T from option prices using the methods of Breeden and Litzenberger (1978), this occurs at values of T reﬂecting traded option maturities T − t at time t; these are typically at intervals of a month. In contrast, the statistical density is best estimated at the horizon of daily returns. This time discrepancy is difﬁcult to overcome. There is little one may do about accessing risk neutral densities at maturities below the ﬁrst liquid traded maturity. On the other hand, one may be tempted to construct monthly returns out of daily returns assuming independence and stationarity; however, the considerable evidence in support of correlated squared returns makes these assumptions problematic. For other recent approaches in this direction, the reader is referred to Jackwerth (2000) and Bliss and Panigirtzoglou (2002). The approach we take here is to attempt to observe from options data and time series data the limiting densities as T approaches t. Furthermore, to recover the classical ﬁnancial setting focusing on log returns, we ﬁrst change variables to these magnitudes by making the transformation L = Ut el and subsume the dependence on the current observed level Ut into the subscript t. $ qt,T (l) ≡ qt,T (Ut , Ut el )Ut el $t,T (l) ≡ pt,T (Ut , Ut el )Ut el p These limiting densities may be constructed on normalization by (T − t) as follows $ qt,T (l) T −t $t,T (l) p kP (l) = lim T →t T − t

kQ (l) = lim

T →t

We note importantly that the division by (T − t) is necessary as the numerator in each case goes to zero for l = 0 and goes to inﬁnity for l = 0 (as the limiting measures are Dirac measures at l = 0). Another key and different observation is the fact that, unfortunately, for continuous processes both limits remain zero for l = 0. For discontinuous processes in contrast, the situation is different; in this wide collection, we choose for tractability the class of purely discontinuous L´evy processes. For these pure jump processes, the above limits are well deﬁned for all l = 0 and converge to the L´evy measures deﬁned by kQ (l), and kP (l), respectively. The statistical L´evy measure, kP (l), has the heuristic interpretation of the expected number of jumps of size l in log returns per unit time. Analogously, kQ (l) is the futures price of a contract paying at unit time the dollar number of jumps of size l that occur in this period. Apart from these horizon matching considerations, the use of L´evy processes in modeling asset returns, both statistically and risk neutrally, has a number of other well noted advantages. First, from the statistical perspective, it is well known that kurtosis levels in short period returns are substantially above 3, arguing for non-Gaussian distributions. L´evy processes easily accommodate a much richer structure of moments for short horizon returns, including negative skewness when needed. Risk neutrally, these processes easily capture short maturity skews that are prominent in options data. The transition from the statistical to the risk neutral probability is also less constrained, as, in principle, all moments may be altered, unlike the diffusion case where local volatilities must remain the same. For further details on the applications of L´evy processes in ﬁnance, we refer the reader to Schoutens (2003).

54

Exotic Option Pricing and Advanced L´evy Models

From the perspective of studying the ratio of the risk neutral to the statistical L´evy measure, it is useful to work with processes that have simple analytical forms for this entity and are capable of both providing a good ﬁt to the data and of synthesizing the high activity levels observed in the markets. This leads us to ‘inﬁnite activity’ L´evy processes (see Geman et al. (2001)) that have a sufﬁciently rich parametric structure to capture at least the ﬁrst four moments of the local motions. A particularly attractive example is provided by the CGMY model (Carr et al. (2002)) with further properties described in the next section. Other candidates include the Normal Inverse Gaussian model of Barndorff-Nielsen (1998), the Meixner process studied by Schoutens and Teugels (1998), and the generalized hyperbolic model (Barndorff-Nielsen (1977), Eberlein and Prause (1998) and Prause (1999)). The next section presents the details of the CGMY model employed in this current study. This section is followed by estimation details presented in Section 3.3, for both the statistical analysis and the inference of the risk neutral process. In comparing the two probabilities at the instantaneous level, we consider explicitly here the structure of returns on securities paying the market gap risk. These are securities that pay a dollar whenever there is a large up or down move of a prespeciﬁed size. Section 3.4 presents the results for ﬁve world equity indexes (USA, UK, Germany, Spain and Japan) showing that world-wide tail gap risk securities for both positive and negative moves are insurance-based with expected negative rates of return reﬂecting the presence of insurance premia, while the central part of the return distribution represents investment where positive rates of return reﬂect the expected risk compensation. We anticipate that market participants taking long positions protect themselves by buying downside gap risk claims and pay the requisite insurance charge for this service. On the other hand, participants short the market protect themselves by buying upside gap risk claims and pay the insurance charge on this side. The relative strength of the long side to the short side is then reﬂected in the larger premia for downside gap risk claims, as compared to the comparable upside gap risk claims. We see in the structure of the change of measure density the ways in which investment risk and insurance protection complement each other in the ﬁnancial markets of the world.

3.2

CGMY MODEL DETAILS

The general idea is to model the statistical and risk neutral log price relative over an interval, X(t + h) − X(t) = ln(S(t + h)/S(t)), as the increment of a purely discontinuous L´evy process. Such processes have independent and identically distributed increments over non-overlapping intervals of equal length with inﬁnitely divisible densities. They are characterized by the L´evy-Khintchine decomposition for their characteristic exponents, ψ(u) by E exp(iuX(t)) = exp (−tψ(u)) ∞ 1 + iux1|x|≤a − eiux k(x) dx ψ(u) = iγ u +

(3.1)

−∞

where γ is called the drift coefﬁcient and k(x) is the L´evy density that integrates x 2 in a neighborhood of 0. The processes may, in general, have inﬁnite variation in that the limiting sum of absolute changes in the log price over smaller and smaller time intervals tends to inﬁnity. In the special case of a ﬁnite limit, we have ﬁnite variation and the characteristic

Risks in Returns: A Pure Jump Perspective

55

exponent then has the representation ψ(u) = iγ u +

∞

−∞

1 − eiux k(x) dx.

In the ﬁnite variation case, the process for X(t) may be written as the difference of two increasing processes X(t) = Xp (t) − Xn (t) where the increasing processes Xp (t), Xn (t) have characteristic exponents, ψp (u), ψn (u)

∞

1 − eiux k(x) dx

ψp (u) = iγp u +

0 ∞

ψn (u) = iγn u +

1 − eiux k(−x) dx

0

γ = γp − γn For the inﬁnite variation case, we add to the difference of two increasing compound Poisson processes Xpa , Xna with characteristic exponents ψpa (u)

= iγp u +

λap

ψna (u) = iγn u + λan λap =

∞

a

λan =

−a

−∞

∞ a ∞

a

1 − eiux fpa (x) dx

1 − eiux fna (x) dx

k(x) dx ; fpa (x) =

k(x) , x>a λap

k(x) dx ; fna (x) =

k(−x) , x>a λan

γ = γp − γn the limit as ε tends to zero, of the compensated jump compound Poisson martingale Xε (t) X (t) = ε

Xs 1ε 0, is given by: φV G (u; C, G, M) =

GM GM + (M − G)iu + u2

C .

This distribution is inﬁnitely divisible and one can deﬁne the VG-process X(V G) = {Xt(V G) , t ≥ 0} as the process which starts at zero, has independent and stationary incre(V G) ments and where the increment Xs+t − Xs(V G) over the time interval [s, s + t] follows a VG(Ct, G, M) law. In Madan et al. [19], it was shown that the VG-process may also be expressed as the difference of two independent Gamma processes, which is helpful for simulation issues (see Section 4.4.2).

Model Risk for Exotic and Moment Derivatives

CIR Stochastic Clock. solves the SDE:

73

Carr et al. [8] use as the rate of time change the CIR process that

1/2

dyt = κ(η − yt ) dt + λyt

dWt ,

where W = {Wt , t ≥ 0} is a standard Brownian motion. The characteristic function of Yt (given y0 ) is explicitly known (see Cox et al. [12]): ϕCI R (u, t; κ, η, λ, y0 ) = E[exp(iuYt )|y0 ] =

exp(κ 2 ηt/λ2 ) exp(2y0 iu/(κ + γ coth(γ t/2))) (cosh(γ t/2) + κ sinh(γ t/2)/γ )2κη/λ2

,

where γ = Gamma-OU Stochastic Clock.

κ 2 − 2λ2 iu.

The rate of time change is now a solution of the SDE: dyt = −λyt dt + dzλt ,

(4.5)

where the process z = {zt , t ≥ 0} is as in equation (4.4) a compound Poisson process. In the Gamma-OU case, the characteristic function of Yt (given y0 ) can be given explicitly. ϕ −OU (u; t, λ, a, b, y0 ) = E[exp(iuYt )|y0 ]

b λa −1 −λt b log − iut . = exp iuy0 λ (1 − e ) + iu − λb b − iuλ−1 (1 − e−λt ) Time-Changed L´evy Process. Let Y = {Yt , t ≥ 0} be the process we choose to model our business time (remember that Y is the integrated process of y). Let us denote by ϕ(u; t, y0 ) the characteristic function of Yt given y0 . The (risk-neutral) price process S = {St , t ≥ 0} is now modelled as follows: St = S0

exp((r − q)t) exp(XYt ), E[exp(XYt )|y0 ]

(4.6)

where X = {Xt , t ≥ 0} is a L´evy process. The factor exp((r − q)t)/E[exp(XYt )|y0 ] puts us immediately into the risk-neutral world by a mean-correcting argument. Basically, we model the stock price process as the ordinary exponential of a time-changed L´evy process. The process incorporates jumps (through the L´evy process Xt ) and stochastic volatility (through the time change Yt ). The characteristic function φ(u, t) for the log of our stock price is given by: φ(u, t) = E[exp(iu log(St ))|S0 , y0 ] = exp(iu((r − q)t + log S0 ))

ϕ(−iψX (u); t, y0 ) , ϕ(−iψX (−i); t, y0 )iu

(4.7)

74

Exotic Option Pricing and Advanced L´evy Models

where ψX (u) = log E[exp(iuX1 )]; ψX (u) is called the characteristic exponent of the L´evy process, Since we consider two L´evy processes (VG and NIG) and two stochastic clocks (CIR and Gamma-OU), we will ﬁnally end up with four resulting models abbreviated as VGCIR, VG-OU , NIG-CIR and NIG-OU . Because of (time)-scaling effects, one can set y0 = 1, and scale the present rate of time change to one. More precisely, we have that the characteristic function φ(u, t) of equation (4.7) satisﬁes: √ φNIG-CIR (u, t; α, β, δ, κ, η, λ, y0 ) = φNIG-CIR (u, t; α, β, δy0 , κ, η/y0 , λ/ y0 , 1), φNIG− OU (u, t; α, β, δ, λ, a, b, y0 ) = φNIG− OU (u, t; α, β, δy0 , λ, a, by0 , 1), √ φVG-CIR (u, t; C, G, M, κ, η, λ, y0 ) = φVG-CIR (u, t; Cy0 , G, M, κ, η/y0 , λ/ y0 , 1), φVG− OU (u, t; C, G, M, λ, a, b, y0 ) = φVG− OU (u, t; Cy0 , G, M, λ, a, by0 , 1).

Actually, this time-scaling effect lies at the heart of the idea of incorporating stochastic volatility through making time stochastic. Here, it comes down to the fact that instead of making the volatility parameter (of the Black–Scholes model) stochastic, we are making the parameter δ in the NIG case and the parameter C in the VG case stochastic (via the time). Note that this effect does not only inﬂuence the standard deviation (or volatility) of the processes; the skewness and the kurtosis are also now ﬂuctuating stochastically.

4.3 CALIBRATION Carr and Madan [7] developed pricing methods for the classical vanilla options which can be applied in general when the characteristic function of the risk-neutral stock price process is known. Let α be a positive constant such that the αth moment of the stock price exists. For all stock price models encountered here, typically a value of α = 0.75 will do ﬁne. Carr and Madan then showed that the price C(K, T ) of a European call option with strike K and time to maturity T is given by: exp(−α log(K)) +∞ exp(−iv log(K))(v) dv, (4.8) C(K, T ) = π 0 where (v) = =

exp(−rT )E[exp(i(v − (α + 1)i) log(ST ))] α 2 + α − v 2 + i(2α + 1)v exp(−rT )φ(v − (α + 1)i, T ) . α 2 + α − v 2 + i(2α + 1)v

(4.9) (4.10)

Using Fast Fourier Transforms, one can compute within a second the complete option surface on an ordinary computer. We apply the above calculation method in our calibration procedure and estimate the model parameters by minimizing the difference between market prices and model prices in a least-squares sense.

Model Risk for Exotic and Moment Derivatives

75

The data set consists of 144 plain vanilla call option prices with maturities ranging from less than one month up to 5.16 years. These prices are based on the implied volatility surface of the Eurostoxx 50 index, having a value of 2461.44 on October 7th, 2003. The volatilities can be found in Table 4.1. For the sake of simplicity and to focus on the essence of the stochastic behaviour of the asset, we set the risk-free interest rate equal to 3 percent and the dividend yield to zero. Table 4.1 Implied volatility surface data (Eurostoxx 50 index; October 7th, 2003) Strike

Maturity (year fraction) 0.0361

1081.82 1212.12 1272.73 1514.24 1555.15 1870.30 1900.00 2000.00 2100.00 2178.18 2200.00 2300.00 2400.00 2499.76 2500.00 2600.00 2800.00 2822.73 2870.83 2900.00 3000.00 3153.64 3200.00 3360.00 3400.00 3600.00 3626.79 3700.00 3800.00 4000.00 4070.00 4170.81 4714.83 4990.91 5000.00 5440.18

0.3175 0.3030 0.2990 0.2800 0.2650 0.2472 0.2471

0.2000

0.3105 0.3076 0.2976 0.2877 0.2800 0.2778 0.2678 0.2580 0.2493 0.2493 0.2405

1.1944

2.1916

4.2056

5.1639

0.3804 0.3667 0.3603 0.3348 0.3305 0.2973 0.2946 0.2858 0.2775 0.2709 0.2691 0.2608 0.2524 0.2446 0.2446 0.2381 0.2251 0.2240 0.2213 0.2198 0.2148 0.2113 0.2103 0.2069 0.2060

0.3451 0.3350 0.3303 0.3116 0.3084 0.2840 0.2817 0.2739 0.2672 0.2619 0.2604 0.2536 0.2468 0.2400 0.2400 0.2358 0.2273 0.2263 0.2242 0.2230 0.2195 0.2141 0.2125 0.2065 0.2050 0.1975 0.1972 0.1964 0.1953 0.1931

0.3150 0.3082 0.3050 0.2920 0.2899 0.2730 0.2714 0.2660 0.2615 0.2580 0.2570 0.2525 0.2480 0.2435 0.2435 0.2397 0.2322 0.2313 0.2295 0.2288 0.2263 0.2224 0.2212 0.2172 0.2162 0.2112 0.2105 0.2086 0.2059 0.2006 0.1988 0.1961 0.1910 0.1904 0.1903

0.3137 0.3073 0.3043 0.2921 0.2901 0.2742 0.2727 0.2676 0.2634 0.2600 0.2591 0.2548 0.2505 0.2463 0.2463 0.2426 0.2354 0.2346 0.2328 0.2321 0.2296 0.2258 0.2246 0.2206 0.2196 0.2148 0.2142 0.2124 0.2099 0.2050 0.2032 0.2008 0.1957 0.1949 0.1949 0.1938

76

Exotic Option Pricing and Advanced L´evy Models 1600

1400

1200

Option price

1000

800

600

400

200

0 1000

Figure 4.1

1500

2000

2500

3000 3500 Strike

4000

4500

5000

5500

Calibration of the NIG-CIR model (Eurostoxx 50 index; October 7th, 2003)

Contrary to the approach described in Hirsa et al. [15], we search for a single global set of parameters per model which we ﬁt (and which captures smile information) across the full range of maturities in the data set. This global parameter set can then be used to price path-dependent derivatives (e.g. payoffs at multiple points during its lifetime or moment derivatives; see Sections 4.6 and 4.7). This is in contrast with the parameter set resulting from a ﬁtting procedure at a single maturity date, which can in principle only be used to price option payoffs occurring at that speciﬁc maturity. The results of the global calibration are visualized in Figures 4.1 and 4.2 for the NIGCIR and the BN–S model, respectively. The other models give rise to completely similar ﬁgures. Here, the circles are the market prices and the plus signs are the analytical prices (calculated via equation (4.8) using the respective characteristic functions and obtained parameters). In Table 4.2, one ﬁnds the risk-neutral parameters for the different models. For comparative purposes, one computes several global measures of ﬁt. We consider the root mean square error (rmse), the average absolute error as a percentage of the mean price (ape), the average absolute error (aae) and the average relative percentage error (arpe): ' ( ( (Market price − Model price)2 rmse = ) number of options options

ape =

|Market price − Model price| 1 mean option price number of options options

Model Risk for Exotic and Moment Derivatives

1600

1400

1200

Option price

1000

800

600

400

200

0 1000

1500

2000

2500

3000 3500 Strike

4000

4500

5000

5500

Figure 4.2 Calibration of the BN–S model (Eurostoxx 50 index; October 7th, 2003)

Table 4.2 Risk-neutral parameters for the different models HEST σ02 = 0.0654, κ = 0.6067, η = 0.0707, θ = 0.2928, ρ = −0.7571 HESJ σ02 = 0.0576, κ = 0.4963, η = 0.0650, θ = 0.2286, ρ = −0.9900, µj = 0.1791, σj = 0.1346, λ = 0.1382 BN–S ρ = −4.6750, λ = 0.5474, b = 18.6075, a = 0.6069, σ02 = 0.0433 VG-CIR C = 18.0968, G = 20.0276, M = 26.3971, κ = 1.2145, η = 0.5501, λ = 1.7913, y0 = 1 VG-OU C = 6.1610, G = 9.6443, M = 16.0260, λ = 1.6790, a = 0.3484, b = 0.7664, y0 = 1 NIG-CIR α = 16.1975, β = −3.1804, δ = 1.0867, κ = 1.2101, η = 0.5507, λ = 1.7864, y0 = 1 NIG-OU α = 8.8914, β = −3.1634, δ = 0.6728, λ = 1.7478, a = 0.3442, b = 0.7628, y0 = 1

77

78

Exotic Option Pricing and Advanced L´evy Models

aae = arpe =

|Market price − Model price| number of options options |Market price − Model price| 1 number of options Market price options

In Table 4.3, an overview of these measures of ﬁt are given. Table 4.3 Global ﬁt error measures for the different models Model HEST HESJ BN–S VG-CIR VG-OU NIG-CIR NIG-OU

rmse

ape

aae

arpe

3.0281 2.8101 3.5156 2.3823 3.4351 2.3485 3.2737

0.0048 0.0045 0.0056 0.0038 0.0056 0.0038 0.0054

2.4264 2.2469 2.8194 1.9337 2.8238 1.9194 2.7385

0.0174 0.0126 0.0221 0.0106 0.0190 0.0099 0.0175

4.4 SIMULATION In this section, we describe in detail how the particular processes presented in Section 4.2, can be implemented in practice in a Monte Carlo simulation pricing framework. For this, we ﬁrst discuss the numerical implementation of the four building block processes which drive them. This will be followed by an explanation of how one assembles a time-changed L´evy process. 4.4.1 NIG L´evy process To simulate a NIG process, we ﬁrst describe how to simulate NIG(α, β, δ) random numbers. The latter can be obtained by mixing Inverse Gaussian (IG) random numbers and standard Normal numbers in the following manner. An IG(a, b) random variable X has a characteristic function given by: E[exp(iuX)] = exp(−a −2ui + b2 − b). First, simulate IG(1, δ α 2 − β 2 ) random numbers ik , for example, by using the Inverse Gaussian generator of Michael, Schucany and Haas (see Devroye [13]). Then sample a sequence of standard Normal random variables uk . NIG random numbers nk are then obtained via: nk = δ 2 β ik + δ ik uk . Finally, the sample paths of a NIG(α, β, δ) process X = {Xt , t ≥ 0} in the time points tn = nt, n = 0, 1, 2, . . . can be generated by using the independent NIG(α, β, δt) random numbers nk as follows: X0 = 0,

Xtk = Xtk−1 + nk ,

k ≥ 1.

Model Risk for Exotic and Moment Derivatives

79

4.4.2 VG L´evy process Since a VG process can be viewed as the difference of two independent Gamma processes, the simulation of a VG process becomes straightforward. A Gamma process with parameters a, b > 0 is a L´evy process with Gamma(a, b) distributed increments, i.e. following a Gamma distribution with mean a/b and variance a/b2 . A VG process X(V G) = {Xt(V G) , t ≥ (2) (1) = 0} with parameters C, G, M > 0 can be decomposed as Xt(V G) = G(1) t − Gt , where G (1) (2) (2) {Gt , t ≥ 0} is a Gamma process with parameters a = C and b = M and G = {Gt , t ≥ 0} is a Gamma process with parameters a = C and b = G. The generation of Gamma numbers is quite standard. Possible generators are Johnk’s gamma generator and Berman’s gamma generator [13]. 4.4.3 CIR stochastic clock The simulation of a CIR process y = {yt , t ≥ 0} is straightforward. Basically, we discretize the SDE: 1/2

dyt = κ(η − yt ) dt + λyt

dWt ,

y0 ≥ 0,

where Wt is a standard Brownian motion. Using a ﬁrst-order accurate explicit differencing scheme in time, the sample path of the CIR process y = {yt , t ≥ 0} in the time points t = nt, n = 0, 1, 2, . . . , is then given by: 1/2 √ ytn = ytn−1 + κ(η − ytn−1 )t + λytn−1 t vn , where {vn , n = 1, 2, . . . } is a series of independent standard Normally distributed random numbers. For other more involved simulation schemes, like the Milstein scheme, resulting in a higher-order discretization in time, we refer to J¨ackel [17]. 4.4.4 Gamma-OU stochastic clock Recall that for the particular choice of an OU-Gamma process, the subordinator z = {zt , t ≥ 0} in (equation (4.3)) is given by the compound Poisson process (equation (4.4)). To simulate a Gamma(a, b)-OU process y = {yt , t ≥ 0} in the time points tn = nt, n = 0, 1, 2, . . . , we ﬁrst simulate in the same time points a Poisson process N = {Nt , t ≥ 0} with intensity parameter aλ. Then (with the convention that an empty sum equals zero) ytn = (1 − λt)ytn−1 +

Ntn

xk exp(−λt u˜ k ),

k=Ntn−1 +1

where u˜ k is a series of independent uniformly distributed random numbers and xk can be obtained from your preferred uniform random number generator via xk = − log(uk )/b. 4.4.5 Path generation for time-changed L´evy process The explanation of the building block processes above allow us next to assemble all the parts of the time-changed L´evy process simulation puzzle. For this one can proceed through the following ﬁve steps [22]:

80

(i) (ii) (iii) (iv) (v)

Exotic Option Pricing and Advanced L´evy Models

simulate the rate of time change process y = {y tt , 0 ≤ t ≤ T }; calculate from (i) the time change Y = {Yt = 0 ys ds, 0 ≤ t ≤ T }; simulate the L´evy process X = {Xt , 0 ≤ t ≤ YT }; calculate the time changed L´evy process XYt , for 0 ≤ t ≤ T ; calculate the stock price process using equation (4.6). The mean correcting factor is calculated as: exp((r − q)t) exp((r − q)t) = . E[exp(XYt )|y0 ] ϕ(−iψX (−i); t, 1)

4.5 PRICING OF EXOTIC OPTIONS As evidenced by the quality of the calibration on a set of European call options in Section 4.3, we can hardly discriminate between the different processes on the basis of their smileconform pricing characteristics. We therefore put the models further to the test by applying them to a range of more exotic options. These range from digital barriers, one-touch barrier options, lookback options and ﬁnally cliquet options with local as well as global parameters. These ﬁrst-generation exotics with path-dependent payoffs were selected since they shed more light on the dynamics of the stock processes. At the same time, the pricings of the cliquet options are highly sensitive to the forward smile characteristics induced by the models. 4.5.1 Exotic options Let us consider contracts of duration T , and denote the maximum and minimum process, respectively, of a process Y = {Yt , 0 ≤ t ≤ T } as MtY = sup{Yu ; 0 ≤ u ≤ t} and mYt = inf{Yu ; 0 ≤ u ≤ t},

0 ≤ t ≤ T.

4.5.1.1 Digital barriers We ﬁrst consider digital barrier options. These options remain worthless unless the stock price hits some predeﬁned barrier level H > S0 , in which case they pay at expiry a ﬁxed amount D, normalized to 1 in the current settings. Using risk-neutral valuation, assuming no dividends and a constant interest rate r, the time t = 0 price is therefore given by: digital = e−rT EQ [1(MTS ≥ H )], where the expectation is taken under the risk-neutral measure Q. Observe that with the current deﬁnition of digital barriers their pricing reﬂects exactly the chance of hitting the barrier prior to expiry. The behaviour of the stock after the barrier has been hit does not inﬂuence the result, in contrast with the classic barrier options deﬁned below. 4.5.1.2 One-touch barrier options For one-touch barrier call options, we focus on the following four types: • The down-and-out barrier call is worthless unless its minimum remains above some ‘low barrier’ H , in which case it retains the structure of a European call with strike K. Its

Model Risk for Exotic and Moment Derivatives

81

initial price is given by: DOB = e−rT EQ [(ST − K)+ 1(mST > H )] • The down-and-in barrier is a normal European call with strike K, if its minimum went below some ‘low barrier’ H . If this barrier was never reached during the lifetime of the option, the option remains worthless. Its initial price is given by: DIB = e−rT EQ [(ST − K)+ 1(mST ≤ H )] • The up-and-in barrier is worthless unless its maximum crossed some ‘high barrier’ H , in which case it obtains the structure of a European call with strike K. Its price is given by: UIB = e−rT EQ [(ST − K)+ 1(MTS ≥ H )] • The up-and-out barrier is worthless unless its maximum remains below some ‘high barrier’ H , in which case it retains the structure of a European call with strike K. Its price is given by: UOB = e−rT EQ [(ST − K)+ 1(MTS < H )] 4.5.1.3 Lookback options The payoff of a lookback call option corresponds to the difference between the stock price level at expiry ST and the lowest level it has reached during its lifetime. The time t = 0 price of a lookback call option is therefore given by: LC = e−rT EQ [ST − mST ]. Clearly, of the three path-dependent options introduced so far, the lookback option depends the most on the precise path dynamics. 4.5.1.4 Cliquet options Finally, we also test the proposed models on the pricing of cliquet options. These still are very popular options in the equity derivatives world which allow the investor to participate (partially) in the performance of an underlying over a series of consecutive time periods [ti , ti+1 ] by ‘clicking in’ the sum of these local performances. The latter are measured relative to the stock level Sti attained at the start of each new subperiod, and each of the local performances is ﬂoored and/or capped to establish whatever desirable mix of positive and/or negative payoff combination. Generally, on the ﬁnal sum an additional global ﬂoor (cap) is applied to guarantee a minimum (maximum) overall payoff. This can all be summarized through the following payoff formula:

min capglob , max ﬂoorglob ,

N i=1

Sti − Sti−1 min caploc , max ﬂoorloc , Sti−1

82

Exotic Option Pricing and Advanced L´evy Models

Observe that the local ﬂoor and cap parameters effectively border the relevant ‘local’ price ranges by centering them around the future, and therefore unknown, spot levels Sti . The pricing will therefore depend in a non-trivial subtle manner on the forward volatility smile dynamics of the respective models, further complicated by the global parameters of the contract. For an in-depth account of the related volatility issues, we refer to Wilmott [24]. 4.5.2 Exotic option prices We price all exotic options through Monte Carlo simulation. We consistently average over 1 000 000 simulated paths. All options have a lifetime of three years. In order to check the accuracy of our simulation algorithm, we simulated option prices for all European calls available in the calibration set. All algorithms gave a very satisfactory result, with pricing differences with respect to their analytic calibration values of less than 0.5 percent. An important issue for the path-dependent lookback, barrier and digital barrier options above, is the frequency at which the stock price is observed for purposes of determining whether the barrier or its minimum level have been reached. In the numerical calculations below, we have assumed a discrete number of observations, namely at the close of each trading day. Moreover, we have assumed that a year consists of 250 trading days. In Figure 4.3, we present simulation results with models for the digital barrier call option as a function of the barrier level (ranging from 1.05S0 to 1.5S0 ). As mentioned before, aside from the discounting factor e−rT , the premiums can be interpreted as the chance of hitting the barrier during the option lifetime. In Figures 4.4–4.6, we show prices for all

0.8

0.7

Option price

0.6

0.5

0.4

0.3

1.05

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR 1.1

1.15

1.2

1.25 1.3 1.35 Digital as percentage of spot

1.4

1.45

Figure 4.3 Digital barrier prices (Eurostoxx 50 index; October 7th, 2003)

1.5

Model Risk for Exotic and Moment Derivatives

500

450

Option price

400

350

300

250

200 0.5

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR 0.55

Figure 4.4

0.6

0.65

0.7 0.75 0.8 0.85 Barrier as percentage of spot

0.9

0.95

DOB prices (Eurostoxx 50 index; October 7th, 2003)

350 NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR

300

Option price

250

200

150

100

50

0 0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

Barrier as percentage of spot

Figure 4.5

DIB prices (Eurostoxx 50 index; October 7th, 2003)

0.95

83

84

Exotic Option Pricing and Advanced L´evy Models

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR

180 160 140

Option price

120 100 80 60 40 20 0 1.05

1.1

1.15

1.2

1.25

1.3

1.35

1.4

1.45

1.5

Barrier as percentage of spot

Figure 4.6 UOB prices (Eurostoxx 50 index; October 7th, 2003) 520 500 480

Option price

460 440 420 400 380 360 340 320 1.05

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR 1.1

1.15

1.2 1.25 1.3 1.35 Barrier as percentage of spot

1.4

1.45

Figure 4.7 UIB prices (Eurostoxx 50 index; October 7th, 2003)

1.5

Model Risk for Exotic and Moment Derivatives

85

Table 4.4 Exotic option prices H/S0

NIG-OUT

VG-CIR

VG-OUT

HEST

HESJ

BN–S

NIG-CIR

724.80 511.80 293.28 391.17 448.10 479.83 496.95 505.24 509.10 510.75 511.40 511.67

713.49 509.33 318.35 402.24 452.97 481.74 496.80 504.05 507.21 508.53 509.06 509.24

844.51 510.88 173.85 280.79 359.05 414.65 452.76 477.37 492.76 501.74 506.46 508.91

845.18 510.89 174.64 282.09 360.99 416.63 454.33 479.12 494.25 502.84 507.41 509.51

771.28 509.89 230.25 352.14 423.21 461.82 481.85 492.62 498.93 503.17 505.93 507.68

730.84 512.21 284.10 387.83 446.52 479.77 496.78 505.38 509.34 511.09 511.80 512.08

LC Call DOB DOB DOB DOB DOB DOB DOB DOB DOB DOB

0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

722.34 509.76 300.25 396.80 451.61 481.65 497.00 504.31 507.53 508.88 509.43 509.64

DIB DIB DIB DIB DIB DIB DIB DIB DIB DIB

0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

209.51 112.95 58.14 28.11 12.76 5.45 2.23 0.88 0.33 0.12

218.51 120.62 63.69 31.96 14.84 6.55 2.70 1.04 0.39 0.13

190.98 107.08 56.35 27.59 12.53 5.28 2.11 0.79 0.26 0.09

337.03 230.09 151.83 96.24 58.13 33.51 18.12 9.14 4.42 1.98

336.25 228.80 149.90 94.26 56.56 31.77 16.64 8.05 3.48 1.38

279.61 157.72 86.65 48.04 28.01 17.24 10.94 6.69 3.94 2.19

228.10 124.37 65.68 32.43 15.42 6.83 2.87 1.11 0.40 0.13

UIB UIB UIB UIB UIB UIB UIB UIB UIB UIB

1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

509.32 506.68 500.33 489.05 472.47 450.54 423.62 393.01 359.77 325.25

511.52 509.80 505.21 496.50 482.84 463.62 439.32 410.46 378.05 343.46

508.84 506.11 499.56 488.30 471.39 449.23 422.32 391.36 357.80 322.79

510.78 500.90 507.08 501.04 490.73 475.30 454.77 428.96 399.24 365.57

510.81 510.00 507.28 501.31 490.93 474.86 452.47 424.09 389.56 350.68

509.73 508.38 504.28 495.95 482.66 464.48 441.48 414.98 385.50 354.90

511.98 510.37 505.93 497.41 483.94 465.16 441.00 412.16 380.04 345.79

UOB UOB UOB UOB UOB UOB UOB UOB UOB UOB

1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

0.44 3.08 9.43 20.71 37.29 59.22 86.14 116.75 149.98 184.50

0.27 2.00 6.59 15.29 28.95 48.17 72.47 101.33 133.74 168.33

0.49 3.22 9.77 21.03 37.94 60.10 87.00 117.96 151.52 186.53

0.103 0.979 3.80 8.96 20.15 35.58 56.10 81.93 111.65 145.31

0.08 0.89 3.61 9.85 19.96 36.03 58.42 86.80 121.33 160.21

0.13 1.48 5.58 13.91 27.20 45.38 68.39 94.88 124.36 154.96

0.23 1.84 6.27 14.80 28.26 47.04 71.21 100.04 132.16 166.41

DIG DIG

1.05 1.1

0.7995 0.7201

0.8064 0.7334

0.7909 0.7120

0.8218 0.7478

0.8189 0.7421

0.8173 0.7360

0.8118 0.7380

(continued overleaf )

86

Exotic Option Pricing and Advanced L´evy Models Table 4.4 (continued ) H/S0

DIG DIG DIG DIG DIG DIG DIG DIG

NIG-OUT

VG-CIR

VG-OUT

HEST

HESJ

BN–S

NIG-CIR

0.6458 0.5744 0.5062 0.4418 0.3816 0.3264 0.2763 0.2321

0.6628 0.5940 0.5273 0.4630 0.4021 0.3456 0.2940 0.2474

0.6382 0.5678 0.5003 0.4363 0.3767 0.3217 0.2722 0.2280

0.6762 0.6069 0.5408 0.4769 0.4169 0.3603 0.3087 0.2610

0.6685 0.5971 0.5290 0.4637 0.4012 0.3426 0.2877 0.2374

0.6580 0.5836 0.5138 0.4493 0.3893 0.3355 0.2870 0.2446

0.6670 0.5977 0.5308 0.4668 0.4059 0.3490 0.2975 0.2510

1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

Table 4.5 Lookback option prices for the different models HEST

HESJ

BN–S

VG-CIR

VG-OU

NIG-CIR

844.51

845.19

771.28

724.80

713.49

730.84

NIG-OU 722.34

one-touch barrier options (as a percentage of the spot). The strike K was always taken equal to the spot S0 . For reference, we summarize in Table 4.4 all option prices for the above discussed exotics. One can check that the barrier results agree well with the identity DIB + DOB = vanilla call = UIB + UOB, suggesting that the simulation results are well converged. Lookback prices are presented in Table 4.5. Consistently over all of the ﬁgures the Heston prices suggest that this model (for the current calibration) results in paths dynamics that are more volatile, breaching more frequently the imposed barriers. The results for the L´evy models with stochastic time change seem to move in pairs, with the choice of stochastic clock dominating over the details of the L´evy model upon which the stochastic time change is applied. The ﬁrst couple, VG- and NIG- display very similar results, overall showing the least volatile path dynamics, whereas the VG-CIR and NIG-CIR prices consistently fall midway of the pack. Finally, the OU- results without stochastic clock typically fall between the Heston and the VG-CIR and NIG-CIR prices. Besides these qualitative observations, it is important to note the magnitude of the observed differences. Lookback prices vary over about 15 percent and the one-touch barriers over 200 percent, whereas for the digital barriers we found price differences of over 10 percent. For the cliquet options, the prices are shown in Figures 4.8 and 4.9 for two different combinations. The numerical values can be found in Tables 4.6 and 4.7. These results are in-line with the previous observations. Variations of over 40 percent are noted.

4.6 PRICING OF MOMENT DERIVATIVES These derivatives depend on the realized higher moments of the underlying. More precisely, their payoff is a function of powers of the (daily) log-returns and allows to cover different kinds of market shocks. Variance swaps were already created to cover changes in the volatility regime. Besides the latter, skewness and kurtosis also play an important role. To protect against a wrongly estimated skewness or kurtosis, moment derivatives of higher order can

Model Risk for Exotic and Moment Derivatives

87

0.18

Option price

0.16

0.14

0.12

0.1

NIG-CIR NIG-OUGamma VG-OUGamma VG-CIR HESJ HEST BN–S

0.08

0.06

0

0.02

0.04

0.06

0.08 0.1 0.12 Global floor

0.14

0.16

0.18

0.2

Figure 4.8 Cliquet prices: caploc = 0.08; ﬂoloc = −0.08; capglo = +∞; N = 3; t1 = 1; t2 = 3 (Eurostoxx 50 index; October 7th, 2003) 0.16 0.15

Option price

0.14 0.13 0.12 0.11 0.1 NIG-CIR NIG-OUGamma VG-OUGamma VG-CIR HESJ HEST BN–S

0.09 0.08

−0.05

0

0.05 Global floor

0.1

0.15

Figure 4.9 Cliquet Prices: caploc = 0.05; ﬂoloc = −0.03; capglo = +∞; T = 3; N = 6; ti = i/2 (Eurostoxx 50 index; October 7th, 2003)

88

Exotic Option Pricing and Advanced L´evy Models

Table 4.6 Cliquet prices: caploc = 0.08; f lo−loc = 0.08; capglo = +∞; f loglo ∈ [0, 0.20]; N = 3; t1 = 1; t2 = 2; t3 = 3 f loglo

NIG-CIR

NIG-OUT

VG-OU

VG-CIR

HESJ

HEST

BN–S

0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.20

0.0785 0.0817 0.0850 0.0885 0.0922 0.0960 0.1000 0.1042 0.1086 0.1144 0.1203 0.1264 0.1327 0.1391 0.1456 0.1523 0.1591 0.1661 0.1732 0.1805 0.1880

0.0837 0.0866 0.0897 0.0930 0.0964 0.1000 0.1037 0.1076 0.1117 0.1174 0.1232 0.1292 0.1353 0.1415 0.1478 0.1543 0.1610 0.1677 0.1747 0.1817 0.1889

0.0835 0.0865 0.0896 0.0928 0.0963 0.0998 0.1036 0.1075 0.1116 0.1173 0.1231 0.1291 0.1352 0.1414 0.1478 0.1543 0.1610 0.1678 0.1747 0.1818 0.1890

0.0785 0.0817 0.0850 0.0885 0.0921 0.0960 0.1000 0.1042 0.1085 0.1144 0.1203 0.1264 0.1327 0.1391 0.1456 0.1523 0.1591 0.1661 0.1733 0.1806 0.1880

0.0667 0.0704 0.0743 0.0783 0.0825 0.0868 0.0913 0.0959 0.1008 0.1072 0.1137 0.1204 0.1272 0.1342 0.1412 0.1485 0.1558 0.1633 0.1709 0.1787 0.1866

0.0683 0.0719 0.0757 0.0796 0.0837 0.0879 0.0923 0.0969 0.1017 0.1080 0.1145 0.1211 0.1279 0.1348 0.1418 0.1489 0.1562 0.1637 0.1712 0.1789 0.1868

0.0696 0.0731 0.0767 0.0805 0.0845 0.0887 0.0930 0.0976 0.1024 0.1085 0.1149 0.1214 0.1280 0.1348 0.1418 0.1489 0.1561 0.1635 0.1711 0.1788 0.1867

Table 4.7 Cliquet prices: ﬂoloc = −0.03; caploc = 0.05; capglo = +∞; T = 3; N = 6; ti = i/2 f loglo

NIG-CIR

NIG-OU

VG-OU

VG-CIR

HESJ

HEST

BN–S

−0.05 −0.04 −0.03 −0.02 −0.01 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15

0.0990 0.0997 0.1005 0.1015 0.1028 0.1044 0.1060 0.1079 0.1099 0.1121 0.1145 0.1171 0.1204 0.1239 0.1277 0.1317 0.1361 0.1406 0.1456 0.1508 0.1567

0.1092 0.1098 0.1104 0.1112 0.1124 0.1137 0.1152 0.1168 0.1185 0.1205 0.1226 0.1250 0.1280 0.1312 0.1346 0.1382 0.1421 0.1463 0.1508 0.1556 0.1611

0.1131 0.1137 0.1144 0.1151 0.1162 0.1175 0.1189 0.1204 0.1221 0.1240 0.1260 0.1283 0.1311 0.1342 0.1375 0.1410 0.1448 0.1488 0.1531 0.1576 0.1630

0.1001 0.1008 0.1017 0.1026 0.1039 0.1054 0.1071 0.1089 0.1109 0.1131 0.1154 0.1180 0.1213 0.1248 0.1286 0.1326 0.1368 0.1414 0.1462 0.1514 0.1573

0.0724 0.0734 0.0745 0.0757 0.0776 0.0798 0.0821 0.0847 0.0874 0.0904 0.0937 0.0971 0.1016 0.1063 0.1113 0.1165 0.1220 0.1278 0.1339 0.1403 0.1474

0.0762 0.0771 0.0781 0.0762 0.0811 0.0831 0.0853 0.0877 0.0904 0.0932 0.0963 0.0996 0.1039 0.1084 0.1132 0.1183 0.1237 0.1293 0.1352 0.1415 0.1484

0.0788 0.0796 0.0805 0.0815 0.0831 0.0849 0.0869 0.0891 0.0915 0.0942 0.0972 0.1004 0.1045 0.1088 0.1135 0.1185 0.1238 0.1294 0.1353 0.1415 0.1484

Model Risk for Exotic and Moment Derivatives

89

be useful. Recent studies by Nualart and Schoutens [20] [21] and Corcuera et al. [10] [11] suggest that functionals of powers of returns seem the natural choice to complete the market. It was shown that allowing trade in the power-assets of all orders in an incomplete L´evy market leads to a complete market. Power assets are strongly related to the realized higher moments and they mainly coincide in a discrete time framework [11]. 4.6.1 Moment swaps Consider a ﬁnite set of discrete times {t0 = 0, t1 , . . . , tn = T } at which the path of the underlying is monitored. We denote the price of the underlying at these points, i.e. Sti , by Si for simplicity. Typically, the ti correspond to daily closing times and Si is the closing price at day i. Note that then: log(Si ) − log(Si−1 ),

i = 1, . . . , n,

correspond to the daily log-returns. Next, we deﬁne the moment swaps. The kth-moment swap is a contract where the parties agree to exchange at maturity:

n n

k Si (k) k log , MOMS = N × (log(Si ) − log(Si−1 )) = N × Si−1 i=1

i=1

where N is the nominal amount. A special case of these swaps is the second moment swap, better known as the Variance Swap. The non-centred payoff function in that case is given by: n

2 VS = N × (log(Si ) − log(Si−1 )) . i=1

Basically, this contract swaps ﬁxed (annualized) variance by the realized variance (second moment) and as such provides protection against unexpected or unfavourable changes in volatility. Higher moment swaps provide the same kind of protection. The MOMS(3) is related to realized skewness and provides protection against changes in the symmetry of the underlying distribution. MOMS(4) derivatives are linked to realized kurtosis and provide protection against the unexpected occurrences of very large jumps, or in other words, changes in the tail behaviour of the underlying distribution. 4.6.2 Moment options Related to the above discussed swaps, we deﬁne the associated options on the realized kth moment. More precisely, a moment option of order k, pays out at maturity T :

+ n (log(Si /Si−1 ))k − K . i=1

The price of these options under risk-neutral valuation is given by: % n

+ & (k) k . (log(Si /Si−1 )) − K MOMO (K, T ) = exp(−rT )EQ i=1

90

Exotic Option Pricing and Advanced L´evy Models

Note that since odd moments can be negative, the strike price for these options can range over the whole real line. 4.6.3 Hedging moment swaps We focus on hedging the moment swaps which are written on the future price as underlying. The price process of the future is given by F = {Ft = exp((r − q)(T − t))St }; we write Fi = Fti . In line with the results obtained by Carr and Lewis [6], ﬁrst consider the following (Taylor-like) expansion of the kth power of the logarithmic function: (log(x))2 (log(x))k = k! x − 1 − log(x) − 2!

3 (log(x)) (log(x))k−1 − − ··· − + O((x − 1)k+1 ) . 3! (k − 1)! Substituting x by Fi /Fi−1 leads to (log(Fi /Fi−1 ))k = k!

(log(Fi /Fi−1 ))j Fi − log(Fi /Fi−1 ) − Fi−1 j! k−1

j =2

+ O((Fi /Fi−1 )k+1 ) , where Fi = Fi − Fi−1 . Summing over i gives a decomposition of the MOMS(k) (on a future) payoff: MOMS(k) = N ×

n i=1

= N k!

n i=1

(log(Fi /Fi−1 ))k

Fi − log(Fi /Fi−1 ) − Fi−1

k−1 (log(Fi /Fi−1 ))j j =2

j!

+ O((Fi /Fi−1 )k+1 ) = −N k!(log(FT ) − log(F0 ))

n

n k−1 Fi k! (j ) k+1 MOMS + O +N k! −N (Fi /Fi−1 ) Fi−1 j! i=1

j =2

i=1

(4.11)

Thus, up to (k + 1)th-order terms the sum of the kth powered log-returns decomposes into the payouts from: • −k! log-contracts on the future with payoff log(FT ) − log(F0 ); i ) in futures; • a self-ﬁnancing dynamic strategy (k! ni=1 FF i−1 • a series of moment contracts of order strictly smaller than k.

Model Risk for Exotic and Moment Derivatives

91

The log-contract can be hedged by a dynamic trading strategy in combination with a static position in bonds, European vanilla call and put options maturing at time T. More precisely, ﬁrst note that for any L > 0: log(FT ) − log(F0 ) =

1 (FT − F0 ) − u(FT ) + u(F0 ), L

(4.12)

for u(x) =

x−L − log(x) + log(L) . L

Moreover Carr and Lewis [6] show that:

L

u(FT ) = u(ST ) =

1 (K − ST )+ dK + K2

0

n

Since FT − F0 = implies:

i=1 Fi ,

MOMS

(k)

+∞

L

1 (ST − K)+ dK. K2

(4.13)

substituting equation (4.13) into equations (4.12) and (4.11)

+∞ 1 1 + + ≈ N k! (K − ST ) dK + (ST − K) dK. 2 K2 0 K L

n k−1 1 k! 1 Fi − u(F0 ) − N +N k! MOMS(j ) . − Fi−1 L j! L

j =2

i=1

4.6.4 Pricing of moments swaps We calculate under the different models, the risk-neutral expectation: EQ MOMS(k) . We consistently average over 1 000 000 simulated paths. All options have a lifetime of 1 year. In Table 4.8, we clearly see how the price differences are even more pronounced as compared to the exotic option pricings discussed in Section 4.5.2.

Table 4.8 Moment swaps (N = 10 000) for the different models Order e−rT EQ MOMS(2) e−rT EQ MOMS(3) e−rT EQ MOMS(4)

HEST 623.89

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

804.60

557.55

628.85

557.75

641.71

−0.0807

−312.58

−21.03

−74.91

−21.69

−88.82

0.6366

322.40

7.8698

33.89

8.554

47.99

92

Exotic Option Pricing and Advanced L´evy Models 800 Heston BN–S VG-CIR VG-GAM NIG-CIR NIG-GAM

700

600

Price

500

400

300

200

100

0 100

200

300

400

500 600 K (bp)

700

800

900

1000

Figure 4.10 Moment option of order 2 (N = 10 000) 100 Heston BN–S VG-CIR VG-GAM NIG-CIR NIG-GAM

90 80 70

Price

60 50 40 30 20 10 0 −0.01 −0.008 −0.006 −0.004 −0.002

0 K

0.002

0.004

0.006

Figure 4.11 Moment option of order 3 (N = 10 000)

0.008

0.01

Model Risk for Exotic and Moment Derivatives

93

4.6.5 Pricing of moments options Next, we calculate the prices of moment call option, MOMO(k) , paying out at maturity T : n

+ (log(Si /Si−1 )) − K k

,

i=1

where the price of these call moment options is by the risk-neutral valuation: MOMO (K, T ) = exp(−rT )EQ (k)

% n

& (log(Si /Si−1 )) − K) k

+

.

i=1

We plot in Figures 4.10–4.11 the price for moment options of order 2 and 3; corresponding values for these options and fourth order moment option prices can be found in Tables 4.9–4.11.

The disparity between the models is ampliﬁed. The L´evy models with stochastic timechange seem again to move in the same pairs as in agreement with the results in Section 4.5.2, but now only up to the third-order moment option. The BN–S model has very pronounced second- and fourth-order moment option prices, while HEST drops (in absolute value) to very low values for the fourth-order moment option when compared to the other models.

4.7 CONCLUSIONS We have looked at different models, all reﬂecting non-normal returns and stochastic volatility. Empirical work has generally supported the need for both ingredients. We have demonstrated the clear ability of all proposed processes to produce a very convincing ﬁt to a market-conform volatility surface. At the same time, we have shown that this calibration could be achieved in a timely manner by using a very fast computational Table 4.9

Moment option data of order 2 (N = 10 000) for the different models

K (bp)

HEST

100 200 300 400 500 600 700 800 900 1000

302.3301 219.1410 156.7058 109.5242 75.9747 52.4440 37.0312 25.9978 17.4472 11.2276

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

491.4817 436.4667 394.0581 357.8177 326.6175 300.0791 277.1135 256.4670 238.1357 221.1757

212.0101 152.6484 110.6050 80.4430 58.2142 42.2200 30.0486 21.2445 14.8481 10.4386

183.3647 121.4477 83.4256 58.7503 42.1646 31.0541 23.5572 17.9608 14.1807 11.1815

249.068 186.381 140.753 106.257 80.235 60.917 46.407 36.175 28.402 24.19

161.5099 100.5233 67.6916 47.7753 35.2828 26.9061 20.2450 15.0211 11.2782 8.4718

94

Exotic Option Pricing and Advanced L´evy Models Table 4.10 Moment option data of order 3 (N = 10 000) for the different models

K −0.010 −0.009 −0.008 −0.007 −0.006 −0.005 −0.004 −0.003 −0.002 −0.001 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.010

HEST

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

98.0459 88.2341 78.4223 68.6105 58.7987 48.9869 39.1751 29.3686 19.5673 9.8393 0.7274 0.0008 0 0 0 0 0 0 0 0 0

79.9869 71.3740 62.8212 54.3490 46.0100 37.7781 29.6576 21.7306 14.0430 6.6657 0.0997 0 0 0 0 0 0 0 0 0 0

89.5862 79.9497 70.3448 60.8159 51.3447 41.9813 32.7639 23.7993 15.3273 7.5947 1.9022 0.8012 0.4866 0.3267 0.2293 0.1614 0.1052 0.0648 0.0416 0.0268 0.0170

84.6786 75.1442 65.6663 56.2683 46.9840 37.8554 28.9710 20.3739 12.3057 5.3112 0.8162 0.2915 0.1559 0.0938 0.0486 0.0322 0.0224 0.0126 0.0028 0 0

87.4962 78.1320 68.8121 59.5498 50.3831 41.3352 32.4824 23.8966 15.6840 8.1994 2.4520 1.1520 0.6438 0.4213 0.2819 0.1998 0.1325 0.0873 0.0578 0.0287 0.0091

82.2679 72.8796 63.5948 54.3998 45.3475 36.4821 27.8108 19.4166 11.4694 4.4442 0.1462 0.0173 0.0074 0 0 0 0 0 0 0 0

Table 4.11 Moment option data of order 4 (N = 10 000) for the different models K

HEST

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

0.0001 0.0002 0.0003 0.0004 0.0005 0.0006 0.0007 0.0008 0.0009 0.0010

0.0781 0.0259 0.0120 0.0065 0.0033 0.0011 0.0001 0 0 0

35.7465 35.4823 35.2471 35.0274 34.8220 34.6307 34.4525 34.2810 34.1127 33.9479

1.9322 1.5977 1.3603 1.1821 1.0428 0.9281 0.8328 0.7506 0.6790 0.6202

2.3416 2.0015 1.7655 1.5879 1.4542 1.3386 1.2403 1.1552 1.0805 1.0160

5.2360 4.8754 4.6077 4.3995 4.2249 4.0755 3.9403 3.8221 3.7211 3.6303

3.6095 3.3309 3.1274 2.9601 2.8158 2.6878 2.5750 2.4746 2.3819 2.2979

procedure based on FFT. Note that an almost identical calibration means that at the timepoints of the maturities of the calibration data set the marginal distribution is ﬁtted accurately to the risk-neutral distribution implied by the market. If we have different models all leading to such almost perfect calibrations, all models have almost the same marginal distributions. It should, however, be clear that even if at all time-points 0 ≤ t ≤ T marginal distributions among different models coincide, this does not imply that exotic prices should also be

Model Risk for Exotic and Moment Derivatives

95

the same. This can be seen from the following discrete-time example. Let n ≥ 2 and X = {Xi , i = 1, . . . , n} be an iid sequence and let {ui , i = 1, . . . , n} be an independent sequence which randomly varies between ui = 0 and 1. We propose two discrete (be it unrealistic) stock price models, S (1) and S (2) , with the same marginal distributions: Si(1) = ui X1 + (1 − ui )X2 and Si(2) = Xi . The ﬁrst process ﬂips randomly between two states X1 and X2 , both of which follow the distribution of the iid sequence, and so do all of the marginals at the time points i = 1, . . . , n. The second process changes value in all time-points. The values are independent of each other and all follow again the same distribution of the iid sequence. In both cases, all of the marginal distributions (at every i = 1, . . . , n) are the same (as the distribution underlying the sequence X). It is clear, however, that the maximum and minimum of both processes behave completely different. For the ﬁrst process, the maximal maxj ≤i Si(1) = max(X1 , X2 ) and the minimal process minj ≤i Si(1) = min(X1 , X2 ) for i being large enough, whereas for the second process there is much more variation possible and it clearly leads to other distributions. In summary, it should be clear that equal marginal distributions of a process do not at all imply equal marginal distributions of the associated minimal or maximal process. This explains why matching European call prices do not lead necessarily to matching exotic prices. It is the underlying ﬁne-grain structure of the process that will have an important impact on the path-dependent option prices. We have illustrated this by pricing exotics by Monte Carlo simulation, showing that price differences for one-touch barriers of over 200 percent are no exception. For lookback call options, a price range of more than 15 percent among the models was observed. A similar conclusion was valid for the digital barrier premiums. Even for cliquet options, which only depend on the stock realizations over a limited amount of time-points, prices vary substantially among the models. Moment derivatives amplify pricing disparity. At the same time, the presented details of the Monte Carlo implementation should allow the reader to embark on his/her own pricing experiments. The conclusion is that great care should be taken when employing attractive ‘fancy-dancy’ models to price (or even more important, to evaluate hedge parameters for) exotics. As far as we know, no detailed study about the underlying path structure of assets has been carried out yet. Our study motivates such a deeper investigation.

ACKNOWLEDGEMENTS The ﬁrst author is a Postdoctoral Fellow of the Fund for Scientiﬁc Research, Flanders, Belgium (FWO – Vlaanderen). We thank Marc Jeannin for his devoted programming work.

REFERENCES [1] Bakshi, G., Cao, C. and Chen, Z. (1997), “Empirical performance of alternative option pricing models”, The Journal of Finance, LII(5), 2003–2049.

96

Exotic Option Pricing and Advanced L´evy Models

[2] Barndorff-Nielsen, O.E. and Shephard, N. (2001), “Non-Gaussian Ornstein–Uhlenbeck-based models and some of their uses in ﬁnancial economics”, Journal of the Royal Statistical Society, B, 63, 167–241. [3] Barndorff-Nielsen, O.E., Nicolata, E. and Shephard, N. (2002), “Some recent developments in stochastic volatility modelling”, Quantitative Finance, 2, 11–23. [4] Bertoin, J. (1996), L´evy Processes, Cambridge Tracts in Mathematics, 121, Cambridge University Press, Cambridge, UK. [5] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities”, Journal of Political Economy, 81, 637–654. [6] Carr, P. and Lewis, K. (2004), “Corridor variance swaps”, Risk Magazine, 17(2), 67–72. [7] Carr, P. and Madan, D. (1998), “Option valuation using the fast fourier transform”, Journal of Computational Finance, 2, 61–73. [8] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2001), Stochastic Volatility for L´evy Processes, Pr´epublications du Laboratoire de Probabilit´es et Mod`eles Al´eatoires, 645, Universit´es de Paris 6 and Paris 7, Paris, France. [9] Clark, P. (1973), “A subordinated stochastic process model with ﬁnite variance for speculative prices”, Econometrica, 41, 135–156. [10] Corcuera, J.M., Nualart, D. and Schoutens, W. (2005a), “Completion of a L´evy market by powerjump assets”, Finance and Stochastics, 9, 109–127. [11] Corcuera, J.M., Nualart, D. and Schoutens, W. (2005b), “Moment derivatives and L´evy-type market completion”, in A.E. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 169–193. [12] Cox, J., Ingersoll, J. and Ross, S. (1985), “A Theory of the term structure of interest rates”, Econometrica, 53, 385–408. [13] Devroye, L. (1986), Non-Uniform Random Variate Generation, Springer-Verlag, New York. [14] Heston, S. (1993), “A closed-form solution for options with stochastic volatility with applications to bond and currency options”, Review of Financial Studies, 6, 327–343. [15] Hirsa, A., Courtadon, G. and Madan, B.D. (2003), “The effect of model risk on the valuation of barrier options”, Journal of Risk Finance, Winter, 1–8. [16] Hull, J. and Suo, W. (2001), A Methodology for Assessing Model Risk and its Application to the Implied Volatility Function Model, Working Paper, University of Toronto, Toronto, ON, Canada, July. [17] J¨ackel, P. (2002), Monte Carlo Methods in Finance, Wiley, Chichester, UK. [18] Knudsen, Th. and Nguyen-Ngoc, L. (2003), Pricing European Options in a Stochastic VolatilityJump-Diffusion Model, DBQuant Working Paper, Deutsche Bank, London, UK; Journal of Financial and Quantitative Analysis, to be published. [19] Madan, D.B., Carr, P. and Chang, E.C. (1998), “The variance gamma process and option pricing”, European Finance Review, 2, 79–105. [20] Nualart, D. and Schoutens W. (2000), “Chaotic and predictable representations for L´evy processes”, Stochastic Processes and their Applications, 90, 109–122. [21] Nualart, D. and Schoutens W. (2001), “Backwards stochastic differential equations and feynman–Kac formula for L´evy processes with applications in ﬁnance”, Bernoulli, 7, 761–776.

Model Risk for Exotic and Moment Derivatives

97

[22] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [23] Schoutens, W., Simons E. and Tistaert, J. (2004), “A perfect calibration! Now what?”, Wilmott Magazine, March. [24] Wilmott, P. (2002), “Cliquet options and volatility models”, Wilmott Magazine, December.

5 Symmetries and Pricing of Exotic Options in L´evy Models Ernst Eberlein and Antonis Papapantoleon University of Freiburg, Freiburg, Germany Abstract Standard models fail to reproduce observed prices of vanilla options because implied ´ volatilities exhibit a term structure of smiles. We consider time-inhomogeneous Levy processes to overcome these limitations. Then the scope of this paper is two-fold. On the one hand, we apply measure changes in the spirit of Geman et al., to simplify the valuation problem for various options. On the other hand, we discuss a method for the valuation of ´ models. European options and survey valuation methods for exotic options in Levy

5.1

INTRODUCTION

The efforts to calibrate standard Gaussian models to the empirically observed volatility surfaces very often do not produce satisfactory results. This phenomenon is not restricted to data from equity markets, but it is observed in interest rate and foreign exchange markets as well. There are two basic aspects to which the classical models cannot respond appropriately: the underlying distribution is not ﬂexible enough to capture the implied volatilities either across different strikes or across different maturities. The ﬁrst phenomenon is the so-called volatility smile and the second one the term structure of smiles; together they lead to the volatility surface, a typical example of which can be seen in Figure 5.1. One way to improve the calibration results is to use stochastic volatility models; let us just mention Heston (1993) for a very popular model, among the various stochastic volatility approaches. A fundamentally different approach is to replace the driving process. L´evy processes offer a large variety of distributions that are capable of ﬁtting the return distributions in the real world and the volatility smiles in the risk-neutral world. Nevertheless, they cannot capture the term structure of smiles adequately. In order to take care of the change of the smile across maturities, one has to go a step further and consider time-inhomogeneous L´evy processes – also called additive processes – as the driving processes. For term structure models this approach was introduced in Eberlein et al. (2004) and further investigated in Eberlein and Kluge (2004), where cap and swaption volatilities were calibrated quite successfully. As far as plain vanilla options are concerned, a number of explicit pricing formulas is available for L´evy-driven models, one of which is also discussed in this article. The situation is much more difﬁcult in the case of exotic options. The aim of this paper is to derive symmetries and to survey valuation methods for exotic options in L´evy models. By symmetries, we mean a relationship between pricing formulae for options of different Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

100

Exotic Option Pricing and Advanced L´evy Models

14

Implied volatility (%)

13.5 13 12.5 12 11.5 11 10.5 10 10 20 30 40 50 60 70 80 Delta (%) or strike 90

1

2

3

4

5

6

8 7 Maturity

9

10

Figure 5.1 Implied volatilities of vanilla options on the Euro/Dollar rate: spot, 0.93; date, 5 November 2001. Data available at http://www.Mathfinance.de/FF/sampleinputdata.txt

type. Such a relation is of particular interest if it succeeds to derive the value of a complex payoff from that of a simpler one. A typical example is Theorem 5.1 (see below), where a ﬂoating strike Asian or lookback option can be priced via the formula for a ﬁxed strike Asian or lookback option. Moreover, some symmetries are derived in situations where a put-call parity is not available. The discussion here is rather general as far as the class of time-inhomogeneous L´evy processes is concerned. For implementation of these models, a very convenient class are the processes generated by the Generalized Hyperbolic distributions (cf. Eberlein and Prause (2002)). The paper is organized as follows: in the next section, we present time-inhomogeneous L´evy processes, the asset price model and some useful results. In Section 5.3, we describe a method for exploring symmetries in option pricing. The next section contains symmetries and valuation methods for vanilla options while exotic options are tackled in the following section. Finally, in Section 5.6 we present symmetries for options depending on two assets.

5.2

MODEL AND ASSUMPTIONS

Let (, F, F, IP) be a complete stochastic basis in the sense of Jacod and Shiryaev (2003, I.1.3). Let T ∈ R+ be a ﬁxed time horizon and assume that F = FT . We shall consider T ∈ [0, T ]. The class of uniformly integrable martingales is denoted by M; for further notation, we refer the reader to Jacod and Shiryaev (2003). Let D = {x ∈ Rd : |x| > 1}. Following Eberlein et al. (2004), we use as driving process L a timeinhomogeneous L´evy process, more precisely, L = (L1 , . . . , Ld ) is a process with

Exotic Options in L´evy Models

101

independent increments and absolutely continuous characteristics, in the sequel abbreviated PIIAC. The law of Lt is described by the characteristic function t* 1 iu, bs − u, cs u IE eiu,Lt = exp 2 0 + + (eiu,x − 1 − iu, x)λs (dx) ds, (5.2.1) Rd

where bt ∈ Rd , ct is a symmetric non-negative deﬁnite d × d matrix and λt is a L´evy measure on Rd , i.e. it satisﬁes λt ({0}) = 0 and Rd (1 ∧ |x|2 )λt (dx) < ∞ for all t ∈ [0, T ]. The Euclidean scalar product on Rd is denoted by ·, ·, the corresponding norm by | · | while · denotes a norm on the set of d × d matrices. The transpose of a matrix or vector v is denoted by v and 1 denotes the unit vector, i.e. 1 = (1, . . . , 1) . The process L has c`adl`ag paths and F = (Ft )t∈[0,T ] is the ﬁltration generated by L; moreover, L satisﬁes Assumptions (AC) and (EM) given below. Assumption (AC). Assume that the triplets (bt , ct , λt ) satisfy

T

*

|bt | + ct +

Rd

0

+ (1 ∧ |x|2 )λt (dx) dt < ∞.

Assumption (EM). Assume there exists a constant M > 1, such that the L´evy measures λt satisfy

T 0

expu, xλt (dx)dt < ∞,

∀u ∈ [−M, M]d .

D

Under these assumptions, L is a special semimartingale and its triplet of semimartingale characteristics (cf. Jacod and Shiryaev (2003, II.2.6)) is given by t t t Bt = bs ds, Ct = cs ds, ν([0, t] × A) = λs (dx)ds, (5.2.2) 0

0

0

A

where A ∈ B(Rd ). The triplet of semimartingale characteristics (B, C, ν) completely characterizes the distribution of L. Additionally, L is exponentially special (cf. Kallsen and Shiryaev (2002) pp. 2.12–2.13). We model the asset price process as an exponential PIIAC St = S0 exp Lt 1

d

(5.2.3)

with (S 1 , . . . , S d ) = (S01 eL , . . . , S0d eL ), where the superscript i refers to the i -th coordinate, i ≤ d. We assume that IP is a risk neutral measure, i.e. the asset prices have mean rate i of return µi r − δ i and the auxiliary processes Sti = eδ t Sti , once discounted at the rate r, i are IP-martingales. Here, r is the risk-free rate and δ is the dividend yield of the i -th asset. Notice that ﬁniteness of IE[ ST ] is ensured by Assumption (EM).

102

Exotic Option Pricing and Advanced L´evy Models

The driving process L has the canonical decomposition (cf. Jacod and Shiryaev (2003), II.2.38 and Eberlein et al. (2004)) Lt =

t

t

bs ds +

0

0

cs1/2 dWs +

t Rd

0

x(µL − ν)(ds, dx)

(5.2.4)

1/2

where, ct is a measurable version of the square root of ct , W a IP-standard Brownian motion on Rd , µL the random measure of jumps of the process L and ν(dt, dx) = λt (dx)dt is the IP-compensator of the jump measure µL . Because S is modelled under a risk neutral measure, the drift characteristic B is completely determined by the other two characteristics (C, ν) and the rate of return of the asset. Therefore, the i -th component of Bt has the form Bti =

t

(r − δ i )ds −

0

1 2

t

(cs 1)i ds −

0

t 0

i

Rd

(ex − 1 − x i )ν(ds, dx).

(5.2.5)

In a foreign exchange context, δ i can be viewed as the foreign interest rate. In general, markets modelled by exponential time-inhomogeneous L´evy processes are incomplete and there exists a large class of risk neutral (equivalent martingale) measures. An exception occurs in interest rate models driven by L´evy processes, where – in certain cases – there is a unique martingale measure; we refer to Theorem 6.4 in Eberlein et al. (2004). Eberlein and Jacod (1997) provide a characterization of the class of equivalent martingale measures for exponential L´evy models in the time-homogeneous case; this was later extended to general semimartingales in Gushchin and Mordecki (2002). In this article, we do not dive into the theory of choosing a martingale measure; we rather assume that the choice has already taken place. We refer to Eberlein and Keller (1995), Kallsen and Shiryaev (2002) for the Esscher transform, Frittelli (2000), Fujiwara and Miyahara (2003) for the minimal entropy martingale measure and Bellini and Frittelli (2002) for minimax martingale measures, to mention just a small part of the literature on this subject. A unifying exposition – in terms of f-divergences – of the different methods for selecting an equivalent martingale measure can be found in Goll and R¨uschendorf (2001). Alternatively, one can consider the choice of the martingale measure as the result of a calibration to the smile of the vanilla options market. Hakala and Wystup (2002) describe the calibration procedure in detail; we refer to Cont and Tankov (2004) for a numerically stable calibration method for L´evy driven models. Remark 2.1. In the above setting, we can easily incorporate dynamic interest rates and dividend yields (or foreign and domestic rates). Let Dt denote the domestic and Ft the foreign savings account, respectively; then, they can have the form Dt = exp

t

rs ds 0

and

t

Ft = exp

δs ds 0

and equation (5.2.5) has a similar form, taking rs and δs into account. Remark 2.2. The PIIAC L is an additive process, i.e. a process with independent increments, which is stochastically continuous and satisﬁes L0 = 0 a.s. (Sato (1999) Deﬁnition 1.6).

Exotic Options in L´evy Models

103

Remark 2.3. If the triplet (bt , ct , λt ) is not time-dependent, then the PIIAC L becomes a (homogeneous) L´evy process, i.e. a process with independent and stationary increments (PIIS). In that case, the distribution of L is described by the L´evy triplet (b, c, λ), where λ is the L´evy measure and the compensator of µL becomes a product measure of the form ν = λ ⊗ λ\1 , where λ\1 denotes the Lebesgue measure. In that case, equation (5.2.1) takes the form IE[exp(iu, Lt )] = exp[t · ψ(u)] where 1 ψ(u) = iu, b − u, cu + 2

Rd

(eiu,x − 1 − iu, x)λ(dx)

(5.2.6)

which is called the characteristic exponent of L. Lemma 2.4. For ﬁxed t ∈ [0, T ], the distribution of Lt is inﬁnitely divisible with L´evy triplet (b , c , λ ), given by

b :=

t

bs ds, 0

t

c :=

cs ds, 0

λ (dx) :=

t

λs (dx) ds.

(5.2.7)

0

(The integrals should be understood componentwise.) Proof. We refer to the proof of Lemma 1 in Eberlein and Kluge (2004). Remark 2.5. The PIIACs L1 , . . . , Ld are independent, if and only if, the matrices Ct are diagonal and the L´evy measures λt are supported by the union of the coordinate axes; this follows directly from Exercise 12.10 in Sato (1999) or I.5.2 in Bertoin (1996) and Lemma 2.4. Describing the dependence is a more difﬁcult task; we refer to M¨uller and Stoyan (2002) for a comprehensive exposition of various dependence concepts and their applications. We also refer to Kallsen and Tankov (2004), where a L´evy copula is used to describe the dependence of the components of multidimensional L´evy processes. Remark 2.6. Assumption (EM) is sufﬁcient for all our considerations, but is in general too strong. In the sequel, we will replace (EM), on occasion, by the minimal necessary assumptions. From a practical point of view though, it is not too restrictive to assume (EM), since all examples of L´evy models we are interested in, e.g. the Generalized Hyperbolic model (cf. Eberlein and Prause (2002)), the CGMY model (cf. Carr et al. (2002)) or the Meixner model (cf. Schoutens (2002)), possess moments of all order. We can relate the ﬁniteness of the g-moment of Lt for a PIIAC L and a submultiplicative function g, with an integrability property of its compensator measure ν. For the notions of the g-moment and submultiplicative function, we refer to Deﬁnitions 25.1 and 25.2 in Sato (1999). Lemma 2.7. (g-Moment). Let g be a submultiplicative, locally bounded, measurable function on Rd . Then the following statements are equivalent T

g(x)ν(dt, dx) < ∞ (2) IE g(LT ) < ∞.

(1)

0

D

104

Exotic Option Pricing and Advanced L´evy Models

Proof. The result follows from Theorem 25.3 in Sato (1999) combined with Lemma 6 in Eberlein and Kluge (2004). Now, since g(x) = expu, x is a submultiplicative function, we immediately get the following equivalence concerning Assumption (EM). Corollary 2.8. Let M > 1 be a constant. Then the following statements are equivalent T

expu, xν(dt, dx) < ∞, ∀u ∈ [−M, M]d D (2) IE expu, LT < ∞, ∀u ∈ [−M, M]d . (1)

0

We can describe the characteristic triplet of the dual of a one-dimensional PIIAC in terms of the characteristic triplet of the original process. First, we introduce some necessary notation and the next lemma provides the result. Notation. We denote by −λt the L´evy measure deﬁned by −λt ([a, b]) := λt ([−b, −a]) for a, b ∈ R, a < b, t ∈ R+ . Thus, −λt is a non-negative measure and the mirror image of the original measure with respect to the vertical axis. For a compensator of the form ν(dt, dx) = λt (dx)dt, we denote by −ν the (non-negative) measure, deﬁned as −ν(dt, dx) := −λt (dx)dt. Whenever we use the symbol “−” in front of a L´evy measure or a compensator, we will refer to measures deﬁned as above. Lemma 2.9 (dual characteristics). Let L be a PIIAC, as described above, with characteristic triplet (B, C, ν). Then L := −L is again a PIIAC with characteristic triplet (B , C , ν ), where B = −B, C = C and ν = −ν. Proof. From the L´evy-Khintchine representation we have that t* + iuLt cs 2 ϕLt (u) = IE e ibs u − u + (eiux − 1 − iux)λs (dx) ds. = exp 2 R 0 We get immediately ϕ−Lt (u) = ϕLt (−u) t* + cs 2 ibs (−u) − u + (ei(−u)x − 1 − i(−u)x)λs (dx) ds = exp 2 R 0 t* + cs 2 = exp i(−bs )u − u + (eiu(−x) − 1 − iu(−x))λs (dx) ds. 2 R 0 (AC). Hence, we can conThen bt = −bt , ct = ct , and λt = −λt clearly satisfy Assumption t t clude that L is also a PIIAC and has characteristics Bt = 0 bs ds = −Bt , Ct = 0 cs ds = Ct and ν (dt, dx) = λt (dx)dt = −ν(dt, dx).

Exotic Options in L´evy Models

5.3

105

GENERAL DESCRIPTION OF THE METHOD

In this section, we give a brief and general description of the method we shall use to explore symmetries in option pricing. The method is based on the choice of a suitable num´eraire and a subsequent change of the underlying probability measure; we refer to Geman et al. (1995) who pioneered this method. The discounted asset price process, corrected for dividends, serves as the num´eraire for a number of cases, in case the option payoff is homogeneous of degree one. Using the num´eraire, evaluated at the time of maturity, as the Radon–Nikodym derivative, we form a new measure. Under this new measure, the num´eraire asset is riskless while all other assets, including the savings account, are now risky. In case the payoff is homogeneous of higher degree, say α ≥ 1, we have to modify the asset price process so that it serves as the num´eraire. As a result, the asset dynamics under the new measure will depend on α as well. We consider three cases for the driving process L and the asset price process(es): (P1): L = L1 is a (1-d) PIIAC, L2 = k is constant and S 1 = S01 exp L1 , S 2 = exp L2 = K; (P2): L = L1 is a (1-d) PIIAC, S 1 = S01 exp L1 and S 2 = h(S 1 ) is a functional of S 1 ; (P3): L = (L1 , L2 ) is a 2-dimensional PIIAC and S i = S0i exp Li , i = 1, 2. Consider a payoff function f : R+ × R+ → R+

(5.3.1)

which is homogeneous of degree α ≥ 1, that is, for κ, x, y ∈ R∗+ f (κx, κy) = κ α f (x, y); for simplicity we assume that α = 1 and later – in the case of power options – we will treat the case of a more general α. According to the general arbitrage pricing theory (Delbaen and Schachermayer (1994, 1998)), the value V of an option on assets S 1 , S 2 with payoff f is equal to its discounted expected payoff under an equivalent martingale measure. Throughout this paper, we will assume that options start at time 0 and mature at T ; therefore we have V = e−rT IE f ST1 , ST2 .

(5.3.2)

We choose asset S 1 as the num´eraire and express the value of the option in terms of this num´eraire, which yields % & f ST1 , ST2 V −rT $= = e IE V S01 S01 % & 2

−rT S 1 e S 1 T = e−δ T IE f 1, T1 . ST e−δ 1 T S01

(5.3.3)

106

Exotic Option Pricing and Advanced L´evy Models

Deﬁne a new measure $ IP via the Radon–Nikodym derivative e−rT ST1 d$ IP = = ηT . dIP e−δ 1 T S01

(5.3.4)

After the change of measure, the valuation problem, under the measure $ IP, becomes + * $ = e−δ 1 T $ (5.3.5) IE f 1, ST1,2 V where we deﬁne the process S 1,2 :=

S2 . S1

loc IP ∼ IP The measures IP and $ IP are related via the density process ηt = IE[ηT |Ft ]; therefore $ and we can apply Girsanov’s theorem for semimartingales (cf. Jacod and Shiryaev (2003, IP. III.3.24)); this will allow us to determine the dynamics of S 1,2 under $ After some calculations, which depend on the particular choice of L2 or S 2 , we can transform the original valuation problem into a simpler one.

5.4 VANILLA OPTIONS These results are motivated by Carr (1994), where a symmetry relationship between European call and put options in the Black and Scholes (1973) and Merton (1973) model was derived. This result was later extended by Carr and Chesney (1996) to American options for the Black–Scholes case and for general diffusion models; see also McDonald and Schroder (1998) and Detemple (2001). This relationship has an intuitive interpretation in foreign exchange markets (cf. Wystup (2002)). Consider the Euro/Dollar market; then a call option on the Euro/Dollar exchange rate St with payoff (ST − K)+ has time-t value Vc (St , K; rd , re ) in dollars and Vc (St , K; rd , re )/St in euros. This euro-call option can also be viewed as a dollar-put option on the Dollar/Euro rate with payoff K(K −1 − ST−1 )+ and time-t value KVp (K −1 , ST−1 ; re , rd ) in euros. Since the processes S and S −1 have the same (Black–Scholes) volatility, by the absence of arbitrage opportunities, their prices must be equal. 5.4.1 Symmetry For vanilla options, the setting is that of (P1): L1 = L is the driving R-valued PIIAC with triplet (B, C, ν), S 1 = exp L1 = S and L2 = k, such that S 2 = ek = K, the strike price of the option. In accordance with the standard notation, we will use σs2 instead of cs , which corresponds to the volatility in the Black–Scholes model. Therefore, the characteristic C in equation t (5.2.2) has the form Ct = 0 σs2 ds. We will prove a more general version of Carr’s symmetry, namely a symmetry relating power options; the payoff of the power call and put option, respectively, is α α and (K − ST )+ (ST − K)+ where α ∈ N (more generally α ∈ R). We introduce the following notation for the value of a power call option with strike K and power index α α Vc (S0 , K, α; r, δ, C, ν) = e−rT IE (ST − K)+

Exotic Options in L´evy Models

107

where the asset is modelled as an exponential PIIAC according to equations (5.2.3)–(5.2.5) and x + = max{x, 0}. Similarly, for a power put option we set α Vp (S0 , K, α; r, δ, C, ν) = e−rT IE (K − ST )+ . Of course, for α = 1 we recover the European plain vanilla option and the power index α will be omitted from the notation. Assumption (EM) can be replaced by the following weaker assumption, which is the minimal condition necessary for the symmetry results to hold. Let D+ = D ∩ R+ and D− = D ∩ R− . Assumption (M). The L´evy measures λt of the distribution of Lt satisfy 0

T

|x|λt (dx)dt < ∞

T

xeαx λt (dx)dt < ∞.

and

D−

0

D+

Theorem 4.1. Assume that (M) is in force and the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the power call and put option via the following symmetry: ∗

Vc (S0 , K, α; r, δ, C, ν) = K α S0α CT eαCT Vp (S0−1 , K, α; δ, r, C, −f ν)

(5.4.1)

where the constants C and C∗ are given by equations (5.4.3) and (5.4.10), respectively (see ∗ below), K = K −1 e−CT and f (x) = eαx . Proof. First, we note that [e(δ−r)t St ]α = S0α exp(α(δ − r)t + αLt ) is not a IP-martingale; we denote by Lα the martingale part of the exponent; hence Lαt =

t

ασs dWs +

0

t 0

R

αx(µL − ν)(ds, dx).

Since Lα is exponentially special, with Theorem 2.18 in Kallsen and Shiryaev (2002) we have that its exponential compensator, denoted CLα , has the form CLαt =

1 2

0

t

α 2 σs2 ds +

t 0

R

(eαx − 1 − αx)ν(ds, dx)

and exp(Lα − CLα ) ∈ M. The price of the power call option expressed in units of the num´eraire yields −rT + α $c := Vc = e V α α IE (ST − K) S0 S0 −rT α α * +α e ST K −1 + −1 (K = e−δT IE − S ) T e−δT S0α

108

Exotic Option Pricing and Advanced L´evy Models

= e−δT K α IE exp (δ − r)T + α * α

× exp LαT − CLT =e

−δT

K CT IE exp α

T

0

(K −1 − ST−1 )+

LαT

−

CLαT

*

bs ds + CLαT

+α

(K

−1

−

ST−1 )+

+α (5.4.2)

where, by using equations (5.2.5) and (5.2.2), we have that log CT = (δ − r)T + αBT + CLαT = (α − 1)(r − δ)T + + 0

T

R

α(α − 1) 2

T 0

σs2 ds

(eαx − αex + α − 1)ν(ds, dx).

(5.4.3)

Deﬁne a new measure $ IP via its Radon–Nikodym derivative d$ IP = exp LαT − CLαT = ηT dIP and the valuation problem (equation (5.4.2)) becomes α $c = e−δT K α CT I$ $−$ V E (K ST )+

(5.4.4)

(5.4.5)

$ = K −1 and $ St := St−1 . where K Since the measures IP and $ IP are related via the density process (ηt ), which is a positive loc martingale with η0 = 1, we immediately deduce that $ IP ∼ IP and we can apply Girsanov’s theorem for semimartingales (cf. Jacod and Shiryaev (2003) III.3.24). The density process can be represented in the usual form $ dIP Ft = exp Lαt − CLαt ηt = IE dIP t t = exp ασs dWs + αx(µL − ν)(ds, dx) 0

1 − 2

0

0 t

α 2 σs2 ds −

R

t 0

R

(eαx − 1 − αx)ν(ds, dx) .

(5.4.6)

Consequently, we can identify the tuple (β, Y ) of predictable processes β(t) = α and Y (t, x) = exp(αx) which characterizes the change of measure. From Girsanov’s theorem, combined with Theorem II.4.15 in Jacod and Shiryaev (2003), we deduce that a PIIAC remains a PIIAC under the measure $ IP, because the processes β and Y are deterministic and the resulting characteristics under $ IP satisfy Assumption (AC).

Exotic Options in L´evy Models

109

$ t As a consequence of Girsanov’s theorem for semimartingales, we infer that Wt = Wt − $ $ ν = Y ν is the IP compensator of the jumps of L. 0 ασs ds is a IP-Brownian motion and $ Furthermore, as a corollary of Girsanov’s theorem, we can calculate the canonical decomposition of L under $ IP; t t t $ $ σs dWs + x(µL − $ ν)(ds, dx) (5.4.7) bs ds + Lt = 0

0

0

R

where $t = B

1 t 2 $ σ ds bs ds = (r − δ)t + α − 2 0 s 0 t (e−αx − e(1−α)x + x)$ ν(ds, dx) +

t

R

0

(5.4.8)

$ C,$ and hence, its triplet of characteristics is (B, ν). Deﬁne its dual process, L := −L $ C, −$ ν). The canonical and by Lemma 2.9, we get that its triplet is (B , C , ν ) = (−B, decomposition of L is t t t $ σs dWs + x(µL − ν )(ds, dx) (5.4.9) bs ds + Lt = − 0

0

0

R

IP-martingale for α = 1. and we can easily deduce that e(r−δ)t St is not a $ Adding the appropriate terms, we can re-write L as L := C∗ + L, where · · σs2 ds − (e−αx − e(1−α)x + 1 − e−x )$ ν(ds, dx) C∗ = (1 − α) 0

0

R

(5.4.10)

and L is such that e(r−δ)t S t is a $ IP-martingale. The characteristic triplet of L is (B − −1 ∗ C , C, ν ) and S t = S0 exp Lt . Therefore, we can conclude the proof α $−$ $c = e−δT K α CT $ IE (K ST )+ V +α * $ − eC∗T S T )+ = e−δT K α CT $ IE (K α ∗ = e−δT K α CT eαCT $ IE (K − S T )+ ∗

∗

$ −CT = K −1 e−CT . where K = Ke Setting α = 1 in the previous theorem, we immediately get a symmetry between European plain vanilla call and put options. Corollary 4.2. Assuming that (M) is in force and the asset price evolves as an exponential PIIAC, we can relate the European call and put option via the following symmetry: (5.4.11) Vc (S0 , K; r, δ, C, ν) = KS0 Vp S0−1 , K −1 ; δ, r, C, −f ν where f (x) = ex .

110

Exotic Option Pricing and Advanced L´evy Models

This symmetry relating European and also American plain vanilla call and put options, in exponential L´evy models, was proved independently in Fajardo and Mordecki (2003). Schroder (1999) proved similar results in a general semimartingale model; however, using a L´evy or PIIAC as the driving motion allows for the explicit calculation of the distribution under the new measure. A different symmetry, again relating European and American call and put options, in the Black–Scholes model, was derived by Peskir and Shiryaev (2002), where they use the mathematical concept of negative volatility; their main result states that Vc (ST , K; σ ) = Vp (−ST , −K; −σ ).

(5.4.12)

See also the discussion – and the corresponding cartoon – in Haug (2002). In this framework, one can derive symmetry relationships between self-quanto and European plain vanilla options. This result is, of course, a special case of Theorem 6.4; nevertheless, we give a short proof since it simpliﬁes considerably because the driving process is one-dimensional. The payoff of the self-quanto call and put option is ST (ST − K)+

ST (K − ST )+ ,

and

respectively. Introduce the following notation for the value of the self-quanto call option Vqc (S0 , K; r, δ, C, ν) = e−rT IE ST (ST − K)+ and similarly, for the self-quanto put option we set Vqp (S0 , K; r, δ, C, ν) = e−rT IE ST (K − ST )+ . Assumption (EM) can be replaced by the following weaker assumption, which is the minimal condition necessary for the symmetry results to hold. Assumption (M ). The L´evy measures λt of the distribution of Lt satisfy

T 0

|x|λt (dx)dt < ∞

T

e2x λt (dx)dt < ∞.

and

D−

0

D+

Theorem 4.3. Assume that the asset price evolves as an exponential PIIAC and (M ) is in force. We can relate the self-quanto and European plain vanilla call and put options via the following symmetry: ∗

Vqc (S0 , K; r, δ, C, ν) = S0 eCT Vc (S0 , K ∗ ; δ, r, C, f ν) C∗T

Vqp (S0 , K; r, δ, C, ν) = S0 e Vp (S0 , K ∗ ; δ, r, C, f ν)

(5.4.13) (5.4.14)

∗

where C ∗ is given by equation (5.4.16) (see below), K ∗ = Ke−CT and f (x) = ex . Proof. Expressing the value of the self-quanto call option in units of the num´eraire as described in Section 5.3, we deﬁne a new measure $ IP via its Radon–Nikodym derivative

Exotic Options in L´evy Models

111

given by equation (5.3.4) and the original valuation problem becomes $qc = e−δT $ IE (ST − K)+ . V

(5.4.15)

Now it sufﬁces to calculate the characteristic triplet of L under $ IP. Arguing as in the proof of Theorem 4.1, the density process η has the form of equation (5.4.6) for α = 1; hence, the tuple (β, Y ) of predictable processes that describes the change of measure is β(t) = 1 and Y (t, x) = exp(x). Therefore, L has the canonical decomposition under $ IP

t

Lt =

$ bs ds +

0

t

$s + σs dW

t

0

R

0

x(µL − $ ν)(ds, dx)

where σ2 $ bt = r − δ + t + 2

R

(e−x − 1 + x)ex λt (dx).

IP-martingale, but if we deﬁne L∗ as Notice that e(r−δ)t eLt is not a $ L∗t := (δ − r)t +

t

− 0

t 0

σt2 ds − 2

$s + σs dW

R

0

t 0

t

R

x(µL − $ ν)(ds, dx)

(ex − 1 − x)ex ν(ds, dx)

∗

then e(r−δ)t eLt ∈ M. Next, we re-express L as L = L∗ + C ∗ , where CT∗

= exp 2(r − δ)T +

T 0

σs2 ds

T

+ 0

(e + e x

R

−x

− 2)e ν(ds, dx) . x

(5.4.16)

By re-arranging the terms in equation (5.4.15), the result follows. 5.4.2 Valuation of European options We outline a method for the valuation of vanilla options, based on bilateral Laplace transforms, that was developed in the PhD thesis of Sebastian Raible (see Raible (2000) Chap. 3). The method is extremely fast and allows for the valuation not only of plain vanilla European derivatives, but also of more complex payoffs, such as digital, self-quanto and power options; in principle, every European payoff can be priced by using this method. Moreover, a large variety of driving processes can be handled, including L´evy and additive processes. The main idea of Raible’s method is to represent the option price as a convolution of two functions and consider its bilateral Laplace transform; then, by using the property that the Laplace transform of a convolution equals the product of the Laplace transforms of the factors, we arrive at two Laplace transforms that are easier to calculate analytically than the original one. Inverting this Laplace transform yields the option price.

112

Exotic Option Pricing and Advanced L´evy Models

A similar method, in Fourier space, can be found in Lewis (2001). See also Carr and Madan (1999) for some preliminary results that motivated this research. Lee (2004) uniﬁes and generalizes the existing Fourier-space methods and develops error bounds for the discretized inverse transforms. We ﬁrst state the necessary assumptions regarding the distribution of the asset price process and the option payoff respectively. (L1): Assume that ϕLT (z), the extended characteristic function of LT , exists for all z ∈ C with z ∈ I1 ⊃ [0, 1]. (L2): Assume that IPLT , the distribution of LT , is absolutely continuous w.r.t. the Lebesgue measure λ\1 with density ρ. (L3): Consider a European-style payoff function f (ST ) that is integrable. (L4): Assume that x → e−Rx |f (e−x )| is bounded and integrable for all R ∈ I2 ⊂ R. In order to price a European option with payoff function f (ST ), we proceed as follows. −rT −rT V = e IE[f (ST )] = e f (ST )dIP = e−rT = e−rT

R

R

f (S0 ex )dIPLT (x) f (S0 ex )ρ(x)dx

(5.4.17)

because of absolute continuity. Deﬁne ζ = − log S0 and g(x) = f (e−x ), and then −rT g(ζ − x)ρ(x)dx = e−rT (g ∗ ρ)(ζ ) (5.4.18) V =e R

which is a convolution at point ζ . Applying bilateral Laplace transforms on both sides of equation (5.4.18) and using Theorem B.2 in Raible (2000), we get LV (z) = e−rT e−zx (g ∗ ρ)(x)dx R

= e−rT

R

e−zx g(x)dx

= e−rT Lg (z)Lρ (z)

R

e−zx ρ(x)dx (5.4.19)

Lh (z) denotes the bilateral Laplace transform of a function h at z ∈ C, i.e. Lh (z) := where −zx h(x)dx. The Laplace transform of g is very easy to compute analytically and the e R Laplace transform of ρ can be expressed as the extended characteristic function ϕLT of LT . By numerically inverting this Laplace transform, we recover the option price. The next theorem gives us an explicit expression for the price of an option with payoff function f and driving PIIAC L. Theorem 4.4. Assume that (L1)–(L4) are in force and let g(x) := f (e−x ) denote the modiﬁed payoff function of an option with payoff f (x) at time T . Assume that I1 ∩ I2 = ∅

Exotic Options in L´evy Models

113

and choose an R ∈ I1 ∩ I2 . Letting V (ζ ) denote the price of this option, as a function of ζ := − log S0 , we have eζ R−rT V (ζ ) = 2π

R

eiuζ Lg (R + iu)ϕLT (iR − u)du,

(5.4.20)

whenever the integral on the right-hand side exists. Proof. The claim can be proved by using the arguments of the proof of Theorem 3.2 in Raible (2000); there, no explicit statement is made about the driving process L; hence, it directly transfers to the case of a time-inhomogeneous L´evy process. Remark 4.5. In order to apply this method, validity of the necessary assumptions has to be veriﬁed. (L1), (L3) and (L4) are easy to certify, while (L2) is the most demanding one. Let us mention that the distributions underlying the most popular L´evy processes, such as the Generalized Hyperbolic L´evy motion (cf. Eberlein and Prause (2002)), possess a known Lebesgue density. Remark 4.6. The method of Raible for the valuation of European options can be applied to general driving processes that satisfy Assumptions (L1)–(L4). Therefore, it can also be applied to stochastic volatility models based on L´evy processes that have attracted much interest lately; we refer to Barndorff-Nielsen and Shephard (2001), Eberlein et al. (2003) and Carr et al. (2003) for an account of different models. 5.4.3 Valuation of American options The method of Raible, presented in the previous section, can be used for pricing several types of European derivatives, but not path-dependent ones. The valuation of American options in L´evy-driven models is quite a hard task and no analytical solution exists for the ﬁnite horizon case. For perpetual American options, i.e. options with inﬁnite time horizon, Mordecki (2002) derived formulae in the general case in terms of the law of the extrema of the L´evy process, using a random walk approximation to the process. He also provides explicit solutions for the case of a jump-diffusion with exponential jumps. Alili and Kyprianou (2005) recapture the results of Mordecki by making use of excursion theory. Boyarchenko and Levendorskiˇı (2002c) obtained formulae for the price of the American put option in terms of the Wiener–Hopf factors and derive some more explicit formulae for these factors. Asmussen et al. (2004) ﬁnd explicit expressions for the price of American put options for L´evy processes with two-sided phase-type jumps; the solution uses the Wiener–Hopf factorization and can also be applied to regime-switching L´evy processes with phase-type jumps. For the valuation of ﬁnite time horizon American options, one has to resort to numerical methods. Denote by x = ln S the log price, τ = T − t the time to maturity and v(τ, x) = f (ex , T − τ ) the time-t value of an option with payoff function g(ex ) = φ(x). One approach is to use numerical schemes for solving the corresponding partial integro-differential inequality (PIDI), ∂v − Av + rv ≥ 0 ∂τ

in (0, T ) × R

(5.4.21)

114

Exotic Option Pricing and Advanced L´evy Models

subject to the conditions a.e. in [0, T ] × R v(τ, x) ≥ φ(x), ∂v − Av + rv = 0, in (0, T ) × R (v(τ, x) − φ(x)) ∂τ v(0, x) = φ(x)

(5.4.22)

where σ 2 d2 v σ 2 dv + Av(x) = r − δ − 2 dx 2 dx 2 dv v(x + y) − v(x) − (ey − 1) (x) λ(dy) + dx R

(5.4.23)

is the inﬁnitesimal generator of the transition semigroup of L; see Matache et al. (2003, 2005) for all of the details and numerical solution of the problem using wavelets. Almendral (2004) solves the problem numerically by using implicit–explicit methods in case the CGMY is the driving process. Equation (5.4.21) is a backward PIDE in spot and time to maturity; Carr and Hirsa (2003) develop a forward PIDE in strike and time of maturity and solve it by using ﬁnite-difference methods. Another alternative is to employ Monte Carlo methods adapted for optimal stopping problems, such as the American option; we refer here to Rogers (2002) or Glasserman (2003). K¨ellezi and Webber (2004) constructed a lattice for L´evy-driven assets and applied it to the valuation of Bermudan options. Levendorskiˇı (2004) develops a non-Gaussian analog of the method of lines and uses Carr’s randomization method in order to formulate an approximate algorithm for the valuation of American options. Chesney and Jeanblanc (2004) revisit the perpetual American problem and obtain formulae for the optimal boundary when jumps are either only positive or only negative. Using these results, they approximate the ﬁnite horizon problem in a fashion similar to Barone-Adesi and Whaley (1987). Empirical tests show that this approximation provides good results only when the process is continuous at the exercise boundary.

5.5 EXOTIC OPTIONS The work on this topic follows along the lines of Henderson and Wojakowski (2002); they proved an equivalence between the price of ﬂoating and ﬁxed strike Asian options in the Black–Scholes model. We also refer to Vanmaele et al. (2002) for a generalization of these results to forward-start options and discrete averaging in the Black–Scholes model. 5.5.1 Symmetry For exotic options, the setting is that of (P2): L1 = L is the driving R-valued PIIAC with triplet (B, C, ν), S 1 = S01 exp L1 = S and S 2 = h(S) is a functional of S. The most prominent candidates for functionals are the maximum, the minimum and the (arithmetic) average; let 0 = t1 < t2 < · · · < tn = T be equidistant time points, and then the resulting processes, in case of discrete monitoring, are 1 Sti . n n

MT = max Sti , 0≤ti ≤T

NT = min Sti 0≤ti ≤T

and "T =

i=1

Exotic Options in L´evy Models

115

Table 5.1 Types of payoffs for Asian and lookback options Option type

Asian payoff

Lookback payoff

Fixed strike call Fixed strike put Floating strike call Floating strike put

("T − K)+ (K − "T )+ (ST − "T )+ ("T − ST )+

(MT − K)+ (K − NT )+ (ST − NT )+ (MT − ST )+

Therefore, we can exploit symmetries between ﬂoating and ﬁxed strike Asian and lookback options in this framework; the different types of payoffs of the Asian and lookback option are summarized in Table 5.1. We introduce the following notation for the value of the ﬂoating strike call option, be it Asian or lookback Vc (ST , h(S); r, δ, C, ν) = e−rT IE (ST − h(S)T )+ and similarly, for the ﬁxed strike put option we set Vp (K, h(S); r, δ, C, ν) = e−rT IE (K − h(S)T )+ ; a similar notation will be used for the other two cases. Now we can state a result that relates the value of ﬂoating and ﬁxed strike options. Notice that because stationarity of the increments plays an important role in the proof, the result is valid only for L´evy processes. Theorem 5.1. Assuming that the asset price evolves as an exponential L´evy process, we can relate the ﬂoating and ﬁxed strike Asian or lookback option via the following symmetry: Vc ST , h(S); r, δ, σ 2 , λ = Vp S0 , h(S); δ, r, σ 2 , −f λ Vp h(S), ST ; r, δ, σ 2 , λ = Vc h(S), S0 ; δ, r, σ 2 , −f λ

(5.5.1) (5.5.2)

where f (x) = ex . Proof. We refer to the proof of Theorems 3.1 and 4.1 in Eberlein and Papapantoleon (2005). The minimal assumptions necessary for the results to hold are also stated there. Remark 5.2. These results also hold for forward-start Asian and lookback options, for continuously monitored options, for partial options and for Asian options on the geometric and harmonic average; see Eberlein and Papapantoleon (2005) for all of the details. Note that the equivalence result is not valid for in-progress Asian options. 5.5.2 Valuation of barrier and lookback options The valuation of barrier and lookback options for assets driven by general L´evy processes is another hard mathematical problem. The difﬁculty stems from the fact that (a) the distribution of the supremum or inﬁmum of a L´evy process is not known explicitly, and (b) the

116

Exotic Option Pricing and Advanced L´evy Models

overshoot distribution associated with the passage of a L´evy process across a barrier is also not known explicitly. Various authors have treated the problem in the case where the driving process is a spectrally positive/negative L´evy process; see, for example, Rogers (2000), Sch¨urger (2002) and Avram et al. (2004). Kou and Wang (2003, 2004) have derived explicit formulas for the values of barrier and lookback options in a jump diffusion model where the jumps are double-exponentially distributed; they make use of a special property of the exponential distribution, namely the memoryless property, which allows them to explicitly calculate the overshoot distribution. Lipton (2002) derives similar formulas for the same model, making use of ﬂuctuation theory. Fluctuation theory and the Wiener–Hopf factorization of L´evy processes play a crucial role in every attempt to derive closed form solutions for the value of barrier and lookback options in L´evy-driven models. Introduce the notation Mt = sup Ls

Nt = inf Ls

and

0≤s≤t

0≤s≤t

and let θ denote a random variable exponentially distributed with parameter q, independent of L. Then, the celebrated Wiener–Hopf factorization of the L´evy process L states that IE[exp(izLθ )] = IE[exp(izMθ )] · IE[exp(izNθ )]

(5.5.3)

or equivalently q(q − ψ(z))−1 = ϕq+ (z) · ϕq− (z),

z ∈ R,

(5.5.4)

where ψ denotes the characteristic exponent of L. The functions ϕq+ and ϕq− have the following representations ϕq+ (z) ϕq− (z)

= exp = exp

*

∞

t *

−1 −qt

e

0 ∞

t 0

−1 −qt

e

∞

dt dt

(eizx − 1)µt (dx)

+ (5.5.5)

0 0 −∞

(eizx − 1)µt (dx)

+ (5.5.6)

where µt (dx) = IPLt (dx) is the probability measure of Lt . These results were ﬁrst proved for L´evy processes in Bingham (1975) – where an approximation of L´evy processes by random walks is employed – and subsequently by Greenwood and Pitman (1980) – where excursion theory is applied. See also the recent books by Sato (1999, Chapter 9) and Bertoin (1996, Chapter VI) respectively, for an account of these two methods. Building upon these results, various authors have derived formulae for the valuation of barrier and lookback options; Boyarchenko and Levendorskiˇı (2002a) apply methods from potential theory and pseudodifferential operators to derive formulae for barrier and touch options, while Nguyen-Ngoc and Yor (2005) use a probabilistic approach based on excursion theory. Recently, Nguyen-Ngoc (2003) takes a similar probabilistic approach, motivated from Carr and Madan (1999) and derives quite simple formulae for the value of barrier and lookback options, which can be numerically evaluated with the use of Fourier inversion algorithms in two and three dimensions.

Exotic Options in L´evy Models

117

More speciﬁcally, let us denote by Vc (MT , K; T ) the price of a ﬁxed strike lookback option with payoff (MT − K)+ , where MT = max0≤t≤T St and S is an exponential L´evy α,γ process. Choose γ > 1 and α > 0 such that IE[e2L1 ] < er+α and set Vc (MT , K; T ) = −αT −γ k Vc (MT , K; T ) where k = log(K/S0 ). Then, we have the following result. e Proposition 5.3. If k > 0, then for all q, u > 0 we have: ∞ ∞ e−qT dT e−uk Vcα,γ (MT , S0 ek ; T )dk 0

0

1 1 + [ϕ + (i(z − 1)) + (z − 1)ϕq+r+α (−i) − z] = S0 q + r + α z(z − 1) q+r+α

(5.5.7)

where z = u + γ . Proof. We refer to the proof of Proposition 3.9 in Nguyen-Ngoc (2003). The formula for the value of the ﬂoating strike lookback option is – as one could easily foresee – a lot more complicated than equation (5.5.7). Using the symmetry result of Theorem 5.1, this case can be dealt with via a change of the L´evy triplet and strike in the previous proposition. The Wiener–Hopf factors are not known explicitly in the general case and numerical computation could be extremely time-consuming. Boyarchenko and Levendorskiˇı (2002b) provide some more efﬁcient formulas for the Wiener–Hopf factors of – what they call – regular L´evy processes of exponential type (RLPE); for the deﬁnition refer to Section 1.2.2 in the above-mentioned reference. Given that L is an RLPE, ϕq+ (z) has an analytic continuation on the half plane z > ω and * z +∞+iω ln(q + ψ(u)) + du . (5.5.8) ϕq+ (z) = exp 2π i −∞+iω u(z − u) The family of RLPEs contains many popular – in mathematical ﬁnance – L´evy motions such as the Generalized Hyperbolic and Variance Gamma models (see Boyarchenko and Levendorskiˇı (2002b)). Discretely monitored options have received much less attention in the literature than their continuous time counterparts. Borovkov and Novikov (2002) use Fourier methods and Spitzer’s identity to derive formulae for ﬁxed strike lookback options. Various numerical methods have been applied for the valuation of barrier and lookback options in L´evy-driven models. Cont and Voltchkova (2005a, 2005b) study ﬁnite-difference methods for the solution of the corresponding PIDE (see also Matache et al. (2004)). Ribeiro and Webber (2003, 2004) have developed fast Monte Carlo methods for the valuation of exotic options in models driven by the Variance Gamma (VG) and Normal Inverse Gaussian (NIG) L´evy motions; their method is based on the construction of Gamma and Inverse Gaussian bridges, respectively, to speed up the Monte Carlo simulation. The recent book of Schoutens (2003) contains a detailed account of Monte Carlo methods for L´evy processes, also allowing for stochastic volatility. 5.5.3 Valuation of Asian and basket options An explicit solution for the value of the arithmetic Asian or basket option is not known in the Black–Scholes model and, of course, the situation is similar for L´evy models. The difﬁculty

118

Exotic Option Pricing and Advanced L´evy Models

is that the distribution of the arithmetic sum of log-normal random variables – more generally, random variables drawn from some log r inﬁnitely divisible distribution – is not known in closed-form. Veˇceˇr and Xu (2004) formulated a PIDE for all types of Asian options – including inprogress options – in a model driven by a process with independent increments (PII) or, more generally, a special semimartingale. Their derivation is based on the construction of a suitable self-ﬁnancing trading strategy to replicate the average and then a change of num´eraire – which is essentially the one we use – in order to reduce the number of variables in the equation. Their PIDE is relatively simple and can be solved by using numerical techniques such as ﬁnite-differences. Albrecher and Predota (2002, 2004) use moment-matching methods to derive approximate formulae for the value of Asian options in some popular L´evy models such as the NIG and VG models; they also derive bounds for the option price in these models. See also the survey paper by Albrecher (2004) for a detailed account of the above mentioned results. Hartinger and Predota (2002) apply Quasi Monte-Carlo methods for the valuation of Asian options in the Hyperbolic model. Their method can be extended to the class of Generalized Hyperbolic L´evy motions, which contains the VG motion as a special case; see Eberlein and von Hammerstein (2004). Benhamou (2002), building upon the work of Carverhill and Clewlow (1992), uses the Fast Fourier transform and a transformation of dependent variables into independent ones, in order to value discretely monitored ﬁxed strike Asian options. As he points out, this method can be applied when the return distribution is fat-tailed, with L´evy processes being prominent candidates. Henderson et al. (2004) derive an upper bound for in-progress ﬂoating strike Asian options in the Black–Scholes model, using the symmetry result of Henderson and Wojakowski (2002) and valuation methods for ﬁxed strike ones. Their pricing bound relies on a modeldependent symmetry result and a model-independent decomposition of the ﬂoating-strike Asian option into a ﬁxed-strike one and a vanilla option. Therefore, given the symmetry result of Theorem 5.1, their general methodology can also be applied to L´evy models. Albrecher et al. (2004) derive static super-hedging strategies for ﬁxed strike Asian options in L´evy models; these results were extended to L´evy models with stochastic volatility in Albrecher and Schoutens (2005). The method is based on super-replicating the Asian payoff with a portfolio of plain vanilla calls, using the following upper bound + n n Stj − nK ≤ (Stj − nKj )+ (5.5.9) j =1

j =1

and then optimizing the hedge, i.e. the choice of Kj s, using results from co-monotonicity theory. Similar ideas appear in Hobson et al. (2004) for the static super-hedging of basket options. The payoff of the basket option is super-replicated by a portfolio of plain vanilla calls on each individual asset, using the upper bound

+ n n + i wi ST − K ≤ (5.5.10) wi STi − li K n

i =1

i =1

where li ≥ 0 and i =1 li = 1; subsequently, the portfolio is optimized using co-monotonicity theory. Moreover, no distribution is assumed about the asset dynamics, since all of the

Exotic Options in L´evy Models

119

information needed is the marginal distributions which can be deduced from the volatility smile; we refer to Breeden and Litzenberger (1978). This is also observed by Albrecher and Schoutens (2005).

5.6 MARGRABE-TYPE OPTIONS In this section, we derive symmetry results between options involving two assets – such as Margrabe or Quanto options – and European plain vanilla options; therefore, we generalize results by Margrabe (1978) and Fajardo and Mordecki (2003) to the case of timeinhomogeneous L´evy processes. Schroder (1999) provides similar results for semimartingale models; the advantage of using a L´evy process or a PIIAC instead of a semimartingale as the driving motion, is that the distribution of the asset returns under the new measure can be deduced from the distribution of the returns of each individual asset under the risk-neutral measure. For Margrabe-type options, the setting is that of (P3): L = (L1 , L2 ) is the driving 2 R -valued PIIAC with triplet (B, C, ν) and S = (S 1 , S 2 ) is the asset price process. For convenience, we set i = 1, 2, (5.6.1) Sti = S0i exp (r − δ i )t + Lit , modifying the characteristic triplet (B, C, ν) accordingly. With Theorem 25.17 in Sato (1999) and Lemma 2.4, Assumption (EM) guarantees the existence of the moment generating function MLt of Lt for u ∈ Cd such that u ∈ [−M, M]d . Furthermore, for u ∈ Cd with u ∈ [−M, M]d , we have that MLt (u) = ϕLt (−iu) = IE eu,Lt t* 1 u, bs + u, cs u = exp 2 0 + + (eu,x − 1 − u, x)λs (dx) ds. (5.6.2) Rd

The next result will allow us to calculate the characteristic triplet of a one-dimensional process, deﬁned as a scalar product of a vector with the d-dimensional process L, from the characteristics of L under an equivalent change of probability measure. Proposition 6.1. Let L be a d-dimensional PIIAC with triplet (B, C, ν) under IP, let u, v be loc IP ∼ IP, with density vectors in Rd and v ∈ [−M, M]d . Moreover, let $ d$ IP ev,LT = . dIP IE[ev,LT ] := u, L is a $ Then, the one-dimensional process L IP-PIIAC and its characteristic triplet is (B, C, ν) with 1 u, x ev,x − 1 λs (dx) bs = u, bs + u, cs v + v, cs u + 2 Rd cs = u, cs u λs = T (κs )

120

Exotic Option Pricing and Advanced L´evy Models

where T is a mapping T : Rd → R, such that x → T (x) = u, x, and κs is a measure deﬁned by ev,x λs (dx). κs (A) = A

Proof. Because the density process (ηt ) is given by ηt = ev,Lt IE[ev,Lt ]−1 , by using equation (5.6.2) we get $ IE ezu,Lt = IE ezu,Lt ηt * −1 + = IE ezu,Lt ev,Lt IE ev,Lt −1 = IE ezu+v,Lt IE ev,Lt t* 1 zu + v, bs + zu + v, cs (zu + v) = exp 2 0 + + (ezu+v,x − 1 − zu + v, x)λs (dx) ds Rd

t

× exp 0

* 1 − v, bs + v, cs v 2 + + (ev,x − 1 − v, x)λs (dx) ds Rd

t* , 1 = exp z u, bs + u, cs v + v, cs u 2 0 v,x - 1 u, x e − 1 λs ds + z2 u, cs u + 2 Rd + zu,x e − 1 − zu, x ev,x λs (dx) ds. + Rd

If we write κs for the measure on Rd given by ev,x λs (dx), κs (A) =

(5.6.3)

(5.6.4)

A

A ∈ B(Rd ) and T for the linear mapping T : Rd → R given by T (x) = u, x, then we get for the last term in the exponent of equation (5.6.3) v,x zu,x zy − 1 − zu, x e λs (dx) = e e − 1 − zy T (κs )(dy) Rd

R

by the change-of-variable formula. The resulting characteristics satisfy Assumption (AC), and thus the result follows. The valuation of options depending on two assets modelled by a two-dimensional PIIAC can now be simpliﬁed – using the technique described in Section 5.3 and Proposition 6.1 – to

Exotic Options in L´evy Models

121

the valuation of an option on a one-dimensional asset. Subsequently, this option can be priced by using bilateral Laplace transforms, as described in Section 5.4.2. The payoff of a Margrabe option, or option to exchange one asset for another, is + 1 ST − ST2 and we denote its value by Vm (S01 , S02 ; r, δ, C, ν) = e−rT IE

*

ST1 − ST2

+ +

where δ = (δ 1 , δ 2 ). The payoff of the Quanto call and put option is + ST1 ST2 − K

and

+ ST1 K − ST2 ,

respectively, and we will use the following notation for the value of the Quanto call option * + + Vqc (S01 , S02 , K; r, δ, C, ν) = e−rT IE ST1 ST2 − K and similarly for the Quanto put option * + + . Vqp (S01 , S02 , K; r, δ, C, ν) = e−rT IE ST1 K − ST2 The different variants of the Quanto option traded in Foreign Exchange markets are explained in Musiela and Rutkowski (1997). The payoff of a cash-or-nothing and a two-dimensional asset-or-nothing option is 1l{ST >K}

and

ST1 1l{S 2 >K} . T

The holder of a two-dimensional asset-or-nothing option receives one unit of asset S 1 at expiration, if asset S 2 ends up in the money; of course, this is a generalization of the (standard) asset-or-nothing option, where the holder receives one unit of the asset if it ends up in the money. We denote the value of the cash-or-nothing option by Vcn (S0 , K; r, δ, C, ν) = e−rT IE 1l{ST >K} and the value of the two-dimensional asset-or-nothing option by + * Van (S01 , S02 , K; r, δ, C, ν) = e−rT IE ST1 1l{S 2 >K} . T

Notice that in the ﬁrst case, r, δ, C and ν correspond to a one-dimensional driving process, while in the second case to a two-dimensional one. Theorem 6.2. Let Assumption (EM) be in force and assume that the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the value of a Margrabe and a European plain vanilla option via the following symmetry: ν Vm (S01 , S02 ; r, δ, C, ν) = IE[ST1 ]eCT Vp S02 /S01 , K; δ 1 , r, C,

(5.6.5)

122

Exotic Option Pricing and Advanced L´evy Models

ν) where K = e−CT , C is given by equation (5.6.9) (see below) and the characteristics (C, are given by Proposition 6.1 for v = (1, 0) and u = (−1, 1).

Proof. Expressing the value of the Margrabe option in units of the num´eraire, we get * −rT + 1 2 + $ := Vm = e S I E − S V T T S01 S01 %

+ & ST2 e−rT ST1 ηT1 −δ 1 T 1− 1 IE =e ST e−δ 1 T S01 ηT1 where η1 = IE[exp(L1 )] = IE[expv, L], for v = (1, 0), and by using equation (5.6.1) we get % =

1 e−δ T ηT1 IE

1

eLT ηT1

+ & ST2 1− 1 . ST

(5.6.6)

Deﬁne a new measure $ IP via its Radon–Nikodym derivative 1 eLT d$ IP = 1 dIP IE[eLT ]

and the valuation problem takes the form $ = e−δ 1 T ηT1 $ IE V

*

1− ST

+ +

where, by using equation (5.6.1), we get S2 1 2 S2 2 1 St := t1 = 01 e(δ −δ )t+Lt −Lt =: S0 exp (δ 1 − δ 2 )t + Lt St S0

(5.6.7)

:= L2 − L1 = u, L for u = (−1, 1). The characteristic triplet of L, (B, C, ν) under and L $ IP, is given by Proposition 6.1 for v = (1, 0) and u = (−1, 1). 1 IP-martingale. However, if we deﬁne St is not a $ Observe that e(r−δ )t t 1 t 1 cs ds − Lt := (δ − r)t − (ex − 1 − x) ν(ds, dx) 2 0 R 0 t t $s + cs1/2 dW x(µL − ν)(ds, dx) (5.6.8) + 0

0

R

$ is a $ where W IP-standard Brownian motion and µL is the random measure of jumps of L, 1 t + then e(r−δ )t eLt ∈ M. Therefore, we re-express the exponent of equation (5.6.7) as L Ct where (δ 1 − δ 2 )t = Lt + t t 1 t 2 cs ds + (ex − 1 − x) ν(ds, dx) (5.6.9) bs ds + Ct = (r − δ )t + 2 0 R 0 0

S0 exp Lt . and deﬁne S t :=

Exotic Options in L´evy Models

123

Now the result follows, because

* + + 1− ST 1 + = e−δ T ηT1 I$ E 1 − S T eCT

$ = e−δ 1 T ηT1 I$ E V

IE = e−δ T ηT1 eCT $ 1

+ . e−CT − S T

Theorem 6.3. Let Assumption (EM) be in force and assume that the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the value of a Quanto and a European plain vanilla call option via the following symmetry: ν Vqc (S01 , S02 , K; r, δ, C, ν) = IE[ST1 ]eCT Vp S02 , K; δ 1 , r, C, (5.6.10)

C is given by where K = e−CT , the constant t t 1 t 1 2 cs ds + Ct = (2r − δ − δ )t + (ex − 1 − x) ν(ds, dx) bs ds + 2 0 R 0 0 ν) are given by Proposition 6.1 for v = (1, 0) and u = (0, 1). A and the characteristics (C, similar relationship holds for the Quanto and European plain vanilla put options. Proof. The proof follows along the lines of that of Theorem 6.2. Theorem 6.4. Let Assumption (EM) be in force and assume that the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the value of a two-dimensional asset-or-nothing and a cash-or-nothing option via the following symmetry: ν (5.6.11) Van (S01 , S02 , K; r, δ, C, ν) = IE[ST1 ]Vcn S02 , K; δ 1 , r, C, C is given by where K = Ke−CT , the constant t t 1 t cs ds + (ex − 1 − x) ν(ds, dx) bs ds + Ct = (2r − δ 1 − δ 2 )t + 2 R 0 0 0

ν) are given by Proposition 6.1 for v = (1, 0) and u = (0, 1). A and the characteristics (C, similar relationship holds for the corresponding put options. Proof. The proof follows along the lines of that of Theorem 6.2. Remark 6.5. Notice that the factor IE[ST1 ] is the forward price of the asset S 1 , the num´eraire asset.

ACKNOWLEDGEMENTS We thank Wolfgang Kluge for helpful discussions during the work on these topics. The second named author acknowledges the ﬁnancial support provided through the European Community’s Human Potential Programme under Contract HPRN-CT-2000-00100 DYNSTOCH.

124

Exotic Option Pricing and Advanced L´evy Models

REFERENCES [1] Albrecher, H. (2004), “The valuation of Asian options for market models of exponential L´evy type”, in M. Van Maele, A. De Schepper, J. Dhaene, H. Reynaerts, W. Schoutens and P. Van Goethem (Eds), Handelingen Contact forum – 2nd Actuarial and Financial Mathematics Day, Royal Flemish Academy of Belgium for Science and Arts, Brussels, Belgium, pp. 11–20. [2] Albrecher, H., Dhaene, J., Goovaerts, M. and Schoutens, W. (2005), “Static hedging of Asian options under L´evy models: the comonotonicity approach”, The Journal of Derivatives, 12(3), 63–72. [3] Albrecher, H. and Predota, M. (2002), “Bounds and approximations for discrete Asian options in a variance-gamma model”, Grazer Mathematische Berichte, 345, 35–57. [4] Albrecher, H. and Predota, M. (2004), “On Asian option pricing for NIG L´evy processes”, Journal of Computational and Applied Mathematics 172, 153–168. [5] Albrecher, H. and Schoutens, W. (2005), “Static hedging of Asian options under stochastic volatility models using Fast Fourier Transform”, in A. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 129–147. [6] Kyprianou A.E. and Alili, L. (2005), “Some remarks on ﬁrst passage of L´evy process, the American put and pasting principles”, The Annals of Applied Probability, to be published. [7] Almendral, A. (2005), “Numerical valuation of American options under the CGMY process”, in A. Kyprianou, W. Schoutens, and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 259–276. [8] Asmussen, S., Avram, F. and Pistorius, M.R. (2004), “Russian and American put options under exponential phase-type L´evy models”, Stochastic Processes and their Applications, 109, 79–111. [9] Avram, F., Kyprianou, A. and Pistorius, M.R. (2004), “Exit problems for spectrally negative L´evy processes and applications to (Canadized) Russian options”, Annals of Applied Probability, 14, 215–238. [10] Barndorff-Nielsen, O.E. and Shephard, N. (2001), “Non-Gaussian Ornstein–Uhlenbeck-based models and some of their uses in ﬁnancial economics”, Journal of the Royal Statistical Society, Series B , 63, 167–241. [11] Barone-Adesi, G. and Whaley, R.E. (1987), “Efﬁcient analytic approximation of American option values”, Journal of Finance, 42, 301–320. [12] Bellini, F. and Frittelli, M. (2002), “On the existence of minimax martingale measures”, Mathematical Finance, 12, 1–21. [13] Benhamou, E. (2002), “Fast Fourier Transform for discrete Asian options”, Journal of Computational Finance, 6, 49–68. [14] Bertoin, J. (1996), L´evy Processes, Cambridge University Press, Cambridge, UK. [15] Bingham, N.H. (1975), “Fluctuation theory in continuous time”, Advances in Applied Probability, 7, 705–766. [16] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities”, Journal of Political Economy, 81, 637–654. [17] Borovkov, K. and Novikov, A. (2002), “On a new approach to calculating expectations for option pricing”, Journal of Applied Probability, 39, 889–895. [18] Boyarchenko, S.I. and Levendorskiˇı, S.Z. (2002a), “Barrier options and touch-and-out options under regular L´evy processes of exponential type”, Annals of Applied Probability, 12, 1261–1298. [19] Boyarchenko, S.I. and Levendorskiˇı, S.Z. (2002b), Non-Gaussian Merton–Black–Scholes Theory, World Scientiﬁc, Singapore. [20] Boyarchenko, S.I. and Levendorskiˇı, S.Z. (2002c), “Perpetual American options under L´evy processes”, SIAM Journal on Control and Optimization, 40, 1663–1696.

Exotic Options in L´evy Models

125

[21] Breeden, D. and Litzenberger, R. (1978), “Prices of state-contingent claims implicit in option prices”, Journal of Business, 51, 621–651. [22] Carr, P. (1994), “European put call symmetry”, Working Paper, Cornell University, Ithaca, New York, NY, USA. ´ [23] Carr, P. and Chesney, M. (1996), “American put call symmetry”, Working Paper, Ecole des hautes e´ tudes commerciales de Montreal (HEC), Montreal, QU, Canada. [24] Carr, P., Geman, H. Madan, D.B. and Yor, M. (2002), “The ﬁne structure of asset returns: an empirical investigation”, Journal of Business, 75, 305–332. [25] Carr, P., Geman, H. Madan, D.B. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [26] Carr, P. and Hirsa, A. (2003), “Why be backward? Forward equations for American options”, Risk , 16(1), 103–107. Reprinted in A. Kyprianou, W. Schoutens and P. Wilmott (Eds) (2005), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 237–257. [27] Carr, P. and Madan, D.B. (1999), “Option valuation using the fast Fourier transform”, Journal of Computational Finance, 2, 61–73. [28] Carverhill, A. and Clewlow, L. (1992), “Flexible convolution”, in R. Tompkins (Ed.), From Black–Scholes to Black Holes, Risk Publications, London, UK, pp. 165–171. [29] Chesney, M. and Jeanblanc, M. (2004), “Pricing American currency options in an exponential L´evy model”, Applied Mathematical Finance, 11, 207–225. [30] Cont, R. and Tankov, P. (2004), “Nonparametric calibration of jump-diffusion option pricing models”, Journal of Computational Finance, 7(3), 1–49. [31] Cont, R. and Voltchkova, E. (2005), “Finite difference methods for option pricing in jumpdiffusion and exponential L´evy models”, SIAM Journal on Numerical Analysis, to be published. [32] Cont, R. and Voltchkova, E. (2005), “Integro-differential equations for option prices in exponential L´evy models”, Finance and Stochastics, 9, 299–325. [33] Delbaen, F. and Schachermayer, W. (1994), “A general version of the fundamental theorem of asset pricing”, Mathematische Annalen, 300, 463–520. [34] Delbaen, F. and Schachermayer, W. (1998), “The fundamental theorem of asset pricing for unbounded stochastic processes”, Mathematische Annalen, 312, 215–250. [35] Detemple, J. (2001), “American options: symmetry properties”, in J. Cvitanic, E. Jouini and M. Musiela (Eds), Option Pricing, Interest Rates and Risk Management, Cambridge University Press, Cambridge, UK, pp. 67–104. [36] Eberlein, E. and Jacod, J. (1997), “On the range of options prices”, Finance and Stochastics, 1, 131–140. [37] Eberlein, E., Jacod, J. and Raible, S. (2005), “L´evy term structure models: no-arbitrage and completeness”, Finance and Stochastics, 9, 67–88. [38] Eberlein, E., Kallsen, J. and Kristen, J. (2003), “Risk management based on stochastic volatility”, Journal of Risk , 5(2), 19–44. [39] Eberlein, E. and Keller, U. (1995), “Hyperbolic distributions in ﬁnance”, Bernoulli , 1, 281–299. [40] Eberlein, E. and Kluge, W. (2004), “Exact pricing formulae for caps and swaptions in a L´evy term structure model”, FDM Preprint 86, University of Freiburg, Freiburg, Germany. [41] Eberlein, E. and Papapantoleon, A. (2005), “Equivalence of ﬂoating and ﬁxed strike Asian and lookback options”, Stochastic Processes and their Applications, 105, 31–40. [42] Eberlein, E. and Prause, K. (2002), “The generalized hyperbolic model: ﬁnancial derivatives and risk measures”, in H. Geman, D. Madan, S. Pliska and T. Vorst (Eds), Mathematical FinanceBachelier Congress 2000, Springer-Verlag, Berlin, Germany, pp. 245–267. [43] Eberlein, E. and von Hammerstein, E.-A. (2004), “Generalized hyperbolic and inverse Gaussian distributions: limiting cases and approximation of processes”, in R. Dalang, M. Dozzi and

126

[44]

[45] [46] [47] [48] [49] [50] [51] [52]

[53] [54] [55]

[56] [57] [58] [59] [60] [61] [62] [63] [64]

Exotic Option Pricing and Advanced L´evy Models F. Russo (Eds), Seminar on Stochastic Analysis, Random Fields and Applications IV, Progress in Probability, Vol. 58, Birkh¨auser Verlag, Basel, Switzerland, pp. 221–264. Fajardo, J. and Mordecki, E. (2003), “Duality and derivative pricing with L´evy processes”, Preprint No. 76, Centro de Matem´atica, Universidad de la Rep´ublica del Uruguay, Montevideo, Uruguay. Frittelli, M. (2000), “The minimal entropy martingale measure and the valuation problem in incomplete markets”, Mathematical Finance, 10, 39–52. Fujiwara, T. and Miyahara, Y. (2003), “The minimal entropy martingale measures for geometric L´evy processes”, Finance and Stochastics, 7, 509–531. Geman, H., El Karoui, N. and Rochet, J.-C. (1995), “Changes of num´eraire, changes of probability measures and option pricing”, Journal of Applied Probability, 32, 443–458. Glasserman, P. (2003), Monte Carlo Methods in Financial Engineering, Springer-Verlag, New York, NY, USA. Goll, T. and R¨uschendorf, L. (2001), “Minimax and minimal distance martingale measures and their relationship to portfolio optimization”, Finance and Stochastics, 5, 557–581. Greenwood, P. and Pitman, J. (1980), “Fluctuation identities for L´evy processes and splitting at the maximum”, Advances in Applied Probability, 12, 893–902. Gushchin, A.A. and Mordecki, E. (2002), “Bounds on option prices for semimartingale market models”, Proceedings of the Steklov Mathematical Institute 237, 73–113. Hakala, J. and Wystup, U. (2002), “Heston’s stochastic volatility model applied to foreign exchange options”, in J. Hakala and U. Wystup (Eds), Foreign Exchange Risk, Risk Publications, London, UK, pp. 267–282. Hartinger, J. and Predota, M. (2002), “Pricing Asian options in the hyperbolic model: a fast quasi Monte Carlo approach”, Grazer Mathematische Berichte, 345, 1–33. Haug, E.G. (2002), “A look in the antimatter mirror”, Wilmott Magazine, September, 38–42. Henderson, V., Hobson, D., Shaw, W. and Wojakowski, R. (2004), “Bounds for in-progress ﬂoating-strike Asian options using symmetry”, Annals of Operational Research (Special Issue on Financial Modelling), (to appear). Henderson, V. and Wojakowski, R. (2002), “On the equivalence of ﬂoating and ﬁxed-strike Asian options”, Journal of Applied Probability, 39, 391–394. Heston, S.L. (1993), “A closed-form solution for options with stochastic volatility with applications to bond and currency options”, Review of Financial Studies, 6, 327–343. Hobson, D., Laurence, P. and Wang, T.-H. (2005), “Static-arbitrage upper bounds for the prices of basket options”, Quantitative Finance, to be published. Jacod, J. and Shiryaev, A.N. (2003), Limit Theorems for Stochastic Processes, 2nd Edn, SpringerVerlag, Berlin, Germany. Kallsen, J. and Shiryaev, A.N. (2002), “The cumulant process and Esscher’s change of measure”, Finance and Stochastics, 6, 397–428. Kallsen, J. and Tankov, P. (2004), “Characterization of dependence of multidimensional L´evy ´ processes using L´evy copulas”, Working Paper, Ecole Polytechnique, Paris, France. K¨ellezi, E. and Webber, N. (2004), “Valuing Bermudan options when asset returns are L´evy processes”, Quantitative Finance, 4, 87–100. Kou, S.G. and Wang, H. (2003), “First passage times of a jump diffusion process”, Advances in Applied Probability, 35, 504–531. Kou, S.G. and Wang, H. (2005), “Option pricing under a double exponential jump diffusion model”, Management Science, 50, 1178–1192.

Exotic Options in L´evy Models

127

[65] Lee, R.W. (2004), “Option pricing by transform methods: extensions, uniﬁcation, and error control”, Journal of Computational Finance, 7(3), 50–86. [66] Levendorskiˇı, S.Z. (2004), “Pricing of the American put under L´evy processes”, International Journal of Theoretical and Applied Finance, 7, 303–335. [67] Lewis, A. (2001), “A simple option formula for general jump-diffusion and other exponential L´evy processes”, Working Paper, Optioncity.net. [68] Lipton, A. (2002), “Assets with jumps”, Risk , 15(9), 149–153. [69] Margrabe, W. (1978), “The value of an option to exchange one asset for another”, Journal of Finance, 33, 177–186. [70] Matache, A.-M., Nitsche, P.-A. and Schwab, C. (2003), “Wavelet Galerkin pricing of American options on L´evy driven assets”, SAM Research Report 2003-06, ETH, Zurich, Switzerland. [71] Matache, A.-M., Schwab, C. and Wihler, T.P. (2005), “Fast numerical solution of parabolic integro-differential equations with applications in ﬁnance”, SIAM Journal on Scientiﬁc Computing, to be published. [72] Matache, A.-M., von Petersdorff, T. and Schwab, C. (2004), “Fast deterministic pricing of options on L´evy driven assets”, ESAIM Mathematical Modelling and Numerical Analysis, 38, 37–72. [73] McDonald, R.L. and Schroder, M.D. (1998), “A parity result for American options”, Journal of Computational Finance, 1, 5–13. [74] Merton, R.C. (1973), “Theory of rational option pricing”, Bell Journal of Economics and Management Science, 4, 141–183. [75] Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493. [76] M¨uller, A. and Stoyan, D. (2002), Comparison Methods for Stochastic Models and Risks, Wiley, Chichester, UK. [77] Musiela, M. and Rutkowski, M. (1997), Martingale Methods in Financial Modelling, SpringerVerlag, Berlin, Germany. [78] Nguyen-Ngoc, L. (2003), “Exotic options in general exponential L´evy models”, Pr´epublication No 850, Universit´es Paris 6, Paris, France. [79] Nguyen-Ngoc, L. and Yor, M. (2005), “Lookback and barrier options under general L´evy processes”, in Y. Ait-Sahalia and L.-P. Hansen (Eds), Handbook of Financial Econometrics, North-Holland, Amsterdam, The Netherlands, (to appear). [80] Peskir, G. and Shiryaev, A.N. (2002), “A note on the call-put parity and a call-put duality”, Theory of Probability and its Applications, 46, 167–170. [81] Raible, S. (2000), “L´evy processes in ﬁnance: theory, numerics, and empirical facts”, Ph.D. Thesis, University of Freiburg, Freiburg, Germany. [82] Ribeiro, C. and Webber, N. (2003), “A Monte Carlo method for the normal inverse Gaussian option valuation model using an inverse Gaussian bridge”, Preprint, City University, London, UK. [83] Ribeiro, C. and Webber, N. (2004), “Valuing path-dependent options in the variance-gamma model by Monte Carlo with a gamma bridge”, Journal of Computational Finance, 7, 81–100. [84] Rogers, L.C.G. (2000), “Evaluating ﬁrst-passage probabilities for spectrally one-sided L´evy processes”, Journal of Applied Probability, 37, 1173–1180. [85] Rogers, L.C.G. (2002), “Monte Carlo valuation of American options”, Mathematical Finance, 12, 271–286. [86] Sato, K.-I. (1999), L´evy Processes and Inﬁnitely Divisible Distributions, Cambridge University Press, Cambridge, UK.

128

Exotic Option Pricing and Advanced L´evy Models

[87] Schoutens, W. (2002), “The Meixner process: theory and applications in ﬁnance”, in O.E. Barndorff-Nielsen (Ed.), Mini-Proceedings of the 2nd MaPhySto Conference on L´evy processes, Aarhus, Denmark, pp. 237–241. [88] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [89] Schroder, M. (1999), “Changes of numeraire for pricing futures, forwards and options”, Reviews of Financial Studies, 12, 1143–1163. [90] Sch¨urger, K. (2002), “Laplace transforms and suprema of stochastic processes”, in K. Sandmann and P.J. Sch¨onbucher (Eds), Advances in Finance and Stochastics, Springer-Verlag, Berlin, Germany, pp. 285–294. [91] Vanmaele, M., Dhaene, J., Deelstra, G., Liinev, J. and Goovaerts, M.J. (2005), “Bounds for the price of discrete arithmetic Asian options”, Journal of Computational and Applied Mathematics, to be published. [92] Veˇceˇr, J. and Xu, M. (2004), “Pricing Asian options in a semimartingale model”, Quantitative Finance, 4, 170–175. [93] Wystup, U. (2002), “Vanilla options”, in J. Hakala and U. Wystup (Eds), Foreign Exchange Risk, Risk Publications, London, UK, pp. 3–14.

6 Static Hedging of Asian Options under Stochastic Volatility Models using Fast Fourier Transform ¨ Albrecher Hansjorg Graz University of Technology, Graz, Austria and

Wim Schoutens Katholieke Universiteit Leuven – UCS, Leuven, Belgium Abstract We present a simple static super-hedging strategy for the payoff of an arithmetic Asian option in terms of a portfolio of European options under various stochastic volatility models. Moreover, it is shown that the obtained hedge is optimal in some sense. The strategy is based on stop-loss transforms and comonotonicity theory. The numerical implementation is based on the Fast Fourier transform. We illustrate the hedging performance for several models calibrated to market data and compare the results with other (trivial) static super-hedging strategies.

6.1

INTRODUCTION

The design of efﬁcient hedging strategies for exotic options is a challenging problem that has received a lot of interest during the last few years. In order to serve the needs of investors, increasingly complex ﬁnancial products have been introduced in the market and the pricing and (in particular) hedging of these products is of great importance for assessing the involved risk when trading these instruments. However, many of the hedging techniques currently used in practice rely on market model assumptions that are clearly not sufﬁciently realistic (such as the Black–Scholes model). A common practice in the hedging of exotics is to calibrate the model to vanilla options traded in the market and then derive the corresponding hedging positions for the exotic option. If the model is then recalibrated to the market on the next day, say, then, in order to make the hedging strategy meaningful, the obtained parameter-set should be rather close to the one from the previous day so that only minor adjustments of the hedging portfolio are needed. That is, in addition to a good ﬁt to historical market data, one crucial requirement for a sound market model is its stability in terms of hedging strategies. Empirical studies in that direction indicate that stochastic volatility models outperform classical models like Black-Scholes by far (see e.g. Bakshi et al. (1997) [7]). Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

130

Exotic Option Pricing and Advanced L´evy Models

Apart from that, proposed dynamic hedging strategies with continuously changing positions in the asset (such as delta-hedging) have various deﬁciencies (see e.g. Allen and Padovani (2002) [6]). These are typically based on assumptions like no limit on the frequency of rebalancing, zero transaction costs and full liquidity of the market. However, in practice these assumptions are usually not fulﬁlled and alternatives are asked for. The most favourable situation is the availability of a static hedging strategy for the exotic option, that is an initial hedge portfolio (in terms of the underlying and vanilla options), which will perfectly replicate the payoff at maturity without any portfolio adjustments during the lifetime of the option. For some exotic options (such as barrier, lookback and cliquet options), it is possible to derive semi-static hedging strategies, where portfolio adjustments are only needed at a ﬁnite (and typically small) number of times before maturity (see, for instance, Allen and Padovani (2002) [6] and Carr et al. (1998) [16]). Another alternative is to look for a static super-hedging strategy, which is a portfolio of the underlying and vanilla options that will dominate the payoff of the exotic option without any adjustments during its lifetime. Such a strategy puts a ﬂoor on the maximum loss whatever the subsequent price path will look like and provides a simple way to hedge the product at the expense of a calculable additional cost (namely the difference of the cost of the hedge portfolio and the actual price of the option). At the same time, this strategy enjoys all the advantages of a static hedge: it is less sensitive to the assumption of zero transaction costs (both commissions and the cost of paying individuals to monitor the positions) and does not face the risk of dried-up liquidity when the market makes large moves (opposed to dynamic hedging (see e.g. Carr and Picron (1999) [19] and Carr and Wu (2002) [20]). Semi-static super-hedging strategies for barrier options are discussed in Brown et al. (2001) [14] and Neuberger and Hodges (2002) [32]). This paper focuses on Asian options. Already, the pricing of these products is far from trivial, especially when leaving the Black–Scholes framework (see e.g. Albrecher and Predota (2002, 2004) [5, 4] and Ve˘ce˘r and Xu (2004) [41] or the recent survey by Albrecher (2004) [3]). Moreover, many of the available pricing techniques do not lead to an effective hedging strategy. A delta-hedging strategy for Asian options in a Black–Scholes model based on approximations was discussed in Jacques (1996) [28]. In Albrecher et al. (2003) [2], a simple static super-hedging strategy for arithmetic Asian call options consisting of a portfolio of European options has been derived and optimized using comonotonicity theory. The performance of the resulting strategy has been studied for models for asset price processes following an exponential L´evy model. In the present paper, we extend this approach to stochastic volatility models and investigate the performance of the resulting hedging strategy. As will be illustrated, the hedging error of this simple super-hedging strategy is very small if the option is in the money. For options at and out of the money, this strategy can be quite conservative, but the static nature of the hedge may compensate for parts of the gap. The paper is structured as follows. In Section 6.2, several stochastic volatility models for the asset price process are introduced. In Section 6.3, we present the static super-hedging strategy in detail and illustrate how it can be optimized by comonotonicity techniques. The numerical implementation of the strategy for the various models on the basis of Fast Fourier transforms is discussed in Section 6.4. In Section 6.5, all of the models are calibrated to market data, namely to the same set of vanilla options on the S&P 500, and the performance

Static Hedging of Asian Options

131

of the corresponding hedging strategies is illustrated. Moreover, the issue of model risk is discussed. Since it turns out that the developed hedging strategy only depends on the marginal risk-neutral densities of the asset price process at each averaging day of the Asian option, it can actually be implemented in a completely model-independent setup by estimating the marginal risk-neutral densities directly from the call option surface. This extension is discussed in Section 6.6.

6.2

STOCHASTIC VOLATILITY MODELS

In the sequel, we will brieﬂy introduce various stochastic volatility models, all of which proved their smile-conform pricing abilities, and consider their risk-neutral dynamics. Let S = {St , 0 ≤ t ≤ T } denote the stock price process and φ(u, t) the characteristic function of the random variable log St , i.e. φ(u, t) = E[exp(iu log(St ))]. We assume the stock pays out a continuous dividend; the dividend yield is assumed to be constant and denoted by q. We also have at our disposal a risk-free bank account, paying out a continuously compounded interest rate, which we assume to be constant and denote by r. The price process for the bank-account (bond) is thus given by B = {Bt = exp(rt), t ≥ 0}. The stochastic dynamics of our stock price process will be driven by L´evy processes. A L´evy process X = {Xt , t ≥ 0} is a stochastic process which starts at zero and has independent and stationary increments such that the distribution of the increment is an inﬁnitely divisible distribution (i.e. a distribution for which the characteristic function is also the nth power of another characteristic function, for every integer n). There is a one-to-one correspondence between L´evy processes and inﬁnitely divisible distributions. A subordinator is a nonnegative nondecreasing L´evy process. A general reference for L´evy processes is Bertoin (1996) [12], while for applications in ﬁnance see Schoutens (2003) [38]. 6.2.1 The Heston stochastic volatility model In the Heston Stochastic Volatility model (HEST), the stock price process follows a Black– Scholes stochastic differential equation, in which the volatility behaves stochastically over time: dSt = (r − q)dt + σt dWt , St

S0 ≥ 0,

where the (squared) volatility follows the classical Cox-Ingersoll-Ross (CIR) process: dσt2 = κ(η − σt2 )dt + θ σt dW˜ t ,

σ0 ≥ 0.

Here, W = {Wt , t ≥ 0} and W˜ = {W˜ t , t ≥ 0} are two correlated standard Brownian motions such that Cov[dWt , dW˜ t ] = ρ dt.

132

Exotic Option Pricing and Advanced L´evy Models

The characteristic function φ(u, t) is in this case given by (see Bakshi et al. [7] or Heston (1993) [27]): φ(u, t) = E[exp(iu log(St ))|S0 , σ0 ] = exp(iu(log S0 + (r − q)t)) × exp(ηκθ −2 ((κ − ρθ ui − d)t − 2 log((1 − ge−dt )/(1 − g)))) × exp(σ02 θ −2 (κ − ρθ iu − d)(1 − e−dt )/(1 − ge−dt )), where d = ((ρθ ui − κ)2 − θ 2 (−iu − u2 ))1/2 ,

(6.1)

g = (κ − ρθ ui − d)/(κ − ρθ ui + d).

(6.2)

An extension of HEST introduces jumps in the asset price (Bakshi et al. [7]), while other extensions also allow for jumps in the volatility (see e.g. Knudsen and Nguyen-Ngoc (2003) [29]). Since for these extensions the characteristic function of the log stock price is also available, one can straightforwardly apply the methods described below for these models too. 6.2.2 The Barndorff-Nielsen–Shephard model This class of models, denoted by BN–S, was introduced in Barndorff-Nielsen and Shephard (2000) [10] and has a structure similar to the Heston model. The difference is basically that here the volatility is modelled by an Ornstein–Uhlenbeck (OU) process driven by a subordinator. In this way, jumps are introduced into the volatility process. Volatility can only jump upwards and then will decay exponentially. A co-movement effect between up-jumps in volatility and (down)-jumps in the stock price is also incorporated. The squared volatility now follows a SDE of the form: dσt2 = −λσt2 dt + dzλt ,

(6.3)

where λ > 0 and z = {zt , t ≥ 0} is a subordinator. The risk-neutral dynamics of the log-price Zt = log St are given by dZt = (r − q − λk(−ρ) − σt2 /2)dt + σt dWt + ρdzλt ,

Z0 = log S0 ,

where W = {Wt , t ≥ 0} is a Brownian motion independent of z = {zt , t ≥ 0} and where k(u) = log E[exp(−uz1 )] is the cumulant function of z1 . Note that the parameter ρ introduces the co-movement effect between the volatility and the asset price process. We use the classical and tractable example of the Gamma-OU process (other choices for OU-processes include the inverse Gaussian-OU process, which also leads to a tractable model [Schoutens (2003) [38], Section 7.2.1]). For a Gamma-OU process, z = {zt , t ≥ 0} is a compound-Poisson process: zt =

Nt n=1

xn ,

(6.4)

Static Hedging of Asian Options

133

where N = {Nt , t ≥ 0} is a Poisson process with intensity parameter a, i.e. E[Nt ] = at and {xn , n = 1, 2, . . . } is an independent and identically distributed sequence of exponential random variables with mean 1/b. One then has log E[exp(−uz1 )] = −au(b + u)−1 , and it can be shown that σ 2 has a stationary marginal law that follows a Gamma distribution. The characteristic function of the log price can, in this case, be written in the form (cf. Barndorff-Nielson et al. (2002) [11]) φ(u, t) = E[exp(iu log St )|S0 , σ0 ] = exp iu(log(S0 ) + (r − q − aλρ(b − ρ)−1 )t) × exp −λ−1 (u2 + iu)(1 − exp(−λt))σ02 /2

b − f1 + f2 λt , × exp a(b − f2 )−1 b log b − iuρ where f1 = f1 (u) = iuρ − λ−1 (u2 + iu)(1 − exp(−λt))/2, f2 = f2 (u) = iuρ − λ−1 (u2 + iu)/2. 6.2.3 L´evy models with stochastic time Another way to incorporate stochastic volatility effects into the price process is by making time stochastic. Periods with high volatility can be interpreted as if time runs faster than in periods with low volatility. Applications of stochastic time change to asset pricing go back to Mandelbrot and Taylor (1967) [31] (see also Clark (1973) [21]). We consider the models introduced by Carr et al. (2003) [18]. The L´evy models with stochastic time considered in this paper are built out of two independent stochastic processes. The ﬁrst process is a L´evy process. The behavior of the asset price will then be modelled by the exponential of the L´evy process, suitably timechanged. Typical examples for the generator of the L´evy process are the normal distribution (leading to Brownian motion), the Normal Inverse Gaussian (NIG) distribution (BarndorffNielsen (1995) [8] and Rydberg (1997) [34], the Variance Gamma (VG) distribution (Madan et al. (1998) [30]), the generalized hyperbolic distribution (Eberlein (1999) [25] and Rydberg (1999) [35], the Meixner distribution (Grigelionis (1999) [26], Schoutens and Teugels (1998) [36] and Schoutens (2002) [37]) and the CGMY distribution (Carr et al. (2002) [17]) (see Schoutens (2003) [38] for an overview). We opt to work with the VG and NIG processes for which simulation issues become quite standard. The second process is a stochastic clock that builds in a stochastic volatility effect. The above mentioned (ﬁrst) L´evy process will be subordinated (i.e. time-changed) by this stochastic clock. By deﬁnition of a subordinator, the time needs to increase and the process modelling the rate of time change y = {yt , t ≥ 0} also needs to be positive. The economic time elapsed in t units of calendar time is then given by the integrated process Y = {Yt , t ≥ 0} with t ys ds. (6.5) Yt = 0

134

Exotic Option Pricing and Advanced L´evy Models

Since y is a positive process, Y is an increasing process. We will consider two processes for the rate of time change y: the CIR process (which is continuous) and the Gamma-OU process (which is a jump process). We will ﬁrst discuss NIG and VG and subsequently introduce the stochastic clocks CIR and Gamma-OU. 6.2.3.1 NIG and VG processes The NIG(α, β, δ) distribution with parameters α > 0, |β| < α and δ > 0 has a characteristic function given by φNI G (u; α, β, δ) = exp −δ α 2 − (β + iu)2 − α 2 − β 2 and the VG(C, G, M) distribution with parameters C > 0, G > 0 and M > 0 has a characteristic function given by φV G (u; C, G, M) =

GM GM + (M − G)iu + u2

C .

Since both distributions are inﬁnitely divisible, each of them generates a L´evy process X = {Xt , t ≥ 0} where the increment X1 follows a NIG(α, β, δ) law (VG(C, G, M) law, respectively). The resulting process is called a NIG process (VG process, respectively). Due to convolution properties of these two distributions, increments of arbitrary length again follow the same law with just a change in parameters: An increment of the NIGprocess over the time interval [s, s + t] follows a NIG(α, β, δt) law and the increment of a VG-process over [s, s + t] is VG(Ct, G, M)-distributed (see also Barndorff-Nielsen (1997) [9]). 6.2.3.2 Stochastic clocks CIR Stochastic Clock:. Carr et al. (2003) [18] use as the rate of time change the CIR process that solves the SDE: 1/2

dyt = κ(η − yt )dt + λyt dWt , where W = {Wt , t ≥ 0} is a standard Brownian motion. The characteristic function of Yt (given y0 ) is explicitly known (see Cox et al. (1985) [22]): ϕCI R (u, t; κ, η, λ, y0 ) = E[exp(iuYt )|y0 ] =

exp(κ 2 ηt/λ2 ) exp(2y0 iu/(κ + γ coth(γ t/2))) (cosh(γ t/2) + κ sinh(γ t/2)/γ )2κη/λ2

where γ =

κ 2 − 2λ2 iu.

,

Static Hedging of Asian Options

Gamma-OU Stochastic Clock:. of the SDE:

135

Another choice for the rate of time change is the solution

dyt = −λyt dt + dzλt ,

(6.6)

where the process z = {zt , t ≥ 0} is, as in equation (6.4), a compound Poisson process. In the Gamma-OU case, there is an explicit expression for the characteristic function of Yt (given y0 ): ϕ −OU (u; t, λ, a, b, y0 ) = E[exp(iuYt )|y0 ] = exp iuy0 λ−1 (1 − e−λt ) +

b λa b log − iut . iu − λb b − iuλ−1 (1 − e−λt )

6.2.3.3 Time-changed L´evy process Let Y = {Yt , t ≥ 0} as deﬁned in equation (6.5), be the process modelling our business time. The (risk-neutral) price process S = {St , t ≥ 0} is now modelled as follows: St = S0

exp((r − q)t) exp(XYt ), E[exp(XYt )|y0 ]

(6.7)

where X = {Xt , t ≥ 0} is a L´evy process. The factor exp((r − q)t)/E[exp(XYt )|y0 ] puts us immediately into the risk-neutral world by a mean-correcting argument. Essentially, the stock price process is modelled as the ordinary exponential of a time-changed L´evy process. The process incorporates jumps (through the L´evy process Xt ) and stochastic volatility (through the time change Yt ). The characteristic function φ(u, t) for the logarithm of our stock price is given by: φ(u, t) = E[exp(iu log(St ))|S0 , y0 ] = exp(iu((r − q)t + log S0 ))

ϕ(−iψX (u); t, y0 ) , ϕ(−iψX (−i); t, y0 )iu

(6.8)

where ψX (u) = log E[exp(iuX1 )] and ϕ(u; t, y0 ) denotes the characteristic function of Yt given y0 . Since we consider two L´evy processes (VG and NIG) and two stochastic clocks (CIR and Gamma-OU), we will ﬁnally end up with four resulting models abbreviated as VG-CIR, VG-OU , NIG-CIR and NIG-OU . Due to (time-)scaling effects, one can without loss of generality scale the present rate of time change to 1 (y0 = 1). For more details, see Carr et al. (2003) [18] or Schoutens (2003) [38].

136

Exotic Option Pricing and Advanced L´evy Models

6.3

STATIC HEDGING OF ASIAN OPTIONS

Consider now a European-style arithmetic average call option with strike price K, maturity T and n averaging days 0 ≤ t1 < · · · < tn ≤ T . Then, its price according to a risk-neutral pricing measure Q at time t is given by

+ % n & exp(−r(T − t)) AAt = Stk − nK EQ Ft , n k=1

where {Ft , 0 ≤ t ≤ T } denotes the natural ﬁltration of S. In general, the distribution of the dependent sum nk=1 Stk is not available, which makes pricing and hedging of these products difﬁcult. However, for our super-hedging purposes it sufﬁces to look for an upper bound of the above payoff. Assume for simplicity that t = 0 and that the averaging has not yet started. First note, that for any K1 , . . . , Kn ≥ 0 with K = nk=1 Kk , we have a.s.

+ n n + + Stk − nK = (St1 − nK1 ) + · · · + (Stn − nKn ) ≤ Stk − nKk . k=1

k=1

Hence exp(−rT ) EQ AA0 (K, T ) = n ≤

% n

+ Stk − nK

F0

&

k=1

n * + + exp(−rT ) EQ Stk − nKk F0 n k=1

exp(−rT ) exp(rtk )EC0 (κk , tk ), n n

=

(6.9)

k=1

where EC0 (κk , tk ) denotes the price of a European call option at time 0 with strike κk = nKk and maturity tk . In terms of hedging, this means that we have the following static super-hedging strategy: for each averaging day tk , buy exp(−r(T − tk ))/n European call options at time t = 0 with strike κk and maturity tk and hold these until their expiry. Then put their payoff on the bank account. the upper bound (equation (6.9)) holds for all combinations of κk ≥ 0 that satisfy Since n κ k=1 k = nK, we still have the freedom to choose strike values that ﬁt best to our purposes. The simplest choice is κk = K (k = 1, . . ., n). If q ≤ r, we have EC0 (K, t) ≤ EC0 (K, T ) for every K ≥ 0 and 0 ≤ t ≤ T , and thus this trivial choice shows that the Asian option price is dominated by the price of a European option with the same strike and maturity, i.e. AA0 (K, T ) ≤ EC0 (K, T ) (this trivial hedging strategy of an Asian option in terms of the corresponding European option was already observed in Nielsen and Sandmann (2003) [33]). However, for our super-hedging purposes, we naturally look for that combination of κk s which minimizes the right-hand side of equation (6.9). In the Black–Scholes setting, this optimization problem

Static Hedging of Asian Options

137

was solved in Nielsen and Sandmann (2003) [33] by using Lagrange multipliers. In the general case of arbitrary arbitrage-free market models, this optimal combination can be determined by using stop-loss transforms and the theory of comonotonic risks (for a general introduction to comonotonicity techniques, see Dhaene et al. (2002a, b) [23]). Let F (x) be a distribution function of a non-negative random variable X; then its stop-loss transform F (m) is deﬁned by +∞ F (m) = (x − m)dF (x) = E[(X − m)+ ], m ≥ 0. m

If we write An =

n

Stk

k=1

and FAt n (x) = PQ (An ≤ x|Ft ) for the distribution function under Q of An given the information Ft , then we have AAt =

exp(−r(T − t)) F t (nK). An n

(6.10)

In this way the problem of pricing an arithmetic average option is transformed to calculating the stop-loss transform of a sum of dependent risks. Concretely, we will look at bounds for stop-loss transforms based on comonotonic risks. A positive random vector (X1 , . . . , Xn ) with marginal distribution functions F1 (x1 ), . . . , Fn (xn ) is called comonotone, if for the joint distribution function FX1 ,...,Xn (x1 , . . . , xn ) = min{F1 (x1 ), . . . , Fn (xn )} holds for every x1 , . . . , xn ≥ 0. It immediately follows that the distribution of a comonotone random vector (X1 , . . . , Xn ) with given marginal distributions F1 (x1 ), . . . , Fn (xn ) is uniquely determined. In Simon et al. (2000) [40], it was shown that an upper bound for the stop-loss transform of the sum of arbitrary dependent positive random variables nk=1 Xk with marginal distributions F1 (x1 ), . . . , Fn (xn ) is given by the stop-loss transform of the sum S c = nk=1 Yk , where (Y1 , . . . , Yk ) is the comonotone random vector with marginal distributions F1 (x1 ), . . . , Fn (xn ). Let FS c (x) denote the distribution function of nk=1 Yk ; then it follows from general comonotonicity results (see e.g. Dhaene et al. (2002a, b) [23]) that its inverse is given by FS−1 c (x)

=

n

FX−1 (x), k

x ≥ 0.

(6.11)

k=1

The crucial result for our purposes is now that the stop-loss transform of a sum of comonotonic random variables can be obtained as a sum of the stop-loss transforms of the marginals evaluated at speciﬁed points (cf. Proposition 2 in Simon et al. (2000) [40]). More precisely, FS c (m) =

n

c (m)) , FXk FX−1 (F S k

m ≥ 0,

(6.12)

k=1

given that the marginal distribution functions involved are strictly increasing (which is always the case in our applications). At the same time, n

+ n n ≤ Yk − m E((Yk − mk )+ ) = FXk (mk ) (6.13) FS c (m) = E k=1

k=1

k=1

138

Exotic Option Pricing and Advanced L´evy Models

whenever nk=1 mk = m. Thus, the stop-loss transform of the comonotonic sum given by equation (6.12) represents the lowest possible bound in terms of a sum of stop-loss transforms of the marginal distributions. This fact immediately translates to our setting of an arithmetic Asian option. Let F (xk ; tk ) (k = 1, . . . , n) denote the conditional distribution of Stk under the risk-neutral measure Q (given the information available at time t = 0), i.e. for xk , tk > 0, F (xk ; tk ) = PQ Stk ≤ xk |F0 .

(6.14)

Combining equations (6.9), (6.10), (6.12) and (6.13), the optimal combination of strike prices κk is given by κk = F −1 (FS c (n K); tk ) ,

k = 1, . . . , n.

(6.15)

In this way, we have obtained the optimal static super-hedge in terms of European call options with maturity dates equal to the averaging dates. For the practical determination of the strike prices κk , the distribution function of the comonotone sum FS c (x), as given by equation (6.11) has to be calculated and evaluated at nK. For this purpose, we need to approximate the risk-neutral marginal densities of the stock price at the averaging dates, which can be carried out efﬁciently by using Fast Fourier transforms (cf. Section 6.4.1 below). The κk s are then obtained by evaluating the inverse distribution function of F (x; tk ).

6.4 NUMERICAL IMPLEMENTATION 6.4.1 Characteristic function inversion using FFT For all of the above mentioned models, we have the characteristic function of the log-price process at our disposal. However, in order to determine the strike prices of our optimal hedge portfolio as described in Section 6.3, we need the corresponding density functions. Recall that the characteristic function, φ(u), is the Fourier-transform of the corresponding density function f (x): φ(u) =

+∞ −∞

exp(iux)f (x)dx.

So, we need to apply an inverse Fourier-transformation. Next, we illustrate how this can be done fast and accurately by using the Fast Fourier Transform (FFT). The latter is an efﬁcient algorithm for computing the following transformation of a vector (αk , k = 1, . . . , m) into a vector (βk , k = 1, . . . , m): βk =

m

exp(−i2π(j − 1)(k − 1)/N )αj .

j =1

Typically, m is a power of 2. The number of multiplications of the FFT algorithm is of the order O(m log m) and this is in contrast to the straightforward evaluation of the above sums which give rise to O(m2 ) multiplications.

Static Hedging of Asian Options

139

We follow closely a technique described in Carr and Madan (1998) [15] in the context of option pricing. The classical inverse Fourier transform reads: +∞ 1 exp(−iux)φ(u)du. f (x) = 2π −∞ Since f is real, we can write 1 f (x) = π

+∞

exp(−iux)φ(u)du. 0

Next, we are going to discretize the above integral and apply the trapezoid rule. We take a grid on the real line with grid-length u > 0: uj = (j − 1)u,

j = 1, . . . , N.

One approximately has N−1 1 u 1 exp(−iuj x)φ(uj ) + exp(−iuN x)φ(uN ) f (x) ≈ exp(−iu1 x)φ(u1 ) + π 2 2 j =2

u wj exp(−ix(j − 1)u)φ((j − 1)u) π N

=

j =1

where the weights wj are given by 1 1 , w2 = 1, w3 = 1, . . . , wN−1 = 1, wN = . 2 2 We will calculate the value of the density function f in the points w1 =

xk = −b + x(k − 1),

k = 1, . . . , N

where x = 2b/(N − 1), thus covering the interval [−b, b] with an equally spaced grid. In these points we have u wj exp(i(j − 1)bu) exp(−i(j − 1)(k − 1)ux)φ((j − 1)u). π N

f (xk ) ≈

j =1

If we choose the grid sizes such that ux =

2π , N

then u wj exp(i(j − 1)ub) exp(−i(j − 1)(k − 1)2π/N )φ((j − 1)u). π N

f (xk ) ≈

j =1

This sum can be easily computed by FFT: (f (xk ), k = 1, . . . , N ) is the FFT of the vector wj φ(uj ) exp(iuj b), j = 1, . . . , N . Choosing N as a power of 2 allows very fast computation of the FFT.

140

Exotic Option Pricing and Advanced L´evy Models

6.4.2 Static hedging algorithm In order to set up our hedge portfolio, we have to determine the inverse distribution function of the asset price at each averaging day (cf. equation (6.14)). This is carried out by numerically building up the distribution function from the approximated density function obtained in Section 6.4.1. The inverse is then determined by a bisection method from the corresponding table and linear interpolation between grid points is employed. In our implementation, we used 214 points in the grid for both the densities and the inverse distribution functions, which turns out to be sufﬁcient (in the sense that a further increase does not change the signiﬁcant digits of the results). Next, the inverse of the distribution of the comonotone sum is built up according to equation (6.11) and then itself inverted in the above way. Finally, the strike prices κk of the European options are obtained by evaluating the inverse distribution functions of the marginals according to equation (6.15). This numerical procedure to obtain the strike prices for our hedging strategy is both accurate and very quick (the determination of the entire hedge portfolio takes less than a minute on a normal PC for each of the discussed stochastic volatility models).

6.5 NUMERICAL ILLUSTRATION We give numerical results for an arithmetic Asian call option with a maturity of 1 year and averaging every month (i.e. 12 averaging days). First, the model parameters have to be determined from the market prices of vanilla options. 6.5.1 Calibration of the model parameters Carr and Madan (1998) [15] developed pricing methods for the classical vanilla options which can be applied whenever the characteristic function of the risk-neutral stock price process is known. Using Fast Fourier transforms, one can compute within a second the complete option surface on an ordinary computer. In Schoutens (2003) [38], this method was used to calibrate the models (minimizing the difference between market prices and model prices in a least-squares sense) on a dataset of 77 option on the S&P 500 Index [Schoutens (2003) [38], Appendix C]. The results of the calibration are visualized in Figures 6.1 and 6.2 for the NIG-CIR and the Heston model, respectively. Here, the circles are the market prices and the plus signs are the model prices (calculated through the Carr–Madan formula by using the respective characteristic functions and obtained parameters). For details of the ﬁt, see Schoutens (2003) [38]. The Heston model, which is not covered in Schoutens (2003) [38], gives rise to the following calibration errors: ape = 1.31%,

rmse = 1.0530,

aae = 0.8095,

arpe = 1.90%.

Table 6.1 depicts the calibrated parameters for each of the six discussed stochastic volatility models, while Figure 6.3 shows the corresponding marginal density functions of log(St ) for t ranging from 1 month up to 1 year for all six models obtained by Fast Fourier transform, as described in Section 6.4.1.

6.5.2 Performance of the hedging strategy After the strike prices of the hedge portfolio are determined according to equation (6.15), the price of the hedging strategy is easily determined by using the Carr–Madan call option

Static Hedging of Asian Options

200 180 160 140 120 100 80 60 40 20 0 900

1000

1100

1200

1300

1400

1500

1400

1500

Figure 6.1 Calibration of the NIG-CIR Model

200 180 160 140 120 100 80 60 40 20 0 900

1000

1100

1200

1300

Figure 6.2 Calibration of Heston’s Model

141

142

Exotic Option Pricing and Advanced L´evy Models

14 1 2 3 4 5 6 7 8 9 10 11 12

(a) 12 10 1 8 2 6

3 4 5

4

7

6

8

14

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

10 8

2 3

4

4 5 6

9 10

2

8 9 10 11

7 11

2

12

12

0

0 4

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8 1 2 3 4 5 6 7 8 9 10 11 12

(c)

16 14 12 10

1

8

2

6

4.9

5

7

6

4.2

4.3

4.4

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

4.6

4.7

(e) 12 10 1 8 2 3

4

4 5 6

4.4

4.5

7

10

4.9 T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

8

14 12 10

1

5

6

2

0

2 3 4 5

4.9

4

4.1

4.2

4.3

4.4

4.5

1 2 3 4 5 6 7 8 9 10 11 12

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

9

10

4.6

4.7

(f) 12 10

1

6

2

4

3 4 5

12

4.8

4.9 T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

7

5

8 9 10 11

2 12

0

11

1 2 3 4 5 6 7 8 9 10 11 12

14

8

5

8

6 11

4.8

7

2

9 10

4.7

16

11 12

4.8

4.6

18

4

1 2 3 4 5 6 7 8 9 10 11 12

14

6

4.3

(d)

20

8 9

4.5

4.2

6

2 4.1

4.1

8

4 5

4

4 22

3

4

0

1

6

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

1 2 3 4 5 6 7 8 9 10 11 12

(b) 12

12

0 4

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5

4

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5

Figure 6.3 Marginal density functions for the various stochastic volatility models: (a) NIG-OU ; (b) NIG-CIR; (c) VG-OU ; (d) NIG-CIR; (e) BN–S; (f) Heston

pricing formula for European options and equation (6.9). Tables 6.2–6.7 compare the Monte Carlo simulated price of the Asian option AAMC and the comonotonic super-hedge price AAc , with the prices of two trivial super-hedging strategies, namely the trivial super-hedge using the European option price EC with identical strike and maturity (note that q ≤ r) and the super-hedge equation (6.9) with all κi = K with price AAtr . The strike price is given as a percentage of the spot. For the Monte Carlo price, we used 1 million sample paths. The VG process was simulated as a difference of two Gamma processes (cf. Schoutens (2003)

Static Hedging of Asian Options Table 6.1 Risk-neutral parameters obtained by calibration to vanilla calls on S&P 500 HEST σ02 = 0.0224, κ = 0.5144, η = 0.1094, θ = 0.3354, ρ = −0.7392 BN–S ρ = −1.2606, λ = 0.5783, b = 11.6641, a = 1.4338, σ02 = 0.0145 VG-CIR C = 11.9896, G = 25.8523, M = 35.5344, κ = 0.6020, η = 1.5560, λ = 1.9992, y0 = 1 VG-OU C = 11.4838, G = 23.2880, M = 40.1291, λ = 1.2517, a = 0.5841, b = 0.6282, y0 = 1 NIG-CIR α = 18.4815, β = −4.8412, δ = 0.4685, κ = 0.5391, η = 1.5746, λ = 1.8772, y0 = 1 NIG-OU α = 29.4722, β = −15.9048, δ = 0.5071, λ = 0.6252, a = 0.4239, b = 0.5962, y0 = 1

Table 6.2 Hedging performance in the BN–S model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.6065 11.7478 4.5265 0.9431 0.1385

20.9648 12.3153 5.2411 1.4128 0.2972

21.1889 12.4876 5.2415 1.6417 0.5002

22.8511 14.9462 8.3470 3.8643 1.5736

Table 6.3 Hedging performance in Heston’s model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.2896 11.3823 4.3056 0.6939 0.0368

20.5088 11.8872 5.0132 1.1328 0.1193

20.7022 12.0223 5.0137 1.3568 0.2807

22.0898 14.1997 7.7280 3.2476 0.9834

Table 6.4 Hedging performance in the NIG-OU model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.3713 11.4467 4.4063 0.8751 0.0738

20.6307 11.8830 4.9562 1.2170 0.1566

20.7753 11.9975 4.9566 1.4321 0.3277

22.2822 14.1826 7.6203 3.2497 1.0465

143

144

Exotic Option Pricing and Advanced L´evy Models Table 6.5 Hedging performance in the NIG-CIR model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.2817 11.4069 4.4121 0.9102 0.1506

20.4979 11.8418 4.9588 1.2704 0.2864

20.6808 11.9845 4.9598 1.4781 0.4152

22.0975 14.1909 7.6878 3.2162 1.0910

Table 6.6 Hedging performance in the VG-OU model 100K/S0 80 90 100 110 120

Table 6.7

AAMC

AAc

AAtr

20.3528 11.4380 4.4083 0.9070 0.1061

20.5773 11.8695 4.9561 1.2391 0.1988

20.7447 11.9896 4.9567 1.4559 0.3506

EC 22.2073 14.1938 7.6454 3.2408 1.0433

Hedging performance in the VG-CIR model

100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.3256 11.4374 4.4383 0.9294 0.1615

20.4907 11.8395 4.9605 1.2723 0.2883

20.6766 11.9758 4.9613 1.4793 0.4152

22.1156 14.2022 7.6906 3.2159 1.0898

[38], Section 8.4.2) while the NIG paths were obtained as described in Schoutens (2003) [38] (Section 8.4.5). From Tables 6.2–6.7, we observe that the more in the money the Asian option is, the less is the difference between the option price and the comonotonic hedge. For an option with moneyness of 80% the difference is typically around 1.5%, whereas the classical hedge with the European call leads to a difference of almost 10%. For options out of the money, the difference increases, but is then substantially smaller than the differences for the other two trivial hedges. In view of the easy and cheap way in which this hedge can be implemented in practice, this static super-hedge approach seems to be competitive also in these cases. As a by-product, we observe from the Monte Carlo estimates in Tables 6.2–6.7 that the model risk for Asian option prices can be quite substantial (note that all of the models are calibrated to the same set of vanilla option prices with a quite acceptable ﬁt (the average percentage error of the ﬁt is less than 2% for all the models (cf. Schoutens (2003) [38])), but the resulting marginal densities differ considerably (cf. Figure 6.3) and consequently the Asian option prices can differ quite a lot, especially if the option is out of the money). The issue of model risk for other exotic options has recently been discussed in Schoutens et al. (2004) [39].

Static Hedging of Asian Options

6.6

145

A MODEL-INDEPENDENT STATIC SUPER-HEDGE

Since the hedging strategy introduced in this paper only depends on the risk-neutral marginal distribution functions on each averaging day of the Asian option, it can also be applied in a model-independent framework, if for all of the needed maturities tk the European call prices C(K, tk ) are available for every strike value K. In this case, the risk-neutral density function fStk is given by the second derivative of C(K, tk ) with respect to K (see e.g. Breeden and Litzenberger (1978) [13]): fStk (K) = er tk

∂ 2 C(K, tk ) . ∂K 2

In practice, call prices are available for a limited number of strike values K only, so that one has to use sophisticated statistical techniques to estimate f (Stk ). For a recently developed efﬁcient nonparametric estimation procedure utilizing shape restrictions due to no-arbitrage (such as monotonicity and convexity of the call price as a function of the strike), we refer to A¨it-Sahalia and Duarte (2003) [1]. Once the density f (Stk ) is available for all of the needed maturities tk , the hedge portfolio can be determined in just the same way as described in the above sections.

6.7 CONCLUSIONS We have shown that staticly hedging an Asian option in terms of a portfolio of European options is a simple and quick alternative to other strategies. Moreover, in contrast to most of the existing techniques, this approach is applicable in general market models whenever the risk-neutral density of the asset price distribution or an approximation of it is available. In particular, there is a fast algorithm to determine the hedge portfolio for various stochastic volatility models. Since the proposed hedging strategy is static, it is much less sensitive to the assumption of zero transaction costs and to the hedging performance in the presence of large market movements; no dynamic rebalancing is required. These advantages may sometimes compensate for the gap of the hedging price and the option price even for Asian options that are out of the money.

ACKNOWLEDGEMENTS The authors would like to thank Peter Carr for fruitful discussions on the topic. H. Albrecher acknowledges support from the K.U. Leuven (Fellowship F/04/009) and the Austrian Science Foundation Project S-8308-MAT. W. Schoutens is a Postdoctoral Fellow of the Fund for Scientiﬁc Research, Flanders, Belgium (FWO – Vlaanderen).

REFERENCES [1] A¨it-Sahalia, Y. and Duarte, J. (2003), “Nonparametric option pricing under shape restrictions”, Journal of Econometrics, 116, 9–47. [2] Albrecher, H., Dhaene, J., Goovaerts, M. and Schoutens, W. (2005), “Static hedging of Asian options under L´evy models”, Journal of Derivatives, 12(3), 63–72. [3] Albrecher, H. (2004), “The valuation of Asian options for market models of exponential L´evy type”, in M. Vanmaele, et al. (Eds), Proceedings of the Second Day of Actuarial and Financial Mathematics, Royal Flemish Academy of Arts and Sciences, Brussels, Belgium, pp. 11–20.

146

Exotic Option Pricing and Advanced L´evy Models

[4] Albrecher, H. and Predota, M. (2004), “On Asian option pricing for NIG L´evy processes”, Journal of Computational and Applied Mathematics, 172, 153–168. [5] Albrecher, H. and Predota, M. (2002), “Bounds and approximations for discrete Asian options in a variance-gamma model”, Grazer Mathematische Berichte, 345, 35–57. [6] Allen, S. and Padovani, O. (2002), “Risk management using quasi-static hedging”, Economic Notes, 31, 277–336. [7] Bakshi, G., Cao, C. and Chen, Z. (1997), “Empirical performance of alternative option pricing models”, The Journal of Finance, LII, 2003–2049. [8] Barndorff-Nielsen, O.E. (1995), “Normal inverse Gaussian distributions and the modeling of stock returns”, Research Report No. 300, Department of Theoretical Statistics, Aarhus University, Denmark. [9] Barndorff-Nielsen, O.E. (1997), “Normal inverse Gaussian distributions and stochastic volatility models”, Scandinavian Journal of Statistics, 24, 1–13. [10] Barndorff-Nielsen, O.E. and Shephard, N. (2000), “Modelling by L´evy processes for ﬁnancial econometrics”, in O.E. Barndorff-Nielsen, T. Mikosch and S. Resnick (Eds), L´evy Processes – Theory and Applications, Birkh¨auser, Boston, MA, USA, pp. 283–318. [11] Barndorff-Nielsen, O.E., Nicolata, E. and Shephard, N. (2002), “Some recent developments in stochastic volatility modelling”, Quantitative Finance, 2, 11–23. [12] Bertoin, J. (1996), L´evy Processes, Cambridge Tracts in Mathematics, Vol. 121, Cambridge University Press, Cambridge, UK. [13] Breeden, D. and Litzenberger, R. (1978), “Prices of state-contingent claims implicit in option prices”, Journal of Business, 51, 621–651. [14] Brown, H., Hobson, D. and Rogers, L.C.G. (2001), “Robust hedging of barrier options”, Mathematical Finance, 11, 285–314. [15] Carr, P. and Madan, D. (1998), “Option valuation using the fast Fourier transform”, Journal of Computational Finance, 2, 61–73. [16] Carr, P., Ellis, K. and Gupta, V. (1998), “Static hedging of exotic options”, The Journal of Finance, 53, 1165–1190. [17] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2002), “The ﬁne structure of asset returns: an empirical investigation”, Journal of Business, 75, 305–332. [18] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [19] Carr, P. and Picron J. (1999), “Static hedging of timing risk”, Journal of Derivatives, 6, 57–70. [20] Carr, P. and Wu, L. (2002), “Static hedging of standard options”, Preprint. [21] Clark, P. (1973), “A subordinated stochastic process model with ﬁnite variance for speculative prices”, Econometrica, 41, 135–156. [22] Cox, J., Ingersoll, J. and Ross, S. (1985), “A theory of the term structure of interest rates”, Econometrica, 53, 385–408. [23] Dhaene, J., Denuit, M., Goovaerts, M.J., Kaas, R. and Vyncke, D. (2002a), “The concept of comonotonicity in actuarial science and ﬁnance: theory”, Insurance: Mathematics and Economics, 31, 3–33. [24] Dhaene, J., Denuit, M., Goovaerts, M.J., Kaas, R. and Vyncke D. (2002b), “The concept of comonotonicity in actuarial science and ﬁnance: applications”, Insurance: Mathematics and Economics, 31, 133–161. [25] Eberlein, E. (1999), “Application of generalized hyperbolic L´evy motions to ﬁnance”, in O.E. Barndorff-Nielsen T. Mikosch and S. Resnick (Eds), L´evy Processes: Theory and Applications, Birkh¨auser, Boston, MA, USA, pp. 319–337.

Static Hedging of Asian Options

147

[26] Grigelionis, B. (1999), “Processes of Meixner type”, Lithuanian Mathematics Journal, 39, 33–41. [27] Heston, S. (1993), “A closed-form solution for options with stochastic volatility with applications to bond and currency options”, Review of Financial Studies, 6, 327–343. [28] Jacques, M. (1996), “On the hedging portfolio of Asian options”, ASTIN Bulletin, 26, 165–183. [29] Knudsen, Th. and Nguyen-Ngoc, L. (2003), “Pricing European options in a stochastic volatilityjump-diffusion model, DBQuant Working Paper, Deutsche Bank, London, UK; Journal of Financial and Quantitative Analysis, to be published. [30] Madan, D.B., Carr, P. and Chang, E.C. (1998), “The variance gamma process and option pricing”, European Finance Review, 2, 79–105. [31] Mandelbrot, B.B. and Taylor, H.M. (1967), “On the distribution of stock price differences”, Operations Research, 15, 1057–1062. [32] Neuberger, A. and Hodges, S. (2002), “Rational bounds on the prices of exotic options”, IFA Working Paper 359, London Business School, London, UK. [33] Nielsen, J.A. and Sandmann, K. (2003), “Pricing bounds on Asian options”, Journal of Financial and Quantitative Analysis, 38, 449–473. [34] Rydberg, T. (1997), “The normal inverse Gaussian L´evy process: simulation and approximation”, Communications in Statistics: Stochastic Models, 13, 887–910. [35] Rydberg, T. (1999), “Generalized hyperbolic diffusions with applications in ﬁnance”, Mathematical Finance, 9, 183–201. [36] Schoutens, W. and Teugels, J.L. (1998), “L´evy processes, polynomials and martingales”, Communications in Statistics: Stochastic Models, 14, 335–349. [37] Schoutens, W. (2002), “The Meixner process: theory and applications in ﬁnance”, EURANDOM Report 2002-004, EURANDOM, Eindhoven, The Netherlands. [38] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [39] Schoutens, W., Simons, E. and Tistaert, J. (2004), “A perfect calibration! Now what?”, Wilmott Magazine, March, 66–78. [40] Simon S., Goovaerts, M. and Dhaene, J. (2000), “An easy computable upper bound for the price of an arithmetic Asian option”, Insurance: Mathematics and Economics, 26, 175–183. [41] Ve˘ce˘r, J. and Xu, M. (2004), “Pricing Asian options in a semimartingale model”, Quantitative Finance, 4, 170–175.

7 Impact of Market Crises on Real Options Pauline Barrieu London School of Economics, London, UK and

Nadine Bellamy ´ Universite d’Evry Val d’Essonne, Evry, France Abstract We study the impact of market crises on investment decisions through real option theory. The framework we consider involves a Brownian motion and a Poisson process, with the jumps characterizing the crisis effects. We ﬁrst analyze the consequences of different modelling choices. We then provide the real option characteristics and establish the existence of an optimal discount rate. We also characterize the optimal time to invest and derive some properties of its Laplace Transform (bounds, monotonicity, robustness). Finally, we specify the consequences of some wrong model speciﬁcations on the investment decision.

7.1

INTRODUCTION

Investment has always been a crucial question for ﬁrms. Should a given project be undertaken? In addition, if so, when is it the best time to invest? In order to answer these questions, the neo-classical criterion of Net Present Value (N.P.V.) is still widely used. It consists in investing, if and only if, the sum of the project discounted beneﬁts is higher than the sum of its discounted costs. Such a criterion does, however, have several weaknesses. Among many others, the following facts are often mentioned: • The N.P.V. method does not take into account potential uncertainty of future cash ﬂows. • It uses an explicit calculation for the cost of the risk. • It focuses on present time: the investment decision can only be taken now or never. However, reality is often more complex and ﬂexible including, for instance, optional components for the project: a ﬁrm may have the opportunity (but not the obligation) to undertake the project, not only at a precise and given time, but during a whole period of time (or even without any time limit). In this sense, these characteristics may be related to that of an American call option, with the underlying asset being, for example, the ratio discounted beneﬁts/discounted costs, and the strike level ‘1’. Therefore, the N.P.V. criterion implies that the American option has to be exercised as soon as it is in the money, which is obviously a sub-optimal strategy. Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

150

Exotic Option Pricing and Advanced L´evy Models

The use of a method based on option theory, such as the real option theory, would improve the optimality of the investment decision. Several articles have appeared as benchmarks in this ﬁeld. The seminal studies of Brennan and Schwartz (1985), McDonald and Siegel (1986), Pindyck (1991) or Trigeorgis (1996) are often quoted as they present the fundamentals of this method, using particularly dynamic programming and arbitrage techniques. The literature on real options has been proliﬁc from very technical papers to case studies and manuals for practitioners (see among many references, the book edited by Brennan and Trigeorgis (2000) or that of Schwartz and Trigeorgis (2001)). Such an approach better suits reality by taking into account project optional characteristics such as withdrawal, sequential investment, delocalization, crisis management, etc. In that sense, real option theory leads to a decision criterion that adapts to each particular project assessment. However, real options have also some speciﬁc characteristics compared to ‘classical’ ﬁnancial options. In particular, the ‘risk-neutral’ logic widely used in option pricing cannot apply here: the real options’ underlying asset corresponds to the investment project ﬂows and is generally not quoted on ﬁnancial markets. Any replicating strategy of the option payoff is then impossible. So, the pricing is made under a prior probability measure (the historical probability measure or another measure chosen according to the investor’s expectations and beliefs). Moreover, a speciﬁc feature of a real option framework is the key points of interest for the investor. More precisely, she is interested in: • The cash ﬂows generated by the project. They are represented by the ‘price’ of the real option. Note that the notion of ‘price’ is not so obvious in this framework. It corresponds rather to the value a particular investor gives to this project. However, for the sake of simplicity in the notations, we will use the terminology ‘price’ in the rest of this paper. • But also, the optimal time to invest. This optimal time corresponds to the exercising time of the real option. Therefore, it is important noticing that real options are above all a management tool for decision taking. Once the investment project has been well-speciﬁed, the major concern for the investor is indeed summarized in the following question: ‘When is it optimal to invest in the project?’. In that sense, knowing the value of the option is less important than knowing its optimal exercising time. For that reason, in this paper, we focus especially on the properties of this optimal time. Moreover, real options studies are usually written in a continuous framework for the underlying dynamics. However, the existence of crises and shocks on investment markets generates discontinuities. The impact of these crises on the decision process is then an important feature to consider. This is especially relevant when some technical innovations may lead to instabilities in production ﬁelds. For all of these reasons, this paper is dedicated to the analysis of the exercising time properties in an unstable framework. The modelling of the underlying dynamics involves a mixed-diffusion (made up of Brownian motion and Poisson process). The jumps are negative so as to represent troubles and difﬁculties occurring in the underlying market. In the second section of this paper, we describe the framework of the study and analyze the consequences of different modelling choices. The crisis effect may be expressed via a Poisson process or the compensated martingale associated with it. Of course, there is an obvious relation between these models and they are equivalent from a static point of view. However, when studying the real option characteristics and their sensitivity towards the jump size, these models lead to various outcomes.

Impact of Market Crises on Real Options

151

After analyzing the real option characteristics in the third section, we focus on the discount rate. We prove the existence of an optimal discount rate, considering the maximization of the Laplace transform of the optimal time to invest as a choice criterion. We also characterize the average waiting time. In the ﬁfth section we study the robustness of the element decision characteristics. We ﬁrst specify the robustness of the optimal time to enter the project with respect to the jump size. We establish, in particular, that its Laplace transform is a decreasing function. Then, assuming that the investor only knows the expected value of the random jump size, we prove that this imperfect knowledge leads him/her to undertake the project too early. In the last section, we focus on the impact of a wrong model speciﬁcation, assuming that the investor believes in continuous underlying dynamics. In such a framework, we specify the error made in the optimal investment time. All proofs are presented in the Appendix.

7.2

THE MODEL

7.2.1 Notation In this paper, we consider a particular investor evolving in a universe, deﬁned as a ﬁltered probability space (, F, (Ft ), P). She has to decide whether to undertake a given investment project and, if so, when it is optimal to invest. We assume that the investor has no time limit to take this decision. Consequently, the time horizon we consider is inﬁnite. The investment opportunity value at time t = 0 is then of the form C0 = sup E exp (−µτ ) (Sτ − 1)+ τ ∈ϒ

where E denotes the expectation with respect to the prior probability measure P, ϒ is the set of the (Ft )- stopping times and (St , t ≥ 0) is the process of the proﬁts/costs ratio. It is worthwhile noticing that the discount rate µ is usually different from the instantaneous risk-free rate. We will come back later to the real meaning of discount rate in such a framework and to the problem related to its choice. The proﬁts/costs ratio related to the investment project is characterized by the following dynamics dSt = St − [αdt + σ dWt + ϕdMt ] (A) S0 = s0 where (Wt , t ≥ 0) is a standard (P, (Ft ))-Brownian motion and (Mt , t ≥ 0) is the compensated martingale associated with a (P, (Ft ))-Poisson process N . The Poisson process is assumed to have a constant intensity λ and the considered ﬁltration is deﬁned by Ft = σ (Ws , Ms , 0 ≤ s ≤ t). Equivalently, the process (St , t ≥ 0) may be written in the form: St = s0 exp(Xt ) where (Xt , t ≥ 0) is a L´evy process with the L´evy exponent E (exp (ξ Xt )) = exp (t (ξ ))

152

Exotic Option Pricing and Advanced L´evy Models

with (ξ ) = ξ 2

σ2 σ2 − λ 1 − (1 + ϕ)ξ + ξ α − λϕ − 2 2

(7.2.1)

Hence, we have

σ2 σ2 − ξ2 + λ eiξ ln(1+ϕ) − 1 E (exp (iX1 )) = exp iξ α − λϕ − 2 2 = exp (− (ξ )) Therefore, the L´evy measure associated with the characteristic exponent is expressed in terms of the Dirac measure δ as: ν (dx) = λδln(1+ϕ) (dx) Assumptions In the rest of the paper, the following hypothesis (H) holds. (i) 0 < s0 < 1, (ii) σ > 0 (iii) 0 > ϕ > −1.

(H)

Assumption (i) states that s0 is (strictly) less than 1: this is not a restrictive hypothesis, since the problem we study is a ‘true’ decision problem. In fact, delaying the project realization is only relevant in the case where the proﬁts/costs ratio is less than one. Assumption (iii) states that the jump size is negative as we study a crisis situation. The jump process allows us to take into account falls in the project business ﬁeld. These negative jumps may be induced, for instance, by a brutal introduction of a direct substitute into the market, leading to a decrease in the potential sales. Moreover, we assume that the jump size is greater than −1. This hypothesis, together with the identity 1

St = s0 (1 + ϕ)Nt × e(α−λϕ)t × eσ Wt − 2 σ

2t

ensure that the process S remains strictly positive. We also impose the integrability condition µ > sup (α; 0)

(7.2.2)

There exists an optimal frontier L∗ϕ such that sup E(e−µτ (Sτ − 1)+ ) = E(e

τ ∈ϒ

−µτL∗

ϕ

(SτL∗ − 1)+ ) ϕ

where τL is the ﬁrst hitting time of the boundary L by the process S, deﬁned as τL = inf {t ≥ 0; St ≥ L}

(7.2.3)

(For the proof, see, for instance, Darling et al. (1972) or Mordecki (1999).) Before the proﬁts/costs ratio S reaches the optimal boundary L∗ϕ , it is optimal for the investor not to undertake the investment project and to wait. However, as soon as S goes beyond this threshold, it is optimal for her to invest.

Impact of Market Crises on Real Options

153

7.2.2 Consequence of the modelling choice In the framework previously described, we may work a priori with either of the two following models: dSt = St − [αdt + σ dWt + ϕdMt ] (A) S0 = s0 (B)

dSt = St − [αdt + σ dWt + ϕdNt ] S0 = s0

In the case where all of the parameters are constant, these models are obviously equivalent and writing α = α + λϕ

(7.2.4)

is sufﬁcient to see why. Note that the integrability condition for model (B) is expressed as µ > max (α + λϕ; 0) However, when studying the sensitivity of the different option characteristics with respect to the jump size, choosing (A) or (B) really matters. Indeed, monotonicity properties are signiﬁcantly different in both frameworks, as underlined below. • Let us ﬁrst focus on the optimal time to enter the project, characterized by its Laplace transform deﬁned as E(exp(−µτL∗ϕ )). Considering model (A), if the initial value of the proﬁts/costs ratio is not ‘too small’, the Laplace transform of the optimal investment time is monotonic (this result is proved in Proposition 7.5.1). However, this monotonicity property does not hold any more for model (B) as is illustrated in Figure 7.1, which is done for the following set of parameters: s0 = 0.8;

λ = 0.1;

α = 0.05;

µ = 0.15;

σ = 0.2. 0.162 0.161 0.16 0.159 0.158 0.157 0.156 0.155 0.154

−1

−0.9

−0.8

−0.7

−0.6

−0.5 j

−0.4

−0.3

−0.2

−0.1

0.153 0

Figure 7.1 Laplace transform of the optimal time to invest (model (B))

154

Exotic Option Pricing and Advanced L´evy Models

• We now focus on the investment opportunity value C0 . Proposition 7.2.1 Let us consider model (B). Then, the investment opportunity value is an increasing function of the jump size. Figure 7.2 illustrates Proposition 7.2.1. It represents the variations of the investment value with respect to the jump size for different values of the jump intensity and for the following set of parameters: s0 = 0.8;

α = 0.05;

µ = 0.15;

σ = 0.2.

However, this property of the investment opportunity value does not hold any more when considering model (A). Intuitively, the studied model leads to a double effect of the jump size on the underlying level: ϕ has a positive effect on the underlying by increasing the drift but it also has a negative effect on the underlying by acting on the Poisson process level: dSt = St − ((α − λϕ) dt + σ dWt + ϕdNt ) This double effect explains the differences between models (A) and (B), and in particular accounts for the following result: in setting (A), the maximum value of C0 is not necessarily obtained for ϕ = 0. As a conclusion, it cannot be said that one of these models is better or more relevant than the other one. From a static point of view (with respect to the parameter ϕ), both are mathematically equivalent. In particular, given the condition shown in equation (7.2.4), they lead to the same ﬁrst and second moments for S. However, from a dynamic point of view with respect to the jump size, they are different. In the setting (B), crisis is only detected as the spread between the level of S before and after a shock while on the other hand, in the setting (A), there is an additional effect of the shocks on the drift term of S. Economically speaking, both have their own interests and motivations. However, once a model is chosen, the consequences of this choice must be 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 −1

−0.9

−0.8

−0.7

−0.6

λ = 0.1

Figure 7.2

−0.5 j

−0.4

λ = 10

−0.3

−0.2

λ=1

Investment opportunity value (model (B))

−0.1

0 0

Impact of Market Crises on Real Options

155

kept in mind, especially the implications for the monotonicity properties of the real option characteristics. In this present study, since we are particularly interested in the optimal time to invest, we choose a martingale representation for the stochastic part of SdS−t ; therefore, the model t deﬁned by (A) prevails in the following.

7.3 THE REAL OPTION CHARACTERISTICS In this section, we ﬁrst recall the classical formulae for the optimal time to invest and for the investment opportunity. We denote by kϕ the unique real number deﬁned in terms of the L´evy exponent deﬁned in equation (7.2.1) since it satisﬁes: kϕ > 1 and kϕ = µ Then the optimal proﬁts/costs ratio L∗ϕ satisﬁes: L∗ϕ =

kϕ kϕ − 1

The investment opportunity value at time 0 is given by: C0 =

s0 kϕ

kϕ

1 kϕ − 1

1−kϕ (7.3.1)

and the optimal investment time is characterized by its Laplace transform: E exp −µτL∗ϕ =

kϕ s0 kϕ − 1 kϕ

(7.3.2)

(For detailed proofs, see, among others, Gerber and Shiu (1994), Bellamy (1999) and Mordecki (1999, 2002).) It can be noticed that kϕ , as well as the optimal proﬁts/costs ratio L∗ϕ , depend on ϕ, λ and µ. Remark 1. In the framework we deal with, the so-called principle of smooth pasting is satisﬁed. Such a principle is always satisﬁed in a continuous framework but if the model is driven by discontinuous L´evy processes, this property can fail. In the model we consider, however, the smooth pasting principle still holds (see, for instance, Chan (2003, 2005), Boyarchenko and Levendorskii (2002), Alili and Kyprianou (2004) or Avram et al. (2004)). It is also easy to check that the optimal proﬁts/costs ratio satisﬁes L∗ϕ > 1. This underlines the interest of waiting before undertaking the project, as well as the gain in optimality obtained from considering a real option approach rather than the standard N.P.V. method (see, for instance, Dixit et al. (1993)).

156

Exotic Option Pricing and Advanced L´evy Models Table 7.1 Values of the optimal beneﬁts/costs ratio L∗ϕ as a function of σ and ϕ σ \ϕ

−0.9

−0.5

−0.3

−0.1

0.1 0.2 0.4 0.6

16.25 17.13 18.24 20.23

6.59 6.93 8.19 10.26

4.39 4.74 6.08 8.19

3.29 3.69 5.12 7.29

The value of the optimal ratio may be much greater than the limit value ‘1’. This fact is at variance with the N.P.V. criterion and perfectly illustrates what McDonald and Siegel (1986) have called ‘The value of waiting to invest’. As an illustration, the optimal ratio L∗ϕ is calculated in Table 7.1 for the following set of parameters: µ = 0.15;

λ = 1;

α = 0.1;

s0 = 0.8.

Note that high values for the volatility coefﬁcient σ are also considered in this study. This is relevant since the underlying market related to the investment project may be more highly volatile than traditional ﬁnancial markets (for instance, markets related to new technology).

7.4 OPTIMAL DISCOUNT RATE AND AVERAGE WAITING TIME 7.4.1 Optimal discount rate We now focus on the discount rate µ and present some general comments about its choice, which is indeed crucial in this study. The rate µ does not correspond to the instantaneous risk-free rate, traditionally used in the pricing of standard ﬁnancial options. In fact, in this real option framework, the rate µ characterizes the preference of the investor for the present or her aversion for the future. Choosing the ‘right’ µ is extremely difﬁcult. Many different authors have been interested in this question (among many others, Weitzman (1998)). Some have also proved the existence of a speciﬁc relationship between discount rate and future growth rate (Gollier (2002), Gollier and Rochet (2002) and Kimball (1990)). The optimal choice criterion for the rate µ depends, however, on the considered framework. We present here a relevant criterion for this particular problem, corresponding to the maximization of the Laplace transform of the optimal investment time. Proposition 7.4.1 (i) There exists a unique real number µ strictly positive such that E(exp(− µτL∗µ )) = max E(exp(−µτL∗µ )) µ

The real number µ agrees with an optimal choice of the discount rate µ. (ii) The optimal discount rate µ increases with the jumps intensity and decreases with the jumps size.

Impact of Market Crises on Real Options

157

This optimal discount rate is increasing with the absolute value of the jump size and with the intensity of the jumps. Such a behaviour seems rather logical as the occurrence and the frequency of negative jumps in the future make the value of the project decrease and represent an additional risk for the investor. The more important the jump intensity and size in absolute values are, the more the investor favours the present. Thus, she will choose a higher discount rate. Figure 7.3 shows the variations of the optimal rate µ with respect to ϕ for different values of λ and for the following set of parameters: s0 = 0.8;

α = 0.1;

σ = 0.2.

Remark 2. Other criteria may have been considered in order to choose an optimal rate. For instance, the maximization of C0 could appear as an alternative. However, it is not a relevant criterion, since the function kϕ

s0 kϕ − 1 kϕ µ → C0 = −1 kϕ − 1 kϕ is strictly decreasing. 7.4.2 Average waiting time Another question relative to the best time to invest is, of course, that of the characterization of an average waiting time. If we denote this by Tc , it is deﬁned as the unique element of R∗+ such that: E(exp(− µτL∗µ )) = exp(− µT c ) Hence, Tc corresponds to the average waiting time. In fact, it is the certainty equivalent µ. As of τL when the utility criterion is exponential and the risk aversion coefﬁcient is previously seen, this rate µ can easily be interpreted as a future aversion coefﬁcient (or a present preference coefﬁcient) and Tc may be explicitly determined as: Tc = −

1 ln E(exp(− µτL∗µ )) µ 2.5 2

^ m

1.5 1 0.5

−0.9

−0.8

−0.7

−0.6 λ = 0.2

−0.5

j

−0.4

λ=1

−0.3

−0.2 λ=2

Figure 7.3 Optimal discount rate, µˆ

−0.1

0 0

158

Exotic Option Pricing and Advanced L´evy Models 7 6 5

Tc

4 3 2 1

−0.9

−0.8

−0.7

−0.6

−0.5 j

λ = 0.2

Figure 7.4

−0.4 λ=1

−0.3

−0.2

−0.1

0 0

λ=2

Average waiting time, Tc

From Proposition 7.4.1, we deduce that the average waiting time decreases with respect to the jump intensity, as well as to the absolute value of the jump size. This mathematical property can be economically understood as previously. In fact, jumps induce additional risks, increasing with previous jump intensity and the jump size absolute value. The average waiting time can be related to an exponential utility criterion. Therefore, the investor we consider appears to be risk averse, with an exponential utility function and a risk aversion coefﬁcient of µ. So, in her decision process, she will take into account the expected proﬁt as well as the associated risk. She will tend to reduce the risk induced by the business ﬁeld by entering earlier in the project. Obviously, the more she waits, the greater the probability of jumps and then the risk are. Figure 7.4 highlights this fact. It represents the variations of the average waiting time with respect to the jump size. The graphs are produced for different values of the jump intensity. All of these curves converge to the same point as the jump size tends to zero: this point corresponds to the average waiting time in the model without jump, or, in other words, in an universe without crisis. The following set of parameters has been used: s0 = 0.8;

α = 0.1;

σ = 0.2.

7.5 ROBUSTNESS OF THE INVESTMENT DECISION CHARACTERISTICS All of the different parameters of the model have to be estimated using historical data or strategic anticipations. Every estimation and calibration may lead to an error on the choice of the input parameters. Some stability (or robustness) of the results is an essential condition for a real practical use of a model.

Impact of Market Crises on Real Options

159

7.5.1 Robustness of the optimal time to invest As it has already been underlined, the optimal time to invest is the major concern of the investor. Hence, the robustness of its Laplace transform appears as a key point to be checked. We particularly focus on the study of the sensitivity of this quantity with respect to the jump size. We study the behaviour of the Laplace transform of the optimal time to invest when the jump size is not perfectly known: the investor only knows that there exists ϕ and ϕ such that −1 < ϕ ≤ ϕ ≤ ϕ < 0 We ﬁrst provide a monotonicity result. Proposition 7.5.1 Let s0 be the level deﬁned as s0 = satisﬁes s0 < s0 < 1

k0 k0 −1

exp(− k01−1 ). We assume that s0 (7.5.1)

Then, the Laplace transform of the optimal time to invest is an increasing function of the jump size. Proposition 7.5.1 can be heuristically interpreted as follows: the more the jump size increases (hence decreases in absolute value), the more the investor delays entering the investment project. The maximum waiting time is attained in the lack of jump. Remark 3. The Assumption s0 < s0 amounts to consider investment project only if the initial value is not ‘too small’. From an economic point of view, such an assumption is not very restrictive. In fact, the investor will stop being interested in the project as soon as s0 is below a given threshold. If, for example, we consider the following standard set of parameters α = 0.10; σ = 0.20; µ = 0.15, then we get s0 = 0.276. Note that this level s0 is far from the strike value 1. Figure 7.5 shows the changes in the Laplace transform of the optimal time to invest with respect to ϕ for different values of λ. The following set of parameters is used: s0 = 0.8;

α = 0.10;

σ = 0.20;

µ = 0.15.

The robustness property of the Laplace transform is a straightforward consequence of Proposition 7.5.1. Corollary 7.5.2 We assume that the condition shown in equation (7.5.1) holds and −1 < ϕ ≤ ϕ ≤ ϕ < 0 Then, we have E(exp(−µτL∗ϕ )) ≤ E(exp(−µτL∗ϕ )) ≤ E(exp(−µτL∗ϕ )).

160

Exotic Option Pricing and Advanced L´evy Models 0.14 0.12 0.1 0.08 0.06 0.04 0.02 −1

−0.9

−0.8

−0.7

−0.6

λ = 0.2

−0.5 j

−0.4

λ=1

−0.3

−0.2

−0.1

0 0

λ = 10

Figure 7.5 Laplace transform of the optimal time to invest for different values of λ

This result underlines the model robustness as far as the Laplace transform of the optimal time to invest is concerned. More precisely, if the investor does not know exactly the size of the jump, in other words the impact of the market crisis on the project, but knows, however, some boundaries for it, then she has an idea of the optimal time to enter the project. More precisely, the Laplace transform boundaries are expressed in terms of the boundaries for the market crisis impact. Equivalently, having some control or knowledge of the crisis impact enables the investor to have some control of her optimal time to invest. 7.5.2 Random jump size We now consider the situation where the jump size is an unknown random variable . We focus on the impact that this additional hazard may have on the investor decision. Assuming that the investor estimates the jump size by its expected value E (), we focus on the impact of such an error on her decision. Will she invest too early or too late? In order to answer this questions, we compare the ‘true’ Laplace transform of the optimal time to invest, with the Laplace transform estimated by means of E (). The dynamics of the process of the project are now: dSt = St− (αdt + σ dWt + dMt ) ; S0 = s0 and the investor builds her strategy from StE() where E() dStE() = St− (αdt + σ dWt + E () dMt ) ; S0E() = s0

We assume that the random variable is independent of the ﬁltration generated by the Brownian motion and the Poisson process. Let L∗ be the true optimal beneﬁt–cost ratio. If the investor only knows E (), she estimates this ratio by L∗E() . The next proposition provides a comparison between these two quantities.

Impact of Market Crises on Real Options

161

Proposition 7.5.3 We assume that the condition shown in equation (7.5.1) holds. Then, the wrong speciﬁcation in the model leads the investor to underestimate the optimal proﬁts/costs ratio. Moreover we can pr´ecis the consequences of this error on the decision taking. We assume that the investor undertakes the project when the observed process of the beneﬁts/costs ratio reaches what she supposes to be the optimal level. Therefore, her strategy is determined by the ﬁrst hitting time of L∗E() , instead of the ﬁrst hitting time of L∗ by process S. This proposition can be interpreted as follows: when the investor only knows E (), she tends to undertake the project too early.

7.6 CONTINUOUS MODEL VERSUS DISCONTINUOUS MODEL In this section, we focus on the impact of a wrong model choice. This part extends the previous study of robustness. We suppose that the investor believes in a continuous underlying dynamics for S, while its true dynamics is given by (A). As a consequence, the investor governs her strategy according to the following process: St ($ α dt + $ σ dWt ) d$ St = $ where

$ (A)

S0 = s0 $ $ α =α 2 $ σ = σ 2 + λϕ 2 .

$ on the same These equalities come directly from the calibration of both model (A) and (A) $ data set, leading to the same ﬁrst and second moments for S and S. The volatility parameter of the model without jump is different from that of the model with jumps: the absence of jump in the dynamics is indeed compensated by a higher volatility. In order to obtain the ‘equivalent’ volatility, the right brackets of S and $ S have to be equal. The process $ S is called ‘equivalent process without jump’. We now focus on the impact of such a wrong speciﬁcation on the investment time. To this end, we ﬁrst consider the error in the optimal proﬁts/costs ratio. 7.6.1 Error in the optimal proﬁt–cost ratio $ More precisely, $∗ϕ the optimal proﬁts/costs ratio in the model deﬁned by (A). We denote by L ∗ using the same arguments as presented in Section 7.3, $ Lϕ is given by the following ratio $ L∗ϕ =

$ kϕ $ kϕ − 1

where $ kϕ is the solution of

σ 2 + λϕ 2 2 σ 2 + λϕ 2 $ ψ (k) = k + α− k=µ 2 2

(7.6.1)

162

Exotic Option Pricing and Advanced L´evy Models

Note that this optimal ratio depends on the volatility parameter of the model, or equivalently, on both jump parameters ϕ and λ. For the sake of simplicity, as we are especially interested $∗ϕ . in the sensitivity with respect to the jump size, we use the notation L Proposition 7.6.1 The previous wrong speciﬁcation of the model leads the investor to underestimate the optimal proﬁts/costs ratio, if and only if, σ 2 + λϕ 2 + 2α ≥ µ

(7.6.2)

Note that for the usual values of the parameters, the inequality shown in equation (7.6.2) often holds. For instance, if we consider λ = 1, α = 0.1, σ = 0.2 and µ = 0.15, then σ 2 + λϕ 2 + 2α ≥ µ is true for all ϕ in ] − 1, 0[. As an illustration, the relative error (expressed in percentage) on the optimal proﬁts/costs ratio

$∗ϕ L∗ϕ − L ∗ RE(L , ϕ) = 100 × L∗ϕ is calculated in Table 7.2 for different values of the jump size ϕ and for the standard set of parameters: s0 = 0.8; α = 0.1; σ = 0.20; µ = 0.15; λ = 1. Very naturally, the relative error becomes negligible as the jump size tends to zero. This error is still manageable when the jump size is not too large (up to −0.5). For larger values, however, the relative error becomes quite important to reach more than a third of the value of the ratio when the jump size is maximal. Using the same argument as in the previous section, we can pr´ecis the consequences that this wrong speciﬁcation has on the investor’s strategy. The investor’s waiting time is $∗ϕ instead of L∗ϕ . So, if the condition shown in equation (7.6.2) holds, we determined by L can assert that the error in the model leads the investor to undertake the project too early. This fact is brought to the fore by Figure 7.6. The optimal time to enter the project for a well-informed investor, as well as that of the previous investor, are respectively characterized by the Laplace transforms E(exp(−µτL∗ϕ )) and E(exp(−µτ$ L∗ϕ )). Figure 7.6 represents the variations of these Laplace transforms with respect to the jump size ϕ. This is, carried out for the following values: s0 = 0.8;

α = 0.1;

σ = 0.20;

µ = 0.15;

λ = 1.

As another illustration, the relative error (expressed in percentage) on the Laplace transform of the optimal time to invest

. LT − LT RE(LT , ϕ) = 100 × LT Table 7.2

Relative error on the optimal proﬁts/costs ratio as a function of ϕ

ϕ

−0.995

−0.7

RE(L∗ , ϕ)

38.30

15.81

−0.5 7.11

−0.3 1.87

−0.1 0.08

−0.01 0.01

Impact of Market Crises on Real Options

163

0.12 0.10 0.08 0.06 0.04 0.02 −1

−0.8

−0.6

−0.4 j

−0.2

0

~ LT of the hitting time of L∗

LT in the model with jumps

Figure 7.6 Laplace transform of the optimal time to invest in the model with jump and Laplace transform of the hitting time of the proﬁt–cost ratio, L˜ ∗ Table 7.3 Relative error on the Laplace transform of the optimal time to invest as a function of ϕ ϕ RE(LT , ϕ)

−0.995 −51.77

−0.7

−0.5

−0.3

−0.1

−0.02

−14.68

−5.58

−1.26

−0.04

−0.01

is calculated in Table 7.3 for different values of the jump size ϕ and for the previous set of parameters. The interpretation of these results is very similar to those associated with the relative error on the optimal proﬁts/costs ratio. It can be noticed, however, that for large values of the jump size, the relative error becomes quite important to reach more than a half of the Laplace transform when the jump size is maximal. Hence, the impact of a wrong model speciﬁcation could be important if the investor focuses on the optimal time to invest in the project. 7.6.2 Error in the investment opportunity value In the ‘true’ model with jumps, the investment opportunity value is C0 . If we assume that the investor becomes involved in the project when the ‘true’ process S reaches the level $∗ϕ , then her investment opportunity value is L $∗ϕ − 1)E(exp(−µτ$∗ )) $0 = (L C Lϕ in which

$∗ϕ − 1) × $0 = (L C

where $ kϕ is the solution of equation (7.6.1).

s0 $ L∗ϕ

$kϕ

164

Exotic Option Pricing and Advanced L´evy Models 0.65 0.6 0.55 0.5 0.45 0.4 0.35 −1

−0.9

Figure 7.7

−0.8

−0.7

−0.6

−0.5 j ~ C0

−0.4

−0.3

−0.2

−0.1

0.3 0

C0

Investment value estimated with L˜ ∗ and the optimal investment value

$0 with respect to the jump size ϕ. Of Figure 7.7 represents the variations of C0 and C $∗ϕ differs from the optimal frontier L∗ , we have for any ϕ, course, since L $0 ≤ C0 C $0 comes from a wrong investment time. This loss tends to zero when and the loss C0 − C $∗ϕ tends to the optimal frontier the jump size tends to zero and this fact was expected as L ∗ L when ϕ tends to 0. The curves shown in Figure 7.7 are produced by using the following values: s0 = 0.8; α = 0.1; σ = 0.20; µ = 0.15; λ = 1. As another illustration, the relative error (expressed in percentage) on the investment opportunity value RE(C, ϕ) = 100 ×

$0 C0 − C C0

is calculated in Table 7.4 for different values of the jump size ϕ and for the previous set of parameters. The relative error remains manageable even for large values of the jump size since it is always less than 10%. Therefore, the impact of a wrong model speciﬁcation is relatively not so important if the investor focuses on the value of the investment opportunity. Table 7.4 Relative error on the investment opportunity value as a function of ϕ ϕ

−0.995

RE(C, ϕ)

9.02

−0.7 5.33

−0.5 3.19

−0.3 1.14

−0.1 0.06

−0.02 0.01

Impact of Market Crises on Real Options

165

7.7 CONCLUSIONS In this paper, we study the impact of market crises on investment decision via real option theory. The investment project, modelled by its proﬁts/costs ratio, is characterized by a mixed diffusion process, whose jumps represent the consequences of crises on the investment ﬁeld. After having analyzed the implications of different model choices, we study the real option associated with this investment project. We establish the existence of an optimal discount rate, given a criterion based on this investment time and we characterize the average waiting time. We study in detail the properties of the optimal investment time, through its Laplace transform, and focus, in particular, on its robustness when the underlying dynamics of the project are not well-known or are wrongly speciﬁed. We interpret the results in terms of the investment decision. More precisely, when the investor bases his/her decision on the expected value of the random jump size, he/she tends to undertake the project too early. The same property holds if he/she believes in a continuous dynamics for the underlying project. In this paper, we focus on a single investor. The complexity of reality suggests, however, that different other aspects, in particular, strategic relationships between the economic agents, may play an important role. Investigating more general models involving strategic dimensions and game theory is a topic for future research.

APPENDIX Proof of Proposition 7.2.1

ϕ Let S be deﬁned by model (B). We deﬁne C (ϕ, L) as C (ϕ, L) = (L − 1) × E exp −µτL ϕ where τL = inf {t ≥ 0; St ≥ L}. Hence C0 (ϕ) = C ϕ, L∗ϕ where L∗ϕ is the optimal frontier, that is to say, the optimal beneﬁt–cost ratio. Let ϕ2 and ϕ1 be such that −1 < ϕ2 < ϕ1 < 0. We have C0 (ϕ1 ) = C ϕ1 , L∗ϕ1 ≥ C ϕ1 , L∗ϕ2 Then inequality ϕ1 > ϕ2 leads to ∀t ≥ 0, St (ϕ1 ) ≥ St (ϕ2 ) and consequently ϕ

ϕ

E(exp(−µτL∗ϕ1 )) ≥ E(exp(−µτL∗ϕ2 )) 2

∗

2

∗

Finally, we get C0 (ϕ1 ) ≥ C(ϕ1 , L (ϕ2 )) ≥ C(ϕ2 , L (ϕ2 )) = C0 (ϕ2 ). Proof of Proposition 7.4.1 (i) The function k ∈ ]1, +∞[ →

s0 (k−1) k

ln s0 + ln

k

admits a maximum for k = k, deﬁned by:

k−1 1 + =0 k k−1

(7.A.1)

166

Exotic Option Pricing and Advanced L´evy Models

The study of the L´evy exponent leads to the existence of a unique value of µ, denoted µ k. Moreover, µ satisﬁes: by µ, such that µ > α and kϕ() = E(exp(− µτL∗µ )) = max E(exp(−µτL∗µ )) µ

Assertion (ii) comes from the deﬁnition of µ and the following properties of the L´evy exponent: ∀k ∈ ]1, k[ ∀ϕ ∈ ] − 1, 0[,

λ → (k)

is increasing and ∀k ∈ ]1, k[ ∀λ > 0,

ϕ → (k)

is decreasing. Proof of Proposition 7.5.1 Let k be deﬁned by equation (7.A.1). We have k ⇐⇒ s0 ≥ s0 k0 ≤ where s0 =

1 k0 exp − k0 − 1 k0 − 1

and where k0 is the limit: k0 = lim kϕ . ϕ→0

In order to get the conclusion, it sufﬁces to prove that kϕ is strictly increasing with respect to the jump size ϕ. Let F : ] − 1; 0[ × ]1; +∞[→ R be the function deﬁned as: F (ϕ, k) = (k) − µ where is given by equation (7.2.1). For any (ϕ, k) ∈ ] − 1; 0[ × ]1; +∞[ such that F (ϕ, k) = 0, we can easily check that (k) > 0. Using the implicit function theorem, we get: ∂F ∂k ∂ϕ (ϕ, k) = − ∂F ∂ϕ ∂k (ϕ, k)

and the inequality

∂F ∂ϕ

(ϕ, k) < 0 implies

increasing. Proof of Proposition 7.5.3

∂k > 0. Hence the function ϕ → kϕ is strictly ∂ϕ

We denote by and E() the L´evy exponents of the processes Xt t≥0 and XtE() , t≥0

E() St S and XtE() = ln . where Xt = ln t s0 s0

Impact of Market Crises on Real Options

167

Let k (resp.kE() ) be the unique real number strictly greater than 1 such that (k ) = µ (resp. E() kE() = µ). We have (k) = f (, k) + g(k) (resp. E() (k) = f (E(), k) + g(k)) where f (, k) = λ(1 + )k − λk and g (k) =

σ2 2 σ2 k + α− k − λ. 2 2

The convexity of the function x → f (x, k) for any k > 1, together with Jensen inequality, implies that ∀k > 1, E() (k) ≤ (k) Hence k ≤ kE() and from this last inequality, we conclude L∗ ≥ L∗E() . Proof of Proposition 7.6.1

$t $ be the L´evy exponent of the process X Let

t≥0

$ $t = ln St where X s0

and $ kϕ be the

unique real number such that $ kϕ > 1

$ $ kϕ = µ.

Then, from the equalities $ $ (0) = (0) = 0 and (2) = (2) = σ 2 + λϕ 2 + 2α, we get

$ kϕ ≥ kϕ ,

if and only if, σ 2 + λϕ 2 + 2α ≥ µ and therefore we have

$∗ϕ L∗ϕ ≥ L

REFERENCES [1] Alili, L. and Kyprianou, A. (2004), “Some remarks on ﬁrst passage of L´evy processes, the American put and pasting principles”, Annals of Applied Probability, to be published. [2] Avram, F., Kyprianou, A. and Pistorius, M. (2004), “Exit problems for spectrally negative L´evy processes and applications to Russian, American and Canadized options, Annals of Applied Probability, 14, 215–238. [3] Bellamy, N. (1999), Hedging and Pricing in a Market Driven by Discontinuous Processes, Ph. D. Thesis, Universit´e d’Evry Val d’Essonne, Evry, France.

168

Exotic Option Pricing and Advanced L´evy Models

[4] Boyarchenko, S.I. and Levendorskii, S.Z. (2002), “Perpetual American options under L´evy processes”, SIAM Journal on Control and Optimization, 40, 1663–1696. [5] Brennan, M.J. and Schwartz, E.S. (1985), “Evaluating natural resource investments”, Journal of Business, 58, 135–157. [6] Brennan, M.J. and Trigeorgis, L. (Eds) (2000), Project Flexibility, Agency, and Product Market Competition – New Developments in the Theory and Applications of Real Options, Oxford University Press, New York, NY, USA. [7] Chan, T. (2003), “Some applications of L´evy Processes in insurance and ﬁnance”, Finance, 25, 71–94. [8] Chan, T. (2005), “Pricing perpetual American options driven by spectrally one-sided L´evy processes”, in A.E. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 195–216. [9] Darling, D.A., Ligget, T. and Taylor, H.M. (1972), “Optimal stopping for partial sums”, Annals of Mathematical Statistics, 43, 1363–1368. [10] Dixit, A.K. and Pindyck, R. S. (1993). Investment under Uncertainty, Princeton University Press, Princeton, NJ, USA. [11] Gerber, H.U. and Shiu, E.S. (1994), “Martingale approach to pricing perpetual American options”, ASTIN Bulletin, 24, 195–220. [12] Gollier, C. (2002), “Time horizon and the discount rate”, Journal of Economic Theory, 107, 463–473. [13] Gollier, C. and Rochet, J.C. (2002), “Discounting an uncertain future”, Journal of Public Economics, 85, 149–166. [14] Kimball, M.S. (1990), “Precautionary saving in the small and in the large”, Econometrica, 58, 53–73. [15] McDonald, R. and Siegel, D. (1986), “The value of waiting to invest”, Quarterly Journal of Economics, 101, 707–727. [16] Mordecki, E. (1999), “Optimal stopping for a diffusion with jumps”, Finance and Stochastics, 3, 227–236. [17] Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493. [18] Pindyck, R.S. (1991), “Irreversibility, uncertainty and investment”, Journal of Economic Literature, 29, 1110–1148. [19] Schwartz, E.S. and Trigeorgis, L. (2001), Real Options and Investment under Uncertainty: Classical Readings and Recent Contributions, MIT Press, Cambridge, MA, USA. [20] Trigeorgis, L. (1996), Real Options – Managerial Flexibility and Strategy in Resource Allocation, MIT Press, Cambridge, MA, USA. [21] Weitzman, M.L. (1998), “Why the far-distant future should be discounted at its lowest possible rate”, Journal of Environmental Economics and Management, 36, 201–208.

8 Moment Derivatives and L´evy-type Market Completion Jose´ Manuel Corcuera and David Nualart University of Barcelona, Barcelona, Spain and

Wim Schoutens Katholieke Universiteit Leuven – UCS, Leuven, Belgium Abstract ´ We show how moment derivatives can complete Levy-type markets in the sense that by allowing trade in these derivatives any contingent claim can be perfectly hedged by a dynamic portfolio in terms of bonds, stocks and moment-derivative-related products. Moment derivatives depend on the sum of the powered returns, i.e. the realized moments. Squared log-returns are the basis of the nowadays popular Variance Swaps. Higher-powered returns assess other kinds of important characteristics of the underlying distribution such as skewness and kurtosis. We ﬁrst work under a discrete time setting under which we assume that the returns of the stock price process are independent and identically distributed. Out of the Taylor expansion of the payoff function, we extract the positions one has to take in order to perfectly hedge the claim. We illustrate this by some illustrative examples such as the Trinomial tree model. Next, we comment on the continuous time setting. In this case, a Martingale Representation Property lies at the heart of the completion on the market considered. Results in this ´ market were already obtained in previous work of these authors. A survey exponential Levy of the relevant results are given and the relation and similarities with the discrete setting are discussed.

8.1

INTRODUCTION

In this paper, we consider markets where the returns are independent and identically distributed (iid). Typically, these markets are incomplete, and the purpose of this work is to show a systematic way of completing these markets. We shall complete the market by introducing a series of assets related to the powers of the return process. First we present the procedure in a discrete-time setting with discrete returns, while, secondly we consider more general returns, and ﬁnally we consider the continuous-time setting. In fact this latter case has been considered in Corcuera et al. (2004a) [10] and in such a case the new assets are based on the power-jump processes of the underlying L´evy process. In addition, these new assets can be related with options on the stock (see Balland (2002) [2]) Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

170

Exotic Option Pricing and Advanced L´evy Models

and with contracts on realized variance (Carr and Madan (1998) [7] and Demeterﬁ et al. (1999) [15]) that have found their way into OTC markets and are now traded regularly. Higher order power-jump processes have a similar relationship with which one could call realized skewness and realized kurtosis processes. Contracts on these objects, however, are not common. Carr et al. (2002) [8] and Carr and Lewis (2004) [6] have studied contracts on the quadratic variation processes in a model driven by a so-called Sato process. We give an explicit hedging portfolio for claims whose payoff function depends on the prices of the stock and the new assets at maturity. Then, if we introduce utility functions, we can obtain the optimal terminal wealth with respect to these utilities and by the completeness of the enlarged market we can obtain the optimal portfolio by duplicating the optimal wealth. This has been carried out by Corcuera et al. (2004b) [11], where we also analyze the case where the optimal portfolio consists only in stocks and bonds. This corresponds to complete the market with new assets in such a way that they are superﬂuous, that is, we do not improve the terminal expected utility by including these new assets in our portfolio. This is equivalent to choosing an appropriate risk-neutral or martingale measure (see Kallsen (2000) [18] and Schachermayer (2001) [28]). Moreover, this martingale measure is related to the neutral derivative pricing of Davis (1997) [13].

8.2

MARKET COMPLETION IN THE DISCRETE-TIME SETTING

We start by explaining the ideas in the most simple incomplete discrete market setting: the one-step trinomial market model. Next, we will consider a one-step market model, where the stock can attain m different values, then we will consider the same model but with n time-steps, and ﬁnally we will deal with a general multi-step market. 8.2.1 One-step trinomial market In this model, we assume we have a risk-free bond paying out a constant interest rate r > 0, i.e. the bond has a deterministic value process: B0 = 1 and B1 = 1 + r. We have also a risky asset, a stock, which can move from its initial value S0 > 0 to three different values at time 1. More precisely, we have S1 = S0 (1 + X1 ), where X1 can take the values −1 < x1 < x2 < x3 . It is a classical argument, that in order to avoid arbitrage one should have x1 < r < x3 (by investing in stocks you can lose more, but also gain more, than by investing in bonds). This arbitrage-free market is one of the most simple cases of an incomplete market, in the sense that there exist contingent claims which cannot be hedged by positions in bonds and stocks. We will show that by introducing a moment option (a Variance-Swap-like derivative), the model can be completed. Moreover, we show that the position one has to take, in order to hedge any contingent claim, can just be read off from the Taylor expansion of the payoff function of the claim. Indeed, suppose we allow also trade in a contingent claim, paying out X12 at time 1. We will refer to this derivative as the MOM (2) derivative. Let us denote the price of this contingent claim at time zero by z2 . In order to exclude arbitrage, there must be an equivalent martingale measure, making the discounted values of all traded securities martingales. Denoting the risk-neutral probabilities that X1 attains the value xi by qi > 0, i = 1, 2, 3, we must have

Moment Derivatives and L´evy-type Market Completion

171

q1 + q2 + q3 = 1 q1 x1 + q2 x2 + q3 x3 = r q1 x12 + q2 x22 + q3 x32 = z2 (1 + r), where the ﬁrst equation is ensuring that we have a probability measure, the second equation makes the risk-neutral return on the stock equal to r and the third equation ﬁxes the price of the MOM (2) derivative at z2 . This system of equations can be rewritten in matrix form as 1 1 1 q1 1 . $ · q = x1 x2 x3 · q2 = r z2 (1 + r) q3 x12 x22 x32 Since $ is a Vandermonde matrix, det($) = 0 and $ is invertable. So, the system has exactly one solution, namely: 1 . q = $−1 · r z2 (1 + r) If this solution satisﬁes q ∈ (0, 1)3 , i.e. the qi s can be seen as probabilities, we have no arbitrage. Moreover, since the solution is unique, we have also, by the second fundamental theorem of asset pricing, that the market is complete. Note that to choose an arbitrage price z2 is equivalent to choosing a risk-neutral probability q. Consider now a general contingent claim, with payoff function G(X1 ) and develop this function into powers of X1 : G(X1 ) = a0 + a1 X1 + a2 X12 . Since X1 can take only three possible values, the series is cut off after the quadratic term. In order to hedge this claim one needs to carry out the following: • Invest (a0 − a1 )(1 + r)−1 into bond. • Buy a1 /S0 units of stock, for a total price a1 . • Buy a2 units of MOM (2) derivatives, for a total price a2 z2 . At time t = 1, we have the following: • The money invested in bond has grown to a0 − a1 . • We sell the a1 /S0 stocks, giving us a1 (1 + X1 ) of money. • The MOM (2) derivatives each pay out X12 . This leads to a total payout of a2 X12 . In total we thus end up with a0 + a1 X1 + a2 X12 = G(X1 ) of money, exactly the payout of the contingent claim considered. In order to set up this strategy we needed a0 − a1 + a1 + a2 z2 1+r of money, which in order to avoid arbitrage must be the initial price of the contingent claim with payoff function G(X1 ).

172

Exotic Option Pricing and Advanced L´evy Models

8.2.2 One-step ﬁnite markets The above situation can easily be generalized to a (one-step) setting, where the random variable X1 can take a ﬁnite number m of possible values −1 < x1 < · · · < xm , with x1 < r < xm , to avoid arbitrage. For m ≥ 3, the market is an incomplete market; there exist contingent claims which cannot be hedged by holding positions in bonds and stocks alone. Assume trade is allowed into moment derivatives with payoff functions MOM (k) = X1k ,

k = 2, . . . , m − 1.

So, besides investing in bonds and stocks, one can invest also into m − 2 other derivatives, i.e. the MOM (k) ’s moment derivatives. Note that payoff functions and initial prices can be negative. For example, in the case of a negative payoff, the holder must pay the corresponding amount to the issuer. Let zk , k = 2, 3, . . . , m − 1, be the initial price of the MOM (k) derivative. In order to exclude arbitrage, there must be, as above, an equivalent martingale measure, making the discounted values of all traded securities martingales. Denoting the risk-neutral probabilities that X1 attains the value xi by qi > 0, i = 1, . . . , m, we must have q1 + · · · + qm = 1

(8.1)

q1 x1 + · · · + qm xm = r 2 = z2 (1 + r) q1 x12 + · · · + qm xm

.. . m−1 q1 x1m−1 + · · · + qm xm = zm−1 (1 + r),

where the ﬁrst equation is ensuring that we have a probability measure, the second equation makes the risk-neutral return on the stock equal to r and the other equations ﬁx the prices of the MOM (k) derivatives at zk , k = 2, . . . , m − 1. With obvious notation (as above), the system has exactly one solution, namely: 1 r (8.2) q = $−1 · z2 (1 + r) . .. . zm−1 (1 + r) If this solution satisﬁes q ∈ (0, 1)m , i.e. the qi s can be seen as probabilities, we have noarbitrage. Moreover, since the solution is unique, we have also that the market is complete. Since X1 can now take m possible values, the payoff of a contingent claim G(X1 ) can now be written into a Taylor expansion up to degree m − 1: G(X1 ) = a0 + a1 X1 + a2 X12 + · · · + am−1 X1m−1 . Completely analogous as in the trinomial setting, the hedging of this contingent claim can be carried out by performing the following:

Moment Derivatives and L´evy-type Market Completion

173

• Invest (a0 − a1 )(1 + r)−1 into bond. • Buy a1 /S0 units of stock, for a total price a1 . • For each k = 2, 3, . . . , m − 1, buy ak MOM (k) derivatives for a price ak zk . At time t = 1, we have the following: • The money invested in bond has grown to a0 − a1 . • We sell the a1 /S0 stocks, giving us a1 (1 + X1 ) of money. • For each k = 2, 3, . . . , m − 1, each MOM (k) derivatives pays out X1k . This leads to a k total payout of m−1 k=2 ak X1 . k In total, we thus end up with a0 + a1 X1 + m−1 k=2 ak X1 = G(X1 ) of money, exactly the payout of the contingent claim considered. In order to set up this strategy, we needed a0 − a1 ak zk + a1 + 1+r m−1 k=2

of money, which in order to avoid arbitrage must be the initial price of the contingent claim with payoff function G(X1 ). 8.2.3 Multi-step ﬁnite markets In this model, we consider a generalization of the above model, taking into account n timesteps. We assume that we have a risk-free bond paying out an interest rate r, i.e. the bond has a deterministic value process: B0 = 1 and Bi = (1 + r)i , i = 1, . . . , n. We have also a risky asset, a stock, which has the following price process S0 > 0,

Si = Si−1 (1 + Xi ) = S0 (1 + X1 ) · · · (1 + Xi ),

i = 1, . . . , n.

We assume the Xi s are deﬁned on a stochastic basis {, F, P , F}, where F = {Fi }ni=1 is a ﬁltration that describes how the information about the security prices is revealed to the investors. We will suppose that F0 = {∅, }, Fi = σ (S1 , . . . , Si ), i = 1, . . . , n and F = Fn . In addition, we will assume that the Xi s are iid and can attain m possible values −1 < x1 < · · · < xm ,with x1 < r < xm to avoid arbitrage. This arbitrage-free market is again an incomplete market. We will show that by introducing into the market, at each time-step, moment derivatives which mature one time-step later and payoff some power of the return the stock makes over that time-step, the model can be completed. Assume at time t = i − 1, i = 1, . . . , n trade is allowed into, at this time newly introduced, moment derivatives (MOM(k) i ) which mature at time T = i and have a payoff function k MOM(k) i = Xi ,

k = 2, . . . , m − 1.

So, besides investing in bonds and stocks, one can invest at time zero also into the MOM (k) 1 , k = 2, . . . , m − 1 derivatives. These derivatives mature at time T = 1. At this time, a set of m − 2 new derivatives are introduced into the market; these derivatives MOM (k) 2 , k = 2, . . . , m − 1, mature at time T = 2, etc.

174

Exotic Option Pricing and Advanced L´evy Models

The Xi s are iid with respect to P ; in consequence, any possible path of the stock’s prices has non-null P -probability. Then, if the system shown by equation (8.2) satisﬁes q ∈ (0, 1)m , we can ﬁnd a risk-neutral probability Q such that the prices of the kth moment derivatives at their initiation are equal to say zk , independently on the step time. This means that for each time i = 1, . . . , n we take Q(Xi = xj |Fi−1 ) = qj . Note that under Q the Xi s are also iid. Thus, for each i = 1, . . . , n, we have price of MOM (k) i at time i − 1 = zk ,

k = 2, . . . , m − 1.

By the model described in Section 8.2.2, any payoff function G = G(X1, X2, . . . , Xn ) at time t = n can be hedged by a portfolio built at t = n − 1, having ﬁxed the value of (X1 , . . . , Xn−1 ) = (x1 , . . . , xn−1 ). In fact, we can write G(x1 , . . . , xn−1 , Xn ) =

m−1

bn(k) (x1 , . . . , xn−1 )Xnk ,

k=0

and the value of this portfolio at time t = n − 1 will be Vn−1 (x1 , . . . , xn−1 ) =

bn(0) (x1 , . . . , xn−1 ) − bn(1) (x1 , . . . , xn−1 ) 1+r + bn(1) (x1 , . . . , xn−1 ) +

m−1

bn(k) (x1 , . . . , xn−1 )zk .

k=2

Then, we can replicate G(x1 , . . . , xn−2 , Xn−1 , Xn ) by a portfolio built at t = n − 2, by duplicating Vn−1 (x1 , . . . , xn−2 , Xn−1 ). Finally, by backward induction we have that any contingent claim can be hedged by a self-ﬁnancing portfolio. 8.2.4 Multi-step markets with general returns With the same notation as in the previous case, let us assume that the Laplace transform of (X1 , X2 , . . . , Xn ) is deﬁned in an open neighborhood of the origin (under Q); then the polynomials are dense in L2 (Fn , Q). So, for any contingent claim, G = G(X1 , X2 , . . . , Xn ) ∈ L2 (Q), if we are in the trading time n − 1 with (X1 , X2 , . . . , Xn−1 ) = (x1 , x2 , . . . , xn−1 ), we can write L2

G = lim

l→∞

l

bn(k,l) (x1 , x2 , . . . , xn−1 )Xnk .

k=0

and by backward induction we can replicate G by a self-ﬁnancing portfolio (see Corcuera et al. (2005) [10] for more details). 8.2.5 Power-return assets Another way of completing the market, is by allowing trade in the so-called power-return assets. To simplify the exposition, we shall work under the ﬁnite market setting. We thus have a risk-free bond paying out an interest rate r, i.e. the bond has a deterministic value

Moment Derivatives and L´evy-type Market Completion

175

process: B0 = 1 and Bi = (1 + r)i , i = 1, . . . , n. We have also a risky asset, a stock, which is the following price process S0 > 0,

Si = Si−1 (1 + Xi ) = S0 (1 + X1 ) · · · (1 + Xi ),

i = 1, . . . , n,

and where the Xi s are iid (with respect to P ) and can attain m different values −1 < x1 < · · · < xm , and x1 < r < xm . Assume now, that in this market m − 2 new assets are introduced with price process Hi(k)

= (1 + r)i

i

Xjk − µk i ,

k = 2, . . . , m − 1,

j =1

where µk ∈ R. Let us make a few remarks on these assets. The asset with price process Hi(k) will be refereed to as the kth order power-return asset. Remark 1 (Arbitrage) To avoid arbitrage by the introduction of these power-return assets, some conditions are necessary on the constants µk . Classical theory says that to have an arbitrage-free market, there must exist an equivalent martingale measure, under which all of the discounted prices process of the assets are martingales. This condition translates into the existence of probabilities 0 < qi < 1, such that q1 + · · · + qm = 1

(Condition H)

q1 x1 + · · · + qm xm = r 2 = µ2 q1 x12 + · · · + qm xm

.. . m−1 q1 x1m−1 + · · · + qm xm = µm−1 .

Remark 2 The ﬁrst condition forces the qi s to sum up to 1, as probabilities should do. The second condition forces the discounted stock price to be a martingale; the other ones force the discounted power-return asset prices to be martingales. These conditions are almost identical to the conditions in equation (8.1); just replace zk (1 + r) by µk . In fact, if these condition are satisﬁed, it is straightforward to see that the {q1 , . . . , qm } are unique and hence the market is complete. Remark 3 (Relation with MOM (k) derivatives) The two ways of completing the market are very related. To move from the one to the other, one should set zk (1 + r) = µk (as already noted in the previous remark). To exploit the relationship a bit more, we will brieﬂy show how to set up a MOM(k) derivative by an investment strategy in power-return assets. Suppose, that we are at time t = i − 1 and we want to generate at time i a payoff Xik , exactly like the MOM(k) i derivative is doing. In order to achieve this, at time i − 1 one should • invest −(1 + r)−1

i−1 k j =1 Xj

(k) − µk i = −(1 + r)−i Hi−1 + (1 + r)−1 µk in bond;

(k) • buy (1 + r)−i power-return assets of order k, for the total price of (1 + r)−i Hi−1 .

176

Exotic Option Pricing and Advanced L´evy Models

In order to set up this portfolio, an amount (at time i − 1) of (1 + r)−1 µk = zk is needed, exactly the same amount as the time i − 1 price of the MOM(k) i derivative. Next, we will show, under the condition (H), that if trade is allowed in the power-return assets, the market is complete, in the sense that any contingent claim can be perfectly hedged by positions in bond, stock and the power-return assets. Let us consider a general contingent claim which can depend on the complete path followed by the underlying stock, i.e. the claim is characterized by a payoff function: G(X1 , X2 , . . . , Xn ). Write the discounted payoff function in the following form: (1 + r)−n G(X1 , X2 , . . . , Xn ) = M0 +

n

aj (X1 , . . . , Xj −1 )(Xj − r)

(8.3)

j =1

+

n m−1

aj(k) (X1 , . . . , Xj −1 )(Xjk − µk ).

j =1 k=2

Note that the functions aj and aj(k) , k = 2, . . . , m − 1, only depend on X1 , X2 , . . . , Xj −1 and are thus completely known at time t = j − 1; in other words, the aj(k) s are Fj −1 measurable or ‘predictable’. Then, let us consider the martingale Mi = EQ [(1 + r)−n G(X1 , X2 , . . . , Xn )|Fi ],

i = 1, . . . , n,

where Q is the risk-neutral probability deﬁned by Condition (H). Since for j = 1, . . . , n, EQ [Xj |Fj −1 ] = r and EQ [Xjk |Fj −1 ] = µk , k = 2, . . . , m − 1, we have that Mi = M0 +

i

aj (X1 , . . . , Xj −1 )(Xj − r)

j =1

+

i m−1

aj(k) (X1 , . . . , Xj −1 )(Xjk − µk ).

j =1 k=2

We know that the discounted value of any contingent claim is a Q-martingale. Then, EQ [(1 + r)−n G(X1 , X2 , . . . , Xn )] = M0 is the initial price of the claim under consideration and (1 + r)i Mi is the time t = i price of this claim. In order to hedge the claim, one should follow the following self-ﬁnancing strategy. Just before the realization of Si , i = 1, . . . , n take the following positions in, respectively, bonds, stocks, and kth order power-return assets, k = 2, . . . , m − 1: • number of bonds = αi = Mi−1 − (1 + r)−i+1 βi Si−1 − (1 + r)−i+1

m−1

(k) βi(k) Hi−1 ,

k=2

• number of stocks = βi = (1 + r) ai (X1 , . . . , Xi−1 )/Si−1 , i

• number of kth power-jump assets = βi(k) = ai(k) (X1 , . . . , Xi−1 ), k = 2, . . . , m − 1.

Moment Derivatives and L´evy-type Market Completion

177

Note the following • The initial (t = 0) amount needed to set up the initial portfolio is: α 1 B 0 + β 1 S0 +

m−1

β1(k) H0(k) = M0 .

k=2

• Just before the realization of Si , the portfolio value is (1 + r)i−1 Mi−1 . By a straightforward calculation one can see that just after the realization of Si (and before adjusting the portfolio again), the value is given by (1 + r)i Mi . This implies that the portfolio is self-ﬁnancing. Moreover, since the value of the portfolio at time t = n equals (1 + r)n Mn = G(X1 , X2 , . . . , Xn ), the portfolio is replicating the claim. In conclusion, we have that the portfolio (αi , βi , βi(2) , . . . , βi(m−1) ; i = 1, . . . , n) is the selfﬁnancing portfolio which replicates the claim G(X1 , X2 , . . . , Xn ) and has initial value M0 .

8.3

´ THE LEVY MARKET

8.3.1 L´evy processes L´evy processes are the natural continuous time analogs of the sums of iid random variables. Basically, they are processes with the same kind of structure in the increments: stationary and independent. However, not for any general distribution, one can deﬁne such a continuous time stochastic process, where the increments follow the given distribution. We have to restrict ourselves to so-called inﬁnitely divisible distributions (see e.g. Bertoin (1996) [3] or Sato (2000) [27]) Given an inﬁnitely divisible distribution with characteristic function φ(z), one can deﬁne a stochastic process (with c`adl`ag paths), Z = {Zt , t ≥ 0}, called a L´evy process, which starts at zero, has independent and stationary increments and such that the distribution of an increment over [s, s + t], s, t ≥ 0, i.e. Zt+s − Zs , has (φ(z))t as the characteristic function. It is well known that L´evy processes are semimartingales. The function ψ(z) = log φ(z) = log E[exp(izZ1 )] is called the characteristic exponent and satisﬁes the following L´evy–Khintchine formula (see Bertoin (1996) [3]): +∞ c2 2 (exp(izx) − 1 − izx1{|x| 0, and λ > 0 exp(λ|x|)ν(dx) < ∞. (8.5) (−ε,ε)c

This implies that

+∞ −∞

|x|i ν(dx) < ∞,

i ≥ 2,

and that the characteristic function E[exp(iuXt )] is analytic in a neighbourhood of 0 and E[exp(−hZ1 )] < ∞ for all h ∈ (−h1 , h2 ), where 0 < h1 , h2 ≤ ∞. So, all moments of Zt (and Xt ) exist. 8.3.3 Power-jump processes Under our continuous-time setting, the role of the powered returns will be taken by powerjump processes. These are built from the following transformations of Z = {Zt , t ≥ 0}. We set (Zs )i , i ≥ 2, Zt(i) = 0<s≤t

Moment Derivatives and L´evy-type Market Completion

179

where Zs = Zs − Zs− , and for convenience we put Zt(1) = Zt . Note that Zt = 0<s≤t Zs is not necessarily true; it is only true in the bounded variation case (with c necessarily equal to zero). If we deﬁne Xt(i) in an analogous way, we have that Xt(i) = Zt(i) , i ≥ 2. The processes X(i) = {Xt(i) , t ≥ 0}, i ≥ 2, are again L´evy processes and are called the ith-powerjump processes (or the power-jump processes of order i). They jump at the same points as the original L´evy process, but the jumps sizes are the ith power of the jump size of the original L´evy process. We have E[Xt ] = E[Xt(1) ] = tm1 < ∞ and (see Protter (1990), p. 29 [25]) ∞ * + (Xs )i = t E Xt(i) = E x i ν(dx) = mi t < ∞, i ≥ 2. (8.6) 0<s≤t

−∞

We denote by + * Yt(i) = Zt(i) − E Zt(i) = Zt(i) − mi t,

i ≥ 1,

the compensated processes. Using Itˆo’s formula (see Chan (1999) [9] or Protter (1990) [25]) for c`adl`ag semimartingales, one can show that equation (8.4) has an explicit solution

# c2 t (1 + Xs ) exp(−Xs ). St = S0 exp Zt + b − 2 0<s≤t

In order to ensure that St > 0 for all t > 0 almost surely, we need Xt > −1 for all t. We thus need that the L´evy measure ν is supported on a subset of (−1, +∞).

´ 8.4 ENLARGING THE LEVY MARKET MODEL Suppose that we have an equivalent martingale measure Q under which Z remains a L´evy process. Under this measure, the discounted stock price process is a martingale and the ˜ moreprocess Z˜ = {Zt + (b − r)t, t ≥ 0} will be a L´evy process (with L´evy measure ν); over, the process Z˜ is a martingale. Obviously, Z˜ t = Zt and Z˜ t(i) = Zt(i) , i ≥ 2. Let (i) (i) ˜ us consider (based +∞ ion Z) the ith-power-jump processes Y = {Yt , t ≥ 0}. Note that for ˜ and we will require ν˜ to fulﬁl equation (8.5). i ≥ 2, mi = −∞ x ν(dx), We will enlarge the L´evy market with what we will call ith-power-jump assets. More precisely, we will allow trade in assets with price process H (i) = {Ht(i) , t ≥ 0} where Ht(i) = exp(rt)Yt(i) ,

i ≥ 2.

By taking a suitable linear combination of the Y (i) s, one obtains a set of pairwise strongly orthonormal martingales {T (i) , i ≥ 1} (see Protter (1990) [25]). Each T (i) is a linear combination of the Y (j ) , j = 1, 2, . . . , i: T (i) = ci,i Y (i) + ci,i−1 Y (i−1) + · · · + ci,1 Y (1) ,

i ≥ 1.

The constants ci,j can be calculated as described in Nualart and Schoutens (2000) [22]: they correspond to the coefﬁcients of the orthonormalization of the polynomials {x n , n ≥ 0}

180

Exotic Option Pricing and Advanced L´evy Models

with respect to the measure µ(dx) = x 2 ν(dx) + c2 δ0 (dx). The resulting processes T (i) = {Tt(i) , t ≥ 0} are called the orthonormalized ith-power-jump processes. In addition, we will (i) (i) denote their orthonormalized version of Ht(i) by H = {H t , t ≥ 0}, where (i)

H t = exp(rt)Tt(i) ,

i ≥ 2.

Trade in the power-jump assets can be motivated as follows. Consider the 2nd-powerjump asset. This object in some sense measures the volatility of the stock, since it accounts for the squares of the jumps. If one believes that in the future there will be a more volatile environment than the current market’s anticipation, trading the 2nd-power-jump asset can be of interest. In addition, if one would like to cover against periods of high (or low) volatility, they can be useful: Buying 2nd-power-jump assets can cover the possible losses due to such unfavourable periods. The same can be said for the higher order variation assets. Typically, the 3rd-power-jump assets is measuring a kind of asymmetry (cf. skewness) and the 4th-power-jump process is measuring extremal movements (cf. kurtosis). Trade in these assets can be of use if one likes to bet on the realized skewness or realized kurtosis of the stock: one believes that the market is not counting in asymmetry and possible extremal moves correctly. On the other hand, an insurance against a crash can be easily built from the 4th-power-jump (or ith-power jump, i ≥ 4) assets. Note, that clearly the discounted versions of the H (i) are the power-jump processes, and hence martingales: EQ [exp(−rt)Ht(i) |Fs ] = EQ [Yt(i) |Fs ] = Ys(i) ,

0 ≤ s ≤ t.

Hence, the market allowing trade in the bond, the stock and the power-jump assets remains arbitrage-free. 8.4.1 Martingale representation property Our L´evy process Z = {Zt , t ≥ 0} has the Martingale Representation Property (MRP) in terms of the orthonormalized power-jump processes (see also Nualart and Schoutens (2000, 2001) [22] [23]) that is, every square-integrable martingale M = {Mt , t ≥ 0} can be represented as follows: t ∞ t (i) Mt = M0 + hs dZ˜ s + h(i) s dTs , 0

i=2

0

where hs and h(i) s , i ≥ 2 are predictable processes. such that t |hs |2 ds < ∞ E 0

and E

% ∞ t 0 i=2

& 2 |h(i) s |

ds < ∞.

Note the similarity, except for the orthonormalization, between this MRP and equation (8.3).

Moment Derivatives and L´evy-type Market Completion

181

The MRP implies that the market enlarged with the ith-power-jump assets is complete in the sense that for every square-integrable contingent claim X we can set up a sequence of self-ﬁnancing portfolios whose values converge in L2 (Q) to X. These portfolios will consist of ﬁnite number of bonds, stocks and ith-power-jump assets. We will say, for short, that X can be replicated. Note that this notion of completeness is equivalent to the notion of approximately complete of Bj¨ork and co-workers (given in Bj¨ork et al. (1997) [4]). The details of the hedging strategy can be extracted out of the MRP. Consider a squareintegrable contingent claim X ∈ FT with maturity T . Let Mt = EQ [exp(−rT )X|Ft ]. By the MRP given above, if we deﬁne MtN := M0 +

t

hs dZ˜ s +

0

N i=2

t

0

(i) h(i) s dTs .

we have that lim MtN = Mt ,

N→∞

in L2 (Q). Deﬁne the sequence of portfolios (in terms of the orthonormalized ith-powerjump assets) φ N = {φtN = (αtN , βt , βt(2) , βt(3) , . . . , βt(N) ), t ≥ 0}, by N αtN = Mt− − βt St− e−rt − e−rt

N

N ≥2

(i)

βt(i) H t− ,

i=2

βt = e

rt

−1 ht St− ,

βt(i) = h(i) t ,

i = 2, 3, . . . , N.

Here, αtN corresponds to the number of bonds at time t, βt is the number of stocks at that time (i) and βt(i) is the number of assets H , i = 2, 3, . . . , N , one needs to hold at time t. Then, it was shown in Corcuera et al. (2005) [10] that {φ N , N ≥ 2} is the sequence of self-ﬁnancing portfolios which replicates X. In fact, the value VtN of φ N at time t is given by VtN = αtN ert + βt St +

N

(i)

βt(i) H t = ert MtN ,

i=2

and so the sequence of portfolios {φ N , N ≥ 2} is replicating the claim. Moreover, in the case of a contingent claim whose payoff is only a function of the value at maturity of the stock price, i.e. X = f (ST ), one can compute explicitly the sequence of portfolios that replicates the contingent claim. Note that the value of the contingent claim at time t is given by F (t, St ) = exp(−r(T − t))EQ [f (ST )|Ft ]; we call F (t, x) the price function of X.

182

Exotic Option Pricing and Advanced L´evy Models

Denote by D1 the differential operation with respect to the ﬁrst variable, i.e. the time variable, and by D2 the differential operator with respect to the space variable (the second variable – the stock price). Finally, denote by D the following integral operator: DF (t, x) =

+∞

−∞

(F (t, x(1 + y)) − F (t, x) − xyD2 F (t, x)) ν˜ (dy).

If f is Lipschitz and under certain degeneracy conditions (see Chapter 12 in Cont and Tankov (2004) [12]), F ∈ C 1,2 . In this case, we have that, in analogy with the Black–Scholes partial differential equation, in the L´evy market setting F will satisfy a Partial Differential Integral Equation (PDIE). More precisely, the price function (at time t) F (t, x) satisﬁes (see Chan (1999) [9], Nualart and Schoutens (2001) [23] and Raible (2000) [26]): 1 D1 F (t, x) + rxD2 F (t, x) + c2 x 2 D22 F (t, x) + DF (t, x) = rF (t, x). 2

(8.7)

with F (T , ST ) = f (ST ). The sequence of self-ﬁnancing portfolios replicating a contingent claim X, with a payoff only depending on the stock price value at maturity and a price function F (t, x) ∈ C 1,∞ which satisﬁes sup

∞

x 0, t0 > 0, is given at time t by: • number of bonds = αtN = Bt−1 F (t, St− )−St− D2 F (t, St− )−Bt−1

N i D i F (t, S ) St− t− 2

i=2

i!Bt

(i) Ht−

(8.9) • number of stocks = βt = D2 F (t, St− ), • number of ith-power-jump assets = βt(i) =

i D i F (t, S ) St− t− 2 , i!Bt

i = 2, 3, . . . , N .

Remark 4 In the Black–Scholes model, the risk-neutral dynamics of the stock price is given by the stochastic differential equation

1 dSt = r − σ 2 dt + dWt , St 2

S0 > 0,

where W = {Wt , t ≥ 0} is a standard Brownian motion. In this case, all processes H (i) , i ≥ 1 are equal to zero. Hence, it is clear that the market is already complete and that an enlargement is not necessary. Moreover the hedging portfolio is given by F (s,Ss )−SBssD2 F (s,Ss ) number of bonds and D2 F (s, Ss ) number of stocks.

Moment Derivatives and L´evy-type Market Completion

183

8.5 ARBITRAGE We assume our market is already enlarged with the power-jump assets. So, we have chosen constants a (i) , i ≥ 2 and trade is allowed in the bond, the stock and the power-jump assets with price processes Ht(i) = exp(rt)(Xt(i) − a (i) t), i ≥ 2. We investigate whether this enlargement leads to arbitrage or not. For instance, if we choose a (i) and r to equal zero this leads to arbitrage opportunities because all Ht(i) with even i are strictly increasing and starting at zero and trade is allowed in these objects. Actually, the choice of the constants a (i) may prevent arbitrage opportunities. We will discuss below how to make this choice, which is a delicate matter. No arbitrage, in the usual sense and in our portfolios with a ﬁnite number of assets, is implied by the existence of an equivalent martingale measure under which all discounted assets in the market are martingales. This question is related to the moment problem and we will give sufﬁcient conditions to ensure that there exists an equivalent martingale measure (and hence the market is arbitrage free): in continuous time the existence of an equivalent martingale measure is a sufﬁcient but not a necessary condition to ensure no-arbitrage (see Delbaen and Schachermayer (1994) [14]). The problem in its full generality seems to be very hard and challenging. 8.5.1 Equivalent martingale measures In this section, we will describe the measures, equivalent to the canonical (real world) measure under which the discounted stock price process is a martingale and under which Z remains a L´evy process. More precisely, we characterize all structure preserving P equivalent martingale measures Q under which Z remains a L´evy process and the process S˜ = {S˜t = exp(−rt)St , t ≥ 0} is an {Ft }-martingale. Since we are considering a market with ﬁnite horizon T , locally equivalence will be the same as equivalence. We have the following result (see Sato (2000), Theorem 33.1 [27]). Theorem 5 Let Z be a L´evy process with L´evy triplet [α, c2 , ν(dx)] under some probability measure P . Then the following two conditions are equivalent. (a) There is a probability measure Q equivalent to P on Ft for any t ≥ 0, such that Z is a Q-L´evy process with triplet [α, ˜ c˜2 , ν˜ (dx)]. (b) All of the following conditions hold: (i) ν(dx) ˜ = H (x)ν(dx) for some Borel function H : R→ (0, ∞). +∞ (ii) α˜ = α + −∞ x1{|x|≤1} (H (x) − 1)ν(dx) + Gc for some G ∈ R. (iii) c˜ = c. ∞ √ (iv) −∞ (1 − H (x))2 ν(dx) < ∞. The equivalent conditions in the previous theorem imply that the process W˜ = {W˜ t , t ≥ 0} with W˜ t = Wt − Gt is a Brownian motion under Q and also, if ν and ν˜ verify the condition shown in equation (8.5), the process X is a quadratic pure jump L´evy process with Doob–Meyer decomposition

+∞ ˜ Xt = Lt + a + x(H (x) − 1)ν(dx) t, −∞

184

Exotic Option Pricing and Advanced L´evy Models

where L˜ = {L˜ t , t ≥ 0} is a Q-martingale and the L´evy measure is given by ν(dx) ˜ = H (x)ν(dx). We now want to ﬁnd an equivalent martingale measure Q under which the discounted price process S˜ is a martingale. By the above theorem, under such a Q, X has the Doob–Meyer decomposition

+∞ ˜ x(H (x) − 1)ν(dx) t, Xt = Lt + a + −∞

where L˜ = {L˜ t , t ≥ 0} is a Q-martingale. Noting that Lt = L˜ t , we have

c2 S˜t = S0 exp cW˜ t + L˜ t + a + b − r + cG − t 2 +∞

# × exp t x(H (x) − 1)ν(dx) (1 + L˜ s ) exp(−L˜ s ). −∞

0<s≤t

Then, a necessary and sufﬁcient condition for S˜ to be a Q-martingale is the existence of G ∞ √ and H (x), with −∞ (1 − H (x))2 ν(dx) < ∞ such that cG + a + b − r +

+∞ −∞

x(H (x) − 1)ν(dx) = 0.

(8.10)

Remark 6 We remark (see e.g. Eberlein and Jacod (1997) [16]), that if there exists a (nonstructure preserving) locally equivalent martingale measure Q1 under which Z is not a L´evy process, there exists always a (structure preserving) locally equivalent martingale measure Q2 under which Z is a L´evy process. A sufﬁcient condition to guarantee that the enlarged market is free of arbitrage is the existence of an equivalent martingale measure Q making S˜ and the discounted H (i) s martingales. If this measure is structure preserving, the condition that the discounted stock price must be a martingale comes down to the existence of G and H (x) such that equation (8.10) holds. If we also want that the discounted H (i) s, i.e. Xt(i) − a (i) t, be martingales for i ≥ 2, using equation (8.6) together with the fact that the L´evy measure of X under Q is given by H (x)ν(dx), this comes down to +∞ x i H (x)ν(dx) = a (i) , i ≥ 2. (8.11) −∞

The question now is, do there exist G and H (x) such that equations (8.10) and (8.11) hold simultaneously? This question is related to the moment problem: given a series of numbers {µn }, ﬁnd necessary and sufﬁcient conditions for the existence of such a measure with µn as the nth moment. Another point is the uniqueness. A partial result is that if the moment problem has a solution with bounded support, then it will be unique (see Shohat and Tamarkin (1950) [30] or Ahiezer (1965) [1]). We have then the following proposition. Proposition 7 Suppose that ν(dx) has compact support: then, if there is a martingale measure in the market enlarged with the power-jump assets, the martingale measure is unique, structure preserving and the market is complete.

Moment Derivatives and L´evy-type Market Completion

185

Proof. If we have a martingale measure in the enlarged market, there exists, using the same arguments as in Eberlein and Jacod (1997) [16], an H (x) verifying equations (8.10) and (8.11) with H (x) > 0. The measure µ(dx) = x 2 H (x)ν(dx) is ﬁnite and has a bounded support. This implies that H (x) is determined by the condition shown in equation (8.11). On the other hand, since the support is bounded, H (x)ν(dx) veriﬁes equation (8.5) and the model enlarged with the power-jump assets is complete. Finally, since the contingent claim BT 1A with A ∈ F is replicable, the uniqueness of its initial arbitrage price, EQ (1A ), implies the uniqueness of the martingale measure. In general, uniqueness of the martingale measure implies completeness. Proposition 8 If the probability measure that makes the discounted stock price and the power-jump assets martingales is unique, that is, the martingale measure is unique, then the market is complete. Proof. Let Q be a martingale measure. We argue by contradiction. If the market is not complete, there exists a contingent claim X ≥ 0, X ∈ L2 (Q), not identically zero, which is orthogonal to any replicable contingent-claim. Deﬁne Q∗ (dω) = (1 + X)Q(dω). Then Q∗ is a martingale measure different from Q. In fact, for any s ≤ t, and A ∈ Fs , we have EQ∗ (1A (Yt(i) − Ys(i) )) = EQ (1A (Yt(i) − Ys(i) )) + EQ (X1A (Yt(i) − Ys(i) )) = 0, and {Yt(i) , t ≥ 0} are Q∗ -martingales for all i ≥ 2. Clearly, S˜t is also a Q∗ -martingale. 8.5.2 Example: a Brownian motion plus a ﬁnite number of Poisson processes Suppose Zt = cWt +

n

cj Nj,t ,

j =1

where c = 0, W = {Wt , t ≥ 0} a standard Brownian Motion and Nj = {Nj,t , t ≥ 0} are j = 1, . . . , n are independent Poisson processes with intensity aj > 0. The constants cj , assumed to be all different from each other and non-zero. Then, Xt = nj=1 cj Nj,t and E[X1 ] = nj=1 cj aj = a, and n Ht(i) = exp(rt) cji Nj,t − a (i) t , i = 2, 3, . . . . j =1

It is not that hard to see that Ht(i) , for i > n + 1 can be written as a linear combination of the Ht(i) , i = 2, . . . , n + 1 (see L´eon et al. (2002) [19]). In this case, we enlarge the market with only n objects, namely the assets following the price processes Ht(i) , i = 2, . . . , n + 1. In order that an equivalent martingale measure Q exists, we must have the existence of a G and H , such that +∞ x(H (x) − 1)ν(dx) = r − cG − a − b −∞

+∞ −∞

x i H (x)ν(dx) = a (i) ,

i = 2, . . . , n + 1.

186

Exotic Option Pricing and Advanced L´evy Models

The support of H will now be the set {c2 , . . . , cn+1 } and the above equations reduce to n

cj H (cj )aj = r − cG − b

j =1 n

cji H (cj )aj = a (i) ,

i = 2, . . . , n + 1.

j =1

There exists an equivalent martingale measure if the following system of equations for H (cj ), j = 1, . . . , n has a positive solution, i.e. H (cj ) > 0, j = 1, . . . , n.

c12 a1 c3 a1 1 ... c1n+1 a1

c22 a2 c23 a2 ... n+1 c2 a2

. . . cn2 an H (c1 ) . . . cn3 an × H (c2 ) ... ... ... n+1 H (cn ) . . . cn an

a (2) a (3) = ... . a (n+1)

The existence (and uniqueness by Proposition 7) of a positive solution H (cj ), j = 1, . . . , n can be translated into the condition C −1 · a > 0, where C −1 is the inverse of the Vandermonde matrix 1 1 ... 1 c1 c . . . c 2 n C= ... ... ... ... c1n−1 c2n−1 . . . cnn−1

(8.12)

and a is the transpose of [a (2) . . . a (n+1) ]. Note that if all of the ci s are different from each other (as we assumed above), that detC = 0. For the calculation of the inverse of Vandermonde matrices, see Graybill (1983) [17] or Macon and Spitzbart (1958) [20], while for other applications of Vandermonde matrices in ﬁnance see Norberg (1999) [21].

8.6 OPTIMAL PORTFOLIOS Deﬁnition 9 A utility function is a mapping U (x) : R → R ∪ {−∞} which is strictly increasing, continuous on {U > −∞}, of class C ∞ , strictly concave on the interior of {U > −∞} and satisﬁes U (∞) := lim U (x) = 0. x→∞

Denoting by dom(U ) the interior of {U > −∞}, we shall consider only two cases: Case 10 dom(U ) = (0, ∞) in which case U satisﬁes U (0) := lim U (x) = ∞. x→0+

Moment Derivatives and L´evy-type Market Completion

187

Case 11 dom(U ) = R in which case U satisﬁes U (−∞) := lim

x→−∞

U (x) = ∞. 1−p

Typical examples for Case 10 are the so-called HARA utilities, U (x) = x1−p for p ∈ R+ \{0, 1}, and the logarithmic utility U (x) = log(x). A typical example for Case 11 is U (x) = − α1 e−αx . 8.6.1 Optimal wealth Given an initial wealth w0 and an utility function U , we want to ﬁnd the optimal terminal wealth WT , that is, the value of WT that maximizes EP (U (WT )). We will consider the optimization problem

5 WT = w0 . max EP (U (WT )) : EQ BT The corresponding Lagrangian is EP (U (WT )) − λEQ

WT − w0 BT

dQT WT − w0 = EP U (WT ) − λT . dPT BT

Then, the optimal wealth is given by −1 WT = U

λT dQT BT dPT

,

where λT is the solution of the equation EQ

1 −1 U BT

λT dQT BT dPT

= w0 .

(8.13)

It is easy to check the existence and uniqueness of the optimal wealth from the conditions on U . From equation (8.5) and under certain conditions on Q (see Corcuera et al. (2004b) [11]), we can write:

G −1 m (T ) STc eVT , WT = U where

λt − Gc c2 1 2 G m (t) := a+b− t S exp − G t − Bt 0 2 c 2

+∞ G G +t (log H (x) − log (1 + x))H (x) − H (x) + 1 + x ν(dx) c c −∞

188

Exotic Option Pricing and Advanced L´evy Models

and Vt =

+∞ −∞

g(x)(Q((0, t], dx) − tH (x)v(dx)),

with g(x) := log H (x) −

G log (1 + x) . c

It can be shown (see Corcuera et al. (2004b) [11]), that if we consider HARA and exponential utilities we have that the price process of the optimal portfolio is given by Bt EQ WT |Ft = F (t, St , Vt ) BT with

F (t, x1 , x2 ) := φ(t, T ) U

−1

G c x2 m (t) x1 e + χ (t, T ).

(8.14)

We know that under an equivalent martingale measure Q, which is structure preserving, T any random variable WT ∈ L2 (, FT , Q) can be replicated and we have w0 = EQ W BT . Now, by a generalization of equation (8.9) (see Theorem 4 in Corcuera et al. (2004b) [11]) we can ﬁnd the composition of the portfolio with this price process. In fact, we have that the number of stocks and new assets are given, respectively, by G −1

βt =

c Gφ(t, T )m (t) St− eVt− G

c Vt− cU ((U )−1 (m (t) St− e ))

=

Gφ(t, T )U (Wt− ) cSt− U (Wt− )

(8.15)

and βt(i) =

G φ(t, T ) ∂ i −1 c Vt− m S U e H , (t) (y) t− i i!Bt ∂y y=0

i = 2, 3, . . .

8.6.2 Examples

(8.16)

−1 Example 12 Consider U (x) = log x. Then U (x) = U (x) = x1 . Therefore, by solving equation (8.13), we have WT = w0 BT

−1 G dPT = m(T )STc eVT . dQT

Therefore, we have that Bt dPT dPt EQ WT |Ft = w0 Bt EQ |Ft = w0 Bt = Wt BT dQT dQt and the price function of WT at time t is

−1 G c Vt Wt = m (t) St e ,

Moment Derivatives and L´evy-type Market Completion

189

that is, the wealth of the optimal portfolio at time t is the optimal terminal wealth for the period [0, t]; in other words, φ(t, T ) = 1 and χ (t, T ) = 0 in equation (8.14). Now, since U (x) = − x12 , if we apply equation (8.15), we have that βt St− G =− , Wt− c that is, the relative wealth invested in stocks is constant. From equation (8.16), the number of new assets is 1 Wt− ∂ i , i = 2, 3, . . . βt(i) = i!Bt ∂y i H (y) y=0 So, maximization with bonds and stocks corresponds to take 1

H (y) =

1−

G cy

.

where G veriﬁes (see equation (8.10)) cG + a + b − r +

G c

+∞ −∞

y2 ν(dy) = 0. 1 − Gc y

−1 1 1−p Example 13 Consider U (x) = x1−p with p ∈ R+ \ {0, 1}. Then, U (x) = x − p and by solving equation (8.13) we have WT = w0 BT

dPT dQT

EQ

p1

dPT dQT

− p1 G c VT . p1 = m(T )ST e

dPt { dQt , 0 ≤ t ≤ T } is a Q-exponential L´evy process (see Corcuera et al. (2005) [11]), and then 1 dPT p EQ dQ |F t T Bt EQ WT |Ft = w0 Bt

1 BT dPT p EQ dQT

EQ = w0 Bt

dPT ,t dQT ,t

EQ

= w0 Bt

dPt dQt

EQ

p1

dPT ,t dQT ,t

dPt dQt

p1

dPt dQt

p1

dPt dQt

p1

p1 = Wt

|Ft 1

p

190

where

Exotic Option Pricing and Advanced L´evy Models dPT ,t dQT ,t

=

dPT dQT dPt dQt

. That is, again the wealth of the optimal portfolio at time t is the optimal

terminal wealth for the period [0, t], and φ(t, T ) = 1 and χ (t, T ) = 0 in equation (8.14). Now, since U (x) = x −p and U (x) = −px −p−1 , if we apply equation (8.15), we have that βt St− G =− ; Wt− cp and by equation (8.16) the number of new assets is given by Wt− ∂ i (i) − p1 H (y) , i = 2, 3, . . . βt = i!Bt ∂y i y=0 So, we will have an optimal portfolio only based in bonds and stocks, if and only if, H (y) =

1 1−

G cp y

p ,

where G veriﬁes cG + a + b − r +

∞ −∞

y

1 p Gy 1− cp

−1

ν(dy) = 0.

Example 14 Consider the exponential utility function 1 U (x) = − e−αx α −1 with α ∈ (0, ∞). Then, U (x) = − α1 log x and by solving equation (8.13) we have WT = w0 BT +

1 α

G dPT 1 dPT log = − log m(T )STc eVT − EQ log dQT dQT α

Note, that in this case, Wt− is not bounded by below and that there arises the problem of the admissibility of this optimal portfolio (see Kallsen (2000) [18]). In addition, we have that

Bt dPT dPT Bt EQ log WT |Ft = w0 Bt + |Ft − EQ log EQ BT αBT dQT dQT

dPT ,t dPT ,t Bt EQ log = w0 Bt + |Ft − EQ log αBT dQT ,t dQT ,t

dPt dPt + log − EQ log dQt dQt

dPt dPt Bt log = w0 Bt + − EQ log αBT dQt dQt

Bt Bt . Wt + w0 Bt 1 − = BT BT

Moment Derivatives and L´evy-type Market Completion

191

Therefore, in this case φ(t, T ) = BBTt and χ (t, T ) = w0 Bt (1 − BBTt ) in equation (8.14). Now, since U (x) = e−αx and U (x) = −αe−αx , if we apply equation (8.15), we have that BT G βt St− = − , Bt cα that is, the forward value of the wealth invested in stocks is constant. From equation (8.16), the number of new assets is constant: βt(i)

BT ∂ i =− log H (y) , i i!α ∂y y=0

i = 2, 3, . . .

and we obtain the optimal portfolio based only in stocks and bonds by taking H (y) = exp

G y , c

with G verifying

G y − 1 ν(dy) = 0. cG + a + b − r + y exp c −∞

∞

The corresponding martingale measure is then the Esscher measure (see Chan (1999) [9]). Example 15 Consider the quadratic utility U (x) = γ x −

x2 , 2

x γ , it is interesting since the solution of the optimal problem with this utility is the same as that of the solution of the mean-variance portfolio problem if we choose γ =

w0 ((1 + ρ)EQ (ξT ) − (1 + r)) EQ (ξT ) − 1

where ρ > r is a speciﬁed return (see Pliska (1997) [24]).

REFERENCES [1] Ahiezer, N.I. (1965), The Classical Moment Problem, Hafner Publishing Company, New York, NY, USA. [2] Balland, P. (2002), “Deterministic implied volatility models”, Quantitative Finance, 2, 31–44. [3] Bertoin, J. (1996), L´evy Processes, Cambridge Tracts in Mathematics, 121, Cambridge University Press, Cambridge, UK. [4] Bj¨ork, T., Di Masi, G., Kabanov, Y. and Runggaldier, W. (1997), “Towards a general theory of bond markets”, Finance and Stochastics 1, 141–174. [5] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities”, Journal of Political Economy, 81, 637–654. [6] Carr, P. and Lewis, K. (2004), “Corridor variance swaps”, Risk, 17(2), 67–72. [7] Carr, P. and Madan, D. (1998), “Towards a theory of volatility trading”, in R. Jarrow (Ed.), Volatility, Risk Publications, London, UK, pp 417–427. Reprinted (2001) in M. Musiela, E. Jouini and J. Cvitanic (Eds), Option Pricing, Interest Rates and Risk Management, Cambridge University Press, Cambridge, UK, pp. 458–476. [8] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2002), Pricing Options on Realized Variance, Pr´epublications du Laboratoire de Probabilit´es et Mod`eles Al´eatoires, 768, Universit´es de Paris 6 and Paris 7, Paris, France. [9] Chan, T. (1999), “Pricing contingent claims on stocks driven by L´evy processes”, Annals of Applied Probability, 9, 504–528 (1999). [10] Corcuera, J.M., Nualart, D. and Schoutens, W. (2005), “Completion of a L´evy market by powerjump assets”, Finance and Stochastics, 9, 109–127.

Moment Derivatives and L´evy-type Market Completion

193

[11] Corcuera, J.M., Guerra J., Nualart, D. and Schoutens, W. (2004), “Optimal investment in a L´evy market”, Preprint Institut de Matem´atica de la Universitat de Barcelona 249, Universitat de Barcelona, Barcelona, Spain. [12] Cont, R. and Tankov, P. (2004), Financial Modelling with Jump Processes, Chapman & Hall/CRC Press, Boca Raton, FL, USA. [13] Davis, M. (1997), “Option pricing in incomplete markets”, in M.A.H. Dempster and S.R. Pliska (Eds), Mathematics of Derivative Securities, Cambridge University Press, Cambridge, UK, pp. 227–254. [14] Delbaen, F. and Schachermayer, W. (1994), “A general version of the fundamental theorem of asset pricing”, Mathematische Annalen, 300, 463–520. [15] Demeterﬁ, K., Derman, E., Kamal, M. and Zhou, J. (1999), “A guide to volatility and variance swaps”, The Journal of Derivatives, 6(4), 9–32. [16] Eberlein, E. and Jacod, J. (1997), “On the range of option prices”, Finance and Stochastics, 1, 131–140. [17] Graybill, A. (1993), Matrices with Applications to Statistics, 2nd Edn, Wadsworth, Belmont, CA, USA. [18] Kallsen, J. (2000), “Optimal portfolios for exponential L´evy processes”, Mathematical Methods of Operations Research, 51, 357–374. [19] L´eon, J.A., Vives, J., Utzet, F. and Sol´e, J.L. (2002), “On L´evy processes, Malliavin calculus and market models with jumps”, Finance and Stochastics, 6, 197–225. [20] Macon, N. and Spitzbart, A. (1958), “Inverses of Vandermonde matrices”, American Mathematics Monthly, 65, 95–100. [21] Norberg, R. (1999), On the Vandermonde matrix and its role in mathematical ﬁnance Working paper, Laboratory of Actuarial Mathematics (working paper), University of Copenhagen, Denmark. [22] Nualart, D. and Schoutens W. (2000), “Chaotic and predictable representations for L´evy processes”, Stochastic Processes and their Applications, 90, 109–122. [23] Nualart, D. and Schoutens W. (2001), “Backwards stochastic differential equations and Feynman–Kac formula for L´evy processes, with applications in ﬁnance”, Bernoulli, 7, 761–776. [24] Pliska, S. (1997), Introduction to Mathematical Finance, Blackwell, Oxford, UK. [25] Protter, Ph. (1990), Stochastic Integration and Differential Equations, Springer-Verlag, Berlin, Germany. [26] Raible, S. (2000), “L´evy process in ﬁnance: theory, numerics, and empirical facts”, PhD Thesis, University of Freiburg, Freiburg, Germany. [27] Sato, K. (2000), L´evy Processes and Inﬁnitely Divisible Distributions, Cambridge Studies in Advanced Mathematics, 68, Cambridge University Press, Cambridge, UK. [28] Schachermayer, W. (2001), “Optimal investment in incomplete ﬁnancial markets”, in H. Geman, D. Madan, S. Pliska and T. Vorst (Eds), Mathematical Finance – Bachelier Congress 2000, SpringerVerlag, Berlin, Germany, pp. 427–462. [29] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [30] Shohat, J.A and Tamarkin, J.D. (1950), The Problem of Moments, Mathematical Surveys No. 1, American Mathematical Society, New York, NY, USA.

9 Pricing Perpetual American Options Driven by Spectrally One-sided L´evy Processes† Terence Chan Heriot-Watt University, Edinburgh, UK Abstract This paper considers the problem of pricing perpetual American put options on stocks ´ whose price process is the exponential of a Levy process (i.e. a process with stationary independent increments). When the price process has no negative jumps, the problem reduces to one of ﬁnding the law of a relevant ﬁrst-passage time of the process, a problem which has already been well-studied. However, if the price process does have negative jumps, the problem is much more delicate as it involves ﬁnding the joint law of a ﬁrstpassage time and position of the process at that time. This problem is only mathematically tractable under the assumption that the price process has no positive jumps. A renewal ´ process equation for the price is obtained in the case where the jump component of the Levy has ﬁnite variation. In the general case, a simple explicit formula is obtained for the optimal exercise boundary and a formula amenable to efﬁcient numerical computation is obtained for the price of a perpetual put.

9.1

INTRODUCTION

Consider a (non dividend-paying) stock whose price at time t, St , is modelled as St = S0 exp{−Yt }, where Yt is a L´evy process (process with independent stationary increments) of the form Yt = σ Bt + Xt + ct,

Y0 = 0,

(9.1.1)

where Bt is a standard Brownian motion and Xt is a jump process with stationary independent increments. We shall suppose that the market is already risk-neutral, so that for some discount factor δ > 0, e−δt St is a martingale. Of course, such a model is incomplete: there are many equivalent martingale measures and contingent claims cannot be hedged perfectly. However, the purpose of this paper is not to address the problems associated with incompleteness of the market in this model; in particular, it does not deal with the question of how to choose a suitable equivalent martingale measure from the inﬁnitely many available – this problem has been studied in, for example, Chan (1999) and the references cited there. Instead, the present article is concerned with the next step in pricing a contingent claim, namely, once an equivalent martingale measure has been chosen, how to calculate the expected payoff with respect to the chosen martingale measure. It is shown in Chan †

This paper was submitted at the special invitation of the editors. Please see Epilogue on page 215 for details.

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

196

Exotic Option Pricing and Advanced L´evy Models

(1999) that for this model, a L´evy process under the original measure remains a L´evy process under any equivalent martingale measure. We may therefore assume that e−δt St is already a martingale. (However, in the context of perpetual options considered here, it is perhaps worth bearing in mind that while an equivalent martingale measure is equivalent to the actual underlying probability measure over every ﬁnite time interval [0, T ], these two measures are mutually singular over [0, ∞).) In this paper, we consider the problem of pricing perpetual American options. These were ﬁrst studied by Samuelson (1965), in relation to call options. In recent years, there has been renewed interest in this problem: some recent work includes Gerber and Shiu (1994, 1998) and Gerber and Landry (1998). While the implication of Samuelson (1965) seems to be that a few perpetual warrants (i.e. call options) once did exist, the main interest in perpetual put options in recent years arises from their relative mathematical tractability compared with ﬁnitely dated puts and the usefulness of the former in pricing the latter. The problem of pricing an American put option with a ﬁnite maturity presents many difﬁcult mathematical problems because the optimal exercise level depends on the time to maturity and there are no reasonable models for which an explicit pricing formula is known. The article by Gerber and Shiu (1994) contains a long list of references to the literature on American options with ﬁnite maturity. The problem of pricing perpetual American puts is a much more mathematically tractable problem because the optimal exercise level is constant in time. The price of a perpetual American option with payoff function (s) (under the equivalent martingale measure) is simply supτ E[e−δτ (Sτ )] (in the case of payoffs depending only on the stock price at the time of exercise), where the supremum is taken over all stopping times τ . The motivation for studying perpetual American options is three-fold. First, at the most obvious level, perpetual options can be thought of as approximations to options with a long time until expiry. Secondly, the mathematics associated with perpetual options have some applications to the ruin problem for certain generalizations of the classical model in ruin theory. Some recent work that have explored the connection between perpetual American options and ruin theory are Gerber and Landry (1998) and Gerber and Shiu (1997). The third and arguably most compelling motivation for studying perpetual American (put) options lies in their applications to numerical methods for pricing American puts with ﬁnite maturities (not necessarily long-dated). Various numerical approximation schemes developed in recent years involve evaluating perpetual options. MacMillan (1986) and Zhang (1995) consider the difference between the prices of a ﬁnite-maturity American option and a European option and show that under a suitable discretization, this difference can be interpreted as the price of a perpetual American option. Even more strikingly, Carr (1998) presents a recursive algorithm which involves calculating the prices of a sequence of American puts (with different payoffs) expiring at a sequence of independent exponential times. The key observation behind this idea of Carr is that the memoryless property of the exponential distribution reduces the problem with a random exponential time horizon to one with an inﬁnite time horizon and an adjusted discount factor and payoff. To see this, let T (λ) be an exponential time with rate λ, independent of Y ; then the price of an American (λ) option expiring at T (λ) is supτ E[e−δ(τ ∧T ) (Sτ ∧T (λ) )] (where x ∧ y = min(x, y)). Writing (λ) (St ) = (Yt ) for notational convenience, deﬁne U (y) = Ey [e−δT (YT (λ) )] (where Ey denotes expectation given Y0 = y). The strong Markov property at stopping time τ says that Y˜s = Ys+τ − Yτ is independent of {Yu : u ≤ τ } and τ . Hence we may calculate as follows: Ey [e−δ(τ ∧T

(λ) )

(Yτ ∧T (λ) )]

Pricing Perpetual American Options

197

(λ)

= Ey [e−δτ (Yτ )1{τ

1

x ν(dx) = lim xν(x, 1] + x→0

0

ν(x, 1] dx,

(9.2.19)

0

1 so that 0 ν(x, 1] dx < ∞, which in turn implies that limx→0 xν(x, 1] = 0. Integrating by parts and noting that 1 − e−θx ∼ θ x for small x gives

∞

ψ(θ ) = −

(1 − e−θx ) ν(dx) − cθ

0

= −θ

1

e 0

−θx

∞

ν(x, 1] dx −

(1 − e−θ x ) ν(dx) − cθ.

(9.2.20)

1

1 ∞ Since 0 e−θ x ν(x, 1] dx → 0 as θ → ∞ and 1 (1 − e−θ x ) ν(dx) remains bounded as θ → ∞, the result shown in equation (9.2.16) follows.

202

Exotic Option Pricing and Advanced L´evy Models

On the other hand, if Y has inﬁnite variation, then the integral in equation (9.2.19) 1 diverges, and so either 0 ν(x, 1] dx = ∞, or limx→0 xν(x, 1] = ∞, which also implies 1 that 0 ν(x, 1] dx = ∞. Integrating by parts as before gives σ 2θ 2 ψ(θ ) = − bθ − 2 =

σ 2θ 2 −θ 2

1

1

(1 − e

−θx

− θ x) ν(dx) −

0

∞

(1 − e−θ x ) ν(dx) − cθ

1

(e−θx − 1)ν(x, 1] dx −

0

∞

(1 − e−θ x ) ν(dx) + O(θ ).

(9.2.21)

1

1 The same argument as before, only this time 0 (e−θ x − 1)ν(x, 1] dx → ∞ as θ → ∞, shows that equation (9.2.17) holds. (ii) The same sort of argument as used for equation (9.2.19), but this time integrating by parts twice, shows that

1

∞>

1 1

x 2 ν(dx) = 2

0

ν(y, 1] dy dx. x

0

Hence, integrating by parts once more at equation (9.2.21) gives ψ(θ ) =

σ 2θ 2 −θ 2

σ 2θ 2 = + θ2 2 Since

9.3

1 0

e−θ x

1 x

1

(e−θx − 1)ν(x, 1] dx −

0

∞

(1 − e−θ x ) ν(dx) + O(θ )

1

1

e 0

−θx

1

∞

ν(y, 1] dy dx −

x

(1 − e−θ x ) ν(dx) + O(θ ). (9.2.22)

1

ν(y, 1] dy dx → 0 as θ → ∞, the result (equation (9.2.18)) follows.

DESCRIPTION OF THE MODEL, BASIC DEFINITIONS AND NOTATIONS

The price process of a stock is given by St = S0 exp{−Yt } = exp{y − Yt }, where Yt = σ Bt + Xt + ct,

Y0 = 0,

(9.3.1)

is a L´evy process satisfying the basic Assumption 9.2.1. In addition, we assume that the market is risk-neutral – in other words, e−δt St = ey−δt−Yt is a martingale. Since e−Yt −ψ(1)t is a martingale, this requires ψ(1) = δ,

$⇒ φ(δ) = 1.

(9.3.2)

In order to achieve this, we require the drift c to be c = −δ +

σ2 + ψX (1). 2

Note that if Y has ﬁnite variation, then c < 0.

(9.3.3)

Pricing Perpetual American Options

203

We shall consider only options whose payoff is a bounded function of the stock price at the time exercise only. For notational convenience, we shall write the payoff of the option as a function of the logarithm of the price at the time of exercise: thus, if the price of the stock at time of exercise is s, the payoff is given by (w), where w = log s. The expected value of the discounted payoff obtained from exercising the option at a stopping time τ is given by E[e−δτ (y − Yτ )]. The price of such an option is then maxτ E[e−δτ (y − Yτ )]. We shall consider only those options whose payoff (w) is a decreasing function of w (i.e. options of ‘put’ type). For such options, it is clear from the form of the expected payoff and the fact that Y has stationary independent increments that the optimal time of exercise is a stopping time of the form τL = inf{t ≥ 0 : St ≤ L} for some L. It is equally clear that the optimal value of the exercise level L cannot depend on the initial stock price S0 . Writing L = ea , we see that under our model, τL = Ty−a = inf{t ≥ 0 : Yt ≥ y − a}.

(9.3.4)

For a ﬁxed choice of a, let V (y, a) denote the expected value of the discounted payoff obtained from exercising the option at time Ty−a : V (y, a) = E[e−δTy−a (y − Y (Ty−a ))].

(9.3.5)

The price of such an option is then maxa V (y, a). The problem is then to ﬁnd the optimal exercise level a which maximizes V (y, a). Notice that there is a small but important difference between the deﬁnition of Tx as shown in equation (9.3.4) and that of T˜x as in equation (9.2.13): Y (Tx ) ≥ x whereas Y (T˜x ) > x. It is easy to see that Tx = T˜x almost surely except when x = 0 and Y has ﬁnite variation; in the latter case, T0 = 0 by deﬁnition whereas, because c < 0 when Y has ﬁnite variation, Y will not become positive immediately (nor will it hit 0 again immediately), so that T˜0 > 0 almost surely and by letting x ↓ 0 in equation (9.2.15) we can obtain the joint law of T˜0 and Y (T˜0 ). We have chosen the deﬁnition of Tx so as to ensure that V (a, a) = (a), which is consistent with the fact that if the initial stock price S0 = ey is at (or below) the chosen exercise level L = ea , the option is exercised immediately at time 0, resulting in a payoff (y). However, it must be emphasized that V (a, a) = (a) is purely a consequence of the deﬁnition of Ty−a , which provides a neat way of expressing the expected payoff associated with a chosen exercise strategy – in particular, the statement that V (a, a) = (a) is not the same as the continuous junction condition discussed in Gerber and Shiu (1998), which says that ˜

V (a+, a) = lim V (y, a) = E[e−δ T0 (a − Y (T˜0 ))] = (a), y→a+

(9.3.6)

From the deﬁnition of Ty−a , we already have V (a−, a) = limy→a− V (y, a) = (a), and so if equation (9.3.6) holds, the function y → V (y, a) would be continuous at y = a. If

204

Exotic Option Pricing and Advanced L´evy Models

Y has inﬁnite variation, then since T0 = T˜0 = 0 almost surely, V (y, a) is actually jointly continuous in y and a and the continuous junction condition (equation (9.3.6)) is always satisﬁed; on the other hand, if Y has ﬁnite variation, then in general V (a+, a) = (a). However, as we shall see in Section 9.5, there is a unique a ∗ for which lim V (y, a ∗ ) = (a ∗ )

y→a ∗ +

and this a ∗ turns out to be the optimal exercise level (when Y has ﬁnite variation). The ∗ meaning of this special a ∗ is that, when the initial stock price is at the level L∗ = ea , the ∗ normal rule is to exercise the option immediately, resulting in a payoff (a ); however, we would get the same payoff if we waited until the stock price has actually fallen below L∗ – the extra advantage in having S(T˜0 ) < L∗ is exactly counter-balanced by the discount ˜ factor e−d T .

9.4 A RENEWAL EQUATION APPROACH TO PRICING Throughout this section, in addition to the basic Assumption 9.2.1, we make the following assumption. Assumption 9.4.1 The jump component X of Y has ﬁnite variation: thus, equation (9.2.10) holds. We derive a renewal equation for V (y, a), the expected value of the payoff from exercising a perpetual option at level a. This is essentially the same renewal equation as obtained by Gerber and Landry (1998) in the case where X is a compound Poisson process; we show that there is an easy generalization to any jump process with ﬁnite variation. Theorem 9.4.1 Suppose Assumptions 9.2.1 and 9.4.1 hold. Let β=−

2 (ψX (1) − δ) σ2

and let 2 h(s) = 2 e−βs , σ

γ (s) = e

s

∞

e−x ν(dx),

s

and

z

g(z) = 0

2 h(z − s)γ (s) ds = 2 ez σ

z

e

−(β+1)(z−s)

∞

e−x ν(dx) ds.

(9.4.1)

s

0

Then, V (y, a), for y > a, satisﬁes

y−a

V (y, a) =

V (y − z, a)g(z) dz + e−β(y−a) (a)

0

+

∞ y−a

(y − z)g(z) dz − e−β(y−a)

0

∞

(a − z)g(z) dz.

(9.4.2)

Pricing Perpetual American Options

205

Proof. We shall approximate X by a compound Poisson process and use the corresponding result of Gerber and Landry (1998). Thus, if ν is the L´evy measure of X, let Xn (t) be a compound Poisson process with jump rate λn = ν(n−1 , ∞) and jump size distribution Pn (dx) = λ−1 n 1{x>1/n} ν(dx). Note that λn Pn (dx) = 1{x>1/n} ν(dx). Let Yn (t) = σ Bt + Xn (t) + cn t, where the drift parameter cn satisﬁes equation (9.3.3) with X replaced by Xn . Then, Yn is precisely the process used by Gerber and Landry (1998) to model the logarithm of the stock price St . Let βn = −

2 (ψXn (1) − δ) σ2

and let

2 hn (s) = 2 e−βn s , σ

γn (s) = λn e

∞

s

e−x Pn (dx),

s

and

z

gn (z) = 0

2λn hn (z − s)γn (s) ds = 2 ez σ

z

e

−(βn +1)(z−s)

∞

e−x Pn (dx) ds.

(9.4.3)

s

0

Deﬁne Vn (y, a) exactly as in equation (9.3.5) but with Y replaced by Yn . Then Gerber and Landry (1998) showed that Vn satisﬁes

y−a

Vn (y, a) =

Vn (y − z, a)gn (z) dz + e−βn (y−a) (a)

0

+

∞

(y − z)gn (z) dz − e

y−a

−βn (y−a)

∞

(a − z)gn (z) dz

(9.4.4)

0

for y > a. To ﬁnish the proof, we only have to let n → ∞. First, if ψn denotes the Laplace exponent of Yn , it is easy to see that ψn (θ ) → ψ(θ ) for θ ≥ 0. This is equivalent to the weak convergence of Yn to Y under a suitable topology, the J1 -topology of Skorohod (see Billingsley (1968)). Next, let Tx (Yn ) = inf{t ≥ 0 : Yn (t) ≥ x}. As a functional of Yn , the ﬁrst-passage time functional Tx (·) is continuous in the J1 -topology (see Whitt (1971)) and hence Tx (Yn ) converges weakly under J1 to Tx (Y ) = Tx . These facts together imply that Vn (y, a) → V (y, a). In addition, βn → β and hn (s) → h(s), γn (s) → γ (s) pointwise. From the forms of hn and γn , it is easy to see that hn and γn – and hence gn can be bounded by integrable functions and hence gn → g by dominated convergence theorem. Finally, since gn is bounded by an integrable function, letting n → ∞ in equation (9.4.4) and using the dominated convergence theorem for the integrals on the right-hand side gives equation (9.4.2). Note that if X has inﬁnite variation, the integral in the deﬁnition of g diverges; this can be most readily seen if we interchange the order of integration and write 2 g(z) = 2 e−βz σ

∞

e 0

−x

e(β+1) min(x,z) − 1 β +1

ν(dx).

There is a probabilistic explanation for why it is necessary to assume that X has ﬁnite variation in Theorem 9.4.1, which is related to the probabilistic interpretation of the function

206

Exotic Option Pricing and Advanced L´evy Models

gn deﬁned at equation (9.4.3). Let Jn denote the ﬁrst time when Yn attains a record high via a jump: formally Jn = inf{t ≥ 0 : Yn (t) = Sn (t), Yn (t−) = Sn (t−)}, where Sn (t) = sups≤t Yn (s). Consider the joint density function fn (y, t) of Yn (Jn ) and Jn . Then Gerber and Landry (1998) showed that ∞ e−δt fn (y, t) dt. gn (y) = 0

By conditioning on Yn (Jn ), the level of the ﬁrst record high in Yn caused by a jump, Gerber and Landry (1998) derived the renewal equation shown by equation (9.4.4). The same probabilistic interpretations ∞ hold when we let n → ∞: thus, if J = inf{t ≥ 0 : Yt = St , Yt− = St− }, then g(y) = 0 e−δt f (y, t) dt where f (y, t) is the joint density of Y (J ) and J . However, if X has inﬁnite variation, it makes a jump which causes a new record immediately, and so J = 0 and Y (J ) = 0 almost surely. Thus, there is a fundamental obstruction to using the conditioning argument of Gerber and Landry (1998) when X has inﬁnite variation, rather than it being merely a question of certain expressions not behaving well in the limit. To ﬁnd the optimal exercise level a ∗ , we can use the smooth pasting condition, which says that the function y → V (y, a) has a continuous ﬁrst derivative at the optimal boundary a ∗ : lim∗

y→a

∂V (y, a ∗ ) = (a ∗ ). + ∂y

By differentiating the right-hand side of equation (9.4.2), putting y = a and equating this with (a), we obtain an equation for a which is identical to that obtained by Gerber and Landry (1998). In the case of an American put option, we have (a) = (K − ea )+ and the optimal exercise level a ∗ is given by ∗

ea =

σ2 − c −

Kδ ∞ −x ν(dx) 0 xe

(9.4.5)

(where c is given by equation (9.3.3)), which is just the obvious extension of the formula obtained by Gerber and Landry (1998) to the present situation. If σ = 0 (so Y has ﬁnite variation), then since c < 0, the time of the ﬁrst record high caused by a jump is just T˜0 , the time of the ﬁrst jump by Y above its initial level 0. In this case, the same method of approximation by compound Poisson processes, but this time using the results of Gerber and Shiu (1998), shows that V satisﬁes the following simpler renewal equation ∞ y−a V (y − z, a)g(z) ˜ dz + (y − z)g(z) ˜ dz (9.4.6) V (y, a) = y−a

0

where ez g(z) ˜ = c

∞

e−x ν(dx).

z

The function g˜ has a similar probabilistic interpretation: g(y) ˜ = f˜(y, t) is the joint density of Y (T˜0 ) and T˜0 .

∞ 0

e−δt f˜(y, t) dt where

Pricing Perpetual American Options

207

9.5 EXPLICIT PRICING FORMULAE FOR AMERICAN PUTS In this section, we present explicit formulae for the optimal exercise level and value of a perpetual American put option, assuming only Assumption 9.2.1. Recall that the payoff function considered here is (w) = (K − ew )+ and that rθ(x) denotes the resolvent density ∞ of Y ; in other words, the density function of the measure 0 e−θ t P(Yt ∈ dx) dt. The main result is encapsulated in the following theorem. Theorem 9.5.1 Suppose that Assumption 9.2.1 holds and consider a perpetual American put option with payoff (y − YT ) = (K − ey−YT )+ . Then (i) the optimal exercise level is given by ∗

L∗ = e a =

Kδ ψ (1)

(9.5.1)

and for y > a ∗ , the value of a perpetual put option is given by ∞ V (y, a ∗ ) = Kδ rδ (z) dz;

(9.5.2)

y−a ∗

(ii) if Y has inﬁnite variation, the optimal exercise level a ∗ is uniquely determined by the smooth pasting condition lim∗

y→a

∂V (y, a ∗ ) ∗ = (a ∗ ) = −ea ; + ∂y

(9.5.3)

(iii) if Y has ﬁnite variation, V (y, a ∗ ) does not satisfy equation (9.5.3) – instead, the optimal exercise level a ∗ is uniquely determined by the continuity condition ∗

V (a ∗ +, a ∗ ) = lim∗ V (y, a ∗ ) = (a ∗ ) = K − ea . y→a +

(9.5.4)

Proof. (i) Suppose the option is exercised at level a at time Ty−a as described in Section 9.3 and we may assume that a < log K. Then the value function is given by V (y, a) = E[e−δT (y − YT )] = KE[e−δT ] − ey E[e−δT −YT ], where we have put T = Ty−a . For y > a, the terms on the right-hand side above are given by respectively putting θ = δ, η = 0 and θ = δ, η = 1 in equation (9.2.15) and noting the relationship shown in equation (9.3.2): ∞ rδ (z) dz − ψ (1)ea rδ (y − a). (9.5.5) V (y, a) = Kδ rδ (y − a) + y−a

To ﬁnd the optimal value of a so as to maximize V (y, a), we differentiate equation (9.5.5) with respect to a to ﬁnd ∂V (y, a) = (Kδ − ψ (1)ea )(rδ (y − a) − rδ (y − a)) = 0. ∂a

(9.5.6)

208

Exotic Option Pricing and Advanced L´evy Models

The only solution to the above which does not depend on y is a ∗ as given by equation (9.5.1) and upon substitution into equation (9.5.5), we obtain equation (9.5.2). Finally, because the function a → V (y, a) has a discontinuity at a = y when Y has ﬁnite variation, we need to check that the optimal value of V (y, a) is not in fact (y) when y > a ∗ . To this end, we show that V (y, y−) > (y), which implies that the optimal exercise level must be strictly less than y. Letting a → y− in equation (9.5.5) gives

∞ rδ (z) dz − ψ (1)ey rδ (0+). (9.5.7) V (y, y−) = Kδ rδ (0+) + 0

Putting λ = 0 into equation (9.2.14) and using equation (9.3.2) shows that ∞ 1 1 rδ (z) dz = − . δ ψ (1) 0

(9.5.8)

To ﬁnd rδ (0+), we use equation (9.2.14) together with equations (9.2.16) and (9.3.2) to obtain ∞ λ φ (δ)λ −λx − rδ (0+) = lim λ e rδ (x) dx = lim λ→∞ λ→∞ δ − ψ(λ) φ(δ) − λ 0 =

1 1 1 + φ (δ) = + . c c ψ (1)

(9.5.9)

Substituting equations (9.5.9) and (9.5.8) into equation (9.5.7) gives V (y, y−) =

Kδ − ψ (1)ey + K − ey . c

(9.5.10)

Since y > a ∗ , Kδ − ψ (1)ey < 0 and recall that c < 0 when Y has ﬁnite variation. Hence, equation (9.5.10) shows that V (y, y−) > (y) = (K − ey )+ . (ii) Differentiating equation (9.5.2) with respect to y shows that ∂V (y, a ∗ ) = −Kδrδ (y − a ∗ ), ∂y and hence lim

y↓a ∗

∂V (y, a ∗ ) = −Kδrδ (0+). ∂y

(9.5.11)

To ﬁnd rδ (0+), calculating as in equation (9.5.9) but this time using equation (9.2.17) gives ∞ λ ψ∗ (δ)λ −λz − rδ (0+) = lim λ e rδ (z) dz = lim λ→∞ λ→∞ δ − ψ(λ) ψ∗ (δ) − λ 0 = ψ∗ (δ) =

1 ψ (1)

.

(9.5.12)

Substituting this into equation (9.5.11) and using equation (9.5.1) immediately gives equation (9.5.3).

Pricing Perpetual American Options

209

(iii) That a ∗ is uniquely determined by equation (9.5.4) follows immediately from equation (9.5.10). It is also immediately apparent upon substituting equation (9.5.9) into equation (9.5.11) that equation (9.5.3) does not hold. We leave the reader to check that equation (9.5.1) agrees with equation (9.4.5) for the case where X has ﬁnite variation, and also that V (a+, a) = (a) for all a when Y has inﬁnite variation. Furthermore, note that the formula shown in equation (9.5.5) provides an explicit solution to the renewal equation (equation 9.4.2) when the jump component X has ﬁnite variation. Of course, in order to actually evaluate equation (9.5.2), one still has to compute the resolvent density rδ (x) for x > 0, for which there is rarely an explicit formula – unlike the simple formula rθ (x) = φ (θ )eφ(θ)x for x < 0 (e.g. see Section 6 of Bingham (1975)). However, very often – and certainly for all the examples considered in the next section – rδ (x) for x > 0 can be computed easily by Fourier inversion. First, observe that the Fourier transform of rδ (x) is given by ∞ −izx −izx rˆδ (z) = e rδ (x) dx = e e−δt P(Yt ∈ dx) dt =

R ∞

R

e−δt E[e−izYt ] dt =

0

Therefore, if

0 ∞

˜

e−(δ−ψ(z))t dt =

0

1 . ˜ δ − ψ(z)

1 dz < ∞, δ − ψ(z) ˜ R

rδ (x) can be recovered by using the Fourier inversion formula eixz 1 1 ixz rδ (x) = e rˆδ (z) dz = dz, ˜ 2π R 2π R δ − ψ(z)

(9.5.13)

(9.5.14)

(9.5.15)

which can be readily computed using a fast and efﬁcient numerical algorithm like the Fast Fourier Transform (FFT). In particular, equation (9.5.14) is true if σ = 0; however, note that the latter equation implies that Y has inﬁnite variation, since equation (9.2.16) implies ˜ 1/|δ − ψ(z)| dz = ∞ when Y has ﬁnite variation. Of course, one could also try to invert the Laplace transform at equation (9.2.14) but this is generally much harder computationally than Fourier inversion. In principle, other payoff functions besides (w) = (K − ew )+ can be treated in the same way but in practice, this is much harder at the computational level as it entails the inversion of the Laplace transform (equation (9.2.15)).

9.6 SOME SPECIFIC EXAMPLES We consider the following examples of L´evy processes for the jump component X. Gamma process A process is called a Gamma (α, β) process if its L´evy measure is ν(dx) = αx −1 e−βx dx,

x > 0.

(9.6.1)

210

Exotic Option Pricing and Advanced L´evy Models

Note that it has ﬁnite variation. It is well-known that Xt has the following Gamma distribution P(Xt ∈ dx) =

β αt αt−1 −βt x e dx.

(αt)

The L´evy exponent is therefore given by

iθ ψ˜ X (θ ) = −α log 1 + β

(9.6.2)

while its Laplace exponent is given by

θ ψX (θ ) = −α log 1 + . β

(9.6.3)

Stable process: index α ∈ (0, 1) A (spectrally positive) process is called stable of index α if its L´evy measure is ν(dx) = βx −α−1 dx,

x > 0.

Thus, if α ∈ (0, 1), it has ﬁnite variation. Its Laplace exponent is given by ∞ βθ ∞ −θ x −α β (1 − α) α θ . ψX (θ ) = −β (1 − e−θx )x −α−1 dx = − e x dx = − α α 0 0 (9.6.4) It is more convenient if we write the L´evy measure as ν(dx) =

βα x −α−1 dx.

(1 − α)

(9.6.5)

The Laplace exponent is then ψX (θ ) = −βθ α .

(9.6.6)

To ﬁnd the L´evy exponent, ﬁrst note the following identity (see Erd´elyi (1954), Section 2.3) ∞ πα . (9.6.7) x −α sin(θ x) dx = sgn(θ )|θ |α−1 (1 − α) cos 2 0 Hence, integrating by parts, ∞ x −α (cos(θ x) − 1) dx = 0

θ 1−α

∞

x −(α−1) sin(θ x) dx

0

π(α − 1)

(2 − α) cos = |θ |α−1 1−α 2 πα α−1 = |θ | (1 − α) sin . 2

(9.6.8)

Pricing Perpetual American Options

211

To ﬁnd the L´evy exponent, we have ∞ βα ˜ ψX (θ ) = (e−iθx − 1)x −(α+1) dx

(1 − α) 0 ∞

∞ βα = x −(α+1) (cos(θ x) − 1) dx − i x −(α+1) sin(θ x) dx

(1 − α) 0 0

π(α + 1) π(α + 1) βα α α |θ | (−α) sin − i sgn(θ )|θ | (−α) cos =

(1 − α) 2 2 πα πα . (9.6.9) + i sgn(θ ) sin = −β|θ |α cos 2 2 (We have also used the identity (z + 1) = z (z) in the above calculation.) Stable process: index α ∈ (1, 2) This is arguably the most interesting of the examples considered here, as it is the only example where the jump component X has inﬁnite variation. The L´evy measure here is given by ν(dx) = −

βα x −α−1 dx,

(1 − α)

x > 0.

(9.6.10)

(Note that if α ∈ (1, 2), (1 − α) < 0.) Such stable processes are purely discontinuous martingales, which means that in equations (9.2.4) and (9.2.8) ∞ b=− x ν(dx) 1

and we write the Laplace exponent of X as ∞ βα ψX (θ ) = (1 − e−θ x − θ x)x −(α+1) dx

(1 − α) 0 ∞ βθ = (e−θx − 1)x −α dx = βθ α ,

(1 − α) 0

(9.6.11)

where we have integrated by parts and used equation (9.6.4). Note the change in sign from the stable 0 < α < 1 case. To ﬁnd the L´evy exponent, we integrate by parts and use equation (9.6.9): ∞ βα ˜ ψX (θ ) = (1 − e−iθx − iθ x)x −(α+1) dx

(1 − α) 0 ∞ β(iθ ) = (e−iθx − 1)x −α dx

(1 − α) 0 β(iθ ) α−1 πα πα |θ | (1 − α) sin − i sgn(θ )|θ |α−1 (1 − α) cos =

(1 − α) 2 2 π α π α + i sgn(θ ) sin . (9.6.12) = β|θ |α cos 2 2 Again, note the change in sign from the stable 0 < α < 1 case.

212

Exotic Option Pricing and Advanced L´evy Models

Finally, note that there is no such thing as a spectrally positive stable process with α = 1 (other than a deterministic drift t → βt). Numerical examples The main purpose of this section is to illustrate the kind of calculations which can be carried out with these models, rather than to give a detailed comparison of the effects of using different L´evy processes. However, to achieve at least a certain degree of comparability among the different examples below, the parameters are chosen so that σ2 +

1

x 2 ν(dx) = 4

and

ν[1, ∞) = 1

(9.6.13)

0

in all of the examples. The ﬁrst condition in equation (9.6.13) says that the contribution to the volatility (as measured by quadratic variation) coming from the Brownian ﬂuctuations and the small jumps is the same in all of the examples, while the second condition in this equation (9.6.13) says that the rate at which large jumps occur is the same in all of the following examples. Throughout the following examples, we take S0 = 1 – so y = 0 – and δ = 0.1. We then compute the prices of a perpetual American put with strikes K = 0.75, 1 and 1.25 using Theorem 9.5.1. In each case, the integrability condition (equation (9.5.14)) is satisﬁed and the resolvent density is computed by approximating the Fourier integral (equation (9.5.15)) with a suitable discrete Fourier sum which is then computed by an FFT algorithm. The details of how this is carried out are described in the Appendix. We consider the following models: 1. Brownian motion plus Gamma (α,1/2) process: here, α = 1.7864, σ 2 = 3.3555 and according to equations (9.6.3) and (9.3.3), c = −0.3848 and from equation (9.2.8) a∗ ψ (1) = 2.5494, and so the optimal exercise level is L∗ = √ e =2 0.1K/2.5494. 2. Brownian motion plus stable 1/2 process: we take β = π, σ = 3.6667 and according to equations (9.6.6) and (9.3.3), c = −0.0391 and ψ (1) = 2.8196, and so L∗ = 0.1K/2.8196. √ 3. Brownian motion plus stable 3/2 process: we take β = 2 π, σ 2 = 1 and from equations (9.6.11) and (9.3.3) we have c = 3.9449 and ψ (1) = 2.3725, and so L∗ = 0.1K/2.3725. The results are summarized in the Table 9.1.

Table 9.1 Summary of the results obtained from the various models Model

K = 0.75

K = 1.00

K = 1.25

Ex. Lev. L∗ Price Ex. Lev. L∗ Price Ex. Lev. L∗ Price Brownian Motion + Gamma (α,1/2) Brownian Motion + Stable (1/2) Brownian Motion + Stable (3/2)

0.029 0.027 0.032

0.64 0.30 0.64

0.039 0.035 0.042

0.86 0.40 0.86

0.049 0.044 0.053

1.09 0.51 1.08

Pricing Perpetual American Options

213

APPENDIX: USE OF FAST FOURIER TRANSFORM Let {x(n)}N n=1 be a real sequence of length N . The discrete Fourier transform xˆ of x, and its inverse, are given by the relations x(k) ˆ =

N

e−2πi(k−1)(n−1)/N x(n)

(A.1)

n=1

x(n) =

N 1 2πi(k−1)(n−1)/N e x(k) ˆ N

(A.2)

k=1

The Fast Fourier Transform (FFT) is a fast and efﬁcient numerical algorithm for computing equations (A.1) and (A.2). Although not strictly necessary, the FFT is at its most efﬁcient if N is a power of 2. The purpose of this appendix is not to describe the workings of the FFT – there are numerous texts written on the subject and the FFT is also implemented in many standard software packages – rather, the purpose is to describe how to recast the problem of computing the Fourier inversion formula (equation (9.5.15)) into a form equivalent to equation (A.2). ∞ Let f be an integrable function: −∞ |f (x)| dx < ∞. Its Fourier transform is given by ∞ −izx fˆ(z) = −∞ e f (x) dx, and so fˆ(2π z) =

∞ −∞

e−2πizx f (x) dx.

(A.3)

Fix an integer M (ideally a power of 2) and let N = M 2 and = M/N = 1/M. We partition the interval [−M/2, M/2] into steps of length and approximate the Fourier integral (equation (A.3)) by a truncated discrete sum involving the values of f (x), for x = −M/2, −M/2 + . . . , M/2 − as follows: N fˆ 2π [(k − 1) − M/2] = e−2πi[(k−1)−M/2][(n−1)−M/2] f (n − 1) − M/2 . n=1

(A.4) giving an approximation for fˆ(z) for z = −M/2, −M/2 + , . . . , M/2 − . (For simplicity, we have chosen the crudest form of discrete approximation to equation (A.3); however, more sophisticated quadrature rules for the most part involve using weighted averages of sums of the form shown by equation (A.4) and so the method described below can be adapted to handle these more sophisticated approximations.) Rearranging equation (A.4) gives e−iπ((k−1)−N/4) fˆ 2π [(k − 1)/M − M/2] =

N n=1

e−2πi(k−1)(n−1)/N eiπ((n−1)−N/4) f (n − 1)/M − M/2 /M.

(A.5)

214

Exotic Option Pricing and Advanced L´evy Models

If fˆ is also an integrable function, then the Fourier inversion formula holds: ∞ ∞ 1 izx ˆ e f (z) dz = e2πiyx fˆ(2πy) dy. f (x) = 2π −∞ −∞

(A.6)

Approximating the above by a truncated sum over [−M/2, M/2] as before gives N e2πi[(k−1)−M/2][(n−1)−M/2] fˆ 2π [(k − 1) − M/2] , f (n − 1) − M/2 = k=1

which when rearranged gives eiπ((n−1)−N/4) f (n − 1)/M − M/2 M N 1 2πi(k−1)(n−1)/N −iπ((k−1)−N/4) ˆ e e = f 2π [(k − 1)/M − M/2] . (A.7) N k=1

Comparing equations (A.5) and (A.7) with equations (A.1) and (A.2), we see that deﬁning x(n) = M −1 eiπ((n−1)−N/4) f (n − 1)/M − M/2 x(k) ˆ = e−iπ((k−1)−N/4) fˆ 2π [(k − 1)/M − M/2]

(A.8) (A.9)

makes these equations identical. For the purposes of Section 9.6, we have f (x) = rδ (x), whose Fourier transform rˆδ (z) is given by equation (9.5.13). Deﬁning x(k) ˆ as in equation (A.9), we can recover x(n) from equation (A.2) (evaluated using FFT) and then equation (A.8) gives values of rδ (x) for a discrete set of grid-points x ∈ [−M/2, M/2], spaced = 1/M apart. The values of rδ (x) for the grid-points x ∈ [y − a ∗ , M/2] are then used to approximate the integral in equation (9.5.2). The values shown in Table 9.1 are obtained by using M = 29 = 512.

ACKNOWLEDGEMENTS I would like to thank the many people, including several anonymous referees and some of the editors of this volume, who have read and made suggestions for improving various earlier drafts of this article. I am also grateful to the editors for providing the opportunity for its publication.

REFERENCES [1] Bertoin, J. (1996), L´evy Processes, Cambridge University Press, Cambridge, UK. [2] Billingsley, P. (1968), Convergence of Probability Measures, Wiley, New York, NY, USA. [3] Bingham, N.H. (1975), “Fluctuation theory in continuous time”, Advances in Applied Probability, 7, 705–766. [4] Carr, P. (1998), “Randomization and the American put”, Review of Financial Studies, 11, 597–626.

Pricing Perpetual American Options

215

[5] Chan, T. (1999), “Pricing contingent claims on stocks driven by L´evy processes”, Annals of Applied Probability, 9, 504–528. [6] Erd´elyi, A. (1954), Tables of Integral Transforms, Vol. 1, McGraw-Hill, New York, NY, USA. [7] Gerber, H.U. and Landry, B. (1998), “On the discounted penalty at ruin in a jump-diffusion and the perpetual put option”, Insurance: Mathematics and Economics, 22, 263–276. [8] Gerber, H.U. and Shiu, E.S.W. (1994), “Martingale approach to pricing perpetual American options”, ASTIN Bulletin, 24, 195–220. [9] Gerber, H.U. and Shiu, E.S.W. (1997), “From ruin theory to option pricing”, in Joint Day Proceedings Volume of XXVIII International ASTIN Colloquium/7th International AFIR Colloquium, pp. 157–176. [10] Gerber, H.U. and Shiu, E.S.W. (1998), “Pricing perpetual options for jump processes”, North American Actuarial Journal, 2(3), 101–112. [11] MacMillan, L. (1986), “Analytic approximation for the American put option”, Advance Futures Options Research, 1, 119–139. [12] Samuelson, P.A. (1965), “Rational theory of warrant pricing”, Industrial Management Review, 6, 13–33. [13] Whitt, W. (1971), “Weak convergence of ﬁrst passage times”, Journal of Applied Probability, 8, 417–422. [14] Zhang, X. (1995), “Formules quasi-explicites pour les options Am´ericaines dans un mod`ele de diffusion avec sauts”, Mathematics and Computers in Simulation, 38, 151–161.

EPILOGUE† This article is a revised (and hopefully improved) version of an original preprint ﬁrst completed in early 2000. In the period between then and the publication of this present volume, there has been a number of new developments and papers which are related to the results presented here. The purely Brownian model considered in Carr (1998) has been extended by Avram et al. (2002) to models such as that presented here involving spectrally one-sided L´evy processes. Avram et al. (2004) further extend this method to Russian options. Many other authors have also contributed to the recent renewed interest in perpetual options of various kinds and below is a limited bibliography which also lists further papers whose themes are closely related to the present one. Much of this interest has centred on one-sided L´evy processes because of the relative ease in carrying out explicit computations; nevertheless, the more interesting case of two-sided L´evy processes has not been neglected and several authors (see, for example, Asmussen et al. (2004)) have succeeded in performing similar explicit computations in certain special cases.

FURTHER REFERENCES 1. Asmussen, S., Avram, F. and Pistorius, M.R. (2004), “Pricing American and Russian options under spectrally two-sided exponential L´evy models”, Stochastic Processes and their Applications, 109, 79–111. 2. Avram, F., Chan, T. and Usabel, M. (2002), “On the valuation of constant barrier options under spectrally one-sided exponential L´evy models and Carr’s approximation for American puts”, Stochastic Processes and their Applications, 100, 75–107. †

Added at the request of the editors.

216

Exotic Option Pricing and Advanced L´evy Models

3. Avram, F. Kyprianou, A.E. and Pistorious, M.R. (2004), “Exit problems for spectrally negative L´evy processes and applications to Canadized Russian options”, Annals of Applied Probability, 14, 215–238. 4. Boyarchenko, S.I. and Levendorski˘i, S.Z. (2002a), “Perpetual American options under L´evy processes”, SIAM Journal of Control Optimization, 40, 1663–1696. 5. Boyarchenko, S.I. and Levendorski˘i, S.Z. (2002b), “Barrier options and touch-and-out options under regular L´evy processes of exponential type”, Annals of Applied Probability, 12, 1261–1298. 6. Duistermaat, J.J., Kyprianou, A.E. and van Schaik, K. (2005), “Finite expiry Russian options”, Stochastic Processes and their Applications, 115, 609–638. 7. Mordecki, E. (1999), “Optimal stopping for a diffusion with jumps”, Finance and Stochastics, 3, 227–236. 8. Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493.

10 On Asian Options of American Type Goran Peskir University of Aarhus, Aarhus, Denmark† and

Nadia Uys University of the Witwatersrand, Witwatersrand, South Africa Abstract We show that the optimal stopping boundary for the early exercise Asian call option with ﬂoating strike can be characterized as the unique solution of a nonlinear integral equation arising from the early exercise premium representation (an explicit formula for the arbitragefree price in terms of the optimal stopping boundary). The key argument in the proof relies upon a local time-space formula.

10.1 INTRODUCTION According to ﬁnancial theory (see, e.g. Karatzas and Shreve (1998) [7] or Shiryaev (1999) [18]), the arbitrage-free price of the early exercise Asian call option with ﬂoating strike is given as V in equation (10.2.1) below where Iτ /τ denotes the arithmetic average of the stock price S up to time τ . The problem was ﬁrst studied by Hansen and Jørgensen (2000) [5] where approximations to the value function V and the optimal boundary b were derived. The main aim of this present paper is to derive exact expressions for V and b. The optimal stopping problem (equation (10.2.1)) is three-dimensional. When a changeof-measure theorem is applied (as in Shepp and Shiryaev (1994) [16] and Kramkov and Mordecky (1994) [10]) the problem reduces to (equation (10.2.9)) and becomes twodimensional. The problem (equation (10.2.9)) is more complicated than the well-known problems (Peskir (2005, 2003) [12] [13]) since the gain function depends on time in a nonlinear way. From the result of Theorem 3.1 below, it follows that the free-boundary problem (equations (10.2.10)–(10.2.14)) characterizes the value function V and the optimal stopping boundary b in a unique manner. Our main aim, however, is to follow the train of thought initiated by Kolodner (1956) [9] where V is initially expressed in terms of b, and b itself is then shown to satisfy a nonlinear integral equation. A particularly simple approach for achieving this goal in the case of the American put option has been suggested in Kim (1990) [8], Jacka (1991) [6] and Carr et al. (1992) [2] and we will take this up †

Centre for Analytical Finance (funded by the Danish Social Science Research Council) and Network in Mathematical Physics and Stochastics (funded by the Danish National Research Foundation).

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

218

Exotic Option Pricing and Advanced L´evy Models

in the present paper. We will moreover see (as in Peskir (2005, 2003) [12] [13]) that the nonlinear equation derived for b cannot have other solutions. The key argument in the proof relies upon a local time-space formula (see Peskir (2002) [11]). The latter fact of uniqueness may be seen as the principal result of the paper. The same method of proof can also be used to show the uniqueness of the optimal stopping boundary solving nonlinear integral equations derived by Hansen and Jørgensen (2000) [5] and Wu et al. (1999) [19] where this question was not explicitly addressed. These equations arise from the early exercise Asian options (call or put) with ﬂoating strike based on geometric averaging. The early exercise Asian put option with ﬂoating strike can be dealt with analogously to the Asian call option treated here. For ﬁnancial interpretations of the early exercise Asian options and other references on the topic, see Hansen and Jørgensen (2000) [5] and Wu et al. (1999) [19].

10.2 FORMULATION OF THE PROBLEM The arbitrage-free price of the early exercise Asian call option with ﬂoating strike is given by the following expression:

1 + (10.2.1) V = sup E e−rτ Sτ − Iτ τ 0 0 and a ≥ 0 are given and ﬁxed. (Throughout, B = (Bt )t≥0 denotes a standard Brownian motion started at zero.) We recall that T > 0 is the expiration date (maturity), r > 0 is the interest rate and σ > 0 is the volatility coefﬁcient. By the change-of-measure theorem, it follows that: +

+

1 1 −rτ $ = s sup E 1 − Xτ (10.2.4) V = sup E e Sτ 1 − Xτ τ τ 0 0 and x, y ≥ 0, where (s, a) → f (t, s, a) is the probability density function of (St , It ) under $ P with S0 = 1 and I0 = 0 given by: √

2 2 2 2 s r/σ (r + σ 2 /2)2 2 2π f (t, s, a) = 3/2 3 2 √ exp − t − + s) (1 π σ a t σ 2t 2σ 2 σ 2a √

∞ 4π z 4 s 2z2 dz (10.3.4) exp − 2 − 2 cosh(z) sinh(z) sin × σ t σ a σ 2t 0 for s > 0 and a > 0. For a derivation of the right-hand side in equation (10.3.4) see the appendix below. The main result of the paper may be stated as follows. Theorem 3.1 The optimal stopping boundary in the Asian call problem (equation (10.2.9)) can be characterized as the unique continuous increasing solution b : [0, T ] → IR of the nonlinear integral equation: 1−

b(t) = F (T − t, b(t)) t

T −t 1 1 − + r G(u, b(t), b(t + u)) − H (u, b(t), b(t + u)) du t +u t +u 0 (10.3.5)

satisfying 0 < b(t) < t/(1 + rt) for all 0 < t < T . [The solution b satisﬁes b(0+) = 0 and b(T −) = T /(1 + rT ), and the stopping time τb from equation (10.2.17) is optimal in equation (10.2.9).] The arbitrage-free price of the Asian call option (equation (10.2.9)) admits the following ‘early exercise premium’ representation: V (t, x) = F (T − t, x)

T −t 1 1 + r G (u, x, b(t + u)) − H (u, x, b(t + u)) du − t +u t +u 0 (10.3.6) for all (t, x) ∈ [0, T ] × [0, ∞. [Further properties of V and b are exhibited in the proof below.]

Asian Options of American Type

221

Proof. The proof will be carried out in several steps. We begin by stating some general remarks which will be freely used below without further mentioning. 1. The reason that we take the supremum in equations (10.2.1) and (10.2.9) over τ > 0 is that the ratio 1/(t + τ ) is not well deﬁned for τ = 0 when t = 0. Note, however, in equation (10.2.1) that Iτ /τ → ∞ as τ ↓ 0 when I0 = a > 0 and that Iτ /τ → s as τ ↓ 0 when I0 = a = 0. Similarly, note in equation (10.2.9) that Xτ /τ → ∞ as τ ↓ 0 when X0 = x > 0 and Xτ /τ → 1 as τ ↓ 0 when X0 = x = 0. Thus, in both cases the gain process (the integrand in equations (10.2.1) and (10.2.9)) tends to 0 as τ ↓ 0. This shows that in either equation it is never optimal to stop at t = 0. To avoid similar (purely technical) complications in the proof to follow we will equivalently consider V (t, x) only for t > 0 with the supremum taken over τ ≥ 0. The case of t = 0 will become evident (by continuity) at the end of the proof. 2. Recall that it is no restriction to assume that s = 1 and a = x so that Xt = (x + It )/St with I0 = 0 and S0 = 1. We will write Xtx instead of Xt to indicate the dependence on x when needed. It follows that V admits the following representation:

x + Iτ + $ (10.3.7) V (t, x) = sup E 1 − (t + τ ) Sτ 0≤τ ≤T −t for (t, x) ∈ 0, T ] × [0, ∞. From equation (10.3.7) we immediately see that: x → V (t, x) is decreasing and convex on [0, ∞

(10.3.8)

for each t > 0 ﬁxed. 3. We show that V : 0, T ] × [0, ∞ → IR is continuous. For this, using sup(f ) − sup(g) ≤ sup(f − g) and (z − x)+ − (z − y)+ ≤ (y − x)+ for x, y, z ∈ IR, we get:

x + Iτ + y + Iτ + $ −$ E 1− E 1− V (t, x) − V (t, y) ≤ sup (t + τ ) Sτ (t + τ ) Sτ 0≤τ ≤T −t

1 1 ≤ (y − x) E (10.3.9) ≤ (y − x) sup $ (t + τ ) Sτ t 0≤τ ≤T −t for 0 ≤ x ≤ y and t > 0, where in the last inequality we used equation (10.2.8) to deduce that t − (r + σ 2 /2)t) ≤ exp(σ B t − (σ 2 /2)t) and the latter is a martingale under 1/St = exp(σ B $ P. From equation (10.3.9) with equation (10.3.8) we see that x → V (t, x) is continuous at x0 uniformly over t ∈ [t0 − δ, t0 + δ] for some δ > 0 (small enough) whenever (t0 , x0 ) ∈ 0, T ] × [0, ∞ is given and ﬁxed. Thus, to prove that V is continuous on 0, T ] × [0, ∞ it is enough to show that t → V (t, x) is continuous on 0, T ] for each x ≥ 0 given and ﬁxed. For this, take any t1 < t2 in 0, T ] and ε > 0, and let τ1ε be a stopping time such that $ E((1 − (Xtx1 +τ ε )/(t1 + τ1ε ))+ ) ≥ V (t1 , x) − ε. Setting τ2ε = τ1ε ∧ (T − t2 ) we see that 1 V (t2 , x) ≥ $ E((1 − (Xt +τ ε )/(t2 + τ ε ))+ ). Hence we get: 2

2

2

Xtx2 +τ ε +

Xtx1 +τ ε +

1 2 $ $ −E 1− +ε V (t1 , x) − V (t2 , x) ≤ E 1 − t1 + τ1ε t2 + τ2ε Xx ε Xtx1 +τ ε +

t2 +τ2 1 ≤$ E + ε. (10.3.10) − t2 + τ2ε t1 + τ1ε

222

Exotic Option Pricing and Advanced L´evy Models

Letting ﬁrst t2 − t1 → 0 using τ1ε − τ2ε → 0 and then ε ↓ 0 we see that lim sup t2 −t1 →0 (V (t1 , x) − V (t2 , x)) ≤ 0 by dominated convergence. On the other hand, E((1 − (Xtx2 +τ ε )/(t2 + τ2ε ))+ ) ≥ V (t2 , x) − ε. Then let τ2ε be a stopping time such that $ 2 we have: Xtx2 +τ ε +

Xtx1 +τ ε +

2 2 $ $ −E 1− − ε. V (t1 , x) − V (t2 , x) ≥ E 1 − t1 + τ2ε t2 + τ2ε

(10.3.11)

Letting ﬁrst t2 − t1 → 0 and then ε ↓ 0 we see that lim inf t2 −t1 →0 (V (t1 , x) − V (t2 , x)) ≥ 0. Combining the two inequalities we ﬁnd that t → V (t, x) is continuous on 0, T ]. This completes the proof of the initial claim. 4. Denote the gain function by G(t, x) = (1 − x/t)+ for (t, x) ∈ 0, T ] × [0, ∞ and introduce the continuation set C = { (t, x) ∈ 0, T × [0, ∞ | V (t, x) > G(t, x) } and the stopping set S = { (t, x) ∈ 0, T × [0, ∞ | V (t, x) = G(t, x) }. Since V and G are continuous, we see that C is open and S is closed in 0, T × [0, ∞. Standard arguments based on the strong Markov property (cf. Shiryaev (1978) [17]) show that the ﬁrst hitting time τS = inf { 0 ≤ s ≤ T − t | (t + s, Xt+s ) ∈ S } is optimal in equation (10.2.9) as well as that V is C 1,2 on C and satisﬁes equation (10.2.10). In order to determine the structure of the optimal stopping time τS (i.e. the shape of the sets C and S), we will ﬁrst examine basic properties of the diffusion process X solving equation (10.2.6) under $ P. 5. The state space of X equals [0, ∞ and it is clear from the representation (equation (10.2.5)) with equation (10.2.8) that 0 is an entrance boundary point. The drift of X is given by µ(x) = 1 − rx and the diffusion coefﬁcient of X is given by σ (x) = σ x for x ≥ 0. Hence, we see that µ(x) is greater/less than 0, if and only if, x is less/greater than 1/r. This shows that there is a permanent push (drift) of X towards the constant level 1/r (when X is above 1/r the push of X is downwards and when X is below 1/r the push of 2 2 x X is upwards). The scale function of X is given by s(x) = 1 y 2r/σ e2/σ y dy for x > 0, 2 2 and the speed measure of X is given by m(dx) = (2/σ 2 ) x −2(1+r/σ ) e−2/σ x dx on the Borel σ -algebra of 0, ∞. Since s(0) = −∞ and s(∞) = +∞, we see that X is recurrent. 2 ∞ Moreover, since 0 m(dx) = (2/σ 2 )−2r/σ (1 + 2r/σ 2 ) is ﬁnite we ﬁnd that X has an invariant probability density function given by: 2

f (x) =

1 (2/σ 2 )1+2r/σ 2 e−2/σ x 2) 2 2(1+r/σ

(1 + 2r/σ ) x

(10.3.12)

for x > 0. In particular, it follows that Xt /t → 0 $ P-a.s. as t → ∞. This fact has an important consequence for the optimal stopping problem (equation (10.2.9)): if the horizon T is inﬁnite, then it is never optimal to stop. Indeed, in this case letting τ ≡ t and passing to the limit for t → ∞ we see that V ≡ 1 on 0, ∞ × [0, ∞. This shows that the inﬁnite horizon formulation of the problem (equation (10.2.9)) provides no useful information to the ﬁnite horizon formulation (such as in Peskir (2005, 2003) [12] [13], for example). To examine the latter beyond the trivial fact that all points (t, x) with x ≥ t belong to C (which is easily seen by considering the hitting times τε = inf { 0 ≤ s ≤ T − t | Xt+s ≤ (t + s) − ε } and noting that $ Pt,x (0 < τε < T − t) > 0 if x ≥ t with 0 < t < T ), we will examine the gain process in the problem (equation (10.2.9)) using stochastic calculus as follows.

Asian Options of American Type

223

6. Setting α(t) = t for 0 ≤ t ≤ T to denote the diagonal in the state space and applying the local time–space formula (cf. Peskir (2002) [11]) under $ Pt,x when (t, x) ∈ 0, T × [0, ∞ is given and ﬁxed, we get:

s

G(t + s, Xt+s ) = G(t, x) +

Gt (t + u, Xt+u ) du

0

1 s + Gx (t + u, Xt+u ) dXt+u + Gxx (t + u, Xt+u ) dX, Xt+u 2 0 0 1 s α Gx (t + u, α(t + u)+) − Gx (t + u, α(t + u)−) d%t+u + (X) 2 0

s Xt+u 1 − rXt+u I Xt+u < α(t + u) du = G(t, x) + − 2 (t + u) (t + u) 0 s s α d%t+u (X) Xt+u u + 1 I Xt+u < α(t + u) d B (10.3.13) −σ 2 0 t +u 0 t +u

s

where %αt+u (X) is the local time of X on the curve α given by: 1 %αt+u (X) = $ P−lim ε↓0 2ε =$ P−lim ε↓0

1 2ε

u

I α(t + v)−ε < Xt+v < α(t + v) + ε dX, Xt+v

u

σ2 2 X dv (10.3.14) I α(t + v)−ε < Xt+v < α(t + v) + ε 2 t+v

0

0

and d%αt+u (X) refers to the integration with respect to the continuous increasing function u → %αt+u (X). From equation (10.3.13) we respectively read: G(t + s, Xt+s ) = G(t, x) + As + Ms + Ls

(10.3.15)

where A and L are processes of bounded variation (L is increasing ) and M is a continuous (local) martingale. We note, moreover, that s → Ls is strictly increasing only when Xs = α(s) for 0 ≤ s ≤ T − t, i.e. when X visits α. On the other hand, when X is below α then the integrand a(t + u, Xt+u ) of As may be either positive or negative. To determine both regions exactly, we need to examine the sign of the expression a(t, x) = x/t 2 − (1 − rx)/t. It follows that a(t, x) is larger/less than 0, if and only if, x is larger/less than γ (t) where γ (t) = t/(1 + rt) for 0 ≤ t ≤ T . By considering the exit times from small balls in 0, T × [0, ∞ with centre at (t, x) and making use of equation (10.3.13) with the optional sampling theorem (to get rid of the martingale part), upon observing that α(t) < γ (t) for all 0 < t ≤ T so that the local time part is zero, we see that all points (t, x) lying above the curve γ (i.e. x > γ (t) for 0 < t < T ) belong to the continuation set C. Exactly the same arguments (based on the fact that the favourable regions above γ and on α are far away from X) show that for each x < γ (T ) = T /(1 + rT ) given and ﬁxed, all points (t, x) belong to the stopping set S when t is close to T . Moreover, recalling equation (10.3.8) and the fact that V (t, x) ≥ G(t, x) for all x ≥ 0 with t ∈ 0, T ﬁxed, we see that for each t ∈ 0, T there is a point b(t) ∈ [0, γ (t)] such that V (t, x) > G(t, x) for x > b(t) and V (t, x) = G(t, x) for x ∈ [0, b(t)]. Combining it with the previous conclusion on S we ﬁnd that b(T −) = γ (T ) = T /(1 + rT ). (Yet another argument for this identity will be given

224

Exotic Option Pricing and Advanced L´evy Models

below. Note that this identity is different from the identity b(T −) = T used in Hansen and Jørgensen (2000) [5, page 1126].) This establishes the existence of the non-trivial (nonzero) optimal stopping boundary b on a left-neighbourhood of T . We will now show that b extends (continuously and decreasingly) from the initial neighbourhood of T backward in time as long as it visits 0 at some time t0 ∈ [0, T , and later in the second part of the proof below we will deduce that this t0 is equal to 0. The key argument in the proof is provided by the following inequality. Notice that this inequality is not obvious a priori (unlike in Peskir (2005, 2003) [12] [13]) since t → G(t, x) is increasing and the supremum in equation (10.2.9) is taken over a smaller class of stopping times τ ∈ [0, T −t] when t is larger. 7. We show that the inequality is satisﬁed: Vt (t, x) ≤ Gt (t, x)

(10.3.16)

for all (t, x) ∈ C. (It may be noted from equation (10.2.10) that Vt = −(1 − rx)Vx − (σ 2 /2)x 2 Vxx ≤ (1 − rx)/t since Vx ≥ −1/t and Vxx ≥ 0 by equation (10.3.8), so that Vt ≤ Gt holds above γ because (1 − rx)/t ≤ x/t 2 , if and only if, x ≥ t/(1 + rt). Hence, the main issue is to show that equation (10.3.16) holds below γ and above b. Any analytic proof of this fact seems difﬁcult and we resort to probabilistic arguments.) To prove equation (10.3.16), ﬁx 0 < t < t + h < T and x ≥ 0 so that x ≤ γ (t). Let τ = τS (t + h, x) be the optimal stopping time for V (t + h, x). Since τ ∈ [0, T −t −h] ⊆ [0, T − t], we see that V (t, x) ≥ $ Et,x ((1 − Xt+τ /(t + τ ))+ ) so that using the inequality stated prior to equation (10.3.9) above (and the convenient reﬁnement by an indicator function), we get: V (t + h, x) − V (t, x) − G(t + h, x) − G(t, x)

+

x + Iτ + x x x + Iτ $ $ −E 1− − − ≤E 1− (t + h + τ ) Sτ (t + τ ) Sτ t t +h

x + Iτ x + Iτ x + Iτ xh I ≤$ E − ≤1 − (t + τ ) Sτ (t + h + τ ) Sτ (t + h + τ ) Sτ t (t + h)

1 1 x + Iτ x + Iτ xh I − =$ E ≤1 − Sτ t +τ t +h+τ (t + h + τ ) Sτ t (t + h)

x + Iτ h x + Iτ xh I =$ E ≤1 − (t + h + τ ) Sτ t + τ (t + h + τ ) Sτ t (t + h)

x + Iτ x + Iτ h xh E ≤ $ I ≤1 − ≤0 (10.3.17) t (t + h + τ ) Sτ (t + h + τ ) Sτ t (t + h) where the ﬁnal inequality follows from the fact that with Z := (x + Iτ )/((t + h + τ )Sτ ) we have V (t + h, x) = $ E((1−Z)+ ) = $ E((1−Z) I (Z ≤ 1)) = $ P(Z ≤ 1) − $ E(Z I (Z ≤ $ $ 1)) ≥ G(t + h, x) = 1 − x/(t + h) so that E(Z I (Z ≤ 1)) ≤ P(Z ≤ 1) − 1 + x/(t + h) ≤ x/(t + h) as claimed. Dividing the initial expression in equation (10.3.17) by h and letting h ↓ 0 we obtain equation (10.3.16) for all (t, x) ∈ C such that x ≤ γ (t). Since Vt ≤ Gt above γ (as stated following equation (10.3.16) above) this completes the proof of equation (10.3.16).

Asian Options of American Type

225

8. We show that t → b(t) is increasing on 0, T . This is an immediate consequence of equation (10.3.17). Indeed, if (t, x) belongs to C and t0 from 0, T satisﬁes t0 < t1 , then by equation (10.3.17) we have that V (t0 , x) − G(t0 , x) ≥ V (t1 , x) − G(t1 , x) > 0 so that (t0 , x) must belong to C. It follows that b cannot be strictly decreasing thus proving the claim. 9. We show that the smooth-ﬁt condition equation (10.2.12) holds, i.e. that x → V (t, x) is C 1 at b(t). For this, ﬁx a point (t, x) ∈ 0, T × 0, ∞ lying at the boundary so that x = b(t). Then x ≤ γ (t) < α(t) and for all ε > 0 such that x + ε < α(t) we have: G(t, x + ε) − G(t, x) 1 V (t, x + ε) − V (t, x) ≥ =− . ε ε t

(10.3.18)

Letting ε ↓ 0 and using that the limit on the left-hand side exists (since x → V (t, x) is convex), we get the inequality: ∂ +V ∂G 1 (t, x) ≥ (t, x) = − . ∂x ∂x t

(10.3.19)

To prove the converse inequality, ﬁx ε > 0 such that x + ε < α(t), and consider the stopping times τε = τS (t, x + ε) being optimal for V (t, x + ε). Then we have:

+

V (t, x + ε) − V (t, x) 1 $ x + ε + Iτε + x + Iτε ≤ E 1− − 1− ε ε (t + τε ) Sτε (t + τε ) Sτε

1 x + Iτε 1 x + ε + Iτε = −$ E . (10.3.20) E ≤ $ − ε (t + τε ) Sτε (t + τε ) Sτε (t + τε ) Sτε Since each point x in 0, ∞ is regular for X, and the boundary b is increasing, it follows that τε ↓ 0 $ P − a.s. as ε ↓ 0. Letting ε ↓ 0 in equation (10.3.20) we get: ∂ +V 1 (t, x) ≤ − ∂x t

(10.3.21)

by dominated convergence. It follows from equation (10.3.19) and (10.3.21) that (∂ + V /∂x)(t, x) = −1/t implying the claim. 10. We show that b is continuous. Note that the same proof also shows that b(T −) = T /(1 + rT ) as already established above by a different method. Let us ﬁrst show that b is right-continuous. For this, ﬁx t ∈ 0, T and consider a sequence tn ↓ t as n → ∞. Since b is increasing, the right-hand limit b(t+) exists. Because (tn , b(tn )) ∈ S for all n ≥ 1, and S is closed, it follows that (t, b(t+)) ∈ S. Hence by equation (10.2.16) we see b(t+) ≤ b(t). Since the reverse inequality follows obviously from the fact that b is increasing, this completes the proof of the ﬁrst claim. Let us next show that b is left-continuous. Suppose that there exists t ∈ 0, T such that b(t−) < b(t). Fix a point x in b(t−), b(t)] and note by equation (10.2.12) that for s < t we have: x y Vxx (s, z) − Gxx (s, z) dz dy (10.3.22) V (s, x) − G(s, x) = b(s)

b(s)

upon recalling that V is C 1,2 on C. Note that Gxx = 0 below α so that if Vxx ≥ c on R = { (u, y) ∈ C | s ≤ u < t and b(u) < y ≤ x } for some c > 0 (for all s < t close enough

226

Exotic Option Pricing and Advanced L´evy Models

to t and some x > b(t−) close enough to b(t−)) then by letting s ↑ t in equation (10.3.22) we get: 2 x − b(t) >0 (10.3.23) V (t, x) − G(t, x) ≥ c 2 contradicting the fact that (t, x) belongs to D and thus is an optimal stopping point. Hence, the proof reduces to showing that Vxx ≥ c on small enough R for some c > 0. To derive the latter fact we may ﬁrst note from equation (10.2.10) upon using equation (10.3.16) that Vxx = (2/(σ 2 x 2 ))(−Vt − (1 − rx)Vx ) ≥ (2/(σ 2 x 2 ))(−x/t 2 − (1 − rx)Vx ). Suppose now that for each δ > 0 there is s < t close enough to t and there is x > b(t−) close enough to b(t−) such that Vx (u, y) ≤ −1/u + δ for all (u, y) ∈ R (where we recall that −1/u = Gx (u, y) for all (u, y) ∈ R). Then from the previous inequality, we ﬁnd that Vxx (u, y) ≥ (2/(σ 2 y 2 ))(−y/u2 + (1 − ry)(1/u − δ)) = (2/(σ 2 y 2 ))((u − y(1 + ru))/u2 − δ(1 − ru)) ≥ c > 0 for δ > 0 small enough since y < u/(1 + ru) = γ (u) and y < 1/r for all (u, y) ∈ R. Hence, the proof reduces to showing that Vx (u, y) ≤ −1/u + δ for all (u, y) ∈ R with R small enough when δ > 0 is given and ﬁxed. To derive the latter inequality we can make use of the estimate (equation (10.3.20)) to conclude that

1 V (u, y + ε) − V (u, y) (10.3.24) ≤ −$ E ε (u + σε ) Mσε y+ε

where σε = inf { 0 ≤ v ≤ T − u | Xu+v = b(u) } and Mt = sup0≤s≤t Ss . A simple comparison argument (based on the fact that b is increasing) shows that the supremum over all (u, y) ∈ R on the right-hand side of equation (10.3.24) is attained at (s, x + ε). Letting ε ↓ 0 in equation (10.3.24), we thus get:

1 $ (10.3.25) Vx (u, y) ≤ − E (u + σ ) Mσ x for all (u, y) ∈ R where σ = inf { 0 ≤ v ≤ T − s | Xs+v = b(s) }. Since by regularity of X we ﬁnd that σ ↓ 0 $ P-a.s. as s ↑ t and x ↓ b(t−), it follows from equation (10.3.25) that:

1 1 $ (u + σ ) Mσ − u ≤− +δ (10.3.26) Vx (u, y) ≤ − + E u u (u + σ ) Mσ u

for all s < t close enough to t and some x > b(t−) close enough to b(t−). This completes the proof of the second claim, and thus the initial claim is proved as well. 11. We show that V is given by the formula shown in equation (10.3.6) and that b solves equation (10.3.5). For this, note that V satisﬁes the following conditions: V is C 1,2 on C ∪ D

(10.3.27)

Vt + LX V is locally bounded

(10.3.28)

x → V (t, x) is convex

(10.3.29)

t → Vx (t, b(t)±) is continuous.

(10.3.30)

Indeed, the conditions (10.3.27) and (10.3.28) follow from the facts that V is C 1,2 on C and V = G on D upon recalling that D lies below γ so that G(t, x) = 1 − x/t for all

Asian Options of American Type

227

(t, x) ∈ D and thus G is C 1,2 on D. [When we say in condition (10.3.28) that Vt + LX V is locally bounded, we mean that Vt + LX V is bounded on K ∩ (C ∪ D) for each compact set K in [0, T ] × IR+ .] The condition (10.3.29) was established in equation (10.3.8) above. The condition (10.3.30) follows from equation (10.2.12) since according to the latter we have Vx (t, b(t)±) = −1/t for t > 0. Since conditions (10.3.27)–(10.3.30) are satisﬁed, we know that the local time-space formula (cf. Theorem 3.1 in Peskir (2002) [11]) can be applied. This gives: s V (t + s, Xt+s ) = V (t, x) + Vt + LX V (t + u, Xt+u ) I Xt+u = b(t + u) du 0

s

+ 0

=

1 + 2 s

σ Xt+u Vx (t + u, Xt+u ) I Xt+u = b(t + u) dBu

s 0

b Vx (t + u, Xt+u +) − Vx (t + u, Xt+u −) I Xt+u = b(t + u) d%t+u (X)

Gt + LX G (t + u, Xt+u ) I Xt+u < b(t + u) du + Ms

(10.3.31)

0

the ﬁnal equality follows by the smooth-ﬁt condition (10.2.12) and Ms = where s σ X Vx (t + u, Xt+u ) I Xt+u = b(t + u) dBu is a continuous martingale for 0 ≤ s ≤ t+u 0 T − t with t > 0. Noting that (Gt + LX G)(t, x) = x/t 2 − (1 − rx)/t for x < t we see that equation (10.3.31) yields:

s Xt+u 1 − rXt+u I Xt+u < b(t + u) du + Ms . V (t + s, Xt+s ) = V (t, x) + − 2 (t + u) (t + u) 0 (10.3.32) Setting s = T − t, using that V (T , x) = G(T , x) for all x ≥ 0, and taking the $ Pt,x expectation in equation (10.3.32), we ﬁnd by the optional sampling theorem that:

XT + 1− T

T −t Xt+u 1 − rXt+u $ I Xt+u < b(t + u) du. (10.3.33) Et,x − = V (t, x) + (t + u)2 (t + u) 0

$ Et,x

Making use of equations (10.3.1)–(10.3.3) we see that equation (10.3.33) is the formula (10.3.6). Moreover, inserting x = b(t) in equation (10.3.33) and using that V (t, b(t)) = G(t, b(t)) = 1 − b(t)/t, we see that b satisﬁes the equation (10.3.5) as claimed. 12. We show that b(t) > 0 for all 0 < t ≤ T and that b(0+) = 0. For this, suppose that b(t0 ) = 0 for some t0 ∈ 0, T and ﬁx t ∈ 0, t0 . Then, (t, x) ∈ C for all x > 0 as small as desired. Taking any such (t, x) ∈ C and denoting by τS = τS (t, x) the ﬁrst hitting time to S under $ Pt,x , we ﬁnd by equation (10.3.32) that:

Xt+τS + V (t + τS , Xt+τS ) = G(t + τS , Xt+τS ) = 1 − = V (t, x) + Mt+τS t + τS x = 1 − + Mt+τS . (10.3.34) t

228

Exotic Option Pricing and Advanced L´evy Models

Taking the $ Pt,x -expectation and letting x ↓ 0 we get:

Xt+τS + $ Et,0 1 − =1 t + τS

(10.3.35)

Pt,0 (Xt+τS ≥ T ) > 0 we see that the left-hand side of where τS = τS (t, 0). As clearly $ equation (10.3.35) is strictly smaller than 1, thus contradicting the identity. This shows that b(t) must be strictly positive for all 0 < t ≤ T . Combining this conclusion with the known inequality b(t) ≤ γ (t), which is valid for all 0 < t ≤ T , we see that b(0+) = 0 as claimed. 13. We show that b is the unique solution of the nonlinear integral equation (10.3.5) in the class of continuous functions c : 0, T → IR satisfying 0 < c(t) < t/(1 + rt) for all 0 < t < T . (Note that this class is larger than the class of functions having the established properties of b which is, moreover, known to be increasing.) The proof of the uniqueness will be presented in the ﬁnal three steps of the main proof as follows. 14. Let c : 0, T ] → IR be a continuous solution of the equation (10.3.5) satisfying 0 < c(t) < t for all 0 < t < T . We want to show that this c must then be equal to the optimal stopping boundary b. Motivated by the derivation (10.3.31)–(10.3.33) which leads to the formula (10.3.6), let us consider the function U c : 0, T ] × [0, ∞ → IR deﬁned as follows:

XT + 1− T

T −t Xt+u 1 − rXt+u $ I X Et,x − < c(t + u) du − t+u (t + u)2 (t + u) 0

Et,x U (t, x) = $ c

(10.3.36)

for (t, x) ∈ 0, T ] × [0, ∞. In terms of equations (10.3.1)–(10.3.3), note that U c is explicitly given by: U c (t, x) = F (T − t, x)

T −t 1 1 + r G u, x, c(t + u) − H u, x, c(t + u) du − t +u t +u 0 (10.3.37) for (t, x) ∈ 0, T ] × [0, ∞. Observe that the fact that c solves equation (10.3.5) on 0, T means exactly that U c (t, c(t)) = G(t, c(t)) for all 0 < t < T . We will now, moreover, show that U c (t, x) = G(t, x) for all x ∈ [0, c(t)] with t ∈ 0, T . This is the key point in the proof (cf. Peskir (2005, 2003) [12] [13]) that can be derived using a martingale argument as follows. If X = (Xt )t≥0 is a Markov process (with values in a general state space) and we set F (t, x) = Ex (G(XT −t )) for a (bounded) measurable function G with Px (X0 = x) = 1, then the Markov property of X implies that F (t, Xt ) is a martingale under Px for 0 ≤ t ≤ T . T −t Similarly, if we set F (t, x) = Ex ( 0 H (Xu ) du) for a (bounded) measurable function H t with Px (X0 = x) = 1, then the Markov property of X implies that F (t, Xt ) + 0 H (Xu ) du is a martingale under Px for 0 ≤ t ≤ T . Combining these two martingale facts applied to

Asian Options of American Type

the time–space Markov process (t + s, Xt+s ) instead of Xs , we ﬁnd that:

s Xt+u 1 − rXt+u I Xt+u < c(t + u) du − U c (t + s, Xt+s ) − 2 (t + u) (t + u) 0

229

(10.3.38)

is a martingale under $ Pt,x for 0 ≤ s ≤ T − t. We may thus write:

s Xt+u 1 − rXt+u I Xt+u < c(t + u) du = U c (t, x) + Ns − U c (t + s, Xt+s ) − 2 (t + u) (t + u) 0 (10.3.39) where (Ns )0≤s≤T −t is a martingale with N0 = 0 under $ Pt,x . On the other hand, we know from equation (10.3.13) that:

s Xt+u 1 − rXt+u I Xt+u < α(t + u) du − G(t + s, Xt+s ) = G(t, x) + 2 (t + u) (t + u) 0 + Ms + Ls

(10.3.40)

s u is a continuous martingale under whereMs = −σ 0 (Xt+u /(t + u)) I (Xt+u < α(t + u)) d B s α $ Pt,x and Ls = (1/2) 0 d%t+u (X)/(t + u) is an increasing process for 0 ≤ s ≤ T − t. For 0 ≤ x ≤ c(t) with t ∈ 0, T given and ﬁxed, consider the stopping time: σc = inf { 0 ≤ s ≤ T − t | Xt+s ≥ c(t + s) }.

(10.3.41)

Using that U c (t, c(t)) = G(t, c(t)) for all 0 < t < T (since c solves equation (10.3.5) as pointed out above) and that U c (T , x) = G(T , x) for all x ≥ 0, we see that U c (t + σc , Xt+σc ) = G(t + σc , Xt+σc ). Hence from equations (10.3.39) and (10.3.40) using the optional sampling theorem we ﬁnd: Et,x U c (t + σc , Xt+σc ) U c (t, x) = $ σc

Xt+u 1 − rXt+u I X − < c(t + u) du −$ Et,x t+u (t + u)2 (t + u) 0 =$ Et,x G(t + σc , Xt+σc )

Xt+u 1 − rXt+u I Xt+u < c(t + u) du − (t + u)2 (t + u) 0

σc

Xt+u 1 − rXt+u $ − I Xt+u < α(t + u) du = G(t, x) + Et,x (t + u)2 (t + u) 0 σc

Xt+u 1 − rXt+u $ I Xt+u < c(t + u) du − Et,x − (t + u)2 (t + u) 0 −$ Et,x

= G(t, x)

σc

(10.3.42)

since Xt+u < α(t + u) and Xt+u < c(t + u) for all 0 ≤ u < σc . This proves that U c (t, x) = G(t, x) for all x ∈ [0, c(t)] with t ∈ 0, T as claimed.

230

Exotic Option Pricing and Advanced L´evy Models

15. We show that U c (t, x) ≤ V (t, x) for all (t, x) ∈ 0, T ] × [0, ∞. For this, consider the stopping time: τc = inf { 0 ≤ s ≤ T − t | Xt+s ≤ c(t + s) }

(10.3.43)

under $ Pt,x with (t, x) ∈ 0, T ] × [0, ∞ given and ﬁxed. The same arguments as those given following equation (10.3.41) above show that U c (t + τc , Xt+τc ) = G(t + τc , Xt+τc ). Inserting τc instead of s in equation (10.3.39) and using the optional sampling theorem we get: Et,x U c (t + τc , Xt+τc ) = $ Et,x G(t + τc , Xt+τc ) ≤ V (t, x) (10.3.44) U c (t, x) = $ where the ﬁnal inequality follows from the deﬁnition of V proving the claim. 16. We show that c ≥ b on [0, T ]. For this, consider the stopping time: σb = inf { 0 ≤ s ≤ T − t | Xt+s ≥ b(t + s) }

(10.3.45)

under $ Pt,x where (t, x) ∈ 0, T × [0, ∞ such that x < b(t) ∧ c(t). Inserting σb in place of s in equations (10.3.32) and (10.3.39) and using the optional sampling theorem we get: σb

Xt+u 1 − rXt+u $ du (10.3.46) Et,x V (t + σb , Xt+σb ) = G(t, x) + $ Et,x − (t + u)2 (t + u) 0 $ Et,x U c (t + σb , Xt+σb ) σb

Xt+u 1 − rXt+u I X − < c(t + u) du (10.3.47) = G(t, x) + $ Et,x t+u (t + u)2 (t + u) 0 where we also use that V (t, x) = U c (t, x) = G(t, x) for x < b(t) ∧ c(t). Since U c ≤ V it follows from equations (10.3.46) and (10.3.47) that: σb

Xt+u 1 − rXt+u $ I X − ≥ c(t + u) du ≥ 0. (10.3.48) Et,x t+u (t + u)2 (t + u) 0 Due to the fact that b(t) < t/(1 + rt) for all 0 < t < T , we see that Xt+u /(t + u)2 − (1 − rXt+u )/(t + u) < 0 in equation (10.3.48), so that by the continuity of b and c it follows that c ≥ b on [0, T ] as claimed. 17. We show that c must be equal to b. For this, let us assume that there is t ∈ 0, T such that c(t) > b(t). Pick x ∈ b(t), c(t) and consider the stopping time τb from equation (10.2.17). Inserting τb instead of s in equations (10.3.32) and (10.3.39) and using the optional sampling theorem, we get: $ Et,x G(t + τb , Xt+τb ) = V (t, x) (10.3.49) $ Et,x G(t + τb , Xt+τb ) τb

Xt+u 1 − rXt+u c $ I Xt+u < c(t + u) du − (10.3.50) = U (t, x) + Et,x (t + u)2 (t + u) 0

Asian Options of American Type

231

where we also use that V (t + τb , Xt+τb ) = U c (t + τb , Xt+τb ) = G(t + τb , Xt+τb ) upon recalling that c ≥ b and U c = G either below c or at T . Since U c ≤ V we see from equations (10.3.49) and (10.3.50) that: τb

Xt+u 1 − rXt+u $ I X Et,x − < c(t + u) du ≥ 0. (10.3.51) t+u (t + u)2 (t + u) 0 Due to the fact that c(t) < t/(1 + rt) for all 0 < t < T by assumption, we see that Xt+u /(t + u)2 − (1 − rXt+u )/(t + u) < 0 in equation (10.3.51) so that by the continuity of b and c it follows that such a point (t, x) cannot exist. Thus c must be equal to b, and the proof is complete.

10.4 REMARKS ON NUMERICS 1. The following method can be used to calculate the optimal stopping boundary b numerically by means of the integral equation (10.3.5). Note that the formula (10.3.6) can be used to calculate the arbitrage-free price V when b is known. Set ti = ih for i = 0, 1, . . . , n where h = T /n and denote: J (t, b(t)) = 1 −

b(t) − F (T −t, b(t)) t

(10.4.1)

K(t, b(t);t + u, b(t + u)) (10.4.2)

1 1 + r G(u, b(t), b(t + u)) − H (u, b(t), b(t + u)) . (10.4.3) = t +u t +u Then, the following discrete approximation of the integral equation (10.3.5) is valid: J (ti , b(ti )) =

n

K(ti , b(ti ); tj , b(tj )) h

(10.4.4)

j =i+1

for i = 0, 1, . . . , n − 1. Letting i = n−1 and b(tn ) = T /(1 + rT ) we can solve equation (10.4.4) numerically and get a number b(tn−1 ). Letting i = n−2 and using the values of b(tn−1 ) and b(tn ) we can solve equation (10.4.4) numerically and get a number b(tn−2 ). Continuing the recursion, we obtain b(tn ), b(tn−1 ), . . . , b(t1 ), b(t0 ) as an approximation of the optimal stopping boundary b at points 0, h, . . . , T −h, T . It is an interesting numerical problem to show that the approximation converges to the true function b on [0, T ] as h ↓ 0. Another interesting problem is to derive the rate of convergence. 2. To perform the previous recursion, we need to compute the functions F, G, H from equations (10.3.1)–(10.3.3) as efﬁciently as possible. Simply by observing the expressions (10.3.1)–(10.3.4) it is apparent that ﬁnding these functions numerically is not trivial. Moreover, the nature of the probability density function f in expression (10.3.4) presents a further numerical challenge. Part of this probability density function is the Hartman–Watson density discussed in Barrieu et al. (2003) [1]. As t tends to zero, the numerical estimate of the Hartman–Watson density oscillates, with the oscillations increasing rapidly in both amplitude and frequency as t gets closer to zero. Barrieu et al. (2003) [1] mention that this may be a consequence of the fact that t → exp(2π 2 /σ 2 t) rapidly increases to inﬁnity while z → sin(4π z/σ 2 t) oscillates more and more frequently. This rapid oscillation makes accurate estimation of f (t, s, a) with t close to zero very difﬁcult.

232

Exotic Option Pricing and Advanced L´evy Models

The problems when dealing with t close to zero are relevant to pricing the early exercise Asian call option. To ﬁnd the optimal stopping boundary b as the solution to the implicit equation (10.4.4), it is necessary to work backward from T to 0. Thus, to get an accurate estimate for b when b(T ) is given, the next estimate of b(u) must be found for some value of u close to T so that t = T −u will be close to zero. Even if we get an accurate estimate for f , to solve equations (10.3.1)–(10.3.3) we need to evaluate two nested integrals. This is slow computationally. A crude attempt has been made at storing values for f and using these to estimate F, G, H in equations (10.3.1)–(10.3.3) but this method has not produced reliable results. 3. Another approach to ﬁnding the functions F, G, H from equations (10.3.1)–(10.3.3) can be based on numerical solutions of partial differential equations. Two distinct methods are available. Consider the transition probability density of the process X given by: p(s, x; t, y) =

d $ P(Xt ≤ y | Xs = x) dy

(10.4.5)

where 0 ≤ s < t and x, y ≥ 0. Since p(s, x; t, y) = p(0, x; t −s, y), we see that there is no restriction to assume that s = 0 in the sequel. 4. The forward equation approach leads to the initial-value problem: pt = −((1−ry) p)y + (D y p)yy p(0, x; 0+, y) = δ(y −x)

(t > 0, y > 0)

(y ≥ 0)

(10.4.6) (10.4.7)

where D = σ 2/2 and x ≥ 0 is given and ﬁxed (recall that δ denotes the Dirac delta function). Standard results (cf. Feller (1952) [4]) imply that there is a unique non-negative solution (t, y) → p(0, x; t, y) of equations (10.4.6) and (10.4.7). The solution p satisﬁes the following boundary conditions: p(0, x; t, 0+) = 0 (0 is entrance)

(10.4.8)

p(0, x; t, ∞−) = 0 (∞ is normal).

(10.4.9)

The solution p satisﬁes the following integrability condition:

∞

p(0, x; t, y) dy = 1

(10.4.10)

0

for all x ≥ 0 and all t ≥ 0. Once the solution (t, y) → p(0, x; t, y) of equations (10.4.6) and (10.4.7) has been found, the functions F, G, H from equations (10.3.1)–(10.3.3) can be computed using the general formula: $ E0,x (g(Xt )) =

∞

g(y) p(0, x; t, y) dy 0

upon choosing the appropriate function g : IR+ → IR+ .

(10.4.11)

Asian Options of American Type

233

5. The backward equation approach leads to the terminal-value problem: qt = (1−rx) qx + D x 2 qxx q(T , x) = h(x)

(t > 0, x > 0)

(x ≥ 0)

(10.4.12) (10.4.13)

where h : IR+ → IR+ is a given function. Standard results (cf. Feller (1952) [4]) imply that there is a unique non-negative solution (t, x) → q(t, x) of equations (10.4.12) and (10.4.13). Taking x → h(x) to be x → (1−x/T )+ ( with T ﬁxed ), x → x I (x ≤ y) ( with y ﬁxed ), x → I (x ≤ y) ( with y ﬁxed ), it follows that the unique non-negative solution q of equations (10.4.12) and (10.4.13) coincides with F, G, H from equations (10.3.1)–(10.3.3), respectively. (For numerical results of a similar approach, see Rogers and Shi (1995) [14].) 6. It is an interesting numerical problem to carry out either of the two methods described above and produce approximations to the optimal stopping boundary b by using equation (10.4.4). Another interesting problem is to derive the rate of convergence.

APPENDIX In this section we derive the explicit expression for the probability density function f of (St , It ) under $ P with S0 = 1 and I0 = 0 given in equation (10.3.4) above. Let B = (Bt )t≥0 be a standard Brownian motion deﬁned on a probability space (, F , P). With t > 0 and ν ∈ IR given and ﬁxed, recall from Yor (1992, p. 527) [20] that the random t 2(Bs +νs) variable A(ν) ds has the conditional distribution: t = 0 e P A(ν) (10.A.1) t ∈ dy Bt + νt = x = a(t, x, y) dy where the density function a for y > 0 is given by: a(t, x, y) =

2

1 x + π2 1 2x 1 + e + x − exp πy 2 2t 2y

2 ∞ πz ex z dz. cosh(z) sinh(z) sin × exp − − 2t y t 0

(10.A.2)

This implies that the random vector 2(Bt + νt), A(ν) has the distribution: t P 2(Bt + νt) ∈ dx, A(ν) t ∈ dy = b(t, x, y) dx dy where the density function b for y > 0 is given by:

x 1 x − 2νt √ b(t, x, y) = a t, , y √ ϕ 2 2 t 2 t 2

ν2 1 π ν + 1 1 x x− 1+e = + t− √ exp 2t 2 2 2y (2π )3/2 y 2 t

2 ∞ πz ex/2 z cosh(z) sinh(z) sin dz exp − − × 2t y t 0

(10.A.3)

(10.A.4)

234

Exotic Option Pricing and Advanced L´evy Models

√ 2 and we set ϕ(z) = (1/ 2π)e−z /2 for z ∈ IR (for related expressions in terms of Hermite functions, see Dufresne (2001) [3] and Schr¨oder (2003) [15]). t Denoting Kt = αBt + βt and Lt = 0 eαBs +βs ds with α = 0 and β ∈ IR given and ﬁxed, and using that the scaling property of B implies:

2 α y e2(Bs +νs) ds ≤ 4 0 0 (10.A.5) with t = α 2 t/4 and ν = 2β/α 2 , it follows by applying equations (10.A.3) and (10.A.4) that the random vector (Kt , Lt ) has the distribution:

t αBs +βs P αBt + βt ≤ x, e ds ≤ y = P 2(Bt + νt ) ≤ x,

t

P Kt ∈ dx, Lt ∈ dy = c(t, x, y) dx dy

(10.A.6)

where the density function c for y > 0 is given by: 2

α α2 α2 c(t, x, y) = b t, x, y 4 4 4 √ 2

2 2 β β2 1 1 2 2π x = 3/2 3 2 √ exp x− 2 t − 2 1+e + 2+ π α y t α2t α 2 2α α y

∞ 2 x/2 4e 4π z 2z dz. (10.A.7) × exp − 2 − 2 cosh(z) sinh(z) sin α t α y α2t 0 From equations (10.2.8) and (10.2.3) we see that f satisﬁes: 2

1 1 α2 α α2 f (t, s, a) = c(t, log(s), a) = b t, log(s), a s s 4 4 4

(10.A.8)

with α = σ and β = r + σ 2 /2. Hence equation (10.3.4) follows by the ﬁnal expression in equation (10.A.4).

REFERENCES [1] Barrieu, P., Rouault, A. and Yor, M. (2003), “A study of the Hartman–Watson distribution motivated by numerical problems related to Asian options pricing”, Pr´epublication PMA 813, Universit´e Pierre et Marie Curie, Paris, France. [2] Carr, P., Jarrow, R. and Myneni, R. (1992), “Alternative characterizations of American put options”, Mathematical Finance, 2, 78–106. [3] Dufresne, D. (2001), “The integral of geometric Brownian motion”, Advances in Applied Probability, 33, 223–241. [4] Feller, W. (1952), “The parabolic differential equations and the associated semi-groups of transformations”, Annals of Mathematics, 55, 468–519. [5] Hansen, A.T. and Jørgensen, P.L. (2000), “Analytical valuation of American-style Asian options”, Management Science, 46, 1116–1136. [6] Jacka, S.D. (1991), “Optimal stopping and the American put”, Mathematical Finance, 1, 1–14.

Asian Options of American Type

235

[7] Karatzas, I. and Shreve, S.E. (1998), Methods of Mathematical Finance, Springer-Verlag, Berlin, Germany. [8] Kim, I.J. (1990), “The analytic valuation of American options”, Reviews in Financial Studies, 3, 547–572. [9] Kolodner, I.I. (1956), “Free boundary problem for the heat equation with applications to problems of change of phase I. General method of solution”, Communications in Pure and Applied Mathematics, 9, 1–31. [10] Kramkov, D.O. and Mordecky, E. (1994), “Integral option”, Theory of Probability and Applications, 39, 162–173. [11] Peskir, G. (2002), “A change-of-variable formula with local time on curves”, Research Report No. 428, Department of Theoretical Statistics, University of Aarhus, Denmark (30 pp.); Journal of Theoretical Probability, to be published. [12] Peskir, G. (2005), “On the American option problem”, Research Report No. 431, Department of Theoretical Statistics, University of Aarhus, Denmark (13 pp.); Mathematical Finance, 15, 169–181. [13] Peskir, G. (2003), “The Russian option: Finite horizon”, Research Report No. 433, Department of Theoretical Statistics, University of Aarhus, Denmark (16 pp.); Finance and Stochastics, to be published. [14] Rogers, L.C.G. and Shi, Z. (1995), “The value of an Asian option”, Journal of Applied Probability, 32, 1077–1088. [15] Schr¨oder, M. (2003), “On the integral of geometric Brownian motion”, Advances in Applied Probability, 35, 159–183. [16] Shepp, L.A. and Shiryaev, A.N. (1994), “A new look at the Russian option”, Theory of Probability and Applications, 39, 103–119. [17] Shiryaev, A.N. (1978), Optimal Stopping Rules, Springer-Verlag, Berlin, Germany. [18] Shiryaev, A.N. (1999), Essentials of Stochastic Finance (Facts, Models, Theory), World Scientiﬁc, Singapore. [19] Wu, L., Kwok, Y.K. and Yu, H. (1999), “Asian options with the American early exercise feature”, International Journal of Theoretical and Applied Finance, 1, 101–111. [20] Yor, M. (1992), “On some exponential functionals of Brownian motion”, Advances in Applied Probability, 24, 509–531.

11 Why be Backward? Forward Equations for American Options† Peter Carr Bloomberg LP, New York, NY, USA and

Ali Hirsa Caspian Capital Management, LLC, New York, NY, USA Abstract The purpose of this paper is to develop forward equations for standard American options. We assume that the returns on the underlying assets have stationary independent increments, ´ or in other words, that the log price is a Levy process. In all of these models, except for Black–Scholes, the existence of a jump component implies that the backward and forward equations contain an integral in addition to the usual partial derivatives. Despite the computational complications introduced by this term, we use ﬁnite differences to solve these fundamental partial integro-differential equations (PIDEs). Our approach to determining the forward equation for American options is to start with the well-known backward equation ´ processes. In the process of and then exploit the symmetries which essentially deﬁne Levy developing the forward equation, we also determine two hybrid equations of independent interest. To illustrate that our forward PIDE is a viable alternative to the traditional backward approach, we calculate American option values in the diffusion extended VG option pricing model.

11.1 INTRODUCTION Valuing and hedging derivatives consistent with the volatility smile has been a major research focus for over a decade. A breakthrough occurred in the mid-1990s with the recognition that in certain models, European option values satisﬁed forward evolution equations in which the independent variables are the options’ strike and maturity. More speciﬁcally, Dupire (1994) showed that under deterministic carrying costs and a diffusion process for the underlying price, no arbitrage implies that European option prices satisfy a certain partial differential equation (PDE), now called the Dupire equation. Assuming that one could observe European option prices of all strikes and maturities, then this forward PDE can be used to explicitly determine the underlying’s instantaneous volatility as a function of the underlying’s price † The authors thank Dilip Madan and participants of the 2002 ICBI Barcelona conference and the 2002 Risk Boston conference. Errors are our own responsibility. The views expressed are those of the authors and do not necessarily reﬂect the position of Morgan Stanley. A shorter version of this paper was previously published in Risk, 16(1), pp. 103–107, 2003. The extended version is included here with permission of Incisive Media plc.

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

238

Exotic Option Pricing and Advanced L´evy Models

and time. Once this volatility function is known, the value function for European, American and many exotic options can be determined by a wide array of standard methods. As this value function relates theoretical prices of these instruments to the underlying’s price and time, it can also be used to determine many greeks of interest as well. Aside from their use in determining the volatility function, forward equations also serve a second useful purpose. Once one knows the volatility function, either by an explicit speciﬁcation or by a prior calibration, the forward PDE can be numerically solved to efﬁciently value a collection of European options of different strikes and maturities, all written on the same underlying asset. Furthermore, as pointed out in Andreasen (1998), all the greeks of interest satisfy the same forward PDE and hence can also be efﬁciently determined in the same way. Since the original development of forward equations for European options in continuous models, several extensions have been proposed. For example, Esser and Schlag (2002) develop forward equations for European options written on the forward price rather than the spot price. Forward equations for European options in jump diffusion models were developed in Andersen and Andreasen (1999) and extended by Andreasen and Carr (2002). It is straightforward to develop the relevant forward equations for European binary options or for European power options by differentiating or integrating the forward equation for standard European options. Buraschi and Dumas (2001) develop forward equations for compound options.∗ In contrast to the PDE’s determined by others, their evolution equation is an ordinary differential equation whose sole independent variable is the intermediate maturity date. Given the close relationship between compound options and American options, it seems plausible that there might be a forward equation for American options. The development of such an equation has important practical implications since all of the listed options on individual stocks are American-style. The Dupire equation cannot be used to infer the volatility function from market prices of American options, nor can it be used to efﬁciently value a collection of American options of differing strikes and maturities. The purpose of this paper is to develop forward equations for standard American options. This problem is addressed for American calls on stocks paying discrete dividends in Buraschi and Dumas (2001) and it is also considered in a lattice setting in Chriss (1996). We direct our attention to the more difﬁcult problem of pricing continuously exercisable American puts in continuous time models. To do so, we depart from the diffusive models which characterize most of the previous research on forward equations in continuous time. To capture the smile, we assume that prices jump rather than assuming that the instantaneous volatility is a function of stock price and time. Dumas et al. (1998) ﬁnd little empirical support for the Dupire model whereas there is a long history of empirical support for jump-diffusion models.† In particular, we assume that the returns on the underlying asset have stationary independent increments, or in other words that the log price is a L´evy process. Besides the Black and Scholes (1973) model, our framework includes as special cases the variance gamma (VG) model of Madan et al. (1998), the CGMY model of Carr et al. (2002), the ﬁnite moment logstable model of Carr and Wu (2002), the Merton (1976) and Kou (2002) jump-diffusion models, and the hyperbolic models of Eberlein et al. (1998). In all of these models, except for Black–Scholes, the existence of a jump component implies that the backward and forward equations contain an integral in addition to the usual partial derivatives. Despite the ∗ However, their deﬁnition of a compound option is nonstandard in that the critical stock price is speciﬁed in the contract. † For example, three recent papers documenting support for such models are Anderson et al. (2002), Carr et al. (2002) and Carr and Wu (2002).

Forward Equations for American Options

239

computational complications introduced by this term, we use ﬁnite differences to solve both of these fundamental partial integro differential equations (PIDEs). To illustrate that our forward PIDE is a viable alternative to the traditional backward approach, we calculate American option values in the diffusion extended VG‡ option pricing model and ﬁnd very close agreement. Our approach to determining the forward equation for American options is to start with the well-known backward equation and then exploit the symmetries which essentially deﬁne L´evy processes. In the process of developing the forward equation, we also determine two hybrid equations of independent interest. The advantage of these hybrid equations over the forward equation is that they hold in greater generality. Depending on the problem at hand, these hybrid equations can also have large computational advantages over the backward or forward equations when the model has already been calibrated. In particular, the advantage of these hybrid equations over the backward equation is that they are more computationally efﬁcient when one is interested in the variation of prices or greeks across strike or maturity at a ﬁxed time, e.g. market close. The ﬁrst of these hybrid equations has the stock price and maturity as independent variables. The numerical solution of this hybrid equation is an alternative to the backward equation in producing a spot slide, which shows how American option prices vary with the initial spot price of the underlying. If one is interested in understanding how this spot slide varies with maturity, then our hybrid equation is much more efﬁcient than the backward equation. Our second hybrid equation has the strike price and calendar time as independent variables. The numerical solution of this hybrid equation is an alternative to the forward equation in producing an implied volatility smile at a ﬁxed maturity. If one is interested in understanding how the model predicts that this smile will change over time, then our hybrid equation is much more computationally efﬁcient than the forward equation. This second hybrid equation also allows parameters to have a term structure, whereas our forward equation does not.§ Hence, if one needs to efﬁciently value a collection of American options of different strikes in the time-dependent Black–Scholes model, then it is far more efﬁcient to solve our hybrid equation than to use the standard backward equation. The remainder of this paper is structured as follows. The next section introduces our setting and reviews the backward PIDE which governs American option values in this setting. The following section develops the ﬁrst hybrid equation, while the subsequent section develops the second one. The penultimate section develops the forward equation for American options, while the ﬁnal section summarizes and suggests further research.

11.2 REVIEW OF THE BACKWARD FREE BOUNDARY PROBLEM Throughout this article, we focus on (standard) American puts on stocks leaving American calls and other underlyings as an exercise for the reader. We assume perfect capital markets, continuous trading, no arbitrage opportunities, continuous dividend payments and Markovian stock price dynamics under all martingale measures. We further assume that ‡

For details on the use of ﬁnite differences for solving the backward PIDE for American options in the VG model, see Hirsa and Madan (2003). § Note, however, that implied volatility can have a term or strike structure in our L´ evy setting.

240

Exotic Option Pricing and Advanced L´evy Models

the spot interest rate and dividend yield are given by deterministic functions r(t) > 0 and q(t) ≥ 0, respectively. Thus, we assume that under a risk-neutral measure Q, the stock price st satisﬁes the following stochastic differential equation: dst = [r(t) − q(t)]st− dt + σ (st− , t)st− dWt ∞ + st− (ex − 1)[µ(dx, dt) − ν(st− , x, t) dx dt], −∞

(11.1)

for all t ∈ [0, T ]. Thus, the change in the stock price decomposes into three parts. The ﬁrst part is the risk-neutral drift, comprised entirely of the dollar carrying cost of the stock. The second part is the diffusion part, expressed in terms of the instantaneous volatility function σ (S, t). As usual, the term dWt denotes increments of a standard Wiener process deﬁned on the time set [0, T ] and on a complete probability space (, F , Q). The third part is the jump part. The random measure µ(dx, dt) counts the number of jumps of size x in the log price at time t. The Hunt density t ∞ {ν(S, x, t), S > 0, x ∈ , t ∈ [0, T ]} is used to compensate the jump process Jt ≡ 0 −∞ st− (ex − 1)µ(dx, ds), so that the last term in equation (11.1) is the increment of a Q jump martingale.¶ The jump martingale is speciﬁed in such a way that jumps to negative prices are impossible. Since the last two parts are both martingales, we have: t E Q [st |s0 ] = s0 e 0 [r(u)−q(u)]du , where the initial stock price s0 is positive. Consider an American put option on the stock with a ﬁxed strike price K0 > 0 and a ﬁxed maturity date T0 ∈ [0, T ]. Let pt denote the value of the American put at time t ∈ [0, T0 ]. In this general setup, it is not yet known whether the American put value is monotone in S. Hence, we further assume whatever sufﬁcient conditions on the coefﬁcients that are needed so that the put value is monotone in S. Then, for each time t ∈ [0, T0 ], there exists a unique critical stock price, s(t), below which the American put should be exercised early, i.e. if st ≤ s t , then pt = max[0, K0 − st ]

(11.2)

and if st > s t , then pt > max[0, K0 − st ].

(11.3)

The exercise boundary is the time path of critical stock prices, s t , t ∈ [0, T0 ]. This boundary is independent of the current stock price s0 and is bounded above by K0 . It is a smooth, nondecreasing function of time t whose terminal limit is: r(T0 ) . lim s t = K0 min 1, t↑T0 q(T0 ) Right at expiration, the critical stock price is the strike price, i.e. s T0 = K0 . Hence, when q(T0 ) > r(T0 ), there is a discontinuity in the exercise boundary. Figure 11.1 plots the ¶

The function ν(S, x, t) must have the following properties: ν(S, 0, t) = 0,

∞

−∞

(x 2 ∧ 1)ν(S, x, t) dx < ∞.

Forward Equations for American Options

241

Exercise boundary 1 0.9 0.8

Calendar time

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 50

60

70

80 Stock price

90

100

110

Figure 11.1 Exercise boundary in the DEVG model. Critical stock prices are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400 with initial price S0 = 100

exercise boundary in the Diffusion Extended Variance Gamma (DEVG) model. This model extends the pure jump Variance Gamma model of Madan et al. (1998), by adding a diffusion component with constant volatility. The American put value is also a function, denoted p(s, t), mapping its domain D ≡ (s, t) ∈ [0, ∞) × [0, T0 ] into the nonnegative real line. The exercise boundary, s t , t ∈ [0, T0 ], divides this domain D into a stopping region S ≡ [0, s t ] × [0, T0 ] and a continuation region C ≡ (s t , ∞) × [0, T0 ]. Equation (11.2) indicates that in the stopping region, the put value function p(s, t) equals its exercise value, max[0, K0 − S]. In contrast, the inequality expressed in equation (11.3) shows that in the continuation region, the put is worth more ‘alive’ than ‘dead’. The transition between boundaries is smooth in the following sense: lim p(s, t) = K0 − s t ,

t ∈ [0, T0 ]

(11.4)

∂p(s,t) ∂s

t ∈ [0, T0 ].

(11.5)

s↓s t

lim

s↓s t

= −1,

The value matching condition (equation (11.4)) and equation (11.2) imply that the put value is continuous across the exercise boundary. Furthermore, the high contact condition (equation (11.5) and equation (11.2)) further imply that the put’s delta is continuous. Equations (11.4) and (11.5) are jointly referred to as the ‘smooth ﬁt’ conditions.

242

Exotic Option Pricing and Advanced L´evy Models

∂p The partial derivatives, ∂p ∂t , ∂s , and differential equation (PIDE):

∂2p ∂s 2

exist and satisfy the following partial integro

∂p(s, t) σ 2 (s, t)s 2 ∂ 2 p(s, t) ∂p(s, t) + − r(t)p(s, t) + [r(t) − q(t)]s 2 ∂t 2 ∂s ∂s ∞ ∂ p(sex , t) − p(s, t) − p(s, t)s(ex − 1) ν(s, x, t) dx + ∂s −∞

+ 1(s < s t ) r(t)K0 − q(t)s −

∞

6 [p(se , t) − (K0 − se )]ν(s, x, t) dx x

x

= 0.

ln(s t /s)

(11.6) The last term on the left-hand side (LHS) of equation (11.6) is the result of applying the integro-differential operator deﬁned by the ﬁrst two lines to the value p(s, t) = K0 − s holding in the stopping region. The American put value function p(s, t) and the exercise boundary s t jointly solve a backward free boundary problem (FBP), consisting of the backward PIDE (equation (11.6)), the smooth ﬁt conditions (equations (11.4) and (11.5)), and the following boundary conditions: p(s, T0 ) = max[0, K0 − s],

s>0

(11.7)

lim p(s, t) = 0,

t ∈ [0, T0 ]

(11.8)

lim p(s, t) = K0 ,

t ∈ [0, T0 ].

(11.9)

s↑∞ s↓0

These Dirichlet conditions force the American put value to its exercise value along the boundaries. As the efﬁcient implementation of a ﬁnite difference scheme usually requires the use of positive ﬁnite spatial boundaries, our implementation replaces the last two conditions in the target problem by: lim pss (s, t) = 0,

t ∈ [0, T0 ]

(11.10)

lim pss (s, t) = 0,

t ∈ [0, T0 ].

(11.11)

s↑∞ s↓0

Hence, the put gamma is forced to zero along the spatial boundaries. Numerical experimentation suggests that imposition of the zero gamma condition on positive ﬁnite spatial boundaries tends to work better than imposing the Dirichlet conditions. The solution to this alternative speciﬁcation is unique under the further condition that it be continuous along the entire boundary. Figure 11.2 plots American put values in the DEVG model against stock price and time.

11.3 STATIONARITY AND DOMAIN EXTENSION IN THE MATURITY DIRECTION The last section assumed that the strike K and maturity T were ﬁxed at K0 and T0 , respectively. To derive a hybrid FBP for American put values, we ﬁrst extend the domain of the problem to all T ∈ [0, T ], keeping the strike price K ﬁxed at K0 .

Forward Equations for American Options

243

100

American put prices

80 60 40 20 0 1 0.8

400 0.6

Calendar time

300 0.4

200 0.2

100 0

Stock prices

0

Figure 11.2 American put values in the DEVG model against stock price and calendar time. American put values are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400. The value of the American put at the initial stock price of S0 = 100 is $23.9875

Note that the exercise boundary depends on t, r(t), q(t), σ (S, t), ν(S, x, t), T and K0 , but not on s. Suppressing the dependence on r(t), q(t), σ (S, t), ν(S, x, t) and K0 , let s(t; T ) be the function relating the exercise surface to t and T : s t = s(t; T ),

t ∈ [0, T ], T ∈ [0, T ].

The extended continuation region is a three-dimensional region denoted by . This can be pictured as stacking the two-dimensional continuation regions up the Z-axis as T increases from 0. For each T ∈ [0, T ], the union of the two-dimensional continuation region and the two-dimensional stopping region is the plane S > 0, t ∈ [0, T ]. As T increases from zero, the area covered by this plane increases. Thus, the extended domain for the backward PIDE is the wedge S > 0, t ∈ [0, T ], T ∈ [0, T ]. We note that the backward PIDE of the last section holds on this wedge with T0 replaced by T . Let (s, t; T ) be the function solving this backward PIDE: ∂(s, t; T ) σ 2 (s, t)s 2 ∂ 2 (s, t; T ) ∂(s, t; T ) + − r(t)(s, t; T ) + [r(t) − q(t)]s 2 ∂t 2 ∂s ∂s ∞ ∂ (sex , t; T ) − (s, t; T ) − (s, t; T )s(ex − 1) ν(s, x, t) dx + ∂s −∞

244

Exotic Option Pricing and Advanced L´evy Models

+ 1(s < s(t; T )) r(t)K0 − q(t)s − 5 − (K0 − sex )]ν(s, x, t) dx = 0.

∞

[(sex , t; T )

ln(s(t;T )/s)

(11.12)

Now suppose stationarity, i.e. that r(t), q(t), σ (S, t) and ν(S, x, t) are all independent of time t. It follows that the time derivative is just the negative of the maturity derivative: ∂ ∂ (s, t; T ) = − (s, t; T ). ∂t ∂T

(11.13)

Dropping the dependence of r(t), q(t), σ (S, t) and ν(S, x, t) on t and substituting equation (11.13) into equation (11.12) implies that the following relation holds in the extended domain: ∂(s, t; T ) ∂(s, t; T ) σ 2 (s)s 2 ∂ 2 (s, t; T ) + − r(s, t; T ) + (r − q)s ∂T 2 ∂s 2 ∂s ∞ ∂ x x + (se , t; T ) − (s, t; T ) − (s, t; T )s(e − 1) ν(s, x) dx ∂s

−

−∞

+ 1(s < s(t; T )) rK0 − qs −

∞

5 [(sex , t; T ) − (K0 − sex )]ν(s, x) dx = 0.

ln(s(t;T )/s)

(11.14) We note that one can ﬁx t at t0 and just solve the above problem in the s, T plane if desired. In this case, the initial condition is: (s, t0 ; t0 ) = max[0, K0 − s],

s > 0.

(11.15)

The Dirichlet boundary conditions are: lim (s, t0 ; T ) = 0,

s↑∞

lim (s, t0 ; T ) ∼ K0 − s, s↓0

T ∈ [t0 , T ] T ∈ [t0 , T ].

(11.16) (11.17)

Alternatively, these Dirichlet conditions can be replaced by the following zero gamma conditions: lim ss (s, t0 ; T ) = 0,

T ∈ [t0 , T ]

(11.18)

lim ss (s, t0 ; T ) = 0,

T ∈ [t0 , T ].

(11.19)

s↑∞ s↓0

The smooth ﬁt conditions are: lim (s, t0 , T ) = K0 − s(t0 ; T ),

s↓s(t0 ;T )

lim

s↓s(t0 ;T )

∂(s, t0 ; T ) = −1, ∂s

T ∈ [t0 , T ]

T ∈ [t0 , T ].

(11.20) (11.21)

Figure 11.3 plots American put values in the DEVG model against stock price and maturity.

Forward Equations for American Options

245

American put prices

100 80 60 40 20 0 1 0.8

400 0.6

300 0.4

Maturity

200 0.2

100 0

Stock prices

0

Figure 11.3 American put values in the DEVG model against stock price and maturity. American put values are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400

11.4 ADDITIVITY AND DOMAIN EXTENSION IN THE STRIKE DIRECTION The last section assumed that the strike K was ﬁxed at K0 and that r(t), q(t), σ (S, t) and ν(S, x, t) are all independent of time t. To derive a new hybrid PIDE for American put values, we further extend the domain of the problem to all K > 0. We also restore the dependence on t of r(t), q(t), σ (S, t) and ν(S, x, t). On this larger domain, let s(t; T , K) be the function relating the exercise surface to t, T , and K: s t = s(t; T , K),

t ∈ [0, T ],

T ∈ [0, T ],

K > 0.

We note that the backward PIDE (equation (11.12)) holding on the three-dimensional domain of the last section holds on the larger four-dimensional domain with K0 replaced by all K > 0. Let (s, t; K, T ) be the function solving this backward PIDE on the extended four-dimensional domain: ∂(s, t; K, T ) σ 2 (s, t)s 2 ∂ 2 (s, t; K, T ) + [r(t) − q(t)] + ∂t 2 ∂s 2 ∂(s, t; K, T ) − r(t)(s, t; K, T ) ×s ∂s

246

Exotic Option Pricing and Advanced L´evy Models

∞ ∂ x x (se , t; K, T ) − (s, t; K, T ) − (s, t; K, T )s(e − 1) ν(s, x, t) dx + ∂s −∞

+ 1(s < s(t; T , K)) r(t)K − q(t)s −

∞

[(sex , t; K, T )

ln(s(t;T ,K)/s)

5

− (K − se )]ν(s, x, t) dx = 0. x

(11.22)

We now assume that the log price process has independent increments, i.e. is additive or equivalently that σ (S, t) and ν(S, x, t) are both independent of the stock price S. Then, for each ﬁxed t and T , the exercise boundary is a linearly homogeneous function of the strike price: s(t; T , λK) = λs(t; T , K), for all λ ≥ 0. Setting λ =

1 K

implies that: s(t; T , K) = Ks(t; T , 1).

(11.23)

For each ﬁxed s, t and T , the condition s > s(t; T , K) is thus equivalent to the condition sK K < s(t;Ts ,1) = s(t;T ,K) ≡ K(s, t; T ). We refer to the output of this function as the critical strike price. For each ﬁxed s, t and T , the critical strike price is the lowest strike price K at which the put is exercised early. Note that the critical strike price depends on s but is independent of K. For an American put, the critical strike price is bounded above by s. In addition, note that the geometric mean of the two critical prices is just the geometric mean of the stock price and strike price: √ s(t; T , K)K(s, t; T ) = sK. (11.24) The additivity of the log price process implies that the function (s, t; K, T ) is linearly homogeneous in s and K. It follows from Euler’s theorem that: (s, t, K, T ) = s

∂ ∂ (s, t; K, T ) + K (s, t; K, T ). ∂s ∂K

(11.25)

Differentiation with respect to s and K and some obvious algebra establishes that: s2

2 ∂2 2 ∂ (s, t; K, T ) = K (s, t; K, T ). ∂s 2 ∂K 2

(11.26)

Dropping the dependence of σ (S, t) and ν(S, x, t) on S and substituting equations (11.25) and (11.26) into equation (11.22) implies: ∂(s, t; K, T ) σ 2 (t)K 2 ∂ 2 (s, t; K, T ) + ∂t 2 ∂K 2 ∂(s, t; K, T ) − q(t)(s, t; K, T ) − [r(t) − q(t)]K ∂K

Forward Equations for American Options

∞ +

(s, t; Ke

−x

−∞

247

∂ −x (s, t; K, T )K(e − 1) ex ν(x, t) dx , T ) − (s, t; K, T ) − ∂K

+ 1(k > k(s, t; T )) r(t)K − q(t)s −

∞

[(s, t; Ke−x , T )

ln(k(s,t;T )/K)

− (Ke

−x

5

− s)]e ν(x, t) dx = 0. x

(11.27)

We note that one can ﬁx s and T at say s0 and T0 and just solve the above problem in the K, t plane if desired. In this case, the terminal condition is: (s0 , T0 ; K, T0 ) = max[0, K − s0 ],

K > 0.

(11.28)

t ∈ [0, T0 ]

(11.29)

The Dirichlet boundary conditions are: lim (s0 , t; K, T0 ) =∼ K − s0 ,

K↑∞

lim (s0 , t; K, T0 ) = 0,

K↓0

t ∈ [0, T0 ].

(11.30)

Alternatively, these Dirichlet conditions can be replaced by: lim kk (s0 , t; KT0 ) = 0,

t ∈ [0, T0 ]

(11.31)

lim kk (s0 , t; K, T0 ) = 0,

t ∈ [0, T0 ].

(11.32)

K↑∞ K↓0

The smooth ﬁt conditions are: lim K↑K(s,t;T0 )

(s0 , t; K, T0 ) = K(s0 , t; T0 ) − s0 , lim K↑K(s,t;T0 )

∂(s0 ,t;K,T0 ) ∂K

= 1,

t ∈ [0, T0 ]

(11.33)

t ∈ [0, T0 ].

(11.34)

Figure 11.4 plots American put values in the DEVG model against strike price and calendar time. We note that setting jumps to zero reduces the PIDE to a PDE arising in the special case of the time-dependent Black–Scholes model. If one wishes to value American options in this model for multiple strikes and maturities and with ﬁxed time and spot, it is much more efﬁcient to solve the hybrid problem of this section once for each T than it is to solve the usual backward problem once for each K and once for each T , as is usually done.

11.5 THE FORWARD FREE BOUNDARY PROBLEM We now assume that we have both stationarity and additivity. In other words, the log price is a L´evy process and r(t), q(t), σ (S, t) and ν(S, x, t) are all independent of both time t

248

Exotic Option Pricing and Advanced L´evy Models

300

American put prices

250 200 150 100 50 0 1 0.8

400 0.6

300 0.4

Calendar time

200 0.2

100 0

Strike

0

Figure 11.4 American put values in the DEVG model against stock price and calendar time. Critical stock prices are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400. The value of the American put at the initial stock price of S0 = 100 is $23.9875

and the stock price S. Stationarity implies that the function (s, t; K, T ) depends on t and T only through T − t. It thus follows that: ∂ ∂ (s, t; K, T ) = − (s, t; K, T ). ∂t ∂T

(11.35)

Substituting equation (11.35) into equation (11.27) implies:

−

∂(s, t; K, T ) σ 2 K 2 ∂ 2 (s, t; K, T ) ∂(s, t; K, T , ) + − q(s, t; K, T ) − (r − q)K ∂T 2 ∂K 2 ∂K ∞ ∂ −x −x (s, t; Ke , T ) − (s, t; K, T ) − (s, t; K, T )K(e − 1) ex ν(x) dx + ∂K −∞

+ 1(k > k(s, t; T )) rK − qs −

∞

[(s, t; Ke−x , T )

ln(k(s,t;T )/K)

− (Ke

−x

5

− s)]e ν(x) dx = 0. x

(11.36)

Forward Equations for American Options

249

We note that one can ﬁx s and t at say s0 and t0 and just solve the above problem in the K, T plane if desired. In this case, the initial condition is: (s0 , t0 ; K, t0 ) = max[0, K − s0 ],

K > 0.

(11.37)

T ∈ [t0 , T ]

(11.38)

The Dirichlet boundary conditions are: lim (s0 , t0 ; K, T ) ∼ K − S0 ,

K↑∞

lim (s0 , t0 ; K, T ) = 0,

K↓0

T ∈ [t0 , T ].

(11.39)

Alternatively, these Dirichlet conditions can be replaced by: lim kk (s0 , t0 ; K, T ) = 0,

T ∈ [t0 , T ]

(11.40)

lim kk (s0 , t0 ; K, T ) = 0,

T ∈ [t0 , T ].

(11.41)

K↑∞ K↓0

The smooth ﬁt conditions are: lim

K↑K(s,t0 ;T )

(s0 , t0 ; K, T ) = K(s0 , t0 ; T ) − s0 , lim

K↑K(s,t0 ;T )

∂(s0 ,t0 ;K,T ) ∂K

= 1,

T ∈ [t0 , T ]

(11.42)

T ∈ [t0 , T ].

(11.43)

300

American put prices

250 200 150 100 50 0 1 0.8

400 0.6

300 0.4

Maturity

200 0.2

100 0

Strike

0

Figure 11.5 American put values in the DEVG model against stock price and maturity. American put values are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400. The value of the American put at the initial stock price of S0 = 100 is $23.9785

250

Exotic Option Pricing and Advanced L´evy Models Critical strike boundary 1 0.9 0.8 0.7

Maturity

0.6 0.5 0.4 0.3 0.2 0.1 0 100

110

120

130

140

150

160

170

Strike

Figure 11.6 Critical strike boundary in the DEVG model. Critical strike prices are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400

Figure 11.5 plots American put values in the DEVG model against strike price and maturity. The value of the American put at the initial stock price of S0 = 100 is $23.9875 from the backward problem and $23.9785 from the forward problem. The small difference is due to numerical error since the difference gets even smaller as we increase the number of time and spatial steps. Figure 11.6 plots critical strike prices against maturity using the same inputs.

11.6 SUMMARY AND FUTURE RESEARCH We ﬁrst reviewed the backward PIDE governing the arbitrage-free price of an American put option when the underlying spot price process is Markov in itself. By imposing various restrictions on the process, we then derived three new PIDEs for American put values. In particular, by assuming stationarity, we derived a forward PIDE in maturities with spot price still an independent variable. By alternatively assuming that the evolution coefﬁcients for the proportional process are independent of spot, we derived a backward PIDE with the strike price as an independent variable. Finally, by assuming that the log price of the underlying is a L´evy process, we derived the forward PIDE for arbitrage-free American put values. We numerically solved this forward PIDE for the case of the diffusion extended VG model and found very close agreement to the numerical solution of the backward PIDE. A longer version of this paper, downloadable from

Forward Equations for American Options

251

www.math.nyu.edu/research/carrp/papers/pdf, contains an appendix detailing the ﬁnite difference scheme used to numerically solve the forward PIDE for American put options. It is clear how to apply our analysis to American calls or more generally to payoffs which are both monotone and linearly homogeneous in spot and strike. It should be possible to extend our analysis to barrier options in which the payoff is linearly homogeneous in some subset of spot, strike, barrier, or rebate. An open problem is the forward equation for American options when the evolution parameters depend on stock price and/or time. It would also be interesting to extend our univariate approach to additional state variables besides the stock price. If the extra state variable is another asset price, then bivariate American options could be handled. If the extra state variable is a path statistic, then many path-dependent options could be handled. If the extra state variable is the current level of a randomly evolving volatility process, then our approach would encompass stochastic volatility and GARCH models for which there is considerable empirical support. In the interests of brevity, we defer this research to future work.

APPENDIX: DISCRETIZATION OF FORWARD EQUATION FOR AMERICAN OPTIONS This appendix shows how ﬁnite differences can be used to numerically solve the following forward PIDE governing American put values: ∂P (s, t; K, T ) σ 2 2 ∂ 2 P (s, t; K, T ) ∂P (s, t; K, T ) + (r − q)K − K + qP (s, t; K, T ) (11.44) ∂T 2 ∂K 2 ∂K +∞ ∂P (s, t; K, T ) − K(e−y − 1) ey ν(y) dy (11.45) P (s, t; Ke−y , T ) − P (s, t; K, T ) − ∂K −∞ 5 ∞ P (s, t; Ke−y , T ) − (Ke−y − s) ey ν(y) dy = 0 − 1K>K(s,t;T ) rK − qs − ln(K/K(s,t;T ))

(11.46)

We illustrate the solution in the diffusion extended VG model for which the L´evy density has the form: 2 θ2 2 exp(θy/σ ) ν + σ2 |y| . (11.47) ν(y) dy = exp − ν |y| σ Notice that this L´evy density explodes as y approaches zero from either direction. As a result, special measures will have to be taken when approximating the integral containing this L´evy density. One can show that:

+∞ −∞

where: ω≡

(e−y − 1)ey ν(y) dy = ω

1 ln(1 − θ ν − σ 2 ν/2). ν

(11.48)

252

Exotic Option Pricing and Advanced L´evy Models

Dropping the arguments s and t to simplify notation, we can rewrite equation (11.46) as: ∂P (K, T ) σ 2 2 ∂ 2 P (K, T ) ∂P (K, T ) − K + qP (K, T ) + (r − q + ω)K 2 ∂T 2 ∂K ∂K +∞ − (P (Ke−y , T ) − P (K, T ) ey ν(y) dy −∞

− 1K>K(T ) rK − qs −

∞

P (Ke

−y

, T ) − (Ke

−y

5

− s) e ν(y) dy = 0 y

ln(K/K(T ))

By making the change of variable x = ln K we have p(x, T ) = P (K, T ), ∂p ∂P (x, T ) = K (K, T ), ∂x ∂K 2 ∂ 2p ∂p 2∂ P (x, T ) = K (x, T ) − (K, T ), ∂x 2 ∂x ∂K 2 p(x − y, T ) = P (Ke−y , T ),

and hence we obtain the following PIDE for p(x, T ), ∂p σ2 σ 2 ∂ 2 p(x, T ) ∂p + (r − q + (x, T ) − + ω) (x, T ) + qp(x, T ) 2 ∂T 2 ∂x 2 ∂x +∞ − (p(x − y, T ) − p(x, T )) ν˜ (y) dy −∞

− 1x>x(T ) re − qs − x

∞

p(x − y, T ) − (e

x−y

5

− s) ν(y) ˜ dy = 0,

x−x(T )

where:

˜

˜

e −λ p y e−λn |y| 1y>0 + 1yx(Tj ) rK − qe −

∞

xi

xi −x(Tj )

p(xi − y, Tj ) − (K − exi +y ) ν(y) ˜ dy

6 = 0.

Equivalently, we have: (−B − A)pi−1,j +1 + (1 + 2B + qT )pi,j +1 + (−B + A)pi+1,j +1 = +∞ p(xi − y, Tj ) − pi,j ) ν(y) ˜ dy pi,j + T −∞

+ T × 1xi >x(Tj ) re − qs −

∞

xi

xi −x(Tj )

p(xi − y, Tj ) − (e

xi −y

6

− s) ν(y) ˜ dy , (11.52)

T σ2 +ω , A= r −q + 2 2x

where:

σ 2 T , 2 x 2 = (exi − s)+ ,

B= pi,0

x(T0 ) = ln s, and

x(Tj ) = min{xi : pi,j − (exi − s)+ < 0} for j = 1, . . . , M. xi

254

Exotic Option Pricing and Advanced L´evy Models

For the ﬁrst integral on the right-hand side of equation (11.52), we decompose the range of integration into six parts:

+∞

−∞

p(xi − y, Tj ) − pi,j ν˜ (y) dy =

xi −xN

−∞

+

p(xi − y, Tj ) − pi,j ν˜ (y) dy

−x

p(xi − y, Tj ) − pi,j ν(y) ˜ dy

xi −xN

+

0

−x

+x

+

p(xi − y, Tj ) − pi,j ν˜ (y) dy

p(xi − y, Tj ) − pi,j ν˜ (y) dy

0

+

xi −x0

+x

+

+∞

xi −x0

p(xi − y, Tj ) − pi,j ν(y) ˜ dy

p(xi − y, Tj ) − pi,j ν(y) ˜ dy

The six integrals are evaluated as:

0 −x

x 0

p(xi − y, Tj ) − pi,j ν˜ (y) dy ∼ =

p(xi − y, Tj ) − pi,j ν(y) ˜ dy ∼ = −x

xi −xN

=

1 νx λ˜ n 1 νx λ˜ p

˜

(1 − e−λn x )(pi+1,j − pi,j ), ˜

(1 − e−λp x )(pi−1,j − pi,j ).

p(xi − y, Tj ) − pi,j ν˜ (y) dy

N−i−1 1 pi+k,j − pi,j − k(pi+k+1,j − pi+k,j ) ν k=1 " ! × expint(kx λ˜ n ) − expint((k + 1)x λ˜ n )

+

1

N−i−1

λ˜ n νx

k=1

pi+k+1,j − pi+k,j

˜

˜

e−λn kx − e−λn (k+1)x

where:

∞

expint(x) ≡ x

e−t dt t

(11.53)

Forward Equations for American Options

255

is the exponential integral.

xi −x0

p(xi − y, Tj ) − pi,j ν˜ (y) dy

x

=

i−1 , 1 pi−k,j − pi,j − k(pi−k−1,j − pi−k,j ) expint(kxλp ) − expint((k + 1)xλp ) ν k=1

+

i−1 pi−k−1,j − pi−k,j −λp kx e − e−λp (k+1)x . λp νx k=1

xi −xN

−∞

exi expint((N − i)x(λ˜ n − 1)) p(xi − y, Tj ) − pi,j ν(y) ˜ dy = ν −

∞ xi −x0

s + pi,j expint((N − i)x λ˜ n ). ν

1 p(xi − y, Tj ) − pi,j ν(y) ˜ dy = − pi,j expint(ixλp ). ν

The integral inside the Heaviside term in equation (11.52) is treated in the same manner as the other integral. Therefore, we have:

∞

xi −x(Tj )

=

p(xi − y, Tj ) − (exi −y − s) ν˜ (y) dy

i−1 1 pi−k,j − k(pi−k−1,j − pi−k,j ) expint(kx λ˜ p ) − expint((k + 1)x λ˜ p ) ν k=i−l

+

i−1 pi−k−1,j − pi−k,j −λ˜ p kx ˜ e − e−λp (k+1)x λ˜ p νx

k=i−l

−

" 1 ! xi e expint((i − l)x(λ˜ p + 1)) − s expint((i − l)x λ˜ p ) ν

Difference equation Putting all of the pieces together, we obtain the following difference equation at the point (xi , Tj +1 ) Epi−1,j +1 + Fpi,j +1 + Gpi+1,j +1 = pi,j +

T Ri,j + T 1xi >x(Tj ) Hi,j ν

(11.54)

256

Exotic Option Pricing and Advanced L´evy Models

where E = −A − B − Bp , F = 1 + qT + 2B + Bn + Bp +

T (expint(ix λ˜ p ) + expint((N − i)x λ˜ n )), ν

G = A − B − Bn ,

Ri,j =

N−i−1

pi+k,j − pi,j − k(pi+k+1,j − pi+k,j )

k=1

× {expint(kxλn ) − expint((k + 1)xλn )} +

N−i−1 k=1

+

i−1

pi+k+1,j − pi+k,j −λ˜ n kx ˜ (e − e−λn (k+1)x ) λn x

(pi−k,j − pi,j − k(pi−k−1,j − pi−k,j ))

k=1

, × expint(kxλp ) − expint((k + 1)xλp ) +

i−1 pi−k−1,j − pi−k,j −λ˜ p kx ˜ (e − e−λp (k+1)x ) λp x k=1

+ exi expint((N − i)x(λn − 1)) − s expint((N − i)xλn ), Hi,j = rexi − qs −

−

i−1 1 pi−k,j − k(pi−k−1,j − pi−k,j ) expint(kxλp ) − expint((k + 1)xλp ) ν k=i−l i−1 pi−k−1,j − pi−k,j −λ˜ p kx ˜ (e − e−λp (k+1)x ) λp νx

k=i−l

+

1 , xi e expint((i − l)x(λp + 1)) − s expint((i − l)xλp ) , ν

and Bn =

T ˜ (1 − e−λn x ), ˜ νx λn

Bp =

T ˜ (1 − e−λp x ). νx λ˜ p

The initial condition (equation (11.49)) and boundary conditions (equations (11.50) and (11.51)) are discretized in the usual manner. A standard ﬁnite difference solver can then be used to solve the boundary value problem.

Forward Equations for American Options

257

REFERENCES [1] Andersen, L. and Andreasen, J. (1999), “Jumping smiles,” Risk, 12(11), 65–68. [2] Anderson, T.G. Benzoni, L. and Lund, J. (2002), “An empirical investigation of continuous-time equity return models,” Journal of Finance, 57, 1239–1284. [3] Andreasen, J. (1998), “Implied modelling, stable implementation, hedging, and duality,” Working Paper, University of Aarhus, Aarhus, Denmark. [4] Andreasen, J. and Carr, P. (2002), “Put call reversal,” Working Paper, New York University, New York, USA. [5] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities,” Journal of Political Economy, 81, 637–654. [6] Buraschi, A. and Dumas, B. (2001), “The forward valuation of compound options,” Journal of Derivatives, 9(1), 8–17. [7] Carr, P., Geman, H. Madan, D. and Yor, M. (2002), “The ﬁne structure of asset returns: An empirical investigation, Journal of Business, 75, 305–332. [8] Carr, P.P. and Wu, L. (2002), “The ﬁnite moment logstable process and option pricing,” Working Paper, New York University, New York, USA. [9] Chriss, N. (1996), “Transatlantic trees,” Risk, 9(7), 45–48. [10] Dumas, B., Fleming, J. and Whaley, R. (1998), “Implied volatilities: empirical tests,” Journal of Finance, 53, 2059–2106. [11] Dupire, B. (1994), “Pricing with a smile,” Risk, 7(1), 18–20. [12] Eberlein, E., Keller, U. and Prause, K. (1998), “New insights into smile, mispricing, and value at risk: the hyperbolic model,” Journal of Business, 71, 371–406. [13] Esser, A. and Schlag, C. (2002), “A note on forward and backward partial differential equations for derivative contracts with forwards as underlyings,” in J. Hakala and U. Wystup (Eds), Foreign Exchange Risk, Risk Publications, London, pp. 115–124. [14] Hirsa, A. and Madan, D.B. (2003), “Pricing American options under Variance Gamma,” The Journal of Computational Finance, 7(2). [15] Kou, S.G. (2002), “A jump-diffusion model for option pricing,” Management Science, 48, 1086– 1101. [16] Madan, D.B., Carr, P.P. and Chang, E. (1998), “The variance gamma process and option pricing,” European Financial Review, 2, 79–105. [17] Merton, R.C. (1976), “Option pricing when underlying stock returns are discontinuous,” Journal of Financial Economics, 3, 125–144.

12 Numerical Valuation of American Options Under the CGMY Process Ariel Almendral Norwegian Computing Center, Oslo, Norway Abstract American put options written on an underlying stock following a Carr–Madan–Geman–Yor (CGMY) process are considered. It is known that American option prices satisfy a Partial Integro-Differential Equation (PIDE) on a moving domain. These equations are reformulated as a Linear Complementarity Problem, and solved iteratively by an implicit–explicit type of iteration based on a convenient splitting of the Integro-Differential operator. The solution to the discrete complementarity problems is found by the Brennan–Schwartz algorithm and computations are accelerated by the Fast Fourier Transform. The method is illustrated throughout a series of numerical experiments.

12.1 INTRODUCTION In this paper, we propose a numerical method to compute American put options, when the underlying asset is modeled by the Carr–Madan–Geman–Yor (CGMY) process considered in Carr et al. (2002) [8]. Our contribution is to show experimentally that the implicit– explicit method proposed in Cont and Voltchkova (2003) [12] for European options may be successfully applied to the computation of American options under L´evy models. A similar splitting was already proposed in Hirsa and Madan (2004) [13] for the computation of the American price under the Variance Gamma (VG) process (see also, Anon (2004) [3]). Matache et al. (2003) [17] have previously studied the American pricing problem under the CGMY process. They considered a variational inequality formulation combined with a convenient wavelet basis to compress the stiffness matrix. The approach here is different: we essentially work with a formulation as a Linear Complementarity Problem (LCP), and use standard ﬁnite differences. To deal with the singularity of the jump measure at the origin, we ﬁrst approximate the problem by another problem, where small jumps are substituted by a small Brownian component. Next, we solve the approximated problem iteratively, where for each time step one needs to solve tridiagonal linear complementarity problems. The Fast Fourier Transform (FFT) plays also an important role when computing the convolution integrals fast. The sequence of linear complementarity problems are solved with the help of a simple algorithm proposed by Brennan and Schwartz (1977) [6], that works well for the particular case of a put option. We have also veriﬁed numerically the recent results in Alili and Kyprianou (2004) [1] on the smooth-ﬁt principle for general L´evy processes. Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

260

Exotic Option Pricing and Advanced L´evy Models

A statistical study of ﬁnancial time series in Carr et al. (2002) [8] shows that the diffusion component could in most cases be neglected, provided that the remaining part of the process is of inﬁnite activity and ﬁnite variation. We concentrate precisely on the ﬁnite variation case, but also allow for a diffusion component, that may be safely omitted without affecting the pricing algorithm. In Section 12.2, we brieﬂy introduce the CGMY process, the European and American put option problem, and the related PIDEs. For further information on L´evy processes in ﬁnance, we refer the reader to the books by Cont and Tankov (2004) [11] and Schoutens (2003) [20]. An approximation to the equation with a discretization by ﬁnite differences is exposed in Section 12.3 and numerical results are presented in Section 12.4.

´ 12.2 THE CGMY PROCESS AS A LEVY PROCESS A L´evy process is a stochastic process with stationary, independent increments. The L´evy– Khintchine theorem (see Sato (2001) [19]) provides a characterization of L´evy processes in terms of the characteristic function of the process, namely, there exists a measure ν such that, for all z ∈ R and t ≥ 0, E(eizLt ) = exp(tφ(z)), where φ(z) = iγ z −

σ 2 z2 + 2

R

(eizx − 1 − izx1{|x|≤1} )dν(x).

Here σ ≥ 0, γ ∈ R and ν is a measure on R such that ν({0}) = 0 and ∞. Consider a L´evy process {Lt }t≥0 of the form Lt = (r − q + µ)t + σ Wt + Zt ,

R min(1, x

(12.1) 2 )dν(x)

0,

for constants C > 0, G ≥ 0, M ≥ 0 and Y < 2. The process {Zt }t≥0 is known in the literature as the CGMY process (Carr et al. (2002)) [8]; it generalizes a jump-diffusion model by Kou (2002) [15] (Y = −1) and the VG process (Carr et al. (1998)) [10] (Y = 0). The CGMY process is, in turn, a particular case of the Kobol process studied by Boyarchenko and Levendorskiˇı (2002) [5] and Carr et al. (2003) [7], where the constant C is allowed to take on different values on the positive and negative semiaxes. The characteristic function of the CGMY process may be computed explicitly (see Boyarchenko and Levendorskiˇı (2002) [5] and Carr et al. (2002) [8]). In this paper, we

Numerical Valuation of American Options

261

consider only those processes having inﬁnite activity and ﬁnite variation, excluding the VG process, that is, 0 < Y < 1. In such a situation one has σ2 2 φ(z) = (r − q + µ)iz − z 2 , + C (−Y ) (M − iz)Y − M Y + (G + iz)Y − GY .

(12.4)

A market model Let a market consist of one risky asset {St }t≥0 and one bank account {Bt }t≥0 . Let us assume that the asset process {St }t≥0 evolves according to the geometric law St = S0 exp(Lt ),

(12.5)

where {Lt }t≥0 is the L´evy process deﬁned in equation (12.2), and the bank account follows the law Bt = exp(rt). Assume next the existence of some Equivalent Martingale Measure Q (a measure with the same null sets as the market probability, for which the discounted processes {e−(r−q)t St }t≥0 are martingales). In this paper, one works only with a risk-neutral measure Q, where the drift of the L´evy process has been changed. The EMM-condition EQ [St ] = S0 et (r−q) implies φ(−i) = r − q, and so we get the following risk-neutral form for µ: ω := −

, σ2 − C (−Y ) (M − 1)Y − M Y + (G + 1)Y − GY . 2

(12.6)

We keep the same notation for the risk-neutral parameters G and M. The other parameters σ , C and Y are the same in the risk-neutral world (see, e.g. Cont and Tankov (2004) [11] and Raible (2000) [18]). Note that M must be larger than one for ω to be well deﬁned. 12.2.1 Options in a L´evy market 12.2.1.1 European vanilla options Consider a European put option on the asset {St }t≥0 , with time to expiration T , and strike price K. Let us deﬁne the price of a European put option by the formula: v(τ, s) = e−rτ EQ (K − sHτ )+ ,

0 ≤ s < ∞,

0 ≤ τ ≤ T,

(12.7)

where the process {Hτ }τ ≥0 is the underlying risk-neutral process starting at 1, given by Hτ := exp (r − q + ω)τ + σ Wτ + Zτ .

(12.8)

Note that τ means time to expiration T − t. We will not work directly with the asset price s, but rather with its logarithm. Thus, let x = ln s, and deﬁne the new function u(τ, x) := v(τ, ex ).

(12.9)

262

Exotic Option Pricing and Advanced L´evy Models

From a generalization of Ito’s formula it follows that u satisﬁes the following Cauchy problem: uτ − Lu = 0, τ ∈ (0, T ], x ∈ R, u(0, x) = (K − ex )+ , x ∈ R,

(12.10)

where L is an integro-differential operator of the form

σ2 σ2 Lϕ : = ϕxx + r − q − ϕx − rϕ 2 2 ϕ(τ, x + y) − ϕ(τ, x) − (ey − 1)ϕx (τ, x) k(y) dy. +

(12.11)

R

For a derivation of equation (12.10), see Boyarchenko and Levendorskiˇı (2002) [5] or Raible (2000) [18]. 12.2.1.2 American vanilla options Consider an American put option written on the underlying asset {St }t≥0 . The price may be found by solving an optimal stopping problem of the form: * + v(τ, s) = sup EQ e−rτ (K − sHτ )+ .

S

(12.12)

τ ∈ 0,τ

Here S0,τ denotes the set of stopping times taking values in [0, τ ] and {Hτ }τ ≥0 is the process in equation (12.8). The corresponding function u (cf. equation (12.9)) satisﬁes the free-boundary value problem (Boyarchenko and Levendorskiˇı (2002) [5] and Matache et al. (2003) [17]): uτ − Lu = 0, τ > 0, x > $ c(τ ), x c(τ ), u(τ, x) = K − e , τ > 0, x ≤ $ u(τ, x) ≥ (K − ex )+ , τ > 0, x ∈ R, uτ − Lu ≥ 0, τ > 0, x ∈ R, u(0, x) = (K − ex )+ , x ∈ R,

(12.13)

where the operator L is deﬁned in equation (12.11) and the free-boundary is given by , $ c(τ ) = inf x ∈ R | u(τ, x) > (K − ex )+ ,

τ ∈ (0, T ].

(12.14)

The set {x ∈ R | x ≤ $ c(τ )} is the exercise region for the logarithmic prices. Hence, for asset prices s ≤ exp($ c(τ )), the American put should be exercised.

Numerical Valuation of American Options

263

12.3 NUMERICAL VALUATION OF THE AMERICAN CGMY PRICE The function $ c(τ ) is not known a priori, and needs to be found as part of the solution. Thus, rather than solving equation (12.13) directly, it is more convenient to use another formulation as a so-called Linear Complementarity Problem: uτ − Lu ≥ 0 in (0, T ) × R, u≥ψ in [0, T ] × R, (12.15) (uτ − Lu) (u − ψ) = 0 in (0, T ) × R, u(0, x) = ψ(x), where the initial condition is given by ψ(x) := (K − ex )+ .

(12.16)

Note that the dependency on the free-boundary $ c(τ ) has disappeared, but instead we are left with a set of inequalities. The discretization and numerical solution of equation (12.15) is from now on our main goal. The free-boundary is obtained after computing the solution, by making use of equation (12.14). 12.3.1 Discretization and solution algorithm The main idea of the method is to approximate the operator (equation (12.11)) by truncating the integral term close to zero and inﬁnity. The truncation around inﬁnity is harmless, as long as a sufﬁciently large interval is chosen and the price is substituted by the option’s intrinsic value outside the computational domain. However, the truncation around zero gives rise to an artiﬁcial diffusion that must be taken into account. More precisely, the operator L may be split into the sum of two operators: the ﬁrst one containing the Black and Scholes operator and the second accounting for the jumps, namely, L = LBS + LJ . The jump integral part is in turn split into the sum of one operator P for the integration variable in a neighborhood of the origin, and Q for the complementary domain. For P , we use Taylor’s expansion to write the following approximation: (P ϕ)(τ, x) := ϕ(τ, x + y) − ϕ(τ, x) − (ey − 1)ϕx (τ, x) k(y) dy |y|≤

=

|y|≤

ϕ(τ, x + y) − ϕ(τ, x) − yϕx (τ, x) − (ey − 1 − y)ϕx (τ, x) k(y) dy

2 2 $ ϕ)(τ, x) := σ () ϕxx (τ, x) − σ () ϕx (τ, x), ≈ (P 2 2

with the notation: σ () = 2

|y|≤

y 2 k(y)dy.

(12.17)

264

Exotic Option Pricing and Advanced L´evy Models

$ , with a small That is, P has been approximated by a convection–diffusion operator P 2 diffusion coefﬁcient σ (). The operator Q is simply split into a sum, given that this operation is now allowed away from the origin: ϕ(τ, x + y) − ϕ(τ, x) − (ey − 1)ϕx (τ, x) k(y)dy (Q ϕ)(τ, x) := |y|≥

= (J ϕ)(τ, x) − λ()ϕ(τ, x) + ω()ϕx (τ, x), where we have written J for the convolution term, and k(y)dy, λ() = ω() =

|y|≥

|y|≥

(1 − ey )k(y)dy.

(12.18)

(12.19) (12.20)

Remark 12.3.1 These operations have a probabilistic meaning: the pure-jump process has been approximated by a compound Poisson process plus a small Brownian component. As proved in Asmussen and Rosi´nski (2001) [4], this approximation is valid, if and only if, σ ()/ → ∞, as → 0. Note that this condition implies 0 < Y < 1, excluding therefore the VG process and processes with inﬁnite activity. An approximation result in Cont and Voltchkova (2003) [12] states the following. Let $ + Q and u be the solution of the Cauchy problem L := LBS + P uτ − L u = 0, (12.21) u(0, x) = ψ(x), and then there exists a constant C > 0 such that |u(τ, x) − u (τ, x)| < C, for all τ and x. We use here – without proof – the same approximation to numerically solve an American put option. An indication that this approximation works also for American options is shown in Figure 12.1, where one observes that the exercise boundary tends to the theoretical perpetual exercise price, when the time to expiration τ is taken large. The proof of this fact is thus an open problem. Let us focus now on the problem shown in equation (12.15), but with L instead of L. One possible idea to discretize this new problem is to apply Euler’s scheme in time combined with an implicit–explicit iteration in space. Let the time interval [0, T ] be divided into L equal parts, i.e. τj = j τ (j = 0, 1, . . . , L) with τ = T /L and deﬁne the functions uj ≈ u(τj , x). Let operator L be split as L = A + B. We consider the following sequence of problems: j +1 u uj − Auj +1 ≥ d j := + Buj , τ τ uj +1 ≥ ψ, (12.22) j +1

u j +1 j j +1 − Au − d (u − ψ) = 0, τ 0 u = ψ.

Numerical Valuation of American Options

265

(a) 20 18 16 14

Time

12 10 Y = 0.3

8 6 4 2 0

8

8.2

8.4

8.6

8.8 9 9.2 Asset price

9.4

9.6

9.8

10

(b) 20 18 16 14

Time

12 10 Y = 0.7

8 6 4 2 0

6

6.5

7

7.5 8 Asset price

8.5

9

9.5

Figure 12.1 Exercise boundary and perpetual boundary for two different values of the parameter Y , i.e. 0.3 (a) and 0.7 (b): σ = 0; r = 0.1; q = 0; K = 10; T = 20; C = 1; G = 7; M = 9

266

Exotic Option Pricing and Advanced L´evy Models

That is, given the function uj , we compute uj +1 by solving these integro-differential inequalities. A natural choice for the splitting of L is the following: σ 2 + σ 2 () σ 2 + σ 2 () ϕxx + r − q − + ω() ϕx − rϕ Aϕ := (12.23) 2 2 Bϕ := J ϕ − λ()ϕ.

(12.24)

Observe that the integral term is treated explicitly, whereas the differential part is treated implicitly. This method imposes a stability restriction on the time step; see Cont and Voltchkova (2003) [12] for a discussion of this issue for the European case. 12.3.1.1 Spatial discretization of A Consider a computational domain of the form [0, T ] × [xmin , xmax ]. Let ln K ∈ [xmin , xmax ] and deﬁne the uniform spatial grid xi = xmin + ih (i = 0, . . . , N ) where h = (xmax − xmin )/N . Once we have deﬁned the grid, we can discretize A by standard second-order schemes. For the ﬁrst and second derivatives, the central scheme and the standard 3-point scheme are chosen, respectively. Namely, after introducing the notation δ1 (ϕ) := [ϕi+1 − ϕi−1 ]/2h and δ2 (ϕ) := [ϕi+1 − 2ϕi + ϕi−1 ]/ h2 , where ϕi := ϕ(xi ) (i = 0, 1, . . . , N ), we may write (Aϕ)i = βδ2 (ϕ) + γ δ1 (ϕ) − rϕi ,

(12.25)

with the quantities β and γ deﬁned as β :=

σ 2 + σ 2 () , 2

γ := r − q −

σ 2 + σ 2 () + ω(). 2

(12.26) (12.27)

We obtain the following coefﬁcients for the implicit part γ β , + h2 2h 1 2β b= +r + 2, τ h β γ . c=− 2 − h 2h

a=−

(12.28) (12.29) (12.30)

The tridiagonal matrix T associated to the implicit part has constant diagonals: b is on the main diagonal, a is on the subdiagonal and c is on the superdiagonal. From now on, the parameter is taken as the mesh-size h. The artiﬁcial diffusion σ 2 (h) (cf. Matache et al. (2003) [17]) may be approximated by the composite trapezoidal rule on the intervals [−h, 0] and [0, h]. This gives σ 2 (h) ≈

[k(h) + k(−h)] h3 . 2

The quantities λ(h) and ω(h) are approximated in the next section.

(12.31)

Numerical Valuation of American Options

267

12.3.1.2 Spatial discretization of B The discretization of B involves the discretization of J , since Bϕ = J ϕ − λ()ϕ. The discretization of J is explained in detail in Anon (2004) [3]. Brieﬂy, the idea is to truncate the integral to a ﬁnite domain and then apply the composite trapezoidal rule, i.e., Ji := (J ϕ)i =

ϕ(xi + y)k(y)dy

|y|≥h

≈

ϕ(xi + y)k(y)dy h≤|y|≤Mh

≈h

M

ϕi+m km ρm ,

i = 0, 1, . . . , N,

(12.32)

m=−M

where km = k(mh) for m = 0 and we let k0 = 0. The coefﬁcients obtained from applying the trapezoidal rule are: ρm =

1/2 if m ∈ {−M, −1, 1, M}, 1 otherwise.

It is important to substitute ϕ by the payoff function ψ outside the computational domain. The computation of the numbers Ji constitutes the main burden of the method, but thanks to the FFT algorithm, this may be carried out efﬁciently (see next section). However, N must be an even number, and M = N/2, to be able to express this convolution in matrix–vector notation. Finally, we may use the composite trapezoidal rule to compute an approximation to the numbers λ(h) and ω(h) by simply taking ϕ in equation (12.32) as 1 and ey − 1, respectively. 12.3.1.3 Fast convolution by FFT The Fast Fourier Transform is an algorithm that evaluates the Discrete Fourier Transform (DFT) of a vector f = [f0 , f2 . . . , fR−1 ] in O(R log R) operations. The Discrete Fourier Transform is deﬁned as: Fk =

R−1

fn e−i2πnk/R ,

k = 0, 1, . . . , R.

(12.33)

n=0

One of the multiple applications of the DFT is in computing convolutions. Let us ﬁrst introduce the concept of circulant convolution. Let {xm } and {ym } be two sequences with period R. The convolution sequence z := x ∗ y is deﬁned component-wise as zn =

R−1 m=0

xm−n ym .

(12.34)

268

Exotic Option Pricing and Advanced L´evy Models

We now use the FFT to compute the vector [z0 , . . . , zR−1 ]. The periodic structure of x allows the derivation of the following simple relation: Zk = Xk · Yk ,

(12.35)

where X, Y and Z denote the Discrete Fourier Transform of the sequences x, y and z, respectively. That is, the DFT applied to the convolution sequence is equal to the product of the transforms of the original two sequences. The vector [z0 , . . . , zR−1 ] may be recovered by means of the Inverse Discrete Fourier Transform (IDFT): zn =

R−1 1 Zk ei2πkn/R , R

n = 0, 1, . . . , R.

(12.36)

k=0

In the language of matrices, a circulant convolution may be seen as the product of a circulant matrix times a vector. For example, let R = 3, and use the periodicity xk = xk+R to write equation (12.34) as z0 x0 x1 x2 y0 z1 = x2 x0 x1 y1 . (12.37) z2 x1 x2 x0 y2 A circulant matrix is thus a matrix in which each row is a ‘circular’ shift of the previous row. We are interested in the convolution shown in equation (12.32), where the vector ϕ is not periodic. The associated matrix is a so-called Toeplitz matrix, which by deﬁnition is a matrix that is constant along diagonals. A circulant matrix is hence a particular type of Toeplitz matrix. The next idea is to embed a Toeplitz matrix into a circulant matrix. As an example, let M = 1 and N = 2, so that the matrix-vector notation for equation (12.32) reads k1 /2 ϕ1 ϕ0 ϕ−1 ϕ2 ϕ1 ϕ0 k0 . (12.38) ϕ3 ϕ2 ϕ1 k−1 /2 The matrix above may be embedded in a circulant matrix C of size 5 in the following way. (For computational efﬁciency of the FFT algorithm, it is advisable to use a circulant matrix whose size is a power of 2.): ϕ0 ϕ−1 ϕ3 ϕ2 ϕ1 ϕ2 ϕ1 ϕ0 ϕ−1 ϕ3 ϕ2 ϕ1 ϕ0 ϕ−1 (12.39) C = ϕ3 . ϕ−1 ϕ3 ϕ2 ϕ1 ϕ0 ϕ0 ϕ−1 ϕ3 ϕ2 ϕ1 If we deﬁne the vector η := [k1 /2, k0 , k−1 /2, 0, 0]T , then the product (equation (12.38)) is the vector consisting of the ﬁrst three elements in the product Cη. As explained before, a product of a circulant matrix and a vector may be efﬁciently obtained by applying the FFT algorithm.

Numerical Valuation of American Options

269

As a summary, following the ideas explained above, it is possible to compute the convolution (equation (12.32)), with M = N/2, by ‘embedding’ the resulting matrix into a circulant matrix. The product of a circulant matrix and a vector is carried out in three FFT operations, namely, two DFT and one IDFT. In Almendral and Oosterlee (2003) [2], we applied the FFT algorithm in the computation of European options for Merton’s model and Kou’s model, and in Anon (2004) [3] to ﬁnd the American price under the Variance Gamma process. For further details on the computation of convolutions by FFT we refer the reader to Van Loan (1992) [21]. 12.3.1.4 Boundary conditions We used points on the boundary when discretizing the differential operator A. This means that the vector d j needs to be updated. For a put option, this is done by updating the ﬁrst and the last entries of d j as follows: j

j

j

d1 ← d1 − a(K − exmin ),

dN−1 ← 0.

(12.40)

12.3.1.5 Discrete LCP We are now in position to write the discrete inequalities that correspond to the discretization of equation (12.22): T uj +1 ≥ d j , j +1 u ≥ ψ, (12.41) (T uj +1 − d j , uj +1 − ψ) = 0, 0 u = ψ, j

for j = 0, 1, . . . , L − 1. The matrix T has entries given by equations (12.28)–(12.30), di = j j ui /τ + (J uj )i − λ()ui (i = 1, . . . , N − 1) with the update shown in equation (12.40) and ψ is the vector [ψ1 , ψ2 , . . . , ψN−1 ]T , with ψi = ψ(xi ) (cf. Lewis (2001) [16]). The same letter ψ is used to simplify the notation. We proceed now to explain a simple algorithm to solve equation (12.41). 12.3.1.6 Brennan–Schwartz algorithm for a put option Let a tridiagonal matrix

b1 a2 T =

c1 b2 .. .

c2 .. . an−1

..

. bn−1 an

cn−1 bn

(12.42)

and vectors d = [d1 , . . . , dn ]T and ψ = [ψ1 , . . . , ψn ]T be given. Consider the following problem: ﬁnd a vector u satisfying the system T u ≥ d, u ≥ ψ, (12.43) (T u − d, u − ψ) = 0.

270

Exotic Option Pricing and Advanced L´evy Models

The following algorithm to ﬁnd u in equation (12.43) was proposed by Brennan and Schwartz (1977) [6] (for put options) and discussed in detail by Jaillet et al. (1990) [14]: • Step 1: Compute recursively a vector $ b as $ bn = bn , $ bj , bj −1 = bj −1 − cj −1 aj /$

j = n, . . . , 2.

• Step 2: Compute recursively a vector d$ as d$n = dn , bj , d$j −1 = dj −1 − cj −1 d˜j /$

j = n, . . . , 2.

• Step 3: Compute u forward as follows: u1 = max d$1 /b1 , ψ1 , bj , ψ j , uj = max d$j − aj uj −1 /$

j = 2, . . . , n.

We apply these three steps with ai = a, bi = b and ci = c, with a, b and c as in equations (12.28)–(12.30). The splitting proposed in equations (12.23) and (12.24) does not, in general, guarantee the validity of the Brennan–Schwartz algorithm. However, the convection term may be moved to the explicit part of the splitting, so that the conditions of the Brennan– Schwartz algorithm hold Almendral and Oosterlee (2003) [2]. The solutions obtained in both ways are the same, to within the discretization error.

12.4 NUMERICAL EXPERIMENTS In this section, European and American option prices are computed numerically. In the ﬁrst experiment, we compute an European option (problem (12.21)) and compare it with the solution obtained by the Carr–Madan formula in Carr and Madan (1999) [9]; see also the appendix in this paper, formula (12.9). Both solutions are compared in the %∞ -norm, and the results are shown in Table 12.1. A linear convergence rate is observed, and note that the algorithm computes the European price with an error of one cent in about one second. Table 12.1 Linear convergence to exact solution in %∞ -norm and CPU times on a Pentium IV, 1.7 GHz. The parameters are as follows: r = 0; q = 0; K = 10; T = 1; C = 1; G = 7; M = 9; Y = 0.7 N

L

%∞ -error

50 100 200 400

5 10 20 40

0.2675 0.1281 0.0459 0.0160

CPU-time (s) 0.22 0.31 0.34 1.06

Numerical Valuation of American Options

271

A second experiment concerns the veriﬁcation of the theoretical perpetual exercise price against the asymptotic behavior of the free boundary for some large time to expiry. The asymptotic value s ∗ of the American put was veriﬁed with the aid of a formula in Boyarchenko and Levendorskiˇı (2002) [5], (Theorem 3.2 and Theorem 5.1): 6 ∞+iρ ln r + q + φ0 (z) 1 ∗ ∗ dz , (12.44) s = exp(x ) = K exp − 2π −∞+iρ z2 + iz with ρ a positive number (not arbitrary, see Boyarchenko and Levendorskiˇı [5]) and φ0 (z) is given by equation (12.47) (see below). Figure 12.1 above shows two examples of exercise boundaries and their corresponding theoretical asymptotic values. In these examples, ρ = 1 gives the right value. In the next two experiments, we examine the behavior of the option price and free boundary for different values of Y and M. We conclude from Figures 12.2 and 12.3 that the American option price is an increasing function of Y and a decreasing function of M. We mention that the results shown in Figure 12.2 are in accordance with the numerical tests in Matache et al. (2003) [17] (Figure 6). The last test is designed to verify the smooth-ﬁt principle. According to Alili and Kyprianou (2004) [1], the smooth-ﬁt principle holds for perpetual American put options in the bounded variation case considered here, if and only if, the drift r − q + ω is negative, or an additional condition on the jump measure is satisﬁed for zero drift. In Figure 12.4(a), we show the numerical derivative vs at time T = 1, for a set of parameters giving negative drift. In this case, we have smooth-ﬁt. For a second set of parameters chosen such that the drift is positive, we see a discontinuous derivative in Figure 12.4(b) and so there is no smooth-ﬁt.

APPENDIX: ANALYTIC FORMULA FOR EUROPEAN OPTION PRICES We include here the analytic expression given in Lewis (2001) [16] for European options, adapted to the case of a CGMY process: e−rt u(t, x) = 2π

iα+∞

ˆ exp [−izx + tφ0 (−z)] ψ(z)dz,

(12.45)

iα−∞

ˆ where ψ(z) is the generalized Fourier transform of the payoff ψ, which for a put option is given by K iz+1 ˆ ψ(z) =− 2 , z − iz

(12.46)

and the risk-neutral characteristic function φ0 to be used is obtained by substituting µ by ω from equation (12.6) into equation (12.4), i.e. σ2 2 φ0 (z) = (r − q + ω)iz − z 2 , + C (−Y ) (M − iz)Y − M Y + (G + iz)Y − GY .

(12.47)

272

Exotic Option Pricing and Advanced L´evy Models (a)

5

Y = 0.1 Y = 0.3 Y = 0.5 payoff

4.5 4

Option price

3.5 3 2.5 2 1.5 1 0.5 0

(b)

6

8

10 12 Asset price

14

16

5 4.5 4 3.5

Time

3

Y = 0.5

Y = 0.3

Y = 0.1

2.5 2 1.5 1 0.5 0 7.5

8

8.5

9

9.5

10

Asset price

Figure 12.2 (a) Option prices for different values of the parameter Y , i.e. 0.1, 0.3 and 0.7, and (b) the corresponding exercise boundaries: σ = 0; r = 0.1; q = 0; K = 10; T = 5; C = 1; G = 7.8; M = 8.2

Numerical Valuation of American Options (a)

273

M= 5 M= 7 M= 9 payoff

4.5 4

Option price

3.5 3 2.5 2 1.5 1 0.5 0

6

7

8

9

10

11

12

13

14

15

Asset price (b)

5

M=9

4.5 4 3.5 M=7

Time

3 2.5 2 1.5

M=5

1 0.5 0 7.5

8

8.5

9

9.5

10

Asset price

Figure 12.3 (a) Option prices for different values of the parameter M, i.e. 5, 7 and 9, and (b) the corresponding exercise boundaries: σ = 0; r = 0.1; q = 0; K = 10; T = 5; C = 1; G = 7; Y = 0.2

274

Exotic Option Pricing and Advanced L´evy Models (a) −0.1 −0.2

Option Delta

−0.3 −0.4 −0.5 −0.6 −0.7 −0.8 −0.9 6

8

10

12

14

16

Asset price (b) −0.1 −0.2

Option Delta

−0.3 −0.4 −0.5 −0.6 −0.7 −0.8 −0.9 6

7

8

9

10

11

12

Asset price

Figure 12.4 (a) Continuous option Delta for G = 10 and M = 3, and (b) discontinuous option Delta for G = 7 and M = 9: σ = 0; r = 0.1; q = 0; K = 10; T = 1; C = 1

Numerical Valuation of American Options

275

The constant α in equation (12.9) is determined by the region of validity of equation (12.46) together with the strip of regularity of equation (12.47). In this case, we may pick α ∈ (−G, 0). A method using the FFT algorithm was proposed in Carr and Madan (1999) [9] to evaluate an analogous version of equation (12.45).

REFERENCES [1] Alili, L. and Kyprianou, A.E. (2004) “Some remarks on ﬁrst passage of L´evy processes, the American put and pasting principles”, to be published. [2] Almendral, A. and Oosterlee, C.W. (2003), “Numerical valuation of options with jumps in the underlying”, Technical Report, Delft University of Technology, Delft, The Netherlands; to be published. [3] Almendral, A. (2004), “On American options under the Variance Gamma process”, Technical Report, Delft University of Technology, Delft, The Netherlands: to be published. [4] Asmussen, S. and Rosi´nski, J. (2001), “Approximations of small jumps of L´evy processes with a view towards simulation”, Journal of Applied Probability, 38, 482–493. [5] Boyarchenko, S.I. and Levendorski˘ı, S.Z. (2002), Non-Gaussian Merton–Black–Scholes Theory, Advanced Series on Statistical Science and Applied Probability, vol. 9, World Scientiﬁc, River Edge, NJ, USA. [6] Brennan, M.J. and Schwartz, E.S. (1977), “The valuation of the American put option”, Journal of Finance, 32, 449–462. [7] Carr, P., Geman, H., Madan, D.B. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [8] Carr, P.P., Geman, H., Madan, D.B. and Yor, M. (2002), “The ﬁne structure of asset returns: An empirical investigation”, Journal of Business, 75, 305–332. [9] Carr, P.P. and Madan, D.B. (1999), “Option valuation using the Fast Fourier Transform”, Journal of Computational Finance, 2, 61–73. [10] Carr, P.P., Madan, D.B. and Chang, E.C. (1998), “The Variance Gamma process and option pricing”, European Finance Review, 2, 79–105. [11] Cont, R. and Tankov, P. (2004), Financial Modelling with Jump Processes, Chapman & Hall, Boca Raton, FL, USA. [12] Cont, R. and Voltchkova, E. (2003), “A ﬁnite difference scheme for option pricing in jump diffusion and exponential L´evy models”, Technical Report 513, CMAP, Palaiseau, France. [13] Hirsa, A. and Madan, D.B. (2004), “Pricing American options under Variance Gamma”, Journal of Computational Finance, 7. [14] Jaillet, P., Lamberton, D. and Lapeyre, B. (1990), “Variational inequalities and the pricing of American options”, Acta Applied Mathematica, 21, 263–289. [15] Kou, S.K. (2002), “A jump diffusion model for option pricing”, Management Science, 48, 1086– 1101. [16] Lewis, A.L. (2001), “A simple option formula for general jump-diffusion and other exponential L´evy processes”, in Proceedings of the Annual CAP Workshop on Derivative Securities and Risk Management. [17] Matache, A.M., Nitsche, P.A. and Schwab, C. (2003), “Wavelet Galerkin pricing of American options on L´evy driven assets, Working Paper, ETH, Z¨urich, Switzerland. [18] Raible, S. (2000), “L´evy processes in ﬁnance: theory, numerics and empirical Facts”, Ph.D Thesis, Institute f¨ur Mathematische Stochastik, Albert-Ludwigs-Universit¨at Freiburg, Freiburg, Germany.

276

Exotic Option Pricing and Advanced L´evy Models

[19] Sato, K.-I. (2001), “Basic results on L´evy processes”, in L´evy Processes, Birkh¨auser, Boston, MA, USA, pp. 3–37. [20] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [21] van Loan, C. (1992), Computational Frameworks for the Fast Fourier Transform, Frontiers in Applied Mathematics, vol. 10, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, USA.

13 Convertible Bonds: Financial Derivatives of Game Type Jan Kallsen Munich University of Technology, Garching, Germany and

Christoph Kuhn ¨ ¨ Frankfurt am Main, Germany Johann Wolfgang Goethe-Universitat, Abstract A convertible bond is a security that the holder can convert into a speciﬁed number of underlying shares. In addition, very often the issuer can recall the bond, paying some compensation, or force the holder to convert it immediately. Therefore, the pricing problem has also a game-theoretic aspect. When modelling convertible (callable) bonds within the framework of a ﬁrm value model, they can be considered as an example of a standard game contingent claim as long as no dividends are distributed to the equity holders. This article reviews the classical as well as some recent literature in this ﬁeld. Furthermore, we introduce a mathematically rigorous concept of no-arbitrage price processes for these kinds of derivatives, which explicitly incorporates the feature that the contract can be terminated by both counterparties prematurely. We compare this dynamic conception to price derivatives with the static one by Karatzas and Kou (1998) [18].

13.1 INTRODUCTION A game contingent claim (GCC), as introduced in Kifer (2000) [20], is a contract between a seller A and a buyer B which can be terminated by A and exercised by B at any time t ∈ [0, T ] up to a maturity date T when the contract is terminated anyway. More precisely, the contract may be speciﬁed in terms of two stochastic processes (Lt )t∈[0,T ] , (Ut )t∈[0,T ] with Lt ≤ Ut for t ∈ [0, T ) and LT = UT .

(13.1.1)

If A terminates the contract at time t before it is exercised by B, she has to pay B the amount Ut . If B exercises the option before it is terminated by A, he is paid Lt . An example is a put option of game type with constant penalty δ > 0. If S 1 denotes the price process of the underlying and K the strike price, then Lt = (K − St1 )+ and Ut = (K − St1 )+ + δ1{t 0 we deﬁne the recall times σ ε = inf{t ≥ 0 | V t ≥ Ut − ε}. Analogously to Proposition 3.1 in Karatzas and Zamﬁrescu (2003) [19] and Theorem 11 in ε LM, one can show that for all ε > 0 the stopped process V σ is a Q-supermartingale w.r.t. all Q ∈ Me . Note that in LM the payoff processes L and U are supposed to be bounded, but the results still hold under the weaker condition (13.2.1) (see Theorem 1.1 in K¨uhn and Kyprianou (2003b) [23]. By this supermartingale property and L ≤ V we have for all Q ∈ Me , τ ∈ T0 EQ R(τ, σ ε ) ≤ EQ Lτ 1{τ ≤σ ε } + V σ ε 1{σ ε 0 and σ ∈ T0 be a recalling time with hσ := supτ ∈T0 supQ∈Me EQ (R(τ, σ )) ≤ hup + ε. Deﬁne an American contingent claim (ACC) by the c`adl`ag process X = (Xt )t∈[0,T ] , where Xt = Lt 1{t L on [0, T ) × , and U− > L− on [0, T ] × ,

(13.2.10)

which holds true for the American case (Ut = ∞ for t < T ) or for the callable put with constant penalty (see K¨uhn and Kyprianou 2003a [22]). For other sufﬁcient conditions, see Kallsen and K¨uhn (2004) [17]. Proof. Ad ⇒: Step 1: Let S d+1 be an arbitrage-free price process. By condition (2) and the fundamental theorem of asset pricing there exists a probability measure Q ∈ Me (i.e. Q ∼ P and S 1 , . . . , S d are Q-σ -martingales) such that 1{S d+1 =L− } · S d+1 is a Q-σ -supermartingale, −

1{L− <S d+1 0 is the volatility coefﬁcient, and x > 0 is given and ﬁxed. The main purpose of this present paper is to ﬁnd a solution to the following optimal stopping game for the time-homogeneous (strong) Markov process X having the value function: V∗ (x) = inf sup Ex [e−(λ+r)(σ ∧τ ) (G1 (Xσ ) I (σ < τ ) + G2 (Xτ ) I (τ ≤ σ ))] σ

τ

= sup inf Ex [e−(λ+r)(σ ∧τ ) (G1 (Xσ ) I (σ < τ ) + G2 (Xτ ) I (τ ≤ σ ))] τ

σ

(14.2.3)

where Px is a probability measure under which the process X deﬁned in equations (14.2.1) and (14.2.2) starts at some x > 0, the inﬁmum and supremum are taken over all ﬁnite stopping times σ and τ of the process X (i.e. stopping times with respect to (FtX )t≥0 denoting the natural ﬁltration of X: FtX = σ {Xu | 0 ≤ u ≤ t}, t ≥ 0), λ > 0 is a discounting rate, and the functions Gi (x) are deﬁned by: Gi (x) = (x − Li ) I (Li ≤ x < Ki ) + (Ki − Li ) I (x ≥ Ki )

(14.2.4)

The Spread Option Optimal Stopping Game

295

for all x > 0 with some constants Li and Ki such that 0 < Li < Ki , i = 1, 2, as well as L1 < L2 , K1 < K2 and K1 − L1 = K2 − L2 . We will derive sufﬁcient conditions for the existence of a nontrivial closed form solution to the problem (equation (14.2.3)). Note that the existence of a unique value (equation (14.2.3)) was proved in Lepeltier and Mainguenau (1984) [25] and Kifer (2000) [20]. This fact will be re-proved in Theorem 4.1 below under certain conditions on the parameters of the model. It also follows from equation (14.2.3) that the inequalities G2 (x) ≤ V∗ (x) ≤ G1 (x) hold for all x > 0. We will search for optimal stopping times in the problem (equation (14.2.3)) of the following form: σ∗ = inf{t ≥ 0 | Xt ≤ A∗ }

(14.2.5)

τ∗ = inf{t ≥ 0 | Xt ≥ B∗ }

(14.2.6)

for some numbers A∗ and B∗ such that L1 ≤ A∗ ≤ D1 and D2 ≤ B∗ ≤ K2 hold with Di = Li (λ + r)/λ, i = 1, 2 (for an explanation of the latter inequalities see the text following equation (14.4.5) below). In this connection, the points A∗ and B∗ are called optimal stopping boundaries. Note that in this case, A∗ is the largest number from L1 ≤ x ≤ D1 such that V∗ (x) = G1 (x), and B∗ is the smallest number from D2 ≤ x ≤ K2 such that V∗ (x) = G2 (x). The pair of stopping times (σ∗ , τ∗ ) is usually called a saddle point of the optimal stopping game. On a ﬁnancial market, there are investors speculating for a rise of stock prices (so-called ‘bulls’ playing on the increase) and investors speculating for a fall of stock prices (so-called ‘bears’ playing on the decrease), and their strategies on the market are asymmetric (see, e.g. Shiryaev (1999) [31] Chapter I, Section 1c). In order to restrict their losses and gains simultaneously, the investors playing on the increase may turn to a strategy consisting of buying a call option with a strike price L2 and selling a call option with a higher strike price K2 > L2 , while the investors playing on the decrease may turn to a strategy consisting of selling a call option with a strike price L1 and buying a call option with a higher strike price K1 > L1 . Such combinations are called spread options of ‘bull’ and ‘bear’, respectively, and their payoff functions are given by G2 (x) and −G1 (x) from equation (14.2.4), where x denotes the stock price (see Shiryaev (1999) [31] Chapter VI, Section 4e). In this present paper, we consider a contingent claim with arbitrary (random) times of exercise τ and cancellation σ , where according to the conditions of the claim the buyer can choose the exercise time τ and in case τ ≤ σ gets the value G2 (Xτ ) from the seller, and the seller can choose the cancellation time σ and in case σ < τ gives the value G1 (Xσ ) to the buyer. Then, by virtue of the fact that Px is a martingale measure for the given market model (see, e.g. Shiryaev et al. (1994) [29] Section 1, Shiryaev (1999) [31] Chapter VII, Section 3g, and Kifer (2000) [20] Section 3), the value (equation (14.2.3)) may be interpreted as a rational (fair) price of the mentioned contingent claim in the given model. We also observe that from the structure of the problem (equation (14.2.3)) it is intuitively clear that the buyer wants to stop when the process X comes close to L1 (from above) while the seller wants to stop when the process X comes close to K2 (from below) without waiting too long because of the punishment of discounting. Taking into account the arguments stated above, we will call the presented contingent claim a spread game option. Note that the structure of the given option differs from the structure of the game options considered in Kifer (2000) [20] and Kyprianou (2004) [24].

296

Exotic Option Pricing and Advanced L´evy Models

14.3 SOLUTION OF THE FREE-BOUNDARY PROBLEM By means of standard arguments, it is shown that the inﬁnitesimal operator L of the process X acts on an arbitrary function F from the class C 2 on (0, ∞) according to the rule: (LF )(x) = rx F (x) + (θ 2 x 2 /2) F (x)

(14.3.1)

for all x > 0. In order to ﬁnd explicit expressions for the unknown value function V∗ (x) from equation (14.2.3) and the boundaries A∗ and B∗ from equations (14.2.5) and (14.2.6), using the results of general theory of optimal stopping problems for continuous time Markov processes as well as taking into account the results about the connection between optimal stopping games and free-boundary problems (see, e.g. Grigelionis and Shiryaev (1966) [12] and Shiryaev (1978) [30] Chapter III, Section 8; as well as Bensoussan and Friedman (1974, 1977) [3] [4]), we can formulate the following free-boundary problem: (LV )(x) = (λ + r)V (x) V (A+) = A − L1 , V (x) = G1 (x)

for

A<x B

A<x0 B γ1 +1 (γ1 − 1)(γ2 − 1)(B − D2 ) 0. By virtue of the fact that the time spent by the process X at the points L1 , A∗ , B∗ and K2 is of Lebesgue measure zero, from the expression (14.4.2) it therefore follows that the inequalities: e−(λ+r)(σ∗ ∧τ ) G2 (Xσ∗ ∧τ ) ≤ e−(λ+r)(σ∗ ∧τ ) V (Xσ∗ ∧τ ) ≤ V (x) + Mσ∗ ∧τ

(14.4.6)

e−(λ+r)(σ ∧τ∗ ) G1 (Xσ ∧τ∗ ) ≥ e−(λ+r)(σ ∧τ∗ ) V (Xσ ∧τ∗ ) ≥ V (x) + Mσ ∧τ∗

(14.4.7)

are satisﬁed for any ﬁnite stopping times σ and τ of the process X. Let (τn )n∈N be an arbitrary localizing sequence of stopping times for the process (Mt )t≥0 . Then, by using inequalities (14.4.6) and (14.4.7) and taking the expectations with respect to

The Spread Option Optimal Stopping Game

301

Px , by means of the optional sampling theorem (see, e.g. Jacod and Shiryaev (1987) [14] Chapter I, Theorem 1.39) we get: * + Ex e−(λ+r)(σ∗ ∧τ ∧τn ) G1 (Xσ∗ )I (σ∗ < τ ∧ τn ) + G2 (Xτ ∧τn )I (τ ∧ τn ≤ σ∗ ) * + ≤ Ex e−(λ+r)(σ∗ ∧τ ∧τn ) V (Xσ∗ ∧τ ∧τn ) ≤ V (x) + Ex Mσ∗ ∧τ ∧τn = V (x)

(14.4.8)

* + Ex e−(λ+r)(σ ∧τ∗ ∧τn ) G1 (Xσ ∧τn )I (σ ∧ τn < τ∗ ) + G2 (Xτ∗ )I (τ∗ ≤ σ ∧ τn ) * + (14.4.9) ≥ Ex e−(λ+r)(σ ∧τ∗ ∧τn ) V (Xσ ∧τ∗ ∧τn ) ≥ V (x) + Ex Mσ ∧τ∗ ∧τn = V (x) for all x > 0. Hence, letting n go to inﬁnity and using Fatou’s lemma, we obtain that for any ﬁnite stopping times σ and τ the inequalities: * + Ex e−(λ+r)(σ∗ ∧τ ) G1 (Xσ∗ )I (σ∗ < τ ) + G2 (Xτ )I (τ ≤ σ∗ ) * + ≤ V (x) ≤ Ex e−(λ+r)(σ ∧τ∗ ) G1 (Xσ )I (σ < τ∗ ) + G2 (Xτ∗ )I (τ∗ ≤ σ )

(14.4.10)

hold for all x > 0. In order to show that the equalities in expression (14.4.10) are attained at σ∗ and τ∗ from equations (14.2.5) and (14.2.6), let us use the fact that the function V (x) solves the equation (14.3.2) for all A∗ < x < B∗ . In this case, by the expression (14.4.2) and the structure of the stopping times σ∗ and τ∗ , it follows that the equality: e−(λ+r)(σ∗ ∧τ∗ ∧τn ) V (Xσ∗ ∧τ∗ ∧τn ) = V (x) + Mσ∗ ∧τ∗ ∧τn

(14.4.11)

holds, from where, by using the expressions (14.4.6) and (14.4.7), we may conclude that the inequalities: −(K1 − L1 ) ≤ Mσ∗ ∧τ∗ ∧τn ≤ K2 − L2

(14.4.12)

are satisﬁed for all x > 0, where (τn )n∈N is a localizing sequence for (Mt )t≥0 . Hence, letting n go to inﬁnity in the expression (14.4.11) and using the conditions (14.3.3), as well as the obviously fulﬁlled property Px [σ∗ ∧ τ∗ < ∞] = 1 (see, e.g. Shiryaev et al. (1994) [29] Section 8, or Shiryaev (1999), [31] Chapter VIII, Section 2a), by means of the Lebesgue bounded convergence theorem we obtain the equality: * + Ex e−(λ+r)(σ∗ ∧τ∗ ) G1 (Xσ∗ ) I (σ∗ < τ∗ ) + G2 (Xτ∗ ) I (τ∗ ≤ σ∗ ) = V (x)

(14.4.13)

for all x > 0, which together with (14.4.10) directly implies the desired assertion.

302

Exotic Option Pricing and Advanced L´evy Models

14.5 CONCLUSIONS Recall that throughout the paper and particularly in the proof of Theorem 4.1 we have used the assumption that L2 (λ + r)/λ ≤ K2 among others. When the latter condition fails to hold but L1 (λ + r)/λ ≤ K1 holds, let us set B∗ = K2 in equation (14.2.6) and consider the problem (14.2.3) as an optimal stopping problem for the seller. In this case, we can also formulate the free-boundary problem (equations (14.3.2)–(14.3.5)), where L1 ≤ A ≤ D1 and B = K2 with D1 = L1 (λ + r)/λ, and assume that the following condition holds: V (A+) = 1 (smooth ﬁt).

(14.5.1)

By means of the same arguments as in Section 14.3, by using the assumed smoothﬁt condition (14.5.1), it can be shown that the boundary A should satisfy the following equation: γ1 (A − L1 )(K2 /A)γ2 − (K2 − L2 ) γ2 (K2 − L2 ) − (A − L1 )(K2 /A)γ1 + = 1. (14.5.2) A (K2 /A)γ2 − (K2 /A)γ1 A (K2 /A)γ2 − (K2 /A)γ1 In order to ﬁnd sufﬁcient conditions for the existence and uniqueness of solution of the equation (14.5.2) let us deﬁne the function H (A) by: H (A) = [(γ1 − 1)A − γ1 L1 ](K2 /A)γ2 − [(γ2 − 1)A − γ2 L1 ](K2 /A)γ1 + (γ2 − γ1 )(K2 − L2 )

(14.5.3)

for all A such that L1 ≤ A ≤ D1 . By virtue of the fact that for the derivative of the function (14.5.3) the following expression holds:

(γ1 − 1)(γ2 − 1)(A − D1 ) K2 γ2 K2 γ1 K2 , and the conditions (14.5.5) and (14.5.6) are satisﬁed. Then, the value function of the problem (14.2.3) takes the expression (14.4.1) and the optimal stopping times σ∗ and τ∗ have the structure (14.2.5) and (14.2.6) with B∗ = K2 , where the function V (x; A, B) is explicitly given by

The Spread Option Optimal Stopping Game V

303

K1 − L1 = K2 − L2

V∗ (x)

L1 A∗

D1

K1

L2

B∗ = K2 D2 x

Figure 14.2 A computer drawing of the value function V∗ (x) and the optimal stopping boundaries A∗ and K2 V

K1 − L1 = K2 − L2

V∗ (x)

A∗ = L1

D1

K1

L2

B∗ = K2 D2 x

Figure 14.3 A computer drawing of the value function V∗ (x) and the optimal stopping boundaries L1 and K2

equation (14.3.12), and A∗ satisfying the inequalities L1 ≤ A∗ ≤ L1 (λ + r)/λ is determined as a unique solution of the equation (14.5.2) (see Figure 14.2). The veriﬁcation of this assertion can be carried out by means of a slight modiﬁcation of the arguments from the proof of Theorem 4.1, using also the facts that the condition (14.5.6) implies that V (K2 ; D1 , K2 ) < 1 and the function V (K2 ; A, K2 ) is increasing in A on the interval (L1 , D1 ). It is seen that the smooth-ﬁt condition at the point B∗ breaks down in this case. We also note that when the condition (14.5.5) fails to hold, almost the same arguments show that (even when the condition L1 (λ + r)/λ ≤ K1 fails to hold too) the assertion of Proposition 5.1 remains true with A∗ = L1 , while the smooth-ﬁt condition at A∗ also breaks down (see Figure 14.3). Remark 5.2. We also mention that another interesting but difﬁcult question is to present a complete description of the behavior of the optimal stopping boundaries A∗ and B∗ from equations (14.2.5) and (14.2.6) under the changing of the parameters of the model.

ACKNOWLEDGMENTS The author thanks Andreas E. Kyprianou and two anonymous Referees for valuable comments and suggestions.

304

Exotic Option Pricing and Advanced L´evy Models

REFERENCES [1] Alario-Nazaret, M., Lepeltier, J.P. and Marchal, B. (1982), “Dynkin games”, in Stochastic Differential Games (Bad Honnef), Lecture Notes in Control and Information Sciences, Vol. 43, Springer-Verlag, Berlin, Germany, pp. 23–32. [2] Baurdoux, E. and Kyprianou, A. (2004), “Further calculations for Israeli options”, Stochastics and Stochastic Reports, to be published. [3] Bensoussan, A and Friedman, A. (1974), “Non-linear variational inequalities and differential games with stopping times”, Journal of Functional Analysis, 16, 305–352. [4] Bensoussan, A and Friedman, A. (1977), “Nonzero-sum stochastic differential games with stopping times and free-boundary problems”, Transactions of the American Mathematical Society, 231, 275–327. [5] Bismut, J.M. (1977), “Sur un probl`eme de Dynkin”, Zeitschrift f¨ur Wahrscheinlichkeitstheorie und Verwandte Gebiete, 39, 31–53. [6] Cvitani´c, J. and Karatzas, I. (1996), “Backward stochastic differential equations with reﬂection and Dynkin games”, Annals of Probability, 24, 2024–2056. [7] Dynkin, E.B. (1969), “Game variant of a problem on optimal stopping”, Soviet Mathematical Doklady, 10, 270–274. [8] Elbakidze, N.V. (1976), “Construction of the cost and optimal policies in a game problem of stopping a Markov process”, Theory of Probability and its Applications, 16, 163–168. [9] Frid, E.B. (1969), “Optimal stopping of a Markov chain of two persons with opponent interests”, Theory of Probability and its Applications, 14, 746–749. [10] Friedman, A. (1973), “Stochastic games and variational inequalities”, Archive for Rational Mechanics and Analysis, 51, 321–346. [11] Friedman, A. (1976), Stochastic Differential Equations and Applications II, Academic Press, New York, NY, USA. [12] Grigelionis, B.I. and Shiryaev, A.N. (1966), “On Stefan’s problem and optimal stopping rules for Markov processes”, Theory of Probability and its Applications, 11, 541–558. [13] Jacod, J. (1979), Calcul Stochastique et Probl`emes de Martingales, Lecture Notes in Mathematics, Vol. 714, Springer-Verlag, Berlin, Germany. [14] Jacod, J. and Shiryaev, A.N. (1987), Limit Theorems for Stochastic Processes, Springer-Verlag, Berlin, Germany. [15] Kallsen, J. and K¨uhn, C. (2004), “Pricing derivatives of American and game type in incomplete markets”, Finance and Stochastics, 8, 261–284. [16] Kallsen, J. and K¨uhn, C. (2005), “Convertible bonds: ﬁnancial derivatives of game type”, in A.E. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 277–291. [17] Karatzas, I. and Wang, H. (2001), “Connection Between Bounded-Variation Control and Dynkin games”, in J.L. Menaldi et al. (Eds), Optimal Control and Partial Differential Equations (Volume in Honour of Alain Bensoussan), IOS Press, Amsterdam, The Netherlands, pp. 363–373. [18] Kifer, Yu.M. (1971a), “Optimal stopped games”, Theory of Probability and its Applications, 16, 183–188. [19] Kifer, Yu.M. (1971b), “Optimal stopping in games with continuous time”, Theory of Probability and its Applications, 16, 545–550. [20] Kifer, Y. (2000), “Game options”, Finance and Stochastics, 4, 443–463. [21] Krylov, N.V. (1971), “Control of Markov processes and W-spaces”, Izvestija, 5, 233–266. [22] K¨uhn, C. and Kyprianou, A. (2003a), “Callable puts as composite exotic options”, to be published.

The Spread Option Optimal Stopping Game

305

[23] K¨uhn, C. and Kyprianou, A. (2003b), “Pricing Israeli options: a pathwise approach”, to be published. [24] Kyprianou, A. (2004), “Some calculations for Israeli options”, Finance and Stochastics, 8, 73–86. [25] Lepeltier, J.P. and Mainguenau, M.A. (1984), “Le jeu de Dynkin en th´eorie g´en´erale sans l’hypoth`ese de Mokobodski”, Stochastics, 13, 25–44. [26] Morimoto, H. (1984), “Dynkin games and martingale methods”, Stochastics, 13, 213–228. [27] Neveu, J. (1975), Discrete-Parameter Martingales, North-Holland, Amsterdam, The Netherlands. [28] Protter, Ph. (1992), Stochastic Integration and Differential Equations: A New Approach, Applications of Mathematics, Vol. 21, Springer-Verlag, Berlin, Germany. [29] Shiryaev, A.N., Kabanov, Y.M., Kramkov, D.O. and Melnikov, A.V. (1994), “On the pricing of options of European and American types, II. Continuous time”, Theory of Probability and its Applications, 39, 61–102. [30] Shiryaev, A.N. (1978), Optimal Stopping Rules, Springer-Verlag, Berlin, Germany. [31] Shiryaev, A.N. (1999), Essentials of Stochastic Finance, World Scientiﬁc, Singapore. [32] Snell, J.L. (1952), “Applications of martingale systems theorems”, TAMS, 73, 293–312. [33] Stettner, L. (1982), “Zero-sum Markov games with stopping and impulsive strategies”, Applied Mathematics and Optimization, 9, 1–24.

Index ACCs see American contingent claims additive processes see time-inhomogeneous L´evy processes Albrecher, Hansj¨org 129–47 Almendral, Ariel 259–75 American contingent claims (ACCs) 282, 293–4 see also game options American options 5–16, 29, 31–6, 106, 113–14, 149–50, 195–215, 237–56, 259–75, 278–9 Asian options 217–34 CGMY process 238–9, 259–75 forward equations 237–56, 259–76 LCP 259–76 perpetual American options 16, 113–14, 195–215, 271 PIDEs 113–14, 237–56, 259–75 pricing 5–6, 10, 13–16, 29, 31–6, 106, 113–14, 149–50, 195–215, 237–56, 259–75 puts 5, 6, 13–14, 15–16, 31–2, 35–6, 195–215, 217–34, 239–56, 259–76 arbitrage 52, 105–6, 137, 145, 150, 170–92, 217–21, 231, 250–1, 277–86 classical theory 175, 217–18 concepts 175, 183–92, 217–21, 231 game options 277–86 market completion 183–92

arithmetic averages, early exercise Asian options 217–18 Arrow Debreu Securities 64 Asian options 10, 100, 114–19, 129–47, 217–34 see also early exercise. . . American type 217–34 concepts 114–19, 129–47, 217–34 optimal stopping problems 217–34 pricing 114–19, 129–47, 218–34 static super-hedging strategy 129–47 valuations 114–19, 129–47, 218–34 asset-or-nothing options 121–3 at-the-money options 130–45 autocorrelation, squared returns 58–9 average rate call options 39–41 average waiting time, investment decisions 156–65 backward equations 237–47, 293–4 backward free boundary problems 239–47 bankruptcies, convertible bonds 287 Barndorff-Nielsen–Shephard model (BN–S) 9, 31, 54, 67, 70–95, 132–43 barrier options 15–16, 29, 35, 40–8, 80–6, 115–17, 130 Barrieu, Pauline 149–68 basket options 117–19 bear markets 295 Bellamy, Nadine 149–68 Bermudan options 10, 31–2, 35–6, 114

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

308

Index

Bessel function 9–10, 24 beta distribution 37–42 bias correction algorithms 43–8 simulation methods 29, 42–8 bilateral Laplace transforms 111–13, 121–3 binned data, statistical density 58–9 Black-Scholes pricing model assumptions 4–5, 10, 29, 67–9, 106–7, 129, 178, 182 concepts 4–5, 10–14, 29, 67–9, 74, 106–7, 114–19, 129–30, 136–7, 178, 182, 237–9, 247, 263–4, 277, 283, 293–5 Israeli options 13–14, 293–4 Lagrange multipliers 137 SDE 182, 294–5 stochastic-volatility contrasts 129 ‘suicide’ strategies 283 Blumenthal 0–1 law 16 BN–S see Barndorff-Nielsen–Shephard model bonds 52, 169–92, 277–91 contingent claims 52 convertible bonds 277–91 counterparty default 52 Borel function 183, 222 bounded variation, path properties 12, 14–24, 103–4, 112, 136, 178–9, 271, 293–4 Boyarchenko, S.I. 1, 261–2, 271 Brennan, M.J. 150, 259–60, 269–70, 286–7 Brennan–Schwartz algorithm 259–60, 269–70 bridge algorithms concepts 29, 36–48, 117 Monte Carlo simulation 39–42, 117 stratiﬁed sampling 36–42 subordinator representation 37–48 Brownian motion 4–5, 10, 14–17, 30–48, 69–71, 102, 109–11, 122–3, 131–5, 150–1, 160–1, 177–219, 233–4, 259–64, 293–303 see also normal distributions; Wiener processes

market completion 177–92 stable processes 212 BS see Black-Scholes pricing model bull markets 295 c`adl`ag paths 278, 281–5 calibration, model risk 74–8 call options 39–41, 68–95, 106–14, 121–3, 136–45, 149–50, 196, 197–8, 217–34, 259–75, 295–303 callable put options 13–14 caps 99 Carr, Peter 31, 56–63, 68, 73, 74, 90, 106–7, 114, 116, 133–5, 139–40, 145, 170, 196, 215, 237–57, 259–60, 270–1, 275 Carr–Geman–Madan–Yor (CGMY) process 1–2, 4, 7–9, 21–3, 25, 54–65, 71–95, 133–5, 238–9, 259–75 see also generalized tempered stable processes; ‘inﬁnite activity’ L´evy processes American options 238–9, 259–75 change of measure density 54–65 characteristic function 260–2 concepts 1–2, 4, 7–9, 21–3, 25, 54–65, 71–95, 133–5, 238–9, 259–75 European options 271–5 jump models 54–65, 260–75 numerical valuation of American prices 263–71 path properties 21–3, 25, 271 risk-neutral densities 54–65, 261–2 statistical densities 54–65 cash ﬂows, NPV 149–50, 155–6 cash-or-nothing options 121–3 Cauchy sequence 11, 262, 264 CCGMYY processes 7 CGMY process see Carr–Geman–Madan–Yor. . . Chan, Terence 195–216 change of measure density see also Radon–Nikodym derivative concepts 52–65, 99, 105–6, 108–11, 122–3, 217–19 estimation details 57–63

Index

character function inversion, FFT usage 138–9 chi-squared goodness of ﬁt statistic 59 CIR see Cox–Ingersoll–Ross process circulant convolutions, concepts 267–70 classical theory, arbitrage 175, 217–18 cliquet options 68, 81–8, 95, 130 comonotonicity theory, concepts 129–30, 137–47 complete markets 169–92, 280–1 composite trapezoidal rule, spatial discretizations 266–7 compound options 238 compound Poisson process concepts 4–5, 11–12, 17, 30–3, 55–6, 69–70, 132–5, 150–1, 160–1, 177–8, 185–6, 204–6, 264 path properties 17 computer drawings, optimal stopping problems 299, 303 Cont, R. 1, 259–61, 264, 266 contingent claims 51–65, 169–92, 277–89, 294–303 continuation region, exercise boundary 241–5 continuous barrier options 40–8 continuous junction condition, concepts 203–4 continuous-time setting, market completion 169–70, 177–92 continuous/discontinuous models, investment decisions 161–5 continuously reset path-dependent options, valuations 40–8 convertible bonds 277–91 see also game options bankruptcies 287 concepts 277–91 coupon payments 287–9 deﬁnition 277, 287 ﬁrm value 286–9 literature 286–9 perpetual model 286–9 reduced form models 289 convolutions, circulant convolutions 267–70 Corcuera, Jos´e Manuel 169–93

309

correction algorithms, simulation bias 43–8 counterparty default, bonds 52 coupon payments, convertible bonds 287–9 course path properties, L´evy processes 1–2, 10–28 Cox–Ingersoll–Ross process (CIR) 69–95, 131–44 see also stochastic clocks creeping, path properties 1, 14–24 critical stock prices, DEVG model 240–2, 246–50

DAX 61–4 deﬁnitions, L´evy processes 2–4 Delbaen, F. 279–85 delta 130, 241–2, 274 DEVG see Diffusion Extended Variance Gamma DFT see discrete Fourier transform DIB see down-and-in barrier options Diffusion Extended Variance Gamma (DEVG) 241–56 digital barriers 68, 80–1, 82–6, 111 Dirac measure 152, 232 Dirichlet conditions 242, 244–9 discontinuous L´evy processes, real options 155–6 discontinuous martingales, stable processes 211–14 discontinuous models investment decisions 161–5 real options 155–6 discount rates investment decisions 149–66 perpetual American options 197–8, 203–4 discounted payoff function, moment derivatives 176 discrete Fourier transform (DFT) 267–70 discrete LCP, concepts 269–70 discrete-time setting, market completion 169–92 discretely reset path-dependent options, valuations 39–40

310

Index

discretization ﬁnite differences 251–6, 260, 263–75 forward equations 251–6, 263–75 distributional characteristics concepts 4–5, 29, 54–65, 195, 197–202 normal distributions 29–30, 133–5 skewed distributions 29–30, 53–65, 86–95, 169, 180 dividends, convertible bonds 287–9 DOB see down-and-out barrier options Doob–Meyer decomposition 183–4 down-and-in barrier options (DIB) 81–6 down-and-out barrier options (DOB) 80–6 drift 4–5, 17, 33–6, 43–4, 54–65, 102, 240–2, 260–2 Dupire equation 237–9 dynamic hedging 130 dynamic programming 150 dynamic trading strategies 68, 91–5, 130–47 Dynkin’s games 278, 280–1, 288–9, 293–4 see also optimal stopping problems

early exercise Asian options see also Asian options concepts 217–34 numerics 231–3 optimal stopping boundary 217–34 premiums 220–1 pricing 218–34 probability density function 231–4 problem formulation 218–20 proof 220–31 Eberlein, Ernst 31, 99–128, 184–5, 238 efﬁcient markets 52 EMM see equivalent martingale measures enlargement, L´evy market model 179–82 equity indexes 52, 60–5, 130–1, 140–4 equivalent martingale measures (EMM) 183–92, 195–6, 261–2 Esscher measure 191 estimation details, change of measure density 57–63 Euclidean scalar product 101 Euler’s theorem 246–7

European options 29, 31, 35, 67–95, 99, 106–14, 123, 130, 136–45, 237–9, 259, 261–2, 270–5, 278–9, 288–9 calls 68–95, 106–14, 123, 136–45 CGMY 271–5 forward equations 238–9, 270–5 pricing 29, 31, 35, 67–95, 99, 106–14, 123, 130, 136–45, 237–9, 259, 261–2, 270–5 puts 106–14, 261–2, 270–5 Eurostoxx 50 index 75–8, 82–6 exceedance probabilities, barrier options 44–8 excursion theory, concepts 14–15, 113–14, 116 exotic options see also individual types concepts 1–28, 80–95, 114–19 model risk 67–97, 131 path dependency 10, 29, 31–2, 39–48, 67–95, 113, 197 pricing 74–8, 80–95, 99–123, 129–47, 195–215, 218–34 super-hedging strategy 129–47 symmetries 99–123 types 10, 13–14, 15–16, 80–6, 95, 99–100, 114–19 explicit ﬁnite-difference methods 114 exponential L´evy processes 101–23, 130, 178–9, 189–92, 196–7, 264 exponential PIIAC, time-inhomogeneous L´evy processes 101–23 fast Fourier transform (FFT) 74–5, 129–30, 138–45, 209–14, 259–76 Fatou’s lemma 301 FBPs see free boundary problems FFT see fast Fourier transform ﬁnancial mathematics, objectives 29 ﬁne path properties, L´evy processes 1–2, 10–28 ﬁnite expiry American puts 16 ﬁnite markets market completion 172–92 multi-step markets 173–4 one-step markets 172–3 ﬁnite moment logstable model 238–9

Index

ﬁnite variation, L´evy processes 200–13, 260 ﬁnite-difference methods 113–14, 237–57, 259–76 ﬁrm value, convertible bonds 286–9 ﬁrst-passage distributions 195, 197–202 ﬁxed strike Asian options 100, 114–19 ﬁxed strike lookback options 42–8, 100, 115–19 ﬂoating strike Asian options 100, 114–19, 217–34 see also Asian options ﬂoating strike lookback options 42–8, 100, 115–19 ﬂuctuation theory, concepts 10 foreign exchange 102–23 forward equations American options 237–56, 259–76 concepts 232–3, 237–76 DEVG model 241–56 discretization 251–6, 263–75 European options 238–9 hybrid equations 239–56 initial-value problem 232–3, 263 uses 237–8 forward free boundary problems 247–56, 262–71 forward-start options 114–19 Fourier transform methods 2, 31, 56, 57–60, 74–5, 112, 116–17, 129–30, 138–45, 209–14, 259–76 free boundary problems (FBPs) 239–51, 262–71, 293–303 see also optimal stopping. . . concepts 239–51, 262–71, 293–303 solution 296–302 FTSE 61–4 futures prices 51–5 g-moment, PIIAC 103–4 game contingent claim (GCC), concepts 277–89, 294 game options 10, 13–14, 277–303 see also convertible bonds; Israeli. . . arbitrage 277–86 concepts 10, 13–14, 277–303

311

deﬁnition 277–8, 293–4 NFLVR 278–85 optimal stopping problems 278, 280–1, 288–9, 293–4 pricing 277–89 spread game options 293–303 gamma process 1–2, 4, 8–9, 10, 22–3, 25, 32–48, 56–65, 70–95, 117, 132–5, 142–5, 209–14, 242 Gamma-OU stochastic clock 73–4, 79–80, 132–3, 135–44 see also stochastic clocks Gapeev, Pavel 293–305 GARCH models 251 Gaussian processes see also normal distributions concepts 1–2, 9–10, 15, 17, 21, 23–4, 29, 32–48, 53–4 GCC see game contingent claim Geman, H´elyette 51–66, 99, 105, 259 see also Carr–Geman–Madan–Yor process general diffusion model 106 generalized hyperbolic processes see also normal inverse Gaussian. . . concepts 1–2, 5, 9–10, 23–4, 31–48, 54, 71–95, 100–23, 133–44 path properties 23–4 variance gamma process 10 generalized inverse Gaussian distributions (GIG) 33–6 generalized tempered stable processes see also Carr–Geman–Madan–Yor process; truncated. . . ; variance gamma. . . concepts 1–2, 4, 7–9, 19–23, 25, 32–48, 54–65, 71–95, 260–2 path properties 19–23, 25 geometric averages, early exercise Asian options 217–18 geometric Brownian motion 42–3, 71–4, 218–19, 293–303 geometric L´evy model see exponential L´evy processes Gerber, H.U. 196–8, 203–6 German equity indexes, risk-neutral densities 60–5

312

Index

GIG see generalized inverse Gaussian distributions Girsanov’s theorem 106–11 the greeks 238 half lines 6, 16–24 HARA utilities 187–8 Hartman–Watson density 231–2 Heaviside function 255 hedging 29, 57, 68, 90–5, 118–19, 129–47, 169–92 Asian options 129–47 concepts 90–5, 129–47, 169–92 moment derivatives 90–5, 169–92 moment swaps 90–5 static super-hedging strategy 129–47 strategy performance 140–4 Hellinger distances, densities 59 Heston Stochastic Volatility model (HEST) 67–8, 69–95, 131–43 concepts 67–8, 69–70, 82–6, 131–43 jumps 69–70, 132–3 high contact condition, concepts 241–2 Hilbert space 11 Hirsa, Ali 237–57 hitting points concepts 12–14, 17–24, 222–3 path properties 12–14, 17–24 holders, game options 277–91, 293–4 Hunt density 240–2 hybrid equations, forward equations 239–56 hyperbolic process, concepts 9–10, 118–19, 238–9 IBEX 61–4 IDFT see inverse discrete Fourier transform implicit ﬁnite-difference methods 113–14 implicit function theorem 298–301 implied volatilities 29–30, 67–95, 99–123 in-progress Asian options 115 in-the-money options 144–5, 149–50 incomplete markets 169–92, 294 independent increments, L´evy processes 2–4, 100–23, 237–9, 245–7, 260–1

‘inﬁnite activity’ L´evy processes see also Carr–Geman–Madan–Yor process concepts 54–65, 261–2, 264 inﬁnite variation, L´evy processes 200–13 inﬁnitely divisible distributions 2–4, 9, 54–65, 68–74, 131, 177–92 initial-value problem, forward equations 232–3, 263 inner expectations, lattice methods 35–6 instantaneous returns, jump perspectives 58–9 instantaneous volatility 237–56 insurance premiums 54–65 integral equations early exercise Asian options 217–34 PIDEs 113–19, 237–56, 259–75 interest rates models 29–33 risk-free interest rates 75, 174–5, 260–1 simulation methods 29–30, 32–48 inverse discrete Fourier transform (IDFT) 268–70 Inverse Gaussian (IG) random numbers 78 see also normal inverse. . . inverse transform 36–40, 268–70 investment decisions see also real options average waiting time 156–65 concepts 149–65 continuous/discontinuous models 161–5 discount rates 149–66 opportunity value 154–5, 163–6 optimal discount rates 149–50, 156–66 optimal times 150–65, 166, 197, 203–13 optimization 149–66, 197, 203–6 proﬁts/costs ratio 149–65, 167 random jump sizes 160–1, 166–7 relative errors 162–5 robustness checks 158–65 Israeli options 10, 13–14, 293–4 see also game. . . issuers, game options 277–91, 293–4 Itˆo’s formula 218–19, 262, 300 Itˆo –Tanaka–Meyer formula 300

Index

Jacod, J. 278, 300–1 Japanese equity indexes, risk-neutral densities 60–5 joint returns distributions 29–30, 59 jump perspectives concepts 4, 11–12, 51–65, 68–70, 89, 102, 113–17, 149–65, 177–92, 196–8, 238–56, 260–2 HEST 69–70, 132–3 instantaneous returns 58–9 market crises 149–65 optimal discount rates 156–65 perpetual American options 196–8 risks in returns 51–65 jump-diffusion models, concepts 1–2, 4–5, 11–12, 17, 53–65, 116–17, 149–65, 238–56, 260–2 Kallsen, Jan 277–91, 294 Karatzas, I. 277, 280–1, 293 Kifer, Y. 277, 295 KoBoL processes see generalized tempered stable processes Kolmogorov–Smirnov statistic 59 Kolodner, I.I. 217 Kou model 1–2, 5, 260–1 see also jump-diffusion model Kou, S.G. 1–2, 5, 260–1, 277 K¨uhn, Christoph 277–91, 293–4 kurtosis levels 53–4, 86–95, 169, 180 Kyprianou, Andreas E. 1–28, 259, 277, 293–5 Lagrange multipliers 137, 187 Laplace transforms 111–13, 121–3, 149, 151, 153–65, 174, 197, 199–202, 205, 209–14 investment decisions 149, 151, 153–65 relative errors 162–5 lattice methods, concepts 31–3, 35–6 law of a ﬁrst-passage time of the process 195, 197 LC see lookback options LCPs see linear complementary problems Lebesgue measure 112, 200–2, 278, 300–1 Lepeltier, J. 280, 295

313

Levendorskii, S.Z. 1, 261–2, 271 L´evy exponent, concepts 198–214 L´evy measure, concepts 198–214 L´evy processes see also individual classes; stochastic processes bias 42–8 bridge algorithms 29, 36–48, 117 change of measure density 53–65, 99, 105–6, 108–11, 122–3 classes 1–2, 4–10, 17–24, 48, 54, 71–2, 100, 133–5, 195–215, 259–75 concepts 1–48, 53–65, 71–2, 99–123, 131–5, 169–92, 195–215, 237–56, 259–75 deﬁnitions 2–4, 103, 131, 177, 202–4, 260–1 examples 1, 4–25 exponential L´evy processes 101–23, 130, 178–9, 189–92, 196–7, 264 ﬁnite variation 200–13, 260 ﬂuctuation theory 10 geometric L´evy model 178–9 independent increments 2–4, 100–23, 237–9, 245–7, 260–1 ‘inﬁnite activity’ L´evy processes 54–65, 261–2, 264 introduction 1–48 market completion 169–92 model risk 67–97 modelling 1–2, 4–10, 17–24, 30–48, 53–65, 67–97, 202–4 moment derivatives 67–8, 86–95, 169–92 path properties 1–28, 79–80, 103–4, 112, 136, 178–9, 271 problems 10 random walks 116–17 real options 151–5 risk-neutral densities 53–65, 68–78, 89–90, 93–5, 101–23, 131–45, 170–92, 195–7, 202–4, 240–56 simulation methods 29–30, 31–48, 67–8, 72–95, 114, 117–18, 133–44

314

Index

L´evy processes (Continued) stationary independent increments 2–4, 103, 237, 242–56, 260–1 statistical densities 53–65 stochastic time 71–95, 131, 133–5 symmetries 99–123, 237–56 theorems 2–4 time-changed L´evy process 73–4, 78, 79–80, 93–5, 133–44 time-inhomogeneous L´evy processes 99–123, 245–56 L´evy triple, concepts 3–28, 30–3, 100–11, 119–20, 183–4, 278 L´evy–Itˆo composition, concepts 11–12 L´evy–Khintchine formula, concepts 1, 2–4, 6, 7–8, 10–25, 54–65, 104–5, 177–92, 198–202, 260–1 light tails, distributional characteristics 4–5 linear complementary problems (LCPs) 259–76 Lipschitz constant 286 local time-space calculus 217–34 local volatilities 53–4 Loeffen, Ronnie 1–28 log returns 53–4, 67–8, 90–5, 169–92 lookback options (LC) 10, 40–8, 68, 80–6, 95, 100, 115–19, 130 lower half line regularity 6, 16–24 spectrally one-sided processes 6, 17 Madan, Dilip B. 51–66, 74, 116–17, 140, 170, 237, 238, 241, 259, 270–1, 275 see also Carr–Geman–Madan–Yor process Maingueneau, M. 280, 295 management tools, real options 150 Margrabe-type options 119–23 market completion 169–92, 280–1 continuous-time setting 169–70, 177–92 discrete-time setting 169–92 moment derivatives 169–92 market crises, real options 149–65 Markov process 59, 196–7, 219, 222, 228–9, 239–40, 250, 293–9 martingale representation property (MRP) 180–2

martingales 7, 11–12, 30–63, 100–23, 150–1, 169–92, 195–6, 202–4, 211–14, 221–2, 228–31, 240–2, 278–303 compound Poisson process 11–12, 55–6, 150–1, 185–6 early exercise American options 221–2, 228–31 equivalent martingale measures 183–92, 195–6, 261–2 game options 278–89 Meixner processes 7 moment derivatives 169–92 optimal stopping problems 196, 202–4, 211–14, 221–2, 228–31, 240–2, 278–89, 293–303 semi-martingales 33–6, 101–23, 278–89 Matache, A.M. 259, 262, 271 mean-variance mixtures 31, 32–3, 192 Meixner processes concepts 1–2, 5, 6–7, 17–18, 25, 54, 71–2, 103–4, 133–5 path properties 17–18, 25 memoryless property, exponential L´evy processes 116, 196–7 Merton model 1–2, 5 see also jump-diffusion model Miller–Modigliani hypothesis 287 minimal entropy martingale measure 102 minimal martingale measures 192 minimax martingale measures 102 model risk calibration 74–8 exotic options 67–97, 131 model-independent static super-hedges 145 moment derivatives concepts 67–8, 86–95, 169–92 hedging 90–5, 169–92 market completion 169–92 pricing 86–95 moment options (MOMO) 89–95 moment swaps (MOMS) hedging 90–5 pricing 89–95 Mongolian options 10 Monte Carlo simulation

Index

bridge algorithms 39–42, 117 concepts 31–3, 35–6, 39–40, 67–8, 78–95, 114, 117–18, 142–5 NIG 39–42, 78–95, 117–18, 142–4 problems 36 simulation bias 42–8 stratiﬁed sampling 39–42 VG 39–42, 79–95, 117, 142–4 MRP see martingale representation property multi-step ﬁnite markets, market completion 173–4 net present value (NPV) concepts 149–50, 155–6 weaknesses 149 NFLVR see no free lunch with vanishing risk NIG see normal inverse Gaussian processes NIKKEI 61–4 no free lunch with vanishing risk (NFLVR) 278–85 no-arbitrage pricing, game options 279–86 nonlinear integral equations, early exercise Asian options 217–34 normal distributions 29–30, 133–5 see also Brownian motion; Gaussian. . . normal inverse Gaussian processes (NIG) see also generalized hyperbolic processes concepts 1–2, 9–10, 30, 32–48, 54, 71–95, 117–18, 133–44 Monte Carlo simulation 39–42, 78–95, 117–18, 142–4 simulation methods 32–48, 72–95, 117–18, 133–44 NPV see net present value Nualart, David 169–93 numerical approach, simulation methods 33–6, 113–14 one-dimensional driving processes, symmetries 121–3 one-side L´evy processes, two-sided L´evy processes 215 one-step market models 170–92

315

one-touch barriers 68, 80–6, 95 opportunity value, investment decisions 154–5, 163–6 optimal discount rates 149–50, 156–66 optimal portfolios, concepts 186–92 optimal stopping problems 114, 196–215, 217–34, 237–56, 259–76, 278, 280–1, 286–9, 293–303 see also Dynkin’s games; free boundary. . . computer drawings 299, 303 early exercise Asian options 217–34 forward equations 232–3, 237–76 game options 278, 280–1, 288–9, 293–4 perpetual American options 16, 113–14, 195–215, 271 spread game options 293–303 value function 294–303 optimal times, investment decisions 150–65, 166, 197, 203–13 optimal wealth, concepts 187–92 optional sampling theorem 223–31, 301 options see American. . . ; European. . . ; exotic. . . ; game. . . ; real. . . Ornstein Uhlenbeck process (OU) 70–1, 73–95, 132–44 orthonormal martingales 179–82 OU see Ornstein Uhlenbeck process out-of-the-money options 59–63, 130–45 outer expectations, lattice methods 35–6 Papapantoleon, Antonis 99–128 Parisian options 10 partial differential equations (PDEs) 31, 182, 237–57 partial integro-differential equations (PIDEs) 113–19, 237–56, 259–75 American options 113–14, 237–56, 259–75 concepts 113–19, 237–56, 259–75 forward equations 237–56, 259–75 hybrid equations 242–56 partial integro-differential inequality (PIDI) 113–19 passport options 278

316

Index

path dependency, exotic options 10, 29, 31–2, 39–48, 67–95, 113, 197 path properties bounded/unbounded variation 12, 14–24, 103–4, 112, 136, 178–9, 271, 293–4 concepts 1–2, 10–24 L´evy processes 1–28, 79–80, 103–4, 112, 136, 178–9, 271 types 10–24 path variation, path properties 10–24 PDEs see partial differential equations perpetual American options 16, 113–14, 195–215, 271, 293–4 concepts 16, 113–14, 195–215, 271, 293–4 discount rates 197–8, 203–4 jump perspectives 196–8 pricing 113–14, 195–215, 271 renewal equation pricing approach 204–6 spectrally one-sided processes 195–215 perpetual convertible bonds, concepts 286–9 perpetual spread game options 293–303 perpetual warrants 196 Peskir, Goran 217–35 PIDEs see partial integro-differential equations PIDI see partial integro-differential inequality PIIAC see process with independent increments and absolutely continuous characteristics PIIS see process with independent and stationary increments Poisson process, concepts 4–5, 11–12, 17, 30–3, 55–6, 69–71, 132–5, 150–1, 160–1, 177–8, 185–6, 204–6, 264 portfolios, optimal portfolios 186–92 power options concepts 106–11, 174–92 symmetries 106–11 power-jump processes, concepts 178–9 power-return assets, market completion 174–92

premiums creeping 15–17 early exercise Asian options 220–1 insurance premiums 54–65 risks 51–65 pricing see also valuation methods American options 5–6, 10, 13–16, 29, 31–6, 106, 113–14, 149–50, 195–215, 237–56, 259–75 Asian options 114–19, 129–47, 218–34 Black-Scholes pricing model 4–5, 10, 13–14, 29, 67–9, 74, 293 early exercise Asian options 218–34 European options 29, 31, 35, 67–95, 99, 106–14, 123, 130, 136–45, 237–9, 259, 261–2, 270–5 exotic options 74–8, 80–95, 99–123, 129–47, 195–215, 218–34 forward equations 232–3, 237–56 game options 277–89 GCC 277–89 moment derivatives 86–95 perpetual American options 113–14, 195–215, 271 renewal equation approach 204–6 spread game options 293–303 swaps 89–95 symmetries 21–2, 99–123 vanilla options 106–14, 121–3, 261–2 principle of smooth pasting 155–6, 206 process with independent increments and absolutely continuous characteristics (PIIAC), time-inhomogeneous L´evy processes 100–23 process with independent and stationary increments (PIIS) 103, 237, 242–56, 260–1 see also L´evy processes proﬁts/costs ratio, investment decisions 149–65, 167 put options 5, 6, 13–16, 31–2, 35–6, 100, 106–14, 121–3, 195–234, 239–56, 259–79, 284, 288–9 put–call parity 100 p–value, chi-squared goodness of ﬁt statistic 59

Index

qq–plots 59 quadratic utility 191–2 quanto options 110–11, 119–23 Radon–Nikodym derivative 52, 105–6, 108–11, 122–3 see also change of measure density Raible, Sebastian 111–13, 262 random jump sizes, investment decisions 160–1, 166–7 random numbers, simulations 78–9 random walks, L´evy processes 116–17 real options see also investment decisions characteristics 155–6 concepts 149–65 L´evy processes 151–5 management tools 150 market crises 149–65 models 151–5 optimal discount rates 149–50, 156–66 optimal times 150–65, 166 real valued L´evy processes, deﬁnitions 2–4 reduced form models, convertible bonds 289 regular L´evy processes of exponential type (RLPE) 117 regularity of the half line, path properties 6, 16–24 relative errors, Laplace transforms 162–5 renewal equation approach, pricing 204–6 returns jump models 51–65 models 29–33 risks 51–65 simulation methods 29–30, 32–48, 71–95 Ribeiro, C. 37–48 risk management 29, 57 risk-free interest rates 75, 174–5, 260–1 risk-neutral densities CGMY process 54–65, 261–2 concepts 51–65, 68–78, 89–90, 93–5, 101–2, 131–47, 170–92, 195–7, 202–4, 240–2, 261–2 equity indexes 60–5, 68–78 estimation details 57–63, 68–78, 131

317

L´evy processes 53–65, 68–78, 89–90, 93–5, 101–23, 131–45, 170–92, 195–7, 202–4, 240–56, 261–2 stochastic volatility models 131–47 risks jump models 51–65 model risk 67–97, 131 NFLVR 278–85 premiums 51–65 returns 51–65 two-sided features 65 RLPE see regular L´evy processes of exponential type robustness checks, investment decisions 158–65 ruin theory 196 Russian options 6, 10, 15–16, 215 Rydberg, T.H. 31 S&P 500 130–1, 140–4 saddle point, optimal stopping game 295 Sato process 170 Schachermayer, W. 279–85 Schoutens, Wim 1, 7, 53, 54, 67–97, 129–47, 169–93, 260 Schwartz, E.S. 259, 269–70, 286–7 SDE see stochastic differential equations second moment swaps see variance swaps securities, game options 278–89 self-ﬁnancing trading strategies 118, 182 self-quanto options 110–11 semi-heavy tails, distributional characteristics 4–5 semi-martingales 33–6, 101–23, 278–89 Shiryaev, A. 278, 296–7, 301 Simons, Erwin 67–97 simulation methods bias 29, 42–8 bridge algorithms 29, 36–48, 117 concepts 29–30, 32–48, 67–8, 72–95, 118, 133–44 continuously/discretely reset path-dependent options 39–48 L´evy processes 29–30, 31–48, 67–8, 72–95, 114, 117–18, 133–44 Monte Carlo simulation 31–3, 35–6, 39–40, 67–8, 114, 117–18, 142–4

318

Index

simulation methods (Continued) numerical approach 33–6, 113–14 speed-up methods 36–48 Sirbu, M. 286–9 skewed distributions 29–30, 53–65, 86–95, 169, 180 smiles 29–30, 67–95, 99–128, 131–47, 237–9 ‘smooth ﬁt’ conditions, concepts 241–50, 259–60, 271, 293–303 smooth pasting principle 155–6, 206 Spanish equity indexes, risk-neutral densities 60–5 spatial discretizations, CGMY process 266–70 spectrally one-sided processes concepts 1–2, 5, 6, 17–18, 195–215 ﬁrst-passage distributions 195, 197–202 path properties 17–18 perpetual American options 195–215 stable of index 210–14 speed-up methods, simulation methods 36–48 spread game options concepts 293–303 deﬁnition 295 SPX 61–4 squared returns, autocorrelation 58–9 stable of index, spectrally one-sided processes 210–14 static hedging algorithm 140 Asian options 129–30, 136–47 concepts 129–30, 136–47 model-independent super-hedges 145 performance issues 140–4 static positions 68, 91–5, 129–47 stationary independent increments, L´evy processes 2–4, 103, 237, 242–56, 260–1 statistical densities CGMY process 54–65 concepts 51–65 estimation details 57–63 L´evy processes 53–65 stochastic calculus 30–3, 69–71, 79–95, 178–9, 222–31, 240–2, 293, 294–5

stochastic clocks 72–95, 133–44 see also Cox–Ingersoll–Ross. . . ; Gamma-OU. . . stochastic differential equations (SDEs) 30–3, 69–71, 79–95, 178–9, 240–2, 293, 294–5 stochastic processes 1–48, 102–3, 131–5, 178–9, 260–1, 277–89 see also L´evy processes stochastic time, L´evy processes 71–95, 131, 133–5 stochastic volatility 4, 29–33, 58–9, 67–95, 117, 129–47 Black-Scholes contrasts 129 concepts 4, 29–33, 58–9, 67–95, 117, 129–47 models 129–30, 131–47 numerical implementation 138–44 super-hedging strategy 129–47 stocks 52, 60–5, 130–1, 140–4, 169–92, 202–15, 277–89, 295–303 stop-loss transforms, concepts 129–30, 137–47 stopping region, exercise boundary 241–5 stratiﬁed sampling bridge algorithms 36–42 concepts 36, 39–42 Monte Carlo simulation 39–42 sub-optimal strategies 149–50 submultiplicative function, PIIAC 103–4 subordinated Brownian motion 31–3, 34–48 subordinator representation, bridge algorithms 37–48 subordinators, concepts 31–3, 34–48, 68, 70–2, 131–5, 199–200 ‘suicide’ strategies, Black-Scholes pricing model 283 super-hedging strategy, Asian options 129–47 supermartingales, game options 281–6 surveys, valuation methods 99–123 swaps 68, 89–95, 169–71 moment swaps 89–95 pricing 89–95 variance swaps 89, 169–71 swaptions 99

Index

swing-options 278 symmetries 21–2, 99–123, 237–56 concepts 21–2, 99–123 deﬁnition 99–100 exotic options 99–123 Margrabe-type options 119–23 power options 106–11 vanilla options 106–14, 121–3 tails distributional characteristics 4–5, 29, 54–65 insurance claims 57–63 Tankov, P. 1, 260–1 Taylor expansion 90, 169, 170–3, 263–4 term structure of smiles, concepts 99–123 theorems, L´evy processes 2–4 time-changed L´evy process 73–4, 78, 79–80, 93–5, 133–44 time-inhomogeneous L´evy processes concepts 99–123, 245–56 model 100–4 Tistaert, Jurgen 67–97 Toeplitz matrix 268–9 trading strategies dynamic trading strategies 68, 91–5, 130–47 game options 278–89, 295 self-ﬁnancing trading strategies 118, 182 transaction costs 130 Trigeorgis, L. 150 trinomial market model 169–92 truncated stable processes see also generalized tempered. . . concepts 1–2, 4, 19–23 two-agent models 51–65 two-dimensional asset-or-nothing options 121–3 two-dimensional driving processes, symmetries 121–3 two-sided features, risks 65 two-sided L´evy processes, one-side L´evy processes 215 UIB see up-and-in barrier options UK equity indexes, risk-neutral densities 60–5

319

unbounded variation, path properties 12, 14–24 UOB see up-and-out barrier options up-and-in barrier options (UIB) 29, 42–8, 81, 84–6 up-and-out barrier options (UOB) 29, 42–8, 81, 84–6 upper half line, regularity 6, 16–24 USA equity indexes, risk-neutral densities 60–5 utility theory 51–65, 186–92 Uys, Nadia 217–35 valuation methods see also pricing surveys 99–123 value at risk (VaR) 29, 52 value function, optimal stopping game 294–303 value matching condition, exercise boundary 241–2 Vandermonde matrices 186 vanilla options 10, 67, 74–8, 99–100, 106–14, 121–3, 129–30, 140–4, 261–2 pricing 106–14, 121–3, 261–2 symmetries 106–14, 121–3 VaR see value at risk variance gamma process (VG) see also generalized tempered stable processes change of measure densities 56–65 concepts 1–4, 8–10, 22–5, 32–48, 56–65, 71–95, 117, 133–44, 237–9, 241–50, 259–62 DEVG model 241–56 generalized hyperbolic processes 10, 117 Monte Carlo simulation 39–42, 79–95, 117, 142–4 simulation methods 32–48, 72–95, 133–44 variance swaps 89, 169–71 VG see variance gamma process volatility smiles 29–30, 67–95, 99–128, 131–47, 237–9 volatility surface, concepts 99–100 Voltchkova, E. 259, 264, 266

320

Index

waves, Fourier transform methods 2, 31, 56, 57–60, 74–5, 112, 116–17, 129–30, 138–45, 209–14, 259–76 Webber, N. 37–42 Wiener processes 31, 240–2 see also Brownian motion Wiener–Hopf factorization 14–15, 113–14, 116–17

writers see issuers Yor, M. 259 see also Carr–Geman–Madan–Yor process zero-sum Dynkin stopping game see Dynkin’s games

Exotic Option Pricing and Advanced L´evy Models

Edited by

Andreas E. Kyprianou, Wim Schoutens and Paul Wilmott

Copyright 2005

John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England Telephone (+44) 1243 779777

Email (for orders and customer service enquiries): [email protected] Visit our Home Page on www.wiley.com All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1T 4LP, UK, without the permission in writing of the Publisher. Requests to the Publisher should be addressed to the Permissions Department, John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex PO19 8SQ, England, or emailed to [email protected], or faxed to (+44) 1243 770620. This publication is designed to provide accurate and authoritative information in regard to the subject matter covered. It is sold on the understanding that the Publisher is not engaged in rendering professional services. If professional advice or other expert assistance is required, the services of a competent professional should be sought.

Other Wiley Editorial Ofﬁces John Wiley & Sons Inc., 111 River Street, Hoboken, NJ 07030, USA Jossey-Bass, 989 Market Street, San Francisco, CA 94103-1741, USA Wiley-VCH Verlag GmbH, Boschstr. 12, D-69469 Weinheim, Germany John Wiley & Sons Australia Ltd, 42 McDougall Street, Milton, Queensland 4064, Australia John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809 John Wiley & Sons Canada Ltd, 22 Worcester Road, Etobicoke, Ontario, Canada M9W 1L1 Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic books.

British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library ISBN-13 978-0-470-01684-8 ISBN-10 0-470-01684-1 Typeset in 10/12pt Times by Laserwords Private Limited, Chennai, India Printed and bound in Great Britain by Antony Rowe Ltd, Chippenham, Wiltshire This book is printed on acid-free paper responsibly manufactured from sustainable forestry in which at least two trees are planted for each one used for paper production.

Contents

Contributors

xi

Preface

xiii

About the Editors

xvii

About the Contributors

xix

1 L´evy Processes in Finance Distinguished by their Coarse and Fine Path Properties Andreas E. Kyprianou and R. Loeffen 1.1 1.2 1.3

Introduction L´evy processes Examples of L´evy processes in ﬁnance 1.3.1 Compound Poisson processes and jump-diffusions 1.3.2 Spectrally one-sided processes 1.3.3 Meixner processes 1.3.4 Generalized tempered stable processes and subclasses 1.3.5 Generalized hyperbolic processes and subclasses 1.4 Path properties 1.4.1 Path variation 1.4.2 Hitting points 1.4.3 Creeping 1.4.4 Regularity of the half line 1.5 Examples revisited 1.5.1 Compound Poisson processes and jump-diffusions 1.5.2 Spectrally negative processes 1.5.3 Meixner process 1.5.4 Generalized tempered stable process 1.5.5 Generalized hyperbolic process 1.6 Conclusions References

1

1 2 4 5 6 6 7 9 10 10 12 14 16 17 17 17 17 19 23 24 26

vi

Contents

2 Simulation Methods with L´evy Processes Nick Webber 2.1 2.2

Introduction Modelling price and rate movements 2.2.1 Modelling with L´evy processes 2.2.2 Lattice methods 2.2.3 Simulation methods 2.3 A basis for a numerical approach 2.3.1 The subordinator approach to simulation 2.3.2 Applying the subordinator approach 2.4 Constructing bridges for L´evy processes 2.4.1 Stratiﬁed sampling and bridge methods 2.4.2 Bridge sampling and the subordinator representation 2.5 Valuing discretely reset path-dependent options 2.6 Valuing continuously reset path-dependent options 2.6.1 Options on extreme values and simulation bias 2.6.2 Bias correction for L´evy processes 2.6.3 Variation: exceedence probabilities 2.6.4 Application of the bias correction algorithm 2.7 Conclusions References 3 Risks in Returns: A Pure Jump Perspective H´elyette Geman and Dilip B. Madan 3.1 3.2 3.3

Introduction CGMY model details Estimation details 3.3.1 Statistical estimation 3.3.2 Risk neutral estimation 3.3.3 Gap risk expectation and price 3.4 Estimation results 3.4.1 Statistical estimation results 3.4.2 Risk neutral estimation results 3.4.3 Results on gap risk expectation and price 3.5 Conclusions References 4 Model Risk for Exotic and Moment Derivatives Wim Schoutens, Erwin Simons and Jurgen Tistaert 4.1 4.2

Introduction The models 4.2.1 The Heston stochastic volatility model 4.2.2 The Heston stochastic volatility model with jumps

29

29 30 30 31 32 33 34 35 36 36 37 39 40 42 43 44 45 48 48 51

51 54 57 58 59 60 60 61 61 61 63 65 67

67 68 69 69

Contents

4.2.3 The Barndorff-Nielsen–Shephard model 4.2.4 L´evy models with stochastic time 4.3 Calibration 4.4 Simulation 4.4.1 NIG L´evy process 4.4.2 VG L´evy process 4.4.3 CIR stochastic clock 4.4.4 Gamma-OU stochastic clock 4.4.5 Path generation for time-changed L´evy process 4.5 Pricing of exotic options 4.5.1 Exotic options 4.5.2 Exotic option prices 4.6 Pricing of moment derivatives 4.6.1 Moment swaps 4.6.2 Moment options 4.6.3 Hedging moment swaps 4.6.4 Pricing of moments swaps 4.6.5 Pricing of moments options 4.7 Conclusions References 5 Symmetries and Pricing of Exotic Options in L´evy Models Ernst Eberlein and Antonis Papapantoleon 5.1 5.2 5.3 5.4

Introduction Model and assumptions General description of the method Vanilla options 5.4.1 Symmetry 5.4.2 Valuation of European options 5.4.3 Valuation of American options 5.5 Exotic options 5.5.1 Symmetry 5.5.2 Valuation of barrier and lookback options 5.5.3 Valuation of Asian and basket options 5.6 Margrabe-type options References 6 Static Hedging of Asian Options under Stochastic Volatility Models using Fast Fourier Transform Hansj¨org Albrecher and Wim Schoutens 6.1 6.2

Introduction Stochastic volatility models 6.2.1 The Heston stochastic volatility model 6.2.2 The Barndorff-Nielsen–Shephard model 6.2.3 L´evy models with stochastic time

vii

70 71 74 78 78 79 79 79 79 80 80 82 86 89 89 90 91 93 93 95 99

99 100 105 106 106 111 113 114 114 115 117 119 124

129

129 131 131 132 133

viii

Contents

6.3 6.4

Static hedging of Asian options Numerical implementation 6.4.1 Characteristic function inversion using FFT 6.4.2 Static hedging algorithm 6.5 Numerical illustration 6.5.1 Calibration of the model parameters 6.5.2 Performance of the hedging strategy 6.6 A model-independent static super-hedge 6.7 Conclusions References 7 Impact of Market Crises on Real Options Pauline Barrieu and Nadine Bellamy 7.1 7.2

136 138 138 140 140 140 140 145 145 145 149

Introduction The model 7.2.1 Notation 7.2.2 Consequence of the modelling choice 7.3 The real option characteristics 7.4 Optimal discount rate and average waiting time 7.4.1 Optimal discount rate 7.4.2 Average waiting time 7.5 Robustness of the investment decision characteristics 7.5.1 Robustness of the optimal time to invest 7.5.2 Random jump size 7.6 Continuous model versus discontinuous model 7.6.1 Error in the optimal proﬁt–cost ratio 7.6.2 Error in the investment opportunity value 7.7 Conclusions Appendix References

149 151 151 153 155 156 156 157 158 159 160 161 161 163 165 165 167

8 Moment Derivatives and L´evy-type Market Completion Jos´e Manuel Corcuera, David Nualart and Wim Schoutens

169

8.1 8.2

8.3

Introduction Market completion in the discrete-time setting 8.2.1 One-step trinomial market 8.2.2 One-step ﬁnite markets 8.2.3 Multi-step ﬁnite markets 8.2.4 Multi-step markets with general returns 8.2.5 Power-return assets The L´evy market 8.3.1 L´evy processes 8.3.2 The geometric L´evy model 8.3.3 Power-jump processes

169 170 170 172 173 174 174 177 177 178 178

Contents

ix

Enlarging the L´evy market model 8.4.1 Martingale representation property 8.5 Arbitrage 8.5.1 Equivalent martingale measures 8.5.2 Example: a Brownian motion plus a ﬁnite number of Poisson processes 8.6 Optimal portfolios 8.6.1 Optimal wealth 8.6.2 Examples References

179 180 183 183

8.4

9 Pricing Perpetual American Options Driven by Spectrally One-sided L´evy Processes Terence Chan 9.1 9.2

Introduction First-passage distributions and other results for spectrally positive L´evy processes 9.3 Description of the model, basic deﬁnitions and notations 9.4 A renewal equation approach to pricing 9.5 Explicit pricing formulae for American puts 9.6 Some speciﬁc examples Appendix: use of fast Fourier transform References Epilogue Further references

10 On Asian Options of American Type Goran Peskir and Nadia Uys 10.1 Introduction 10.2 Formulation of the problem 10.3 The result and proof 10.4 Remarks on numerics Appendix References 11 Why be Backward? Forward Equations for American Options Peter Carr and Ali Hirsa 11.1 11.2 11.3 11.4 11.5 11.6

Introduction Review of the backward free boundary problem Stationarity and domain extension in the maturity direction Additivity and domain extension in the strike direction The forward free boundary problem Summary and future research

185 186 187 188 192

195

195 198 202 204 207 209 213 214 215 215 217

217 218 220 231 233 234 237

237 239 242 245 247 250

x

Contents

Appendix: Discretization of forward equation for American options References 12 Numerical Valuation of American Options Under the CGMY Process Ariel Almendral

251 257 259

12.1 Introduction 12.2 The CGMY process as a L´evy process 12.2.1 Options in a L´evy market 12.3 Numerical valuation of the American CGMY price 12.3.1 Discretization and solution algorithm 12.4 Numerical experiments Appendix: Analytic formula for European option prices References

259 260 261 263 263 270 271 275

13 Convertible Bonds: Financial Derivatives of Game Type Jan Kallsen and Christoph K¨uhn

277

13.1 Introduction 13.2 No-arbitrage pricing for game contingent claims 13.2.1 Static no-arbitrage prices 13.2.2 No-arbitrage price processes 13.3 Convertible bonds 13.4 Conclusions References

277 279 279 282 286 289 289

14 The Spread Option Optimal Stopping Game Pavel V. Gapeev

293

14.1 Introduction 14.2 Formulation of the problem 14.3 Solution of the free-boundary problem 14.4 Main result and proof 14.5 Conclusions References

293 294 296 299 302 304

Index

307

Contributors Hansj¨org Albrecher Department of Mathematics, Graz University of Technology, Steyrergasse 30, A-8010 Graz, Austria Ariel Almendral Norwegian Computing Center, Gaustadalleen 23, Postbox 114, Blindern, N-0314 Oslo, Norway Pauline Barrieu Statistics Department, London School of Economics, Houghton Street, London, WC2A 2AE, UK Nadine Bellamy Equipe d’Analyse et Probabilit´es, Universit´e d’Evry Val d’Essonne, Rue du P`ere Jarlan, 91025 Evry Cedex, France Peter Carr Bloomberg LP, 731 Lexington Avenue, New York, NY 10022, USA Terence Chan School of Mathematical and Computer Sciences, Heriot-Watt University, Edinburgh, EH14 4AS, Scotland, UK Jos´e Manuel Corcuera Facultat de Matematiques, Universitat de Barcelona, Gran Via de les Corts Catalanes 585, E-08007 Barcelona, Spain Ernst Eberlein Department of Mathematical Stochastics, University of Freiburg, Eckerstraße 1, D-79104, Freiburg, Germany Pavel V. Gapeev Institute of Control Sciences, Russian Academy of Sciences, Profsoyuznaya Str. 65, 117997 Moscow, Russia H´elyette Geman University of Paris Dauphine, Paris, France and ESSEC-Finance Department, 95021 Cergy-Pontoise, France

xii

Contributors

Ali Hirsa Caspian Capital Management, LLC, 745 Fifth Avenue, 28th Floor, New York, NY 10151, USA Jan Kallsen HVB-Institute for Mathematical Finance, Munich University of Technology, D-85747 Garching, Germany Christoph Kuehn Johann Wolfgang Goethe-Universit¨at, Fachbereich Mathematik (Fach 187), D-60054 Frankfurt am Main, Germany Andreas E. Kyprianou School of Mathematical and Computer Sciences, Heriot Watt University, Edinburgh, EH14 4AS, Scotland, UK R. Loeffen Department of Mathematics, University of Utrecht, PO Box 80.010, 3508 TA Utrecht, The Netherlands Dilip B. Madan Department of Finance, Robert H. School of Business, Van Munching Hall, University of Maryland, College Park, MD 20742, USA David Nualart Facultat de Matematiques, Universitat de Barcelona, Gran Via de les Corts Catalanes 585, E-08007 Barcelona, Spain Antonis Papapantoleon Department of Mathematical Stochastics, University of Freiburg, Eckerstraße 1, D-79104, Freiburg, Germany Goran Peskir Department of Mathematical Sciences, University of Aarhus, Ny Munkegade, DK-8000, Aarhus, Denmark Wim Schoutens Katholieke Universiteit Leuven - U.C.S, W. de Croylaan 54, B-3001 Leuven, Belgium Irwin Simons ING SWE, Financial Modelling, Marnixlaan 24, B-1000 Brussels, Belgium Jurgen Tistaert ING SWE, Financial Modelling, Marnixlaan 24, B-1000 Brussels, Belgium Nadia Uys Programme for Advanced Mathematics of Finance, School of Computational and Applied Mathematics, University of the Witwatersrand, Private Bag 3, Witwatersrand 2050, South Africa Nick Webber Warwick Business School, University of Warwick, Coventry CV4 7AL, UK

Preface Since around the turn of the millennium there has been a general acceptance that one of the more practical improvements one may make in the light of the shortfalls of the classical Black–Scholes model is to replace the underlying source of randomness, a Brownian motion, by a L´evy process. Working with L´evy processes allows one to capture distributional characteristics in the stock returns such as semi-heavy tails and asymmetry, as well as allowing for jumps in the price process with the interpretation as market shocks and effects due to trading taking place in business time rather than real time. In addition, L´evy processes in general, as well as having the same properties as Brownian motion in the form of stationary independent increments, have many well understood probabilistic and analytical properties which make them attractive as mathematical tools. At the same time, exotic derivatives are gaining increasing importance as ﬁnancial instruments and are traded nowadays in large quantities in over the counter markets. The consequence of working with markets driven by L´evy processes forces a number of new mathematical challenges with respect to exotic derivatives. Many exotic options are based on the evolving historical path of the underlying. In terms of pricing and hedging, this requires an understanding of ﬂuctuation theory, stochastic calculus and distributional decompositions associated with L´evy processes. This current volume is a compendium of articles, each of which consists of a discursive review and recent research on the topic of Exotic Option Pricing and Advanced L´evy Models written by leading scientists in this ﬁeld. This text is organized as follows. The ﬁrst two chapters can be seen as an introduction to L´evy processes and their applications. The ﬁrst chapter, by A. E. Kyprianou and R. Loeffen, gives a brief introduction to L´evy processes, providing several examples which are commonly used in ﬁnance, as well as examining in more detail some of their ﬁne and coarse path properties. To apply L´evy processes in practice one needs good numerics. In Chapter 2, N. Webber discusses recent progress in the development of simulation methods suitable for most of the widely used L´evy processes. Speed-up methods, bridge algorithms and stratiﬁed sampling are some of the many ingredients. These techniques are applied in the context of the valuation of different kinds of exotic options. In the second part, one can see L´evy-driven equity models at work. In Chapter 3, H. Geman and D. Madan use pure jump models, in particular from the CGMY class, for the evolution of stock prices and investigate in this setting the relationship between the statistical and risk-neutral densities. Statistical estimation is conducted on different world indexes. Their conclusions depart from the standard applications of utility theory to asset pricing which assume a representative agent who is long the market. They argue that one

xiv

Preface

must have at a minimum a two-agent model in which some weight is given to an agent who is short the market. In Chapter 4, W. Schoutens, E. Simons and J. Tistaert calibrate different L´evy-based stochastic volatility models to a real market option surface and price by Monte Carlo techniques a range of exotics options. Although the different models discussed can all be nicely calibrated to the option surface – leading to almost identical vanilla prices – exotic option prices under the different models discussed can differ considerably. This investigation is pushed further by looking at the prices of moment derivatives, a new kind of derivative paying out realized higher moments. Even more pronounced differences are reported in this case. The study reveals that there is a clear issue of model risk and warns of blind use of fancy models in the realm of exotic options. The third part is devoted to pricing, hedging and general theory of different exotics options of a European nature. In Chapter 5, E. Eberlein and A. Papapantoleon consider time-inhomogeneous L´evy processes (or additive processes) to give a better explanation of the so-called ‘volatility smile’, as well as the ‘term structure of smiles’. They derive different kinds of symmetry relations for various exotic options. Their contribution also contains an extensive review of current literature on exotics driven in L´evy markets. In Chapter 6, H. Albrecher and W. Schoutens present a simple static super-hedging strategy for the Asian option, based on stop-loss transforms and comonotonic theory. A numerical implementation is given in detail and the hedging performance is illustrated for several stochastic volatility models. Real options form the main theme of Chapter 7, authored by P. Barrieu and N. Bellamy. There, the impact of market crises on investment decisions is analysed through real options under a jump-diffusion model, where the jumps characterize the crisis effects. In Chapter 8, J.M. Corcuera, D. Nualart and W. Schoutens show how moment derivatives can complete L´evy-type markets in the sense that, by allowing trade in these derivatives, any contingent claim can be perfectly hedged by a dynamic portfolio in terms of bonds, stocks and moment-derivative related products. In the fourth part, exotics of an American nature are considered. Optimal stopping problems are central here. Chapter 9 is a contribution at the special request of the editors. This consists of T. Chan’s original unpublished manuscript dating back to early 2000, in which many important features of the perpetual American put pricing problem are observed for the case of a L´evy-driven stock which has no positive jumps. G. Peskir and N. Uys work in Chapter 10 under the traditional Black–Scholes market but consider a new type of Asian option where the holder may exercise at any time up to the expiry of the option. Using recent techniques developed by Peskir concerning local time–space calculus, they are able to give an integral equation characterizing uniquely the optimal exercise boundary. Solving this integral equation numerically brings forward stability issues connected with the Hartman– Watson distribution. In Chapter 11, P. Carr and A. Hirsa give forward equations for the value of an American put in a L´evy market. A numerical scheme for the VG case for very fast pricing of an American put is given in its Appendix. In the same spirit, A. Almendral discusses the numerical valuation of American options under the CGMY model. A numerical solution scheme for the Partial-Integro-Differential Equation is provided; computations are accelerated by the Fast-Fourier Transform. Pricing American options and their early exercise boundaries can be carried out within seconds. The ﬁnal part considers game options. In Chapter 13, C. K¨uhn and J. Kallsen give a review of the very recent literature concerning game-type options, that is, options in which both holder and writer have the right to exercise. Game-type options are very closely related to convertible bonds and K¨uhn and Kallsen also bring this point forward in their contribution.

Preface

xv

Last, but by far not least, P. Gapeev gives a concrete example of a new game-type option within the Black–Scholes market for which an explicit representation can be obtained. We should like to thank all contributors for working hard to keep to the tempo that has allowed us to compile this text within a reasonable period of time. We would also like to heartily thank the referees, all of whom responded gracefully to the ﬁrm request to produce their reports within a shorter than normal period of time and without compromising their integrity. This book grew out of the 2004 Workshop, Exotic Option Pricing under Advanced L´evy Models, hosted at EURANDOM in The Netherlands. In addition to the excellent managerial and organizational support offered by EURANDOM, it was generously supported by grants from Nederlands Organizatie voor Wetenschappelijk Onderzoek (The Dutch Organization for Scientiﬁc Research), Koninklijke Nederlandse Akademie van Wetenschappen (The Royal Dutch Academy of Science) and The Journal of Applied Econometrics. Special thanks goes to Jef Teugels and Lucienne Coolen. Thanks also to wilmott.com and mathfinance.de for publicizing the event. A. E. Kyprianou, Edinburgh, UK W. Schoutens, Leuven, Belgium P. Wilmott, London, UK

About the Editors Andreas E. Kyprianou Address: School of Mathematical and Computer Sciences, Heriot Watt University, Edinburgh, EH14 4AS, Scotland, UK E-mail: [email protected] Afﬁliation: Heriot Watt University, Scotland, UK Andreas Kyprianou has a degree in mathematics from Oxford University and a PhD in probability theory from Shefﬁeld University. He has held academic positions in the Mathematics and/or Statistics Departments at The London School of Economics, Edinburgh University, Utrecht University and, currently, Heriot Watt University. He has also worked for nearly two years as a research mathematician with Shell International Exploration and Production. His research interests are focused on pure and applied probability with recent focus on L´evy processes. He has taught a range of courses on probability theory, stochastic analysis, ﬁnancial stochastics and L´evy processes on the Amsterdam–Utrecht Masters programme in Stochastics and Financial Mathematics and the MSc programme in Financial Mathematics at Edinburgh University. Wim Schoutens Address: Katholieke Universiteit Leuven – UCS, W. De Croylaan 54, B-3001 Leuven, Belgium E-mail: [email protected] Afﬁliation: Katholieke Universiteit Leuven, Belgium Wim Schoutens has a degree in Computer Science and a PhD in Science (Mathematics). He is a research professor at the Department of Mathematics at the Catholic University of Leuven (Katholieke Universiteit Leuven), Belgium. He has been a consultant to the banking industry and is the author of the Wiley book L´evy Processes in Finance – Pricing Financial Derivatives. His research interests are focused on ﬁnancial mathematics and stochastic processes. He currently teaches several courses related to ﬁnancial engineering in different Master programmes.

xviii

About the Editors

Paul Wilmott Address: ‘Wherever I lay my hat’ E-mail: [email protected] Afﬁliation: Various Paul Wilmott has undergraduate and DPhil degrees in mathematics. He has written over 100 articles on mathematical modeling and ﬁnance, as well as internationally acclaimed books including Paul Wilmott on Quantitative Finance, published by Wiley. Paul has extensive consulting experience in quantitative ﬁnance with leading US and European ﬁnancial institutions. He has founded a university degree course and the popular Certiﬁcate in Quantitative Finance. Paul also manages wilmott.com.

About the Contributors Hansjoerg Albrecher is Associate Professor of Applied Mathematics at the Graz University of Technology. He studied Mathematics in Graz, Limerick and Baltimore, receiving his doctorate in 2001. He held visiting appointments at the Katholieke Universiteit Leuven and the University of Aarhus. Research interests include ruin theory, stochastic simulation and quantitative ﬁnance. Ariel Almendral will take up a research position at the Norwegian Computing Center, starting in August 2005. In 2004, he obtained his PhD from the University of Oslo, Norway. In his thesis he focused on numerical methods for ﬁnancial derivatives in the presence of jump processes, from a differential equation perspective. Parts of his PhD research were carried out at Delft University of Technology, The Netherlands, where he held a postdoctoral position for a year. Pauline Barrieu has been a lecturer in the Department of Statistics at the London School of Economics since 2002, after obtaining a PhD in ﬁnance (doctorate HEC, France) and a PhD in Mathematics (University of Paris 6, France). Her research interests are mainly problems at the interface of insurance and ﬁnance, in particular, optimal design of new types of derivatives and securitization. She also works on quantitative methods for assessing ﬁnancial and non-ﬁnancial risks, on stochastic optimization and environmental economics. Nadine Bellamy is Associate Professor in Mathematics at the University of Evry, France. Her PhD thesis (University of Evry, 1999) deals with hedging and pricing in markets driven by discontinuous processes and her current research interests are related to optimization and real options problems. Dr Peter Carr heads Quantitative Research at Bloomberg LP. He also directs the Masters in Mathematical Finance program at NYU’s Courant Institute. Formerly, Dr Carr was a ﬁnance professor for eight years at Cornell University. Since receiving his PhD in Finance from UCLA in 1989, he has published extensively in both academic and industry-oriented journals. He has recently won awards from Wilmott Magazine for Cutting Edge Research and from Risk Magazine for Quant of the Year. Terence Chan completed his PhD at Cambridge University UK after which he obtained his current position at Heriot-Watt University, Edinburgh. Among his research interests are L´evy processes but he only occasionally dabbles in ﬁnancial mathematics to maintain the illusion that he is doing something of practical use!

xx

About the Contributors

Jos´e Manuel Corcuera is an associate professor since 1997 at the Faculty of Mathematics of the University of Barcelona. His main research interest is in the theoretical aspects of statistics and quantitative ﬁnance. Ernst Eberlein is Professor of Stochastics and Mathematical Finance at the University of Freiburg. He is a co-founder of the Freiburg Center for Data Analysis and Modeling (FDM), an elected member of the International Statistical Institute and at present Executive Secretary of the Bachelier Finance Society. His current research focuses on statistical analysis and realistic modeling of ﬁnancial markets, risk management, as well as pricing of derivatives. Pavel Gapeev was born in Moscow in 1976. He studied and obtained his PhD in Stochastics at Moscow State University in 2001. He is now working as a Senior Researcher at the Institute of Control Sciences, Russian Academy of Sciences in Moscow. He has held a visiting appointment at Humboldt University, Berlin (2001/2002) in addition to some short term research visits to Aarhus, Bochum, Copenhagen, Frankfurt, Helsinki, and Zurich. His main ﬁeld of research is stochastic analysis and its applications into ﬁnancial mathematics, optimal control, optimal stopping, and quickest detection. Apart from mathematics he is interested in arts, sports and travelling, and enjoys playing the violin. H´elyette Geman is a Professor of Finance at the University Paris Dauphine and ESSEC Graduate Business School. She is a graduate of Ecole Normale Superieure in Mathematics, holds a Masters degree in theoretical physics and a PhD in mathematics from the University Pierre et Marie Curie and a PhD in Finance from the University Pantheon Sorbonne. Professor Geman has published more than 60 papers in major ﬁnance journals including the Journal of Finance, Mathematical Finance, Journal of Financial Economics, Journal of Banking and Finance and Journal of Business. Professor Geman’s research includes asset price modelling using jump-diffusions and L´evy processes, commodity forward curve modelling and exotic option pricing for which she won the ﬁrst prize of the Merrill Lynch Awards. She has written a book entitled Commodities and Commodity Derivatives (John Wiley & Sons Ltd, 2005). Ali Hirsa joined Caspian Capital Management as the Head of Analytical Trading Strategy in April 2004. At CCM his responsibilities include design and testing of new trading strategies. Prior to his current position, Ali worked at Morgan Stanley for four years. Ali is also an adjunct professor at Columbia University and New York University where he teaches in the mathematics of ﬁnance program. Ali received his PhD in applied mathematics from University of Maryland at College Park under the supervision of Dilip B. Madan. Jan Kallsen is a Professor of Mathematical Finance at Munich University of Technology. His research interests include pricing and hedging in incomplete markets and the general theory of stochastic processes. Christoph Kuhn is Junior Professor at the Frankfurt MathFinance Institute. He holds a ¨ diploma in Mathematical Economics from the University of Marburg and a PhD in mathematics from Munich University of Technology. His main research interests are pricing and hedging of derivatives in incomplete markets and the microstruture of ﬁnancial markets. Ronnie Loeffen was born in 1981 in the Netherlands and has recently received a Master’s degree in Mathematics at the University Utrecht. The subject of his Master’s thesis was American options on a jump-diffusion model.

About the Contributors

xxi

Dilip B. Madan is Professor of Finance at the Robert H. Smith School of Business, University of Maryland. He is co-editor of Mathematical Finance and served as President of the Bachelier Finance Society 2002–2003. He has been a consultant to Morgan Stanley since 1996. He now also consults for Bloomberg and Caspian Capital. His primary research focus is on stochastic processes as they are applied to the management and valuation of ﬁnancial risks. David Nualart is Professor at the Faculty of Mathematics of the University of Barcelona. His research interests include a variety of topics in stochastic analysis, with emphasis on stochastic partial differential equations, Malliavin calculus and fractional Brownian motion. He is the author of the monograph Malliavin Calculus and Related Topics. Antonis Papapantoleon is a research assistant at the Department for Mathematical Stochastics, University of Freiburg. He received a Diploma in Mathematics from the University of Patras (2000) and an MSc in Financial Mathematics from the University of Warwick (2001). From January to August 2002 he worked at the FX Quantitative Research group of Commerzbank in Frankfurt. Goran Peskir is the Chair in Probability at the School of Mathematics, University of Manchester. In the period 1996–2005 he was an Associate Professor at the Department of Mathematical Sciences, University of Aarhus in Denmark. He is an internationally leading expert in the ﬁeld of Optimal Stopping and author to over sixty papers dealing with various problems in the ﬁeld of probability and its applications (optimal stopping, stochastic calculus, option pricing). Together with Albert Shiryaev he has co-authored the book Optimal Stopping and Free-Boundary Problems. Erwin Simons works in Quantitative Modeling at ING Brussels. After 3 years of frontofﬁce experience in Equity derivatives pricing, over the last year he switched to Interest-Rate derivatives modeling. He holds a PhD in Applied Mathematics from the Catholic University Leuven, von Karman Institute for Fluid Dynamics on the subject of large-scale computing of incompressible turbulent ﬂows. Jurgen Tistaert joined the Credit Risk Management Department of ING Brussels at the end of 1996 where he developed several rating, exposure and risk/performance models. He moved to Financial Markets in 2001, where the team is focusing on the R&D of pricing models for a broad range of derivative products. Before joining ING, he was a research assistant at the Quantitative Methods Group of K.U. Leuven Applied Economics Faculty, where he currently is appointed as a Fellow. Nadia Uys completed her Bachelors in Economic Science, majoring in Mathematical Statistics and Actuarial Science, at the University of the Witwatersrand in 2000, followed with Honours in Advanced Mathematics of Finance in 2001. Her MSc dissertation entitled ‘Optimal Stopping Problems and American Options’ was completed under the supervision of Professor G. Peskir (University of Aarhus) and Mr H. Hulley (Sydney Polytechnic) and received a distinction in 2005. She is currently teaching in the Programme in Advanced Mathematic of Finance at the University of the Witwatersrand and engaging in research toward a PhD under the supervision of Professor F. Lombard (University of Johannesburg). Nick Webber is Director of the Financial Options Research Centre, University of Warwick. Formerly Professor of Computational Finance at Cass Business School, he is interested not

xxii

About the Contributors

only in theoretical ﬁnancial mathematics, but also in methods for the fast evaluation of options prices under a variety of assumptions for returns distributions. As well as work with L´evy processes and numerical methods he has also worked on copulas, credit models and interest rates.

1 L´evy Processes in Finance Distinguished by their Coarse and Fine Path Properties Andreas E. Kyprianou Heriot Watt University, Edinburgh, UK and

R. Loeffen University of Utrecht, The Netherlands Abstract ´ We give a brief introduction to Levy processes and indicate the diversity of this class of stochastic processes by quoting a number of complete characterizations of coarse and ﬁne path properties. The theory is exempliﬁed by distinguishing such properties for ´ Levy processes which are currently used extensively in ﬁnancial models. Speciﬁcally, we treat jump-diffusion models (including Merton and Kou models), spectrally one-sided processes, truncated stable processes (including CGMY and Variance Gamma models), Meixner processes and generalized hyperbolic processes (including hyperbolic and normal inverse Gaussian processes).

1.1

INTRODUCTION

The main purpose of this text is to provide an entr´ee to the compilation Exotic Options and Advanced L´evy Models. Since path ﬂuctuations of L´evy processes play an inevitable role in the computations which lead to the pricing of exotic options, we have chosen to give a review of what subtleties may be encountered there. In addition to giving a brief introduction to the general structure of L´evy processes, path variation and its manifestation in the L´evy–Khintchine formula, we shall introduce classiﬁcations of drifting and oscillation, regularity of the half line, the ability to visit ﬁxed points and creeping. The theory is exempliﬁed by distinguishing such properties for L´evy processes which are currently used extensively in ﬁnancial models. Speciﬁcally, we treat jump-diffusion models (including Merton and Kou models), spectrally one-sided processes, truncated stable processes (including CGMY and variance gamma models), Meixner processes and generalized hyperbolic processes (including hyperbolic and normal inverse Gaussian processes). To support the presentation of more advanced path properties and for the sake of completeness, a number of known facts and properties concerning these processes are reproduced from the literature. We have relied heavily upon the texts by Schoutens (2003) and Cont and Tankov (2004) for inspiration. Another useful text in this respect is that of Boyarchenko and Levendorskii (2002). Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

2

Exotic Option Pricing and Advanced L´evy Models

The job of exhibiting the more theoretical facts concerning path properties have been greatly eased by the existence of the two indispensable monographs on L´evy processes, namely Bertoin (1996) and Sato (1999); see, in addition, the more recent monograph of Applebaum (2004) which also contains a section on mathematical ﬁnance. In the course of this text, we shall also brieﬂy indicate the relevance of the path properties considered to a number of exotic options. In some cases, the links to exotics is rather vague due to the fact that the understanding of pricing exotics and advanced L´evy models is still a ‘developing market’, so to speak. Nonetheless, we believe that these issues will in due course become of signiﬁcance as research progresses.

1.2

´ LEVY PROCESSES

We start with the deﬁnition of a real valued L´evy process followed by the L´evy–Khintchine characterization. Deﬁnition 1 A L´evy process X = {Xt : t ≥ 0} is a stochastic process deﬁned on a probability space (, F, P) which satisﬁes the following properties: (i) (ii) (iii) (iv)

The paths of X are right continuous with left limits almost surely. X0 = 0 almost surely. X has independent increments; for 0 ≤ s ≤ t, Xt − Xs is independent of σ (Xu : u ≤ s). X has stationary increments; for 0 ≤ s ≤ t, Xt − Xs is equal in distribution to Xt−s .

It turns out that there is an intimate relationship between L´evy processes and a class of distributions known as inﬁnitely divisible distributions which gives a precise impression of how varied the class of L´evy processes really is. To this end, let us devote a little time to discussing inﬁnitely divisible distributions. Deﬁnition 2 We say that a real valued random variable has an inﬁnitely divisible distribution if for each n = 1, 2, . . . . there exists a sequence of iid random variables 1 , . . . , n such that d

= 1,n + · · · + n,n d

where = is equality in distribution. Alternatively, we could have expressed this relation in terms of probability laws. That is to say, the law µ of a real valued random variable is inﬁnitely divisible if for each n = 1, 2, . . . there exists another law µn of a real valued random variable such that µ = µ∗n n , the n-fold convolution of µn . The full extent to which we may characterize inﬁnitely divisible distributions is carried out via their characteristic function (or Fourier transform of their law) and an expression known as the L´evy–Khintchine formula. Theorem 3 (L´evy–Khintchine formula) A probability law µ of a real valued random variable is inﬁnitely divisible with characteristic exponent , eiux µ (dx) = e−(u) for u ∈ R, R

L´evy Processes in Finance–Coarse and Fine Path Properties

3

if and only if there exists a triple (γ , σ, ), where γ ∈ R, σ ≥ 0 and is a measure supported on R\{0} satisfying R 1 ∧ x 2 (dx) < ∞, such that 1 (u) = iγ u + σ 2 u2 + (1 − eiux + iux1(|x| 0, Xt is a random variable whose law belongs to the class of inﬁnitely divisible distributions. This follows from the fact that for any n = 1, 2, . . . Xt = Xt/n + (X2t/n − Xt/n ) + · · · + (Xt − X(n−1)t/n )

(1.1)

together with the fact that X has stationary independent increments. Suppose now that we deﬁne for all u ∈ R, t ≥ 0 t (u) = − log E eiuXt then by using equation (1.1) twice we have for any two positive integers m, n that m1 (u) = m (u) = nm/n (u) and hence for any rational t > 0

t (u) = t1 (u) .

(1.2)

If t is an irrational number, then we can choose a decreasing sequence of rationals {tn : n ≥ 1} such that tn ↓ t as n tends to inﬁnity. Almost sure right continuity of X implies right continuity of exp{−t (u)} (by dominated convergence) and hence equation (1.2) holds for all t ≥ 0. In conclusion, any L´evy process has the property that E eiuXt = e−t(u) where (u) := 1 (u) is the characteristic exponent of X1 which has an inﬁnitely divisible distribution. Deﬁnition 5 In the sequel we shall also refer to (u) as the characteristic exponent of the L´evy process. Note that the law of a L´evy process is uniquely determined by its characteristic exponent. This is because the latter characterizes uniquely all one-dimensional distributions of X. From the property of stationary independent increments, it thus follows that the characteristic exponent characterizes uniquely all ﬁnite dimensional distributions which themselves uniquely characterize the law of X.

4

Exotic Option Pricing and Advanced L´evy Models

It is now clear that each L´evy process can be associated with an inﬁnitely divisible distribution. What is not clear is whether given an inﬁnitely divisible distribution, one may construct a L´evy process such that X1 has that distribution. This latter issue is resolved by the following theorem which gives the L´evy–Khintchine formula for L´evy processes. Theorem 6 Suppose that γ ∈ R, σ ≥ 0 and is a measure on R\{0} such that R (1 ∧ |x|2 )(dx) < ∞. From this triple deﬁne for each u ∈ R 1 2 2 (1.3) (u) = iγ u + σ u + [1 − eiux + iux1(|x| 0 is the initial value of the asset and X is a L´evy process. There are essentially four main classes of L´evy processes which feature heavily in current mainstream literature on market modeling with pure L´evy processes (we exclude from the discussion stochastic volatility models such as those of Barndorff–Nielsen and Shephard (2001)). These are the jump-diffusion processes (consisting of a Brownian motion with drift plus an independent compound Poisson process), the generalized tempered stable processes (which include more speciﬁc examples such as Variance Gamma processes and CGMY),

L´evy Processes in Finance–Coarse and Fine Path Properties

5

Generalized Hyperbolic processes and Meixner processes. There is also a small minority of papers which have proposed to work with the arguably less realistic case of spectrally onesided L´evy processes. Below, we shall give more details on all of the above key processes and their insertion into the literature. 1.3.1 Compound Poisson processes and jump-diffusions Compound Poisson processes form the simplest class of L´evy processes in the sense of understanding their paths. Suppose that ξ is a random variable with honest distribution F supported on R but with no atom at 0. Let Xt :=

Nt

t ≥0

ξi ,

i=1

where {ξi : i ≥ 1} are independent copies of ξ and N := {Nt : t ≥ 0} is an independent Poisson process with rate λ > 0. Then, X = {Xt : t ≥ 0} is a compound Poisson process. The fact that X is a L´evy process can easily be veriﬁed by computing the joint characteristic of the variables Xt − Xs and Xv − Xu for 0 ≤ v ≤ u ≤ s ≤ t < ∞ and showing that it factorizes. Indeed, standard facts concerning the characteristic function of the Poisson distribution leads to the following expression for the characteristic exponent of X, (u)) = (1 − eiux )λF (dx) (u) = λ(1 − F R

(u) = E(eiuξ ). Consequently, we can easily identify the L´evy triple via σ = 0 and where F γ = − R xλF (dx) and (dx) = λF (dx). Note that has ﬁnite total mass. It is not difﬁcult to reason that any L´evy process whose L´evy triple has this property must necessarily be a compound Poisson process. Since the jumps of the process X are spaced out by independent exponential distributions, the same is true of X and hence X is pathwise piecewise constant. Up to adding a linear drift, compound Poisson processes are the only L´evy processes which are piecewise linear. The ﬁrst model for risky assets in ﬁnance which had jumps was proposed by Merton (1976) and consisted of the log-price following an independent sum of a compound Poisson process, together with a Brownian motion with drift. That is, Xt = −γ t + σ Bt +

Nt

ξi ,

t ≥0

i=1

where γ ∈ R, {Bt : t ≥ 0} is a Brownian motion and {ξi : i ≥ 0} are normally distributed. Kou (2002) assumed the above structure, the so called jump-diffusion model, but chose the jump distribution to be that of a two-sided exponential distribution. Kou’s choice of jump distribution was heavily inﬂuenced by the fact that analysis of ﬁrst passage problems become analytically tractable which itself is important for the valuation of American put options (see Chapter 11 below). Building on this idea, Asmussen et al. (2004) introduce a jump-diffusion model with two-sided phasetype distributed jumps. The latter form a class of distributions which generalize the two-sided exponential distribution and like Kou’s model, have the desired property that ﬁrst passage problems are analytically tractable.

6

Exotic Option Pricing and Advanced L´evy Models

1.3.2 Spectrally one-sided processes Quite simply, spectrally one-sided processes are characterized by the property that the support of the L´evy measure is restricted to the upper or the lower half line. In the latter case, that is (0, ∞) = 0, one talks of spectrally negative L´evy processes. Without loss of generality we can and shall restrict our discussion to this case unless otherwise stated in the sequel. Spectrally negative L´evy processes have not yet proved to be a convincing tool for modeling the evolution of a risky asset. The fact that the support of the L´evy measure is restricted to the lower half line does not necessarily imply that the distribution of the L´evy process itself is also restricted to the lower half line. Indeed, there are many examples of spectrally negative processes whose ﬁnite time distributions are supported on R. One example, which has had its case argued for in a ﬁnancial context by Carr and Wu (2003) and Cartea and Howison (2005), is a spectrally negative stable process of index α ∈ (1, 2). To be more precise, this is a process whose L´evy measure takes the form (dx) = 1(x 0 and whose parameter σ is identically zero. A lengthy calculation reveals that this process has the L´evy–Khintchine exponent πα (u) = c|u|α 1 + i tan signu . 2 Chan (2000, 2004), Mordecki (1999, 2002) and Avram et al. (2002, 2004), have also worked with a general spectrally negative L´evy process for the purpose of pricing American put and Russian options. In their case, the choice of model was based purely on a degree of analytical tractability centred around the fact that when the path of a spectrally negative process passes from one point to another above it, it visits all other points between them. 1.3.3 Meixner processes The Meixner process is deﬁned through the Meixner distribution which has a density function given by

β(x − µ) (2 cos(β/2))2δ i(x − µ) 2 exp fMeixner (x; α, β, δ, µ) = δ + 2απ (2δ) α α where α > 0, −π < β < π, δ > 0, m ∈ R. The Meixner distribution is inﬁnitely divisible with a characteristic exponent

2δ cos(β/2) Meixner (u) = − log − iµu, cosh(αu − iβ)/2 and therefore there exists a L´evy process with the above characteristic exponent. The L´evy triplet (γ , σ, ) is given by ∞ sinh(βx/α) γ = −αδ tan(β/2) + 2δ dx − µ, sinh(π x/α) 1

L´evy Processes in Finance–Coarse and Fine Path Properties

7

σ = 0 and (dx) = δ

exp(βx/α) dx. x sinh(π x/α)

(1.4)

The Meixner process appeared as an example of a L´evy process having a particular martingale relation with respect to orthogonal polynomials (see Schoutens and Teugels (1998) and Schoutens (2000)). Grigelionis (1999) and Schoutens (2001, 2002) established the use of the Meixner process in mathematical ﬁnance. Relationships between Mexiner distributions and other inﬁnitely divisible laws also appear in the paper of Pitman and Yor (2003). 1.3.4 Generalized tempered stable processes and subclasses The generalized tempered stable process has L´evy density ν := d/dx given by ν(x) =

cp −λp x e 1{x>0} 1+α x p

+

cn eλn x 1{x 0, λn > 0, cp > 0 and cn > 0. These processes take their name from stable processes which have L´evy measures of the form

cp cn (dx) = 1{x>0} + 1{x 0. Stable processes with index α ∈ (0, 1] have no moments and when α ∈ (1, 2) only a ﬁrst moment exists. Generalized tempered stable processes differ in that they have an exponential weighting in the L´evy measure. This guarantees the existence of all moments, thus making them suitable for ﬁnancial modelling where a moment-generating function is necessary. Since the shape of the L´evy measure in the neighbourhood of the origin determines the occurrence of small jumps and hence the small time path behaviour, the exponential weighting also means that on small time scales stable processes and generalized tempered stable processes behave in a very similar manner. Generalized tempered stable processes come under a number of different names. They are sometimes called KoBoL processes, named after the authors Koponen (1995) and Boyarchenko and Levendorskii (2002). Carr et al. (2002, 2003) have also studied this sixparameter family of processes and as a consequence of their work they are also referred to as generalized CGMY processes or, for reasons which will shortly become clear, CCGMYY processes. There seems to be no uniform terminology used for this class of processes at the moment and hence we have simply elected to follow the choice of Cont and Tankov (2004). Since |x|ν(x)dx < ∞ R\(−1,1)

it turns out to be more convenient to express the L´evy–Khintchine formula in the form

(u) = iuγ +

∞

−∞

(1 − eiux + iux)ν(x)dx

(1.5)

8

Exotic Option Pricing and Advanced L´evy Models

where γ = γ − by

R\(−1,1) xν(x)dx

< ∞. In this case, the characteristic exponent is given

(u) = iuγ − Ap − An , where

iu iuc + c (λ − iu) log 1 − p p p λp

iu iu + log 1 − −cp Ap = λ λ p p

iuαp iu αp αp 1− −1+ (−αp )λp cp λp λp

iu −iucn + cn (λn + iu) log 1 + λn

iu iu −cn − + log 1 + An = λ λ n n

αn

iu iuαn 1+ −1− (−αn )λαnn cn λn λn

if αp = 1 if αp = 0 otherwise if αn = 1 if αn = 0 otherwise

(see Cont and Tankov (2004), p. 122). When αp = αn = Y , cp = cn = C, λp = M and λn = G, the generalized tempered stable process becomes the so called CGMY process, named after the authors who ﬁrst introduced it, i.e. Carr et al. (2002). The characteristic exponent of the CGMY process for Y = 0 and Y = 1 is often written as CGMY (u) = −C (−Y )[(M − iu)Y − M Y + (G + iu)Y − GY )] − iuµ,

(1.6)

which is the case for an appropriate choice of γ , namely

Y GY Y MY + iµ. − γ = C (−Y ) M G The properties of the CGMY process can thus be inferred from the properties of the generalized tempered stable process. Note that in this light, generalized tempered stable processes are also referred to as CCGMYY. As a limiting case of the CGMY process, but still within the class of generalized tempered stable processes, we have the variance gamma process. The latter was introduced as a predecessor to the CGMY process by Madan and Seneta (1987) and treated in a number of further papers by Madan and co-authors. The variance gamma process can be obtained by starting with the parameter choices for the CGMY but then taking the limit as Y tends to zero. This corresponds to a generalized tempered stable process with αp = αn = 0. Working with γ = −C/M + C/G + µ, we obtain the variance gamma process with the characteristic exponent

iu iu + log 1 + − iuµ. (1.7) VG (u) = C log 1 − M G The characteristic exponent is usually written as

1 2 2 1 VG (u) = log 1 − iθ κu + σ κu − iuµ, κ 2

L´evy Processes in Finance–Coarse and Fine Path Properties

where

C = 1/κ,

M=

2

θ 2 + 2 σκ − θ σ2

and G =

9

2 θ 2 + 2 σκ + θ σ2

for θ ∈ R and κ > 0. Again, the properties of the variance gamma process can be derived from the properties of the generalized tempered stable process. 1.3.5 Generalized hyperbolic processes and subclasses The density of a generalized hyperbolic distribution is given by λ 1 fGH (x; α, β, λ, δ, µ) = C(δ 2 + (x − µ)2 ) 2 − 4 Kλ− 1 α δ 2 + (x − µ)2 eβ(x−µ) , 2

where

− β 2 )λ/2 C= √ 2π α λ−1/2 δ λ Kλ δ α 2 − β 2 (α 2

and with α > 0, 0 ≤ |β| < α, λ ∈ R, δ > 0 and µ ∈ R. The function Kλ stands for the modiﬁed Bessel function of the third kind with index λ. This distribution turns out to be inﬁnitely divisible with a characteristic exponent GH (u) = − log

α2 − β 2 α 2 − (β + iu)2

λ/2

Kλ (δ α 2 − (β + iu)2 ) − iµu. Kλ (δ α 2 − β 2 )

These facts are non-trivial to prove–see Halgreen (1979) who gives the proofs. The corresponding L´evy measure is rather complicated, being expressed as integrals of special functions. We refrain from offering the L´evy density here on account of its complexity and since we shall not use it in the sequel. Generalized hyperbolic processes were introduced within the context of mathematical ﬁnance by Barndorff-Nielsen (1995, 1998) and Erbelein and Prause (1998). When λ = 1, we obtain the special case of a hyperbolic process and when λ = − 12 , the normal inverse Gaussian process is obtained. Because the modiﬁed Bessel function has a simple form when λ = − 12 , namely K− 1 (z) = 2

π − 1 −z z 2e , 2

the characteristic exponent can be simpliﬁed to NI G (u) = δ

α 2 − (β + iu)2 −

α2 − β 2 .

Eberlein and Hammerstein (2002) investigated some limiting cases of generalized hyperbolic distributions and processes. Because for λ > 0 Kλ ∼

z −λ 1

(λ) 2 2

when z → 0,

10

Exotic Option Pricing and Advanced L´evy Models

we have that

λ 2δ α 2 − β 2 2δ α 2 − (β + iu)2

2 u2 α − (β + iu)2 2βiu = λ log 1 + = λ log − α2 − β 2 α2 − β 2 α2 − β 2 when δ → 0 and for µ = 0. Here we write f ∼ g when u → ∞ to mean that limu→∞ f (u)/ g(u) = 1. So, we see that when δ → 0 and for µ = 0, λ = 1/κ, β = θ/σ 2 and α =

α2 − β 2 GH (u) ∼ − log 2 α − (β + iu)2

λ/2

(2/κ)+(θ 2 /σ 2 ) , σ2

the characteristic exponent of the generalized hyperbolic process converges to the characteristic exponent of the variance gamma process. Because the variance gamma process is obtained by a limiting procedure, its path properties cannot be deduced directly from those of the generalized hyperbolic process. Indeed, we shall see they are fundamentally different processes.

1.4

PATH PROPERTIES

In the following sections, we shall discuss a number of coarse and ﬁne path properties of general L´evy processes. These include path variation, hitting of points, creeping and regularity of the half line. With the exception of the last property, none of the above have played a prominent role in mainstream literature on the modeling of ﬁnancial markets. Initial concerns of L´evydriven models were focused around the pricing of vanilla-type options, that is, options whose value depends on the distribution of the underlying L´evy process at a ﬁxed point in time. Recently, more and more attention has been paid to exotic options which are typically path dependent. Fluctuation theory and path properties of Brownian motion being well understood has meant that many examples of exotic options under the assumptions of the classical Black–Scholes models can and have been worked out in the literature. We refer to objects such as American options, Russian options, Asian options, Bermudan options, lookback options, Parisian options, Israeli or game options, Mongolian options, and so on. However, dealing with exotic options in L´evy-driven markets has proved to be considerably more difﬁcult as a consequence of the more complicated, and to some extent, incomplete nature of the theory of ﬂuctuations of L´evy processes. Nonetheless, it is clear that an understanding of course and ﬁne path properties plays a role in the evaluation of exotics. In the analysis below, we shall indicate classes of exotics which are related to the described path property. 1.4.1 Path variation Understanding the path variation for a L´evy process boils down to a better understanding of the L´evy–Khintchine formula. We therefore give a sketch proof of Theorem 6 which shows that for any given L´evy triple (γ , σ, ) there exists a L´evy process whose characteristic exponent is given by the L´evy–Khintchine formula. Reconsidering the formula for , note that we may write it in the form 1 2 2 iux (1 − e )(dx) (u) = iuγ + σ u + 2 R\(−1,1) iux + (1 − e + ixu)(dx) 0 0, consider the L´evy processes X(3,) deﬁned by (3,) () = Yt − t x(dx), t ≥ 0 (1.8) Xt 0} be the set of points that a L´evy process can hit. We say a L´evy process can hit points if C = ∅. Kesten (1969) and Bretagnolle (1971) give the following classiﬁcation. Theorem 7 Suppose that X is not a compound Poisson process. Then X can hit points if and only if

1 du < ∞. (1.11) 1 + (u) R Moreover, (i) If σ > 0, then X can hit points and C = R. (ii) If σ = 0, but X is of unbounded variation and X can hit points, then C = R. (iii) If X is of bounded variation, then X can hit points, if and only if, d = 0 where d is the drift in the representation (equation (1.10)) of its L´evy–Khintchine exponent . In this case, C = R unless X or −X is a subordinator and then C = (0, ∞) or C = (−∞, 0), respectively.

L´evy Processes in Finance–Coarse and Fine Path Properties

13

The case of a compound Poisson process will be discussed in Section 1.5.1. Excluding the latter case, from the L´evy–Khintchine formula we have that 1 ((u)) = σ 2 u2 + (1 − cos(ux))(dx) 2 R\{0} and

((u)) = γ u +

R\{0}

(− sin(ux) + ux1{|x| 0 we have

p 1 du < ∞ 1 + (u) −p and

−p

−∞

∞ 1 1 du = du 1 + (u) 1 + (u) p

and the question as to whether the integral (equation is ﬁnite or inﬁnite depends (1.11)) 1 on what happens when u → ∞. If, for example, 1+(u) g(u) when u → ∞, then we can use g to deduce whether the integral (equation (1.11)) is ﬁnite or inﬁnite. Note, we use the notation f g to mean that there exists a p > 0, a > 0 and b > 0 such that ag(u) ≤ f (u) ≤ bg(u) for all u ≥ p, This technique will be used quite a lot in the examples we consider later on in the text. An example of an exotic option which in principle makes use of the ability of a L´evy process to hit points is the so-called callable put option. This option belongs to a more general class of exotics called Game or Israeli options, described in Kifer (2000) (see also the review by K¨uhn and Kallsen (2005) in this volume). Roughly speaking, these options have the same structure as American-type options but for one signiﬁcant difference. The writer also has the option to cancel the contract at any time before its expiry. The consequence of the writer cancelling the contract is that the holder is paid what they would have received had they exercised at that moment, plus an additional amount (considered as a penalty for the writer). When the claim of the holder is the same as that of the American put and the penalty of the writer is a constant, δ, then this option has been named a callable put in K¨uhn and Kyprianou (2005) (also an Israeli δ-penalty put option in Kyprianou (2004)). In the latter two papers, the value and optimal strategies of writer and holder of this exotic option have been calculated explicitly for the Black–Scholes market. It turns out there that the optimal strategy of the writer is to cancel the option when the value of the underlying asset hits precisely the strike price, providing that this happens early on enough in the

14

Exotic Option Pricing and Advanced L´evy Models

contract. Clearly, this strategy takes advantage of case (i) of the above theorem. Suppose now for the same exotic option that instead of an exponential Brownian motion we work with an exponential L´evy process which cannot hit points. What would be the optimal strategies of the writer (and hence the holder)? 1.4.3 Creeping Deﬁne for each x ≥ 0 the ﬁrst passage time τx+ = inf{t > 0 : Xt > x}. Here, we work with the deﬁnitions inf ∅ = ∞ and if τx+ = ∞, then Xτx+ = ∞. We say that a L´evy process X creeps upwards if for all x ≥ 0 P (Xτx+ = x) > 0 and that X creeps downwards if −X creeps upwards. Creeping simply means that with positive probability, a path of a L´evy process continuously passes a ﬁxed level instead of jumping over it. A deep and yet enchanting aspect of L´evy processes, excursion theory, allows for the following non-trivial deduction concerning the range of {Xτx+ : x ≥ 0}. With probability one, the random set {Xτx+ : x ≥ 0} ∩ [0, ∞) corresponds precisely to the range of a certain subordinator, killed at an independent exponential time with parameter q ≥ 0. The case that q = 0 should be understood to mean that there is no killing and hence that τx+ < ∞ almost surely for all x ≥ 0. In the obvious way, by considering −X, we may draw the same conclusions for the range of {−Xτx− : x ≥ 0} ∩ [0, ∞) where τx− := inf{t > 0 : Xt < x}. Suppose that κ(u) and κ (u) are the characteristic exponents of the aforementioned subordinators for the ranges of the upward and downward ﬁrst passage processes, respectively. Note, for example, that for u ∈ R κ(u) = q − iau + (1 − eiux )π(dx)

(0,∞)

for some π satisfying 0∞ (1 ∧ x)π(dx) < ∞ and a ≥ 0 (recall that q is the killing rate). It is now clear from Theorem 7 that X creeps upwards, if and only if, a > 0. The so-called Wiener–Hopf factorization tells us where these two exponents κ and κ are to be found: (u) = κ(u) κ (−u).

(1.12)

Unfortunately, there are very few examples of L´evy processes for which the factors κ and κ are known. Nonetheless, the following complete characterization of upward creeping has been established. Theorem 8 The L´evy process X creeps upwards, if and only if, one of the following three situations occurs: (i) X has bounded variation and d > 0 where d is the drift in the representation (equation (1.10)) of its L´evy–Khintchine exponent .

L´evy Processes in Finance–Coarse and Fine Path Properties

(ii) X has a Gaussian component, (σ > 0). (iii) X has unbounded variation, no Gaussian component and 1 x([x, ∞)) dx < ∞. 0 y 0 −x −1 ((−∞, u])dudy

15

(1.13)

This theorem is the collective work of Miller (1973) and Rogers (1984), with the crowning conclusion in case (iii) being given recently by Vigon (2002). As far as collective statements about creeping upwards and downwards are concerned, the situation is fairly straightforward to resolve with the help of the following easily proved lemma. (See Bertoin (1996), p. 16). Lemma 9 Let X be a L´evy process with characteristic exponent (u). (i) If X has ﬁnite variation then lim

u↑∞

(u) = −id u

where d is the drift appearing in the representation (equation (1.10)) of . (ii) For a Gaussian coefﬁcient σ ≥ 0, lim

u↑∞

(u) 1 = σ 2. u2 2

From the above lemma we see, for example, that lim

u↑∞

κ(u) = 0, u

if and only if, X creeps upwards. Consequently, from the Wiener–Hopf factorization (equation (1.12)) the following well-established result holds (see Bertoin (1996), p. 175). Lemma 10 A L´evy process creeps both upwards and downwards, if and only if it has a Gaussian component. There is also a relation between hitting points and creeping. Clearly, a process which creeps can hit points. In the case of bounded variation we see that hitting points is equivalent to creeping upwards or downwards. However, in the case of unbounded variation, it can be that a process does not creep upwards or downwards, but still can hit points. We will see an example of this later on–see Remark 17. A process which hits a point but does not creep over it must therefore do so by jumping above and below that point an inﬁnite number of times before hitting it. When considering the relevance of creeping to exotic option pricing, one need only consider any kind of option involving ﬁrst passage. This would include, for example, barrier options as well as Russian and American put options. Taking the latter case with inﬁnite horizon, the optimal strategy is given by ﬁrst passage below a ﬁxed value of the underlying L´evy process. The value of this option may thus be split into two parts, namely, the premium for exercise by jumping clear of the boundary and the premium for creeping over the boundary. For the ﬁnite expiry case, it is known that the optimal strategy of the holder is

16

Exotic Option Pricing and Advanced L´evy Models

to exercise when the underlying L´evy process crosses a time-varying barrier. In this case, a more general concept of creeping over moving boundaries may be introduced and it would be interesting to know whether the ability to creep over the optimal exercise boundary has any inﬂuence on the continuity or smoothness properties of the boundary as a function of time. 1.4.4 Regularity of the half line For a L´evy process X (which starts at zero) we say that 0 is regular for (0, ∞) (equiv. the upper half line) if X enters (0, ∞) immediately. That is to say, if P(τ (0,∞) = 0) = 1,

where

τ (0,∞) = inf{t > 0 : Xt ∈ (0, ∞)}.

Because of the Blumenthal 0–1 law, the probability P(τ (0,∞) = 0) is necessarily zero or one. When this probability is zero, we say that 0 is irregular for (0, ∞). We also say that 0 is regular for (−∞, 0) (equiv. the lower half line) if −X is regular for the upper half line. The following theorem is the conclusion of a number of works and gives a complete characterization of regularity for the upper half line (see Shtatland (1965), Rogozin (1968) and Bertoin (1997)). Theorem 11 For a L´evy process X, the point 0 is regular for (0, ∞), if and only if, one of the following three situations occurs: (i) X is a process of unbounded variation. (ii) X is a process of bounded variation and d > 0 where d is the drift in the representation (equation (1.10)) of its L´evy–Khintchine exponent . (iii) X is a process of bounded variation, d = 0 (with d as in (ii)) and

1 0

x 0

x(dx) = ∞. (−∞, −y)dy

(1.14)

Regularity of the lower half line has already proved to be of special interest to the pricing of American put options. In Alili and Kyprianou (2004). a perpetual American put is considered where the underlying market is driven by a general L´evy process. Building on the work of Mordecki (1999, 2002), Boyarchenko and Levendorskii (2002a) and Chan (2000, 2004), it is shown that the traditional condition of smooth pasting at the optimal exercise boundary may no longer be taken for granted. Indeed necessary and sufﬁcient conditions are given for no smooth pasting. This condition is quite simply the regularity of (−∞, 0) for 0 (in other words the regularity of the upper half line for −X). It was conjectured in Alili and Kyprianou (2005) that the very same condition would also characterize the appearance of smooth ﬁt for the ﬁnite expiry American put where the boundary is time varying. Indeed, numerical simulations in Matache et al. (2003) and Almendral (2004) support this conjecture. A ﬁnancial interpretation of a non-smooth ﬁt condition has yet to be clariﬁed.

L´evy Processes in Finance–Coarse and Fine Path Properties

1.5

17

EXAMPLES REVISITED

1.5.1 Compound Poisson processes and jump-diffusions Suppose that X is a compound Poisson process. Clearly, X has paths of bounded variation, cannot creep upwards or downwards and is irregular for the upper and lower half lines. Since the L´evy measure is bounded, it is easy to reason that the real and imaginary part 1 of its characteristic (u), and hence 1+(u) , is bounded away from zero and that the integral (equation (1.11)) is inﬁnite. Nonetheless, certain compound Poisson processes can hit points. Take the simple example of a Poisson process. This is a process which hits {0, 1, 2, . . . .}. Other similar examples where the jump distribution is supported on a lattice are possible. This is the reason for the exclusion of compound Poisson processes from Theorem 7. However, it can be said that so long as the jump distribution F is diffuse, a compound Poisson process can hit no point other than 0, its initial holding point. If X is a jump-diffusion then the above properties change drastically. In particular, if the Gaussian component is non-zero then this will dominate the paths of the process. This is because, until the ﬁrst jump, which occurs at arbitrarily large times with positive probability, the process behaves as a Brownian motion with drift. It is clear that paths will be of unbounded variation, there will be regularity for the upper and lower half lines, the process may creep both upwards and downwards and any point can be hit with positive probability. Note the latter fact is a well-known property of Brownian motion and does not require Theorem 7. 1.5.2 Spectrally negative processes By deﬁnition, spectrally negative processes creep upwards and hence can hit points. Therefore, unless there is a Gaussian component present, they cannot creep downwards. It is possible to have such processes of both bounded and unbounded variation according to the 0 ﬁniteness of the integral −1 |x|(dx). Clearly, if it is a process of unbounded variation, then there is regularity for the upper and lower half lines. If it is a process of bounded variation and not the negative of the subordinator, then by reconsidering equation (1.10) we see that necessarily the process must take the form of a strictly positive drift minus a subordinator. Consequently, from Theorem 11 in this case, there is regularity for the upper half line but not for the lower half line. This, in turn, implies that for spectrally negative L´evy processes, regularity of the lower half line coincides with having paths of inﬁnite variation. 1.5.3 Meixner process We begin with a known fact concerning path variation. Proposition 12 The Meixner process is of unbounded variation and hence is regular for the upper and lower half lines. Proof. Denote ν(x) as the density of the L´evy measure (equation (1.4)). We have to prove that (−1,1) |x|ν(x)dx is inﬁnite. For x ∈ (0, 1) we have |x|ν(x) = δ

e(β+π)x/α 1 exp(βx/α) = 2δ 2πx/α ≥ 2δ 2πx/α sinh(π x/α) e −1 e −1

18

Exotic Option Pricing and Advanced L´evy Models

and so

1

1

|x|ν(x)dx ≥

0

2δ 0

showing in particular that

1 e2πx/α

−1

dx

1 (2π/α)x 1 = 2δ log e − 1 − x = ∞ 2π/α 0

(−1,1) |x|ν(x)dx

= ∞.

Proposition 13 A Meixner process cannot hit points and therefore cannot creep. Proof. To see whether the Meixner process can hit points we have to employ the integral given in equation (1.11). In order to use this, we ﬁrst split the characteristic exponent into its real and imaginary part. First note that

1 1 1 1 αu − iβ = cos β cosh αu − sin β sinh αu i. cosh 2 2 2 2 2 Then with

2

2 1 1 1 1 β cosh αu β sinh αu + sin 2 2 2 2 1 1 − sin 2 β sinh 2 αu q = arctan cos 12 β cosh 12 αu r=

cos

and

we have (u) = −2δ log(cos(β/2)) + 2δ log(r) + (2δq − mu)i and hence

r 1 + 2δ log cos(β/2) − (2δq − mu)i 1 = and 2 1 + (u) r 1 + 2δ log cos(β/2) + (2δq − mu)2 r

1 + 2δ log cos(β/2) 1 . = 2 1 + (u) r 1 + 2δ log cos(β/2) + (2δq − mu)2

When u → ∞, then cosh(αu/2) sinh(αu/2) eαu/2 , and so r eαu/2 and log(r/ cos(β/2)) 12 αu when u → ∞. Further, (2δq − mu)2 m2 u2 when u → ∞, because arctan(z) ∈ (− 12 π, 12 π ) for all z ∈ R. So

1 + δαu 1 u−1 when u → ∞. 1 + (u) (1 + δαu)2 + m2 u2 ∞ Because for all p > 0, p u−1 du = ∞, we ﬁnd that the integral (equation (1.11)) is inﬁnite and therefore the Meixner process cannot hit points.

L´evy Processes in Finance–Coarse and Fine Path Properties

19

1.5.4 Generalized tempered stable process We again begin with a known statement concerning path variation. Proposition 14 The generalized tempered stable process has bounded variation, if and only if, αp < 1 and αn < 1. Proof. We have to determine whether the integral given in equation (1.9) is ﬁnite or inﬁnite, where this integral is given by 1 0 1 cp −λp x cn |x|ν(x)dx = e dx + eλn x dx. α αn p −1 0 x −1 (−x) It is clear, however, that this boils down to whether 0 1 −αp x dx + (−x)−αn dx −1

0

is ﬁnite or inﬁnite and the above expression is only ﬁnite when αp < 1 and αn < 1. Proposition 15 In the case of unbounded variation, a generalized tempered stable process creeps upwards, if and only if, αp < αn . Proof. Because the integral given in equation (1.13) is independent of cp and cn , we can assume without loss of generality that cp = cn = 1. In the following calculations, b1 , b2 , . . . are constants in R. Recall from the previous proposition that we have unbounded variation if αp ∈ [1, 2) or αn ∈ [1, 2). We shall therefore prove the result under the additional assumption that there is unbounded variation because αn ∈ [1, 2). Similar arguments then deliver the same conclusions when we assume that there is unbounded variation because αp ∈ [1, 2). For u ∈ (0, 1] we have ([u, ∞)) ∞ e−λp x x −(1+αp ) dx = u

∞ 1 −λp −(1+αp ) e x dx + e−λp x dx = b1 + b2 e−λp u 1 u 1 ∞ ≤ e−λp x dx + x −1 dx = b3 − log(u) 1 u 1 ∞ e−λp x dx + x −(1+αp ) dx = b4 + b5 u−αp

if αp ≤ −1 if αp = 0 if αp ∈ (−1, 2)\{0}

u

1

and for u ∈ [−1, 0) ((−∞, u]) = ≥

u

−∞ u −1

eλn x (−x)−(1+αn ) dx

e−λn (−x)−(1+αn ) dx = b6 (−u)−αn + b7

20

Exotic Option Pricing and Advanced L´evy Models

if αn ∈ [1, 2). Hence y b8 (−y)−αn +1 + b9 + b10 (y + 1) if αn ∈ (1, 2) ((−∞, u]) du ≥ if αn = 1 b11 log(−y) + b12 (y + 1) −1 and then 0 −x

y −1

((−∞, u]) du dy ≥

b13 x −αn +2 + b14 x 2 + b15 x b16 (x log(x) − x) + b17 x 2 + b18 x

if αn ∈ (1, 2) if αn = 1.

We then have for αp ∈ (−1, 2)\{0} and αn ∈ (1, 2) 0 y

x([x, ∞))

−x −1 ((−∞, u]) du dy

≤

b4 x + b5 x −αp +1 −α b13 x n +2 + b14 x 2 + b15 x

=

b4 + b5 x −αp , b13 x −αn +1 + b14 x + b15

for x ∈ (0, 1]. Deﬁne the right side of the above inequality f (x) and note that it is continuous for all x ∈ (0, 1]. When x → 0, then f x αn −1 for αp < 0 and f x αn −1−αp

for αp > 0.

1 Hence, 0 f (x)dx < ∞ for αp < αn and therefore the integral (equation (1.13)) is ﬁnite for these parameter values. When αp ≤ −1 or αp = 0 or αn = 1, there is a similar upper bound for which the same conclusions can be drawn. We have thus so far shown that there is creeping upwards if αp < αn . To prove the ‘only-if’ part, note that the L´evy density of −X is ν(−x) and this density is the same as ν(x) except that the p-parameters and the n-parameters have switched places. So, we can immediately conclude from the previous analysis that X creeps downwards if αp > αn and then from Lemma 10 we see that since there is no Gaussian component, X cannot creep upwards if αp > αn . Now only the case remains when αp = αn ∈ [1, 2). In this case, we can use for u ∈ (0, 1] the lower bound for ((−∞, −u]) as a lower bound for ([u, ∞)) and the upper bound for ([u, ∞)) as an upper bound for ((−∞, −u]) in order to create a lower bound for the integral (equation (1.13)) which turns out to be inﬁnite. Proposition 16 In the case of unbounded variation, a generalized tempered stable process can hit points, unless αp = αn = 1 and cp = cn . Proof. Because this process creeps upwards or downwards when αp = αn , we only have to prove that the process can hit points when αp = αn = α ∈ [1, 2). Let rp = 2 qp = arctan − λup , rn = 1 + λu2 and qn = arctan λun . Then n

Ap (u) = βp (rpα cos(αqp ) − 1) + iβp rpα sin(αqp ) + αu λp αu α α An (u) = βn (rn cos(αqn ) − 1) + iβn rn sin(αqn ) − λn ,

1+

u2 , λ2p

L´evy Processes in Finance–Coarse and Fine Path Properties

21

with βp = (−α)λαp cp and βn = (−α)λαn cn . So, for αp = αn ∈ (1, 2)

1 1 + (u)

=

1 − (Ap (u)) − (An (u)) [1 − (Ap (u)) − (An (u))]2 + [uγ − (Ap (u)) − (An (u))]2

≤

1 − (Ap (u)) − (An (u)) 1 . = [1 − (Ap (u)) − (An (u))]2 1 − (Ap (u)) − (An (u))

1 We see that the above upper bound of 1+(u) as a function of u is continuous and symmetric. So, the question whether the integral of this function from minus inﬁnity to inﬁnity is ﬁnite or inﬁnite depends on how the function behaves when u → ∞. When u → ∞, then rp rn u, qp → − 12 π and qn → 12 π . So, cos(αqp ) = cos(αqn ) → a for u → ∞, where a is a constant smaller than zero. Because (−α) > 0 for α ∈ (1, 2), we have ∞ that 1 − (Ap (u)) − (An (u)) uα when u → ∞. Because for all t > 0 t u−α du < ∞, the integral of the upper bound is ﬁnite and hence this process can hit points when αp = αn ∈ (1, 2). Now, let αp = αn = 1. Then by using the same expressions for rp , rn , qp and qn as above, Ap (u) = cp (λp log(rp ) + uqp ) + icp (u + λp qp − u log(rp )) An (u) = cn (λn log(rn ) − uqn ) + icn (−u + λn qn + u log(rn )) and then when u → ∞, 1 − (Ap (u)) − (An (u)) u if cp = cn uγ − (Ap (u)) − (An (u)) u log(u) uγ − (Ap (u)) − (An (u)) u or uγ − (Ap (u)) − (An (u)) 1 if cp = cn . So, when cp = cn , then

1 1+(u)

∞ t

1 u log2 (u)

and because for large t,

1 1 ∞ =− < ∞, log(u) t u log2 (u)

the process can hit points. For the case where cp = cn , then integral given in equation (1.11) is inﬁnite.

1 1+(u)

u−1 and the

Remark 17 The last two propositions give us an example of a L´evy process which can hit points but cannot creep. Take the example of a CGMY process where the parameter Y ∈ (1, 2). To some extent this is not surprising. As noted earlier, the small time behaviour of generalized tempered stable processes should in principle be similar to the behaviour of stable processes due to the similarities in their L´evy measures in the neighbourhood of the origin. In this sense, the class of CGMY processes mentioned are closely related to a symmetric stable processes of unbounded variation and for this class it is well known that they can hit points but cannot creep. To see the latter fact, note from Lemma 10 that it is clear that a symmetric stable process (or indeed any L´evy process which is symmetric without a Gaussian component) cannot creep upwards nor downwards on account of symmetry. On the

22

Exotic Option Pricing and Advanced L´evy Models

other hand, it is well known (cf. Chapter VIII in Bertoin (1996)) that for a symmetric stable process of index α, (u) = c|u|α for some constant c > 0, and there is unbounded variation when α ∈ (1, 2). It is easily veriﬁed that with this choice of , the integral (equation (1.11)) is ﬁnite. In the case that a generalized tempered stable process has bounded variation, that is, when αp < 1 and αn < 1, we can use the drift d to determine whether the process can hit points, creeps or whether 0 is regular for (0, ∞). However, then we have to know ∞ what the drift looks like. Comparing equations (1.2) and (1.6), we see that d = −γ − −∞ xν(x)dx. The latter integral can be computed explicitly; however, it is easier to use Lemma 9 (i) to identify the drift term since the L´evy–Khintchine formula is polynomial when there is bounded variation. Indeed, it is easy to see by inspection that d = −γ − dp − dn , where

α −1 −cp αp (−αp )λpp cp dp = λp cn αn (−αn )λnαn −1 if cn dn = if − λn

if αp ∈ (−∞, 1)\{0} and

if αp = 0 αn ∈ (−∞, 1)\{0} αn = 0.

Remark 18 For the special cases of the CGMY and variance gamma processes, we note that the representation of the L´evy–Khintchine formula given in equations (1.6) and (1.7) yields in both cases that the drift term d = µ. Proposition 19 When a generalized tempered stable process has bounded variation and drift equal to zero, then 0 is regular for (0, ∞), if and only if, αp ≥ αn and at least one of these two parameters is not smaller than zero. Proof. We can use the integral (equation (1.14)) to determine whether 0 is regular for (0, ∞). For y ∈ (0, 1) we have −y −y c n e λn x c n e λn x dx ≥ dx ((−∞, −y)) = 1+αn 1+αn −∞ (−x) −1 (−x) −y c n e λn b1 y −αn + b2 if αn = 0 dx = ≥ 1+α n b3 log(y) if αn = 0 −1 (−x) and ((−∞, −y)) = ≤

−1 −∞ −1 −∞

=

c n e λn x dx + (−x)1+αn λn x

cn e dx + (−x)1+αn

b4 + b5 y −αn b6 + b7 log(y)

−y −1 −y −1

c n e λn x dx (−x)1+αn cn dx (−x)1+αn

if αn = 0 if αn = 0,

L´evy Processes in Finance–Coarse and Fine Path Properties

where b1 , b2 , . . . are constants in R. So, x b2 x + b8 x −αn +1 ≤ 0 ((−∞, −y))dy ≤ b4 x + b9 x −αn +1

b3 (x log(x) − x) ≤

x 0

23

if αn = 0

((−∞, −y))dy ≤ b6 x + b7 (x log(x) − x)

if αn = 0.

Note that the constants b1 , b2 , . . . have values such that the above upper and lower boundaries are strictly positive for x > 0. Now, when αp < 1, αn < 1 and αn = 0, then x 0

xcp x −1−αp e−λp x cp x −1−αp e−λp x xν(x) ≤ = . −α +1 n b2 x + b8 x b2 + b8 x −αn ((−∞, −y))dy

Let f (x) be the value of the right-hand side of above inequality for x ∈ (0, 1). Then, as x → 0, then f (x) x −1−αp

if αn < 0 and f (x) x αn −αp −1

if αn > 0.

We conclude that the upper bound is ﬁnite if αp < 0 and αn < 0 and if αp < αn and hence therefore 0 is irregular for (0, ∞) in these cases. Because the lower bound on x 0 ((−∞, −y))dy has the same form as the upper bound, we can immediately conclude that in the other cases when αn = 0, 0 is regular for (0, ∞). Now only the case remains when αn = 0. Here we have x 0

cp x −1−αp e−λp x xcp x −1−αp e−λp x xν(x) = . ≤ b3 (x log(x) − x) b3 (log(x) − 1) ((−∞, −y))dy

Let g(x) be the value of the right-hand side of above inequality for x ∈ (0, 1). Then, when x → 0, g(x)

−1 . x 1+αp log(x)

Because for all t < 1, 0

t

−1 dx < ∞, log(x)

x 1+αp

if and only if, αp < 0, the upper bound is ﬁnite in this case and hence the integral (equation (1.14)) is ﬁnite. The lower bound has again the same form as the upper bound and so we conclude that this integral (equation (1.14)) is inﬁnite when αp ≥ 0 and αn = 0. 1.5.5 Generalized hyperbolic process Because the L´evy measure of this process is very complicated, it is very difﬁcult to use this measure to determine whether the process is of ﬁnite of inﬁnite variation. However, this can also be determined by using the characteristic exponent with the help of Lemma 9. We follow the ideas in given Cont and Tankov (2004). Proposition 20 A generalized hyperbolic process is of unbounded variation and has no Gaussian component and hence 0 is regular for the upper and lower half lines.

24

Exotic Option Pricing and Advanced L´evy Models

Proof. Using the properties of the logarithm we have λ λ (u) = − log(α 2 − β 2 ) + log(α 2 − (β + iu)2 ) 2 2 − log Kλ δ α 2 − (β + iu)2 + log Kλ δ α 2 − β 2 − iµu. Let r =

−2βu (α 2 − β 2 + u2 )2 + 4β 2 u2 and q = arctan α 2 −β 2 +u2 . Then √ 1 λ λ (u) = − log(α 2 − β 2 ) + (log(r) + iq) − log Kλ δ re 2 qi 2 2 √ 1 + log Kλ δ α 2 − β 2 − iµu. − log Kλ δ re 2 qi

When u → ∞, r ∼ u2 and q → 0. The modiﬁed Bessel function, Kλ , has the following property: if a → ∞ then Kλ (a + bi) ∼ e−(a+bi) √ 1 √ 1 qi So, Kλ δ re 2 qi ∼ e−δ re 2

π √ 1 qi 2δ re 2

π . 2(a + bi)

and therefore

√ 1 √ √ 1 1 1 qi log Kλ δ re 2 q + log(π ) − log 2δ r ∼ −δ r cos 2 2 2 and

√ 1 √ 1 1 ∼ −δ r sin log Kλ δ re 2 qi q − q 2 4 when u → ∞. So, ((u)) ∼ δu and we conclude from Lemma 9 that the process is of inﬁnite variation and has no Gaussian component. Proposition 21 A generalized hyperbolic process cannot hit points and hence cannot creep. Proof. We have seen from the proof above that ((u)) ∼ δu and that ((u)) ∼ 1 √ 1 δ r sin 2 q + µu, when u → ∞. This implies that 1+(u) u−1 when u → ∞ and therefore the process cannot hit points.

1.6 CONCLUSIONS Let us conclude with some tables with our ﬁndings for some of the more popular models we have mentioned. It will be useful to recall the notation C = {x ∈ R : P (Xt = x for at least one t > 0) > 0}.

L´evy Processes in Finance–Coarse and Fine Path Properties

25

Meixner processes

2δ cos(β/2) Meixner (u) = − log − iµu. cosh((αu − iβ)/2 Path variation: Hitting points: Creeping: Regularity:

Unbounded variation. C = ∅. No creeping upwards or downwards. Always regular for (0, ∞) and (−∞, 0). CGMY processes

CGMY (u) = −C (−Y )[(M − iu)Y − M Y + (G + iu)Y − GY )] − iuµ. Path variation: Hitting points: Creeping: Regularity:

Unbounded variation ⇔ Y ∈ [1, 2). C = ∅ ⇔ Y = 1 or Y ∈ (0, 1) and µ = 0, otherwise C = R. Upwards creeping ⇔ Y ∈ (0, 1) and µ > 0. Downwards creeping ⇔ Y ∈ (0, 1) and µ < 0. Irregular for (0, ∞) ⇔ Y ∈ (0, 1) and µ < 0. Irregular for (−∞, 0) ⇔ Y ∈ (0, 1) and µ > 0. Variance gamma processes

iu iu + log 1 + − iuµ. VG (u) = C log 1 − M G

Path variation: Hitting points: Creeping: Regularity:

Bounded variation C = ∅ ⇔ µ = 0, otherwise C = R. Upwards creeping ⇔ µ > 0, Downwards creeping ⇔ µ < 0 Regular for (0, ∞) ⇔ µ ≥ 0. Regular for (−∞, 0) ⇔ µ ≤ 0. Generalized hyperbolic processes

GH (u) = − log Path variation: Hitting points: Creeping: Regularity:

α2 − β 2 α 2 − (β + iu)2

λ/2

Kλ (δ α 2 − (β + iu)2 ) − iµu. Kλ (δ α 2 − β 2 )

Unbounded variation. C=∅ No creeping upwards or downwards. Always regular for (0, ∞) and (−∞, 0).

ACKNOWLEDGEMENTS We would like to express our thanks to Antonis Papapantoleon and a referee for careful reading of this manuscript.

26

Exotic Option Pricing and Advanced L´evy Models

REFERENCES [1] Alili, L. and Kyprianou, A. (2005), “Some remarks on ﬁrst passage of L´evy process, the American put and pasting principles”, Annals of Applied Probability, to be published. [2] Almendral, A. (2005), “Numerical valuation of American options under the CGMY process”, in A. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 259–276. [3] Applebaum, D. (2004), L´evy Processes and Stochastic Calculus, Cambridge Studies in Advanced Mathematics, Vol. 93, Cambridge University Press, Cambridge, UK. [4] Asmussen, S., Avram, F. and Pistorius, M.R. (2004), “Russian and American put options under exponential phase-type L´evy models”, Stochastic Processes and their Applications, 109, 79–111. [5] Avram F., Chan T. and Usabel M. (2002), “On the valuation of constant barrier options under spectrally one-sided exponential L´evy models and Carr’s approximation for American puts”, Stochastic Processes and their Applications, 100, 75–107. [6] Avram, F., Kyprianou, A.E. and Pistorius, M.R. (2004), “Exit problems for spectrally negative L´evy processes and applications to Russian, American and Canadized options”, Annals of Applied Probability, 14, 215–238. [7] Barndorff-Nielsen, O.E. (1995), “Processes of normal inverse Gaussian type”, Finance and Stochastics, 2, 41–68. [8] Barndorff-Nielsen, O.E. (1998), “Normal inverse Gaussian distributions and the modeling of stock returns”, Research Report No. 300, Department of Theoretical Statistics, Aarhus University, Denmark. [9] Barndorff-Nielsen, O.E. and Shephard, N. (2001), “Non-Gaussian Ornstein–Ulenbeck based models and some of their uses in ﬁnancial econometrics”, Journal of the Royal Statistical Society, B, 63, 167–241. [10] Bertoin, J. (1996), L´evy Processes, Cambridge University Press, Cambridge Tracts in Mathematics, Vol. 121, Cambridge, UK. [11] Bertoin, J. (1997), “Regularity of the half-line for L´evy processes”, Bulletin des Sciences Math´ematiques, 121, 345–354. [12] Boyarchenko, S.I. and Levendorskii, S.Z. (2002), Non-Gaussian Merton–Black–Scholes theory, Advanced Series on Statistical Science and Applied Probability, Vol. 9, World Scientiﬁc, River Edge, NJ, USA. [13] Boyarchenko, S.I and Levendorskii, S.Z. (2002a), “Perpetual American options under L´evy processes”, SIAM Journal on Control and Optimization, 40, 1663–1696. [14] Bretagnolle, J. (1971), “R´esultats de Kesten sur les processus a` accroissements ind´ependants”, in S´eminaire de Probabilit´es V, (Universitaire Strasbourg, Ann´ee Universitaire 1969–1970 ), Lecture Notes in Mathematics, Vol. 191, Springer-Verlag, Berlin, Germany, pp. 21–36 (in French). [15] Carr, P. and Wu, L. (2003), “The ﬁnite moment log stable process and option pricing”, Journal of Finance, 58, 753–778. [16] Carr, P., Geman, H., Madan, D.B. and Yor, M. (2002), “The ﬁne structure of asset returns: an empirical investigation”, Journal of Business, 75, 305–332. [17] Carr, P., Geman, H., Madan, D.B. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [18] Cartea, A. and Howison, A. (2005), “Option pricing with L´evy-stable process”, to be published. [19] Chan, T. (2000), “American options driven spectrally by one sided L´evy processes”, Unpublished manuscript (see also Chapter 9 in this volume). [20] Chan, T. (2004), “Some applications of L´evy processes in insurance and ﬁnance”, Finance, 25, 71–94.

L´evy Processes in Finance–Coarse and Fine Path Properties

27

[21] Cont, R. (2001), “Empirical properties of asset returns: stylized facts and statistical issues”, Quantitive Finance, 1, 223–236. [22] Cont, R. and Tankov, P. (2004), Financial Modeling with Jump Processes, Chapman and Hall/CRC Financial Mathematics Series, Boca Raton, FL, USA. [23] Eberlein, E. and van Hammerstein, E.A. (2004), “Generalized hyperbolic and inverse Gaussian distributions: limiting cases and approximation of processes”, Seminar on Stochastic Analysis, Random Fields and Applications IV, Progress in Probability, Vol. 58, Birkh¨auser Verlag, Basel, Switzerland, pp. 221–264. [24] Eberlein, E. and Prause, K. (2002), “The generalized hyperbolic model: ﬁnancial derivatives and risk measures”, in H. Geman, D. Madan, S. Pliska and T. Vorst (Eds), Mathematical FinanceBachelier Congress 2000, Springer-Verlag, Berlin, Germany, pp. 245–267. [25] Grigelionis, B. (1999), “Processes of Meixner type”, Lithuanian Mathematics Journal, 39, 33–41. [26] Halgreen, C. (1979), “Self-decomposability of the generalized inverse Gaussian and hyperbolic distributions”, Zeitschrift f¨ur Wahrscheinlichkeitstheorie und Verwandte Gebiete, 47, 13–17. [27] Kesten, H. (1969), “Hitting probabilities of single points for processes with stationary independent increments”, in Memoirs of the American Mathematical Society, Vol. 93, American Mathematical Society, Providence, RI, USA, pp. 1–129. [28] Kifer, Y. (2000), “Game options”, Finance and Stochastics, 4, 443–463. [29] Koponen, I. (1995), “Analytic approach to the problem of convergence of truncated L´evy ﬂights towards the Gaussian stochastic process”, Physical Review, E, 52, 1197–1199. [30] Kou, S.G. (2002), “A jump-diffusion model for option pricing”, Management Science, 48, 1086–1101. [31] K¨uhn, C. and Kyprianou, A.E. (2003), “Callable puts as composite exotic options”, to be published. [32] K¨uhn, C. and Kallsen, J. (2005), “Convertible bonds: ﬁnancial derivatives of game type”, in A. Kyprianou, W. Schoutens, and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK. [33] Kyprianou, A.E. (2004), “Some calculations for Israeli options”, Finance and Stochastics, 8, 73–86. [34] Madan, D.P. and Seneta, E. (1987), “The VG for share market returns”, Journal of Business, 63, 511–524. [35] Matache, A.M. Nitsche, P.A. and Schwab, C. (2003), “Wavelet Galerkin pricing of American options on L´evy driven assets”, to be published. [36] Merton, R. (1976), “Option pricing when underlying stock returns are discontinuous”, Journal of Financial Economics, 3, 125–144. [37] Miller, P.W. (1973), “Exit properties of stochastic processes with stationary independent increments”, Transactions of the American Mathematical Society, 178, 459–479. [38] Mordecki, E. (1999), “Optimal stopping for a diffusion with jumps”, Finance and Stochastics, 3, 227–236. [39] Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493. [40] Pitman, J. and Yor, M. (2003), “Inﬁnitely divisible laws associated with hyperbolic functions”, Canadian Journal of Mathematics, 55, 292–330. [41] Rogers, L.C.G. (1984), “A new identity for real L´evy processes”, Annals of the Institute of Henri Poincar´e, 20, 21–34. [42] Rogozin, B.A. (1968), “The local behavior of processes with independent increments”, Teoriga Veroyatnoste¨ui i ee Primeneniya (Theory of Probability and its Applications), 13, 507–512 (in Russian).

28

Exotic Option Pricing and Advanced L´evy Models

[43] Sato, K.-I. (1999), L´evy Processes and Inﬁnitely Divisible Distributions (translated from the 1990 Japanese original, revised by the author), Cambridge Studies in Advanced Mathematics, Vol. 68, Cambridge University Press, Cambridge, UK. [44] Schoutens, W. (2000), Stochastic Processes and Orthogonal Polynomials, Lecture Notes in Statistics, Vol. 146, Springer-Verlag, New York, NY, USA. [45] Schoutens, W. (2001), “The Meixner process in ﬁnance”, EURANDOM Report 2001–2002, EURANDOM, Eindhoven, The Netherlands. [46] Schoutens, W. (2002), “Meixner processes: theory and applications in ﬁnance”, EURANDOM report 2002–2004, EURANDOM, Eindhoven, The Netherlands. [47] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [48] Schoutens, W. and Teugels, J.L. (1998), “L´evy processes, polynomials and martingales”, Communications in Statistics: Stochastic Models, 14, 335–349. [49] Shtatland, E.S. (1965), “On local properties of processes with independent increments”, Teoriga Veroyatnoste¨ui i ee Primeneniya (Theory of Probability and its Applications), 10, 317–322 (in Russian). [50] Vigon, V. (2002), “Votre L´evy rampe-t-il? (Does your Le´vy process creep?)”, Journal of the London Mathematical Society, 2, 65(1), (2), 243–256 (in French).

2 Simulation Methods with L´evy Processes Nick Webber University of Warwick, Coventry, UK Abstract ´ processes are increasingly important tools for modelling asset return’s processes and Levy ´ interest rates. Although when Levy processes are used, direct integration methods are sometimes available to price standard European options, other numerical techniques must generally be employed to price instruments whose pay-offs are either path-dependent or American. This article discusses recent progress made in developing simulation methods suit´ able for some of the most widely used Levy processes. Bridge algorithms are given for the VG and NIG processes and these algorithms are applied to valuing average rate options. We consider the valuation of barrier options. It is shown how simulation bias can be reduced in this case. Once bias is absent, speed-up methods can be applied. Results are presented illustrating the bias reduction achieved for up-and-out and up-and-in barrier options.

2.1

INTRODUCTION

An important objective in ﬁnancial mathematics is to ﬁnd models of asset returns, interest rates and other ﬁnancial processes in order to value and hedge derivative securities, and for VaR and risk management purposes. Once a model has been found it is hoped that the relevant values, hedge ratios, reserves and risk factors can be computed. The standard Black–Scholes assumption, used in many applications and not just for simple options pricing, is that asset returns are normally distributed, and that joint returns distributions are normal. Alas, this assumption is false. Marginal distributions are not normal and joint distributions are not jointly normal (so that, in particular, the joint distribution does not have a Gaussian copula). The side-effects of assuming joint normality are unfortunately not ignorable. Market option implied volatilities have distinct smiles and historical returns distributions are fat tailed and skewed – unlike those predicted by a joint normality assumption. To overcome this problem, a number of different approaches are possible, some of which are discussed below. This article focuses on modelling univariate returns as L´evy processes and the application of simulation methods for option valuation. We are particularly concerned with ﬁnding effective numerical methods that can be used in practice to facilitate the calculation of option values when asset returns processes are L´evy.

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

30

Exotic Option Pricing and Advanced L´evy Models

The next section brieﬂy discusses modelling returns as L´evy processes. In Section 2.3, we discuss an approach towards ﬁnding numerical methods based on the subordinator decomposition of a L´evy process. Sections 2.4 and 2.5 present theory and results related to fast simulation methods, while Section 2.6 applies the methods, together with bias reduction methods, to continuously reset barrier options. The ﬁnal section provides a summary of our conclusions.

2.2

MODELLING PRICE AND RATE MOVEMENTS

Write St for the value at time t ≥ 0 of an asset value, conditional upon S0 . Different approaches to modelling asset price movements are possible, a number of which have been explored in the literature. 1. A standard modelling technique in mathematical ﬁnance is to model S = (St )t≥0 by specifying the SDE it satisﬁes. For example, by specifying SDEs for the asset process and for a stochastic volatility (for instance, Heston (1993) [15]). 2. By specifying the conditional distributions of S. For instance, by (a) giving the conditional distributions F (St | S0 ) themselves, or by (b) giving the densities f (St | S0 ), if they exist (for instance option pricing by log-normal mixtures (see Brigo and Mercurio (2001) [6])). (c) Giving the inverse distribution functions, F −1 (St | S0 ) (for instance, Corrado (2001) [9]). 3. As a time-changed Brownian motion (for instance, Geman et al. (2001) [13]). 4. As a L´evy process determined by its L´evy triple, (a, σ, υ). The process might have a L´evy density k (x), where υ (dx) = k (x) dx. (In practical applications one might approximate the L´evy process as a compound Poisson process.) Some of the many contributions here are cited in the next section. 5. By its time copula (see Bouy´e et al. (2000) [4]). Which ever way one models S, how might one calibrate to prices? It may be possible to use an implied pricing method to ﬁt exactly to an implied volatility surface. More usually, a functional form is speciﬁed, either explicitly or implicitly, and parameter values in the functional form chosen to ﬁt as closely as possible, by some criterion, to prices. We choose here to assume that a L´evy Process is given, whose triple (a, σ, υ) is parameterized, and whose parameter values can be chosen to match to observed prices. 2.2.1 Modelling with L´evy processes Consider an asset price process S = (St )t≥0 . Under the pricing measure with respect to the accumulator account numeraire suppose that St = S0 exp (rt + Lt − ωt)

(2.1)

short rate and ω = ln E exp (L1 ) where L = (Lt )t≥0 is a L´evy process, r is the constant compensates for the drift in L, so that St e−rt t≥0 is a martingale. Equation (2.1) is a

Simulation Methods with L´evy Processes

31

standard assumption in the literature (Madan et al. (1998) [19], Barndorff-Nielsen and Shephard (2000) [2], Eberlein and Keller (1995) [11], etc). Much work has been carried out, for various choices of L, to price options and to calibrate to empirical distributions. Important contributions include Carr et al. (2001) [7], Eberlein (2000) [10], Barndorff-Nielsen (2000) [2] and Rydberg (1999) [25]. There are also a number of interest rate models powered by L´evy processes in the literature (for instance, Eberlein and Raible (1999) [12]). One could consider a model for the short rate r = (rt )t≥0 in which under the pricing measure drt = α (µ (t) − rt ) dt + σ dLt ,

(2.2)

but we do not pursue interest rate models any further here. The problem with equations (2.1) and (2.2) is that, in general, it is hard to price nonvanilla derivatives in models incorporating these processes. For instance, how would one value path-dependent options or Bermudan or American puts? Even for vanilla European options, ﬁnding a price may require a difﬁcult numerical integration of a density function approaching the pathological. In a ‘Black–Scholes world’, where L is a multiple of a Wiener process, even if analytical solutions are unavailable, PDE, Monte Carlo integration and lattice methods can generally be found that provide adequate numerical solutions to many pricing problems. Lattice methods are cheap, ﬂexible and accurate, and can price American and Bermudan options. Monte Carlo methods can often be a long step and very good speed-up methods are available, so that path-dependent options are often quick and easy to price. In a L´evy world, the behaviour of the L´evy measure over short time horizons can cause immense practical problems. Analytic solutions may involve difﬁcult numerical integration (Madan et al. (1998) [19]). Fourier transform methods can work well for European options as long as the time horizon is not too short (Carr and Madan (1999) [8]). Monte Carlo methods can be used, either directly on the asset price process (equation (2.2) or indirectly through a representation of L as a subordinated Brownian motion (Rydberg (1997) [24]), a mean-variance mixture, or though an approximation as a compound Poisson process. However, sampling over a small time step is very hard if a L´evy density becomes unbounded near zero. PDE methods, such as the method of lines, can work for certain processes, but for a general process can prove very hard to use. Lattice methods need very high order branching and again a time step that is not too small. The problem is that on the one hand there are too many small jumps, and on the other there are too many big jumps. 2.2.2 Lattice methods Given the caveats noted above, can a lattice method work at all? The answer turns out to be yes, and hinges on one of the deﬁning properties of L´evy processes: convergence in probability. A process X = (Xt )t≥0 converges in probability if ∀ε > 0, Pr [|Xt − Xs | > ε] → 0 as s → t.

(2.3)

32

Exotic Option Pricing and Advanced L´evy Models

For a process which does not converge in probability, a lattice method might indeed be hard to construct. Consider the following example. Let t ∈ R+ . Deﬁne (Xt )t≥0 as if t ∈ Q then if t ∈ R\Q then

Pr [Xt = 0] = 12 , Pr [Xt = 2] = 12 , Pr [Xt = 0] = 12 ,

(2.4)

Pr [Xt = 1] = 12 .

This process clearly does not converge in probability. A computer only sees (essentially) rational values of t, a set of measure zero and a numerical method would not see the process taking values 1, even though this possibility occurs on a set of measure 1. If a process does converge in probability then, by deﬁnition, as t → 0, the probability of branching further away from a node at Xt than a ﬁxed distance X goes to zero, so that lattice methods cannot immediately be ruled out. Kellezi and Webber (2003) [17] devised a lattice method and applied it to VG and NIG processes. They obtained discrete branching probabilities from several alternative representations of the L´evy processes L. 1. Directly from the density function of Lt . Essentially, this is equivalent to ﬁtting to the characteristic function of the L´evy process. 2. From a representation as a subordinated Brownian motion, when the subordinator representation is known. 3. From the representation as a L´evy triple (a, σ, υ), via an approximation as a compound Poisson process. 4. From the moments of the process. The last possibility can be rejected very rapidly. Even if the moments of the process are known, it is still possible to match only a ﬁnite number of them. In any case, moment matching is equivalent to ﬁtting the characteristic function only at zero. Kellezi and Webber (2004) [17] found it preferable to construct a lattice directly from the density function (known for the examples they give). Nevertheless, their lattice has very high order branching, is relatively slow, still runs into problems when attempting to price American options, and in any case cannot value path-dependent options. Instead of investigating lattice methods any further, this article now turns its attention to simulation methods. Although these may not be usable with Bermudan or American options (although perhaps primal-dual methods could still work), they can value path-dependent options. 2.2.3 Simulation methods It is often not possible to directly and accurately simulate a L´evy process; it may not be possible to sample directly from the distribution of Lt . When it is not possible to simulate directly from some distribution, either a terminal distribution in a long-step Monte Carlo method, or a distribution, exact or otherwise, representing a small time step t, several alternatives are possible. For instance: 1. Represent the distribution as a mean-variance mixture.

Simulation Methods with L´evy Processes

33

2. Express an underlying process as a time-changed Brownian motion.1 3. In the worst case, approximate the L´evy measure as compound Poisson. Given two densities, f (x | α) and g(α), where α ∈ B and g is a density on B, their mixture distribution has the density f (x | y) g (y) dy.

fg (x) =

(2.5)

B

When f depends upon a single parameter α via its mean and variance, fg (x) is a meanvariance mixture. Given such a representation, it may be possible to (i) draw from g to get a y, and then (ii) draw from f (x | y) for x. If a time-change representation exists, so that Xt = wh(t) for a Brownian motion w and a time change h (see below), then it may be possible to (i) sample from h (t) to get a random time τ , and then (ii) sample from wτ . The time-change representation is used in the rest of this article.

2.3

A BASIS FOR A NUMERICAL APPROACH

We exploit the time-change representation of a L´evy process. Let X = (Xt )t≥0 be a onedimensional semi-martingale, and then (Monroe (1978) [20]) X is representable as a timechanged Brownian motion, Xt = wh(t) ,

(2.6)

where w is a Brownian motion, with drift µ and volatility σ , say, and h = (ht )t≥0 is a stochastic time change. A L´evy process L is a semi-martingale and so this representation can be used. In general, h and w need not be independent. However, we assume that they are independent; this assumption is valid for all the processes we consider below. We consider only the case when the time-change h is an increasing L´evy process. Then, X will also be a L´evy process. Since wt = µt + σ zt , for a Wiener process (zt )t≥0 , we can write Lt = µht + σ zh(t) . Both the variance gamma (VG) and the normal inverse Gaussian (NIG) processes have time changes h, whose conditional distributions are taken from a set of generalized inverse Gaussian (GIG) distributions. Write δt = δt. If ht ∼ GIG(δt , λ, γ ) has a GIG distribution, then the density ftGIG of ht is ftGIG (h; δt , λ, γ ) =

λ

1 γ 1 δt2 hλ−1 exp − + γ 2h δt 2Kλ (δt γ ) 2 h

The set of GIG distributions is not closed under convolution. 1

A time-change representation may yield a mean-variance representation.

(2.7)

34

Exotic Option Pricing and Advanced L´evy Models

The VG process. The gamma distribution is the limit of the GIG distribution as δ → 0 and λ → t/v. If ht ∼ (t, tv) is a gamma variate, its density ft conditional on h0 = 0 is ft

t x ν −1 exp − xv (x) = . t ν ν νt

(2.8)

If L is VG we have Lt ≡ LVG (t | σ, v, µ) = µht + σ zh(t)

(2.9)

where ht has the density given by equation (2.8). For the VG process, the compensator ω in equation (2.1) is

1 1 ω = − ln 1 − µν − σ 2 ν . (2.10) ν 2 The NIG process. The inverse Gaussian process is GIG with λ = − 12 . If ht ∼ IG (δt , γ ) is an inverse Gaussian process, then the density ftIG (x) of ht is

2 2 γ δ 3 1 δ t t x− ftIG (x) = √ x − 2 exp − . (2.11) 2 x γ 2π An NIG process L can be written Lt ≡ LNIG (t | µ, θ, δ, γ ) = µt + βht + zh(t)

(2.12)

where ht has the density given by equation (2.11). For the NIG case, the compensator ω is

(2.13) ω = µ + δ γ − α 2 − (β + 1)2 where α 2 = γ 2 + β 2 . 2.3.1 The subordinator approach to simulation Suppose we have a European derivative, with payoff HT at time T and value ct at times t ≤ T , so that cT = HT . In the martingale valuation framework, we have cT ] , ct = Et [

(2.14)

where cT = cT ppTt for a numeraire pt , with expectations Et [•] ≡ E [• | Ft ] taken with respect to the martingale measure associated with pt . Suppose that HT and pt depend on a single-state variable St , which in turn depends upon a L´evy process Lt . In particular, suppose that, as in equation (2.1), under the pricing measure St = S0 exp (rt + Lt − ωt) where r is a constant short rate.

(2.15)

Simulation Methods with L´evy Processes

35

Now, suppose that L has a subordinator representation Lt = wh(t) . Since wt and h (t) are independent, the ﬁltration Ft decomposes as Ft = Ftw × Fth and we can iterate the expectation: cT | Ft ] ct = E [ cT | Ftw × Fth =E =E E cT | Ftw × FTh | Ftw × Fth .

(2.16)

ct = Et [E [ cT | h]]

(2.19)

(2.17) (2.18)

Informally, we can write

where E [ cT | h] represents the expected value of cT , conditional upon knowing the path of h up to time T . The inner and outer expectations can be simulated separately and so a possible procedure to value ct by simulating Lt is to: 1. 2. 3. 4.

Simulate a path {hti }i=1,... ,N of h up to time T . Given {hti }, generate a path for w at times {hti } and set Lti = wh(ti ) . Compute cT from the path of L. Repeat 10 000 times (say) and average.

This procedure can be used to value not just vanilla European options but also options that payoff at a directly determined stopping time, such as barrier options.2 2.3.2 Applying the subordinator approach There are two expectations from equation (2.19). Each could, in principle, be computed by using either a lattice valuation method or a Monte Carlo valuation method.3 Each possible pairing of valuation methods can be assessed for its appropriateness or inappropriateness for valuing path-dependent options (P or NP), for valuing options such as American or Bermudan type which may be exercised early (A or NA), and for its ease of calibration (C or NC). We obtain Table 2.1. Using a Monte Carlo method for the inner expectation and a lattice method for the outer expectation results in the random lattice method (Kuan and Webber (2003) [18]). Table 2.1

Valuation methods

Method

Inner

2 3

MC Lattice

Outer MC

Lattice

P, NA, NC NP, ∼A, C

NP, A, NC NP, A, C

However, not options whose stopping times are determined by optimality conditions, such as American options. It is not immediately clear how a PDE method might be used.

36

Exotic Option Pricing and Advanced L´evy Models

Rydberg (1997) [24] effectively used a Monte Carlo method for both the inner and outer expectation to simulate the NIG process. The procedure was later developed by Ribeiro and Webber (2002, 2003a, b) [21]–[23] who showed how to apply effective speed-up methods and bias-reduction methods. We expound their approach in the remainder of this article.

2.4

´ CONSTRUCTING BRIDGES FOR LEVY PROCESSES

Plain Monte Carlo methods are very slow to achieve acceptable accuracy.4 Various speedup methods need to be employed to produce reasonable computation times. One efﬁcient method is path regularization by stratiﬁed sampling. This ensures that sample points form a less clustered draw from the sample space than a plain Monte Carlo method would produce. If an entire sample path is needed, rather than just a draw from the terminal time, then stratiﬁed sampling has to be used in conjunction with a bridge method. 2.4.1 Stratiﬁed sampling and bridge methods We discuss here the background to stratiﬁed sampling and bridge methods. Good reviews can be found in J¨ackel (2002) [16] and Glasserman (2003) [14]. Stratiﬁed sampling. It is easy to generate a stratiﬁed sample from the unit interval [0, 1]. Let vi ∼ U [0, 1], i = 1, . . . , Q, be a sample of size Q from the uniform distribution U [0, 1], and then ui = i+vQi −1 is a stratiﬁed sample of size Q from U [0, 1]. The set {ui } is guaranteed to have minimal clustering above the scale 1/Q. Stratiﬁed samples can be produced from other distributions by inverse transform. Suppose that X ∼ FX is a random variate with distribution function FX and that FX−1 is computable. Let u ∼ U [0, 1] be a uniform variate, and then FX−1 (u) ∼ FX has distribution FX . Given a stratiﬁed sample ui , i = 1, . . . , Q from U [0, 1], the set FX−1 (ui ), i = 1, . . . , Q, is a stratiﬁed sample from FX . Bridge sampling. Given a L´evy process L, where Lt has distribution Ft at time t (conditional on L0 ), suppose that we have found a sample, Li,N , i = 1, . . . , Q, of LtN from FtN , possibly stratiﬁed. Given a value for L0 at time 0 = t0 , we would like to construct an entire sample path L0 = Li,0 , . . . , Li,N with the correct conditional distributions. This means being able to sample Li,j , at time 0 < tj < tN , conditional upon the values of Li,0 and Li,N . In general, suppose that X ∼ FX , Y ∼ FY and Z ∼ FZ , with densities fX , fY and fZ , respectively, are random variates such that Z = Y + X. For instance, X, Y and Z could be increments in a L´evy process L. Given a draw z of Z, we want to sample from the conditional distribution X | Z. Write fX,Y (x, y) for the joint density of X and Y . Then fX|Z (x) = = 4

fX,Y (x, z − x) , fZ (z)

(2.20)

fX (x) fY (z − x) when X, Y are independent. fZ (z)

(2.21)

Measured by the square root of the second moment of the Monte Carlo estimate. An internally generated estimate of this is the Monte Carlo standard error.

Simulation Methods with L´evy Processes

37

For a L´evy process L, take X ∼ Fti , Y ∼ Ftj , Z ∼ Fti +tj , say, and then fti |ti +tj (x) ≡ fX|Z (x) is the bridge density of L. If the densities ft are known, then the bridge density of L can be computed and perhaps a sampling method can be found. 2.4.2 Bridge sampling and the subordinator representation When Lt = wh(t) has a subordinator representation we could construct a stratiﬁed bridge sample of L by ﬁrst obtaining a stratiﬁed bridge sample of h, and then constructing a stratiﬁed bridge sample of w at the times given by our path for h. The bridge distribution of a Brownian motion is well known and it is easy to sample from it. We need only to know the bridge distribution of the subordinator h, and to be able to sample from it. Ribeiro and Webber (2002, 2003b) [21], [23] ﬁnd the bridge density and a sampling method for the bridge density, for both the gamma and the inverse Gaussian processes. We summarize their results here. Let 0 = h0 < · · · < hN be a series of values at increasing times for the subordinator process h. Given hi and hk at times ti < tk , we want to sample hj at an intermediate time ti < tj < tk . Write z = hk − hi , x = hj − hi , and y = hk − hj , and set τz = tk − ti , τx = tj − ti , and τy = tk − tj . The gamma process. Given hi and hk , the bridge density of x from a gamma process with parameter ν is τ

τy τy τx −1 x ν x ν −1 1 νx + ν 1− fX|Z (x) = . z τνx τνy z z

(2.22)

An algorithm to sample from this density is 1. Generate bj ∼ B(τx /ν, τy /ν), where B is the beta distribution. 2. Set hj = hi + bj (hk − hi ). Existing algorithms to sample (by inverse transform) from the beta distribution are relatively slow. Even so, the numerical results in Section 2.5 below demonstrate the very great speed-ups possible with this sampling method. The inverse Gaussian process. The bridge density of x from an inverse Gaussian process with parameter δ is δ τx τy fX|Z (x) = √ 2π τz

xy z

− 3 2

1 exp − δ 2 2

where y = z − x. An algorithm to sample from this density is 1. Generate q ∼ χ12 .

τy2 τ2 τx2 + − z x y z

,

(2.23)

38

Exotic Option Pricing and Advanced L´evy Models

2. Set λ =

δ 2 τx2 x

and µ = τy /τx , and compute roots s1 and s2 , s1 = µ + s2 =

µ µ2 q − 4µλq + µ2 q 2 , 2λ 2λ

µ2 . s1

(2.24) (2.25)

3. Set s = s1 with probability p, p= or else set s = s2 . 4. Finally, set hj = hi +

1 1+s

µ (1 + s1 ) , (1 + µ) (µ + s1 )

(2.26)

(hk − hi ).

Since sampling from a χ12 distribution, by inverse transform or otherwise, is very fast, sampling from the inverse Gaussian distribution is also very quick. Using the bridge. Given a method for sampling from the subordinator process ht , we adopt the following algorithm. 1. 2. 3. 4.

Given h0 = 0, construct a stratiﬁed sample hi,N of hN . Using the bridge distribution, construct hi,N/2 conditional on h0 and hi,N . Using binary chop, continue to generate a path hi,j for times tj , j = 0, . . . , N . Generate whi,j , and hence Si,j conditional on hi,j .

At each intermediate time it is possible to continue to stratify the sample path. Generating a sample path requires a sequence of uniform variates from which samples with the desired distributions are obtained (by inverse transform). If n uniform variates are needed, this is equivalent to a draw from a unit hypercube of dimension n. To stratify m ≤ n of these draws, one performs a stratiﬁed sample on an m-dimensional hypercube, selecting the remaining n − m draws without stratiﬁcation. In practice, it is only possible to make a fully stratiﬁed draw from a hypercube of dimension 3 or so. To make a stratiﬁed draw from a hypercube of dimension m > 3, low discrepancy sampling is often used. For the VG process it takes one uniform variate to generate (by inverse transform) a draw from the gamma bridge distribution for hi and one more for each normal variate whi ; for the NIG process, it takes two uniform variates to draw for each hi and one more for whi . So, the VG process requires two uniform variates at each time step, and the NIG process requires three. Even if sampling with low discrepancy sequences it may only be possible in practice to sample reliably from a unit hypercube of dimension at most a few dozen. A freely available downloadable code for generating Sobol sequence numbers5 goes up to dimension 39. This means that (with binary chop) for the VG process it is possible to stratify at up to 16 times and for the NIG process one can stratify at up to 8 times. Draws for other times have to be made with ordinary non-stratiﬁed sampling. Even with this restriction, very good speed-ups are possible. 5

See Bratley and Bennett (1988) [5]. Code is downloadable from www.netlib.org/toms/659.

Simulation Methods with L´evy Processes

2.5

39

VALUING DISCRETELY RESET PATH-DEPENDENT OPTIONS

A discretely reset path-dependent option is one in which the option payoff is computed from observations of the underlying asset value at certain discrete times (the reset times). In this section, we value discretely reset path-dependent options when returns to the underlying asset are either VG or NIG processes. We present convergence results (taken from Ribeiro and Webber (2002, 2003b) [21] [23]) and show how true standard errors are improved relative to plain Monte Carlo. When a Monte Carlo method is used with stratiﬁed sampling, successive Monte Carlo estimates are correlated with one another. This means that the internally generated standard error measure is not a true reﬂection of the standard deviation of the Monte Carlo estimate. In the tables below, the true standard deviation is estimated by taking the sample standard deviation of the Monte Carlo estimates obtained from 100 replications of the Monte Carlo procedure. To compare two Monte Carlo methods, we use the efﬁciency gain of one method over the other. Suppose we have two Monte Carlo procedures. Monte Carlo method i, i = 1, 2, gives an estimate with standard deviation σi ina time ti . If the time taken t is proportional to the number of sample paths Q, and if σ is O − 12 in Q, then the efﬁciency gain E1,2 , deﬁned as E1,2 =

t2 σ22 , t1 σ12

(2.27)

is how many times faster method 1 is to achieve the same standard deviation as method 2. In the following tables, K is the number of stratiﬁcation times. K = 0 is plain Monte Carlo with Q = 106 paths. Bridge Monte Carlo uses Q = 104 paths. The benchmark is full low discrepancy with Q = 106 . The initial asset value is S0 = 100, the short rate is r = 0.1 and the option strike is X = 100 with maturity time T = 1. The VG case uses parameter values σ = 0.12136, v = 0.3 and µ = −0.1436 (based upon Madan et al. (1998) [19]). The NIG case uses parameters α = 75.49, β = −4.089, δ = 3 and µ = 0 (based upon Rydberg (1997) [24]). Programmes were written in VBA 6.0 run on a 900 Mhz PC. Tables 2.2 and 2.3 give values, standard deviations and computation times (in seconds) for a discretely reset average rate option when the underlying asset has either VG or NIG returns process. More illuminating are Tables 2.4 and 2.5, which give the efﬁciency gains in each case. In the VG case, speed-ups of up to a factor of about 400 are possible and up to about 200 for the NIG case. Speed-ups tend to improve as both the number of reset dates and the number of stratiﬁcation times increase, although this is not so evident in this example for the NIG case. Here, we have given only a few of the results of Ribeiro and Webber. They investigate many more cases, including discrete barrier and lookback options, demonstrating in each case that worthwhile speed-ups are attainable. Further speed-ups would seem to be possible. In both the VG and NIG cases, the number of stratiﬁcation dates was constrained by the dimension of the low discrepancy sequence generated by the available software. There have also been criticisms of the quality of this generator (J¨ackel (2002) [16]). With a better quality generator, capable of producing low discrepancy sequences of higher dimension, increased speed-ups would be possible.

40

Exotic Option Pricing and Advanced L´evy Models

Table 2.2 Values, standard deviations and computation times for average rate call options: comparison of plain and bridge Monte Carlo methods for the VG case (Ribeiro and Webber (2003b) [23]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

0

6.7720 (0.0064) [85.6]

6.0666 (0.0058) [175.0]

5.7274 (0.0055) [335.4]

5.5497 (0.0053) [647.5]

5.4625 (0.0052) [1277]

5.4075 (0.0052) [5034]

1

6.7993 (0.029) [2.4]

6.0290 (0.025) [3.5]

5.7234 (0.023) [5.5]

5.5242 (0.025) [9.3]

5.4830 (0.029) [17.0]

5.3874 (0.024) [52.2]

2

6.7635 (0.011) [4.0]

6.0741 (0.012) [5.0]

5.7187 (0.014) [6.9]

5.5208 (0.014) [10.8]

5.4761 (0.013) [18.5]

5.376 (0.013) [54.9]

4

6.7594 (—) [13.0]

6.0656 (0.0065) [14.0]

5.7149 (0.0067) [16.0]

5.5510 (0.0072) [19.9]

5.4765 (0.0063) [27.6]

5.3934 (0.0064) [62.9]

8

—

6.0711 (—) [33.8]

5.7283 (0.0029) [35.8]

5.5465 (0.0033) [39.6]

5.4667 (0.0035) [47.4]

5.3997 (0.0035) [83.0]

16

—

—

5.5527 (0.0014) [80.7]

5.4627 (0.0017) [88.1]

5.4008 (0.0017) [123.1]

Benchmark

6.7626 (—) [1297]

6.0702 (—) [3370]

5.7245 (—) [77.3] 5.7250 (—) [7638]

—

—

—

Reproduced by permission of L.D. Donaldson.

In the VG case, the stratiﬁcation algorithm for the gamma bridge distribution is constrained by the method used to generate stratiﬁed beta variates by inverse transform.6 An improved algorithm for the inverse transform of the beta distribution would result in even greater speed-ups in this case.

2.6 VALUING CONTINUOUSLY RESET PATH-DEPENDENT OPTIONS We would like to obtain for continuously reset barrier and lookback options the speed-ups that were found for discretely reset barrier and lookback options by Ribeiro and Webber (2002, 2003b) [21][23]. These same authors (2003a) [22] ﬁnd an approximate method to achieve this result. Particular problems arise when applying numerical methods to continuous barrier options. It turns out that the values of discretely reset barrier options converge only very slowly to values of corresponding continuously reset barrier options as the number of reset dates increases. Since numerical methods are set in discrete time, the values they ﬁnd for continuously reset barrier options may converge only very slowly to the true value. 6

The algorithm uses a root searching method.

Simulation Methods with L´evy Processes

41

Table 2.3 Values, standard deviations and computation times for average rate call options: comparison of plain and bridge Monte Carlo methods for the NIG case (Ribeiro and Webber (2002) [21]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

0, Q = 106

8.5856 (0.0103) [71.6]

7.7892 (0.0094) [141.4]

7.4059 (0.0089) [278.6]

7.2312 (0.0087) [551.8]

7.1286 (0.0086) [1101.4]

7.0698 (0.0086) [4390.1]

1

8.6326 (0.044) [1.9]

7.8205 (0.042) [2.4]

7.3874 (0.048) [4.9]

7.2347 (0.041) [7.7]

7.0763 (0.047) [15.1]

7.0656 (0.044) [58.8]

2

8.5530 (0.021) [1.8]

7.7695 (0.022) [2.1]

7.4282 (0.022) [4.8]

7.2457 (0.026) [7.7]

7.0721 (0.026) [15.0]

7.0497 (0.021) [58.8]

4

8.5695 (—) [1.8]

7.7963 (0.010) [2.0]

7.4181 (0.011) [4.7]

7.2105 (0.011) [7.6]

7.1249 (0.011) [14.9]

7.0430 (0.012) [58.6]

8

—

7.7959 (—) [1.7]

7.4045 (0.0048) [4.3]

7.2121 (0.0059) [7.3]

7.1296 (0.0051) [14.6]

7.0519 (0.0059) [58.4]

Benchmark

8.5807 (—) [169]

7.8072 (—) [169]

—

—

—

—

Table 2.4 Efﬁciency gains for average rate call options: comparison of plain and bridge Monte Carlo methods for the VG case (Ribeiro and Webber (2003b) [23]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

1 2 4 8 16

1.7 7.2 — — —

2.7 8.2 10 — —

3.5 7.5 14 34 —

3.1 8.6 18 42 115

2.4 11 32 60 136

4.5 15 53 134 383

Reproduced by permission of L.D. Donaldson.

Table 2.5 Efﬁciency gains for average rate call options: comparison of plain and bridge Monte Carlo methods for the NIG case (Ribeiro and Webber (2002) [21]) K

4 resets

8 resets

16 resets

32 resets

64 resets

256 resets

1 2 4 8

2.1 9.8 — —

3.0 12.4 58.8 —

2.0 9.9 42.3 219.4

3.2 8.2 42.8 167.1

2.4 8.2 45.7 216.2

2.9 12.2 40.1 157.0

42

Exotic Option Pricing and Advanced L´evy Models

For a Monte Carlo method, this feature is called ‘simulation bias’. A discrete sample path for an underlying asset may not exceed a barrier level, but on the continuous path sampled by the discrete path, the barrier may have been hit in between times observed by the discrete sample path. For lookback options, the maximum (minimum) found along a discrete path will always be less than (more than) the true maximum (minimum) achieved along the continuous path. When the underlying asset has a geometric Brownian motion, simulation bias correction methods are available (Beaglehole et al. (1997) [3] and El Babsiri and Noel (1998) [1]). In this section, following Ribeiro and Webber, we show how these ideas can be extended to asset processes driven by L´evy processes. Only when bias has been removed, or at least signiﬁcantly reduced, does it makes sense to apply speed-up methods such as the bridge methods discussed in previous sections. 2.6.1 Options on extreme values and simulation bias Consider a continuously reset option maturing at time T , and let B be a barrier level. Given a (continuous) sample path {St }t∈[0,T ] for at asset value, set M0,T = max {St }

(2.28)

m0,T = min {St } ,

(2.29)

t∈[0,T ] t∈[0,T ]

An up-and-in barrier option with barrier level B has payoff HT (ω) I{M0,T ≥B } , where HT (ω) is the payoff at time T of the knocked-in option. Similarly, an up-and-out barrier option has payoff HT (ω) I{M0,T t. Denote by pt,T (Ut , L) the statistical density of this event; the corresponding risk neutral density is qt,T (Ut , L). These two densities interact in the determination, for example, of the value at risk in a contingent claim with cash ﬂow cT (L) at time T . The change in value over an interval of length h, of this cash ﬂow on a marked to market basis is the random variable ∞ ∞ −r(T −t+h) −r(T −t) cT (L)qt+h,T (Ut+h , L) dL − e cT (L)qt,T (Ut , L) dL e 0

0

= (Ut , Ut+h ), where it is supposed that interest rates are constant at the continuously compounded rate of r. The risk in this position is assessed by the statistical density pt ,t+h (Ut , Ut+h ) and the value at risk at the 0.95 conﬁdence interval deﬁned as the corresponding quantile of the statistical distribution. We note importantly that the contingent claim example is in fact quite general. It is now recognized explicitly that even bonds are claims contingent on the absence of counterparty default while equity is itself an option in the presence of outstanding bonds, or even otherwise when we take particular note of limited liability. An equally important entity is the ratio of the two densities yt,T (Ut , L) =

qt,T (Ut , L) pt,T (Ut , L)

which is called the change of measure density (or the Radon–Nikodym derivative of the measure q with respect to the measure p). By incorporating the measure change into the valuations above, one may perform all expectations with respect to the statistical measure and write the change in value as ∞ e−r(T −t+h) cT (L)yt+h,T (Ut+h , L)pt+h,T (Ut+h , L) dL −e−r(T −t)

0 ∞

cT (L)yt,T (Ut , L)pt,T (Ut , L) dL 0

= (Ut , Ut+h ) It is clear from these expressions that an understanding of the measure change yt,T (Ut , L) makes important contributions to risk management and investment decisions. The difﬁculty

Risks in Returns: A Pure Jump Perspective

53

however, lies in making observations on the measure change. This is because, although one may extract qt,T from option prices using the methods of Breeden and Litzenberger (1978), this occurs at values of T reﬂecting traded option maturities T − t at time t; these are typically at intervals of a month. In contrast, the statistical density is best estimated at the horizon of daily returns. This time discrepancy is difﬁcult to overcome. There is little one may do about accessing risk neutral densities at maturities below the ﬁrst liquid traded maturity. On the other hand, one may be tempted to construct monthly returns out of daily returns assuming independence and stationarity; however, the considerable evidence in support of correlated squared returns makes these assumptions problematic. For other recent approaches in this direction, the reader is referred to Jackwerth (2000) and Bliss and Panigirtzoglou (2002). The approach we take here is to attempt to observe from options data and time series data the limiting densities as T approaches t. Furthermore, to recover the classical ﬁnancial setting focusing on log returns, we ﬁrst change variables to these magnitudes by making the transformation L = Ut el and subsume the dependence on the current observed level Ut into the subscript t. $ qt,T (l) ≡ qt,T (Ut , Ut el )Ut el $t,T (l) ≡ pt,T (Ut , Ut el )Ut el p These limiting densities may be constructed on normalization by (T − t) as follows $ qt,T (l) T −t $t,T (l) p kP (l) = lim T →t T − t

kQ (l) = lim

T →t

We note importantly that the division by (T − t) is necessary as the numerator in each case goes to zero for l = 0 and goes to inﬁnity for l = 0 (as the limiting measures are Dirac measures at l = 0). Another key and different observation is the fact that, unfortunately, for continuous processes both limits remain zero for l = 0. For discontinuous processes in contrast, the situation is different; in this wide collection, we choose for tractability the class of purely discontinuous L´evy processes. For these pure jump processes, the above limits are well deﬁned for all l = 0 and converge to the L´evy measures deﬁned by kQ (l), and kP (l), respectively. The statistical L´evy measure, kP (l), has the heuristic interpretation of the expected number of jumps of size l in log returns per unit time. Analogously, kQ (l) is the futures price of a contract paying at unit time the dollar number of jumps of size l that occur in this period. Apart from these horizon matching considerations, the use of L´evy processes in modeling asset returns, both statistically and risk neutrally, has a number of other well noted advantages. First, from the statistical perspective, it is well known that kurtosis levels in short period returns are substantially above 3, arguing for non-Gaussian distributions. L´evy processes easily accommodate a much richer structure of moments for short horizon returns, including negative skewness when needed. Risk neutrally, these processes easily capture short maturity skews that are prominent in options data. The transition from the statistical to the risk neutral probability is also less constrained, as, in principle, all moments may be altered, unlike the diffusion case where local volatilities must remain the same. For further details on the applications of L´evy processes in ﬁnance, we refer the reader to Schoutens (2003).

54

Exotic Option Pricing and Advanced L´evy Models

From the perspective of studying the ratio of the risk neutral to the statistical L´evy measure, it is useful to work with processes that have simple analytical forms for this entity and are capable of both providing a good ﬁt to the data and of synthesizing the high activity levels observed in the markets. This leads us to ‘inﬁnite activity’ L´evy processes (see Geman et al. (2001)) that have a sufﬁciently rich parametric structure to capture at least the ﬁrst four moments of the local motions. A particularly attractive example is provided by the CGMY model (Carr et al. (2002)) with further properties described in the next section. Other candidates include the Normal Inverse Gaussian model of Barndorff-Nielsen (1998), the Meixner process studied by Schoutens and Teugels (1998), and the generalized hyperbolic model (Barndorff-Nielsen (1977), Eberlein and Prause (1998) and Prause (1999)). The next section presents the details of the CGMY model employed in this current study. This section is followed by estimation details presented in Section 3.3, for both the statistical analysis and the inference of the risk neutral process. In comparing the two probabilities at the instantaneous level, we consider explicitly here the structure of returns on securities paying the market gap risk. These are securities that pay a dollar whenever there is a large up or down move of a prespeciﬁed size. Section 3.4 presents the results for ﬁve world equity indexes (USA, UK, Germany, Spain and Japan) showing that world-wide tail gap risk securities for both positive and negative moves are insurance-based with expected negative rates of return reﬂecting the presence of insurance premia, while the central part of the return distribution represents investment where positive rates of return reﬂect the expected risk compensation. We anticipate that market participants taking long positions protect themselves by buying downside gap risk claims and pay the requisite insurance charge for this service. On the other hand, participants short the market protect themselves by buying upside gap risk claims and pay the insurance charge on this side. The relative strength of the long side to the short side is then reﬂected in the larger premia for downside gap risk claims, as compared to the comparable upside gap risk claims. We see in the structure of the change of measure density the ways in which investment risk and insurance protection complement each other in the ﬁnancial markets of the world.

3.2

CGMY MODEL DETAILS

The general idea is to model the statistical and risk neutral log price relative over an interval, X(t + h) − X(t) = ln(S(t + h)/S(t)), as the increment of a purely discontinuous L´evy process. Such processes have independent and identically distributed increments over non-overlapping intervals of equal length with inﬁnitely divisible densities. They are characterized by the L´evy-Khintchine decomposition for their characteristic exponents, ψ(u) by E exp(iuX(t)) = exp (−tψ(u)) ∞ 1 + iux1|x|≤a − eiux k(x) dx ψ(u) = iγ u +

(3.1)

−∞

where γ is called the drift coefﬁcient and k(x) is the L´evy density that integrates x 2 in a neighborhood of 0. The processes may, in general, have inﬁnite variation in that the limiting sum of absolute changes in the log price over smaller and smaller time intervals tends to inﬁnity. In the special case of a ﬁnite limit, we have ﬁnite variation and the characteristic

Risks in Returns: A Pure Jump Perspective

55

exponent then has the representation ψ(u) = iγ u +

∞

−∞

1 − eiux k(x) dx.

In the ﬁnite variation case, the process for X(t) may be written as the difference of two increasing processes X(t) = Xp (t) − Xn (t) where the increasing processes Xp (t), Xn (t) have characteristic exponents, ψp (u), ψn (u)

∞

1 − eiux k(x) dx

ψp (u) = iγp u +

0 ∞

ψn (u) = iγn u +

1 − eiux k(−x) dx

0

γ = γp − γn For the inﬁnite variation case, we add to the difference of two increasing compound Poisson processes Xpa , Xna with characteristic exponents ψpa (u)

= iγp u +

λap

ψna (u) = iγn u + λan λap =

∞

a

λan =

−a

−∞

∞ a ∞

a

1 − eiux fpa (x) dx

1 − eiux fna (x) dx

k(x) dx ; fpa (x) =

k(x) , x>a λap

k(x) dx ; fna (x) =

k(−x) , x>a λan

γ = γp − γn the limit as ε tends to zero, of the compensated jump compound Poisson martingale Xε (t) X (t) = ε

Xs 1ε 0, is given by: φV G (u; C, G, M) =

GM GM + (M − G)iu + u2

C .

This distribution is inﬁnitely divisible and one can deﬁne the VG-process X(V G) = {Xt(V G) , t ≥ 0} as the process which starts at zero, has independent and stationary incre(V G) ments and where the increment Xs+t − Xs(V G) over the time interval [s, s + t] follows a VG(Ct, G, M) law. In Madan et al. [19], it was shown that the VG-process may also be expressed as the difference of two independent Gamma processes, which is helpful for simulation issues (see Section 4.4.2).

Model Risk for Exotic and Moment Derivatives

CIR Stochastic Clock. solves the SDE:

73

Carr et al. [8] use as the rate of time change the CIR process that

1/2

dyt = κ(η − yt ) dt + λyt

dWt ,

where W = {Wt , t ≥ 0} is a standard Brownian motion. The characteristic function of Yt (given y0 ) is explicitly known (see Cox et al. [12]): ϕCI R (u, t; κ, η, λ, y0 ) = E[exp(iuYt )|y0 ] =

exp(κ 2 ηt/λ2 ) exp(2y0 iu/(κ + γ coth(γ t/2))) (cosh(γ t/2) + κ sinh(γ t/2)/γ )2κη/λ2

,

where γ = Gamma-OU Stochastic Clock.

κ 2 − 2λ2 iu.

The rate of time change is now a solution of the SDE: dyt = −λyt dt + dzλt ,

(4.5)

where the process z = {zt , t ≥ 0} is as in equation (4.4) a compound Poisson process. In the Gamma-OU case, the characteristic function of Yt (given y0 ) can be given explicitly. ϕ −OU (u; t, λ, a, b, y0 ) = E[exp(iuYt )|y0 ]

b λa −1 −λt b log − iut . = exp iuy0 λ (1 − e ) + iu − λb b − iuλ−1 (1 − e−λt ) Time-Changed L´evy Process. Let Y = {Yt , t ≥ 0} be the process we choose to model our business time (remember that Y is the integrated process of y). Let us denote by ϕ(u; t, y0 ) the characteristic function of Yt given y0 . The (risk-neutral) price process S = {St , t ≥ 0} is now modelled as follows: St = S0

exp((r − q)t) exp(XYt ), E[exp(XYt )|y0 ]

(4.6)

where X = {Xt , t ≥ 0} is a L´evy process. The factor exp((r − q)t)/E[exp(XYt )|y0 ] puts us immediately into the risk-neutral world by a mean-correcting argument. Basically, we model the stock price process as the ordinary exponential of a time-changed L´evy process. The process incorporates jumps (through the L´evy process Xt ) and stochastic volatility (through the time change Yt ). The characteristic function φ(u, t) for the log of our stock price is given by: φ(u, t) = E[exp(iu log(St ))|S0 , y0 ] = exp(iu((r − q)t + log S0 ))

ϕ(−iψX (u); t, y0 ) , ϕ(−iψX (−i); t, y0 )iu

(4.7)

74

Exotic Option Pricing and Advanced L´evy Models

where ψX (u) = log E[exp(iuX1 )]; ψX (u) is called the characteristic exponent of the L´evy process, Since we consider two L´evy processes (VG and NIG) and two stochastic clocks (CIR and Gamma-OU), we will ﬁnally end up with four resulting models abbreviated as VGCIR, VG-OU , NIG-CIR and NIG-OU . Because of (time)-scaling effects, one can set y0 = 1, and scale the present rate of time change to one. More precisely, we have that the characteristic function φ(u, t) of equation (4.7) satisﬁes: √ φNIG-CIR (u, t; α, β, δ, κ, η, λ, y0 ) = φNIG-CIR (u, t; α, β, δy0 , κ, η/y0 , λ/ y0 , 1), φNIG− OU (u, t; α, β, δ, λ, a, b, y0 ) = φNIG− OU (u, t; α, β, δy0 , λ, a, by0 , 1), √ φVG-CIR (u, t; C, G, M, κ, η, λ, y0 ) = φVG-CIR (u, t; Cy0 , G, M, κ, η/y0 , λ/ y0 , 1), φVG− OU (u, t; C, G, M, λ, a, b, y0 ) = φVG− OU (u, t; Cy0 , G, M, λ, a, by0 , 1).

Actually, this time-scaling effect lies at the heart of the idea of incorporating stochastic volatility through making time stochastic. Here, it comes down to the fact that instead of making the volatility parameter (of the Black–Scholes model) stochastic, we are making the parameter δ in the NIG case and the parameter C in the VG case stochastic (via the time). Note that this effect does not only inﬂuence the standard deviation (or volatility) of the processes; the skewness and the kurtosis are also now ﬂuctuating stochastically.

4.3 CALIBRATION Carr and Madan [7] developed pricing methods for the classical vanilla options which can be applied in general when the characteristic function of the risk-neutral stock price process is known. Let α be a positive constant such that the αth moment of the stock price exists. For all stock price models encountered here, typically a value of α = 0.75 will do ﬁne. Carr and Madan then showed that the price C(K, T ) of a European call option with strike K and time to maturity T is given by: exp(−α log(K)) +∞ exp(−iv log(K))(v) dv, (4.8) C(K, T ) = π 0 where (v) = =

exp(−rT )E[exp(i(v − (α + 1)i) log(ST ))] α 2 + α − v 2 + i(2α + 1)v exp(−rT )φ(v − (α + 1)i, T ) . α 2 + α − v 2 + i(2α + 1)v

(4.9) (4.10)

Using Fast Fourier Transforms, one can compute within a second the complete option surface on an ordinary computer. We apply the above calculation method in our calibration procedure and estimate the model parameters by minimizing the difference between market prices and model prices in a least-squares sense.

Model Risk for Exotic and Moment Derivatives

75

The data set consists of 144 plain vanilla call option prices with maturities ranging from less than one month up to 5.16 years. These prices are based on the implied volatility surface of the Eurostoxx 50 index, having a value of 2461.44 on October 7th, 2003. The volatilities can be found in Table 4.1. For the sake of simplicity and to focus on the essence of the stochastic behaviour of the asset, we set the risk-free interest rate equal to 3 percent and the dividend yield to zero. Table 4.1 Implied volatility surface data (Eurostoxx 50 index; October 7th, 2003) Strike

Maturity (year fraction) 0.0361

1081.82 1212.12 1272.73 1514.24 1555.15 1870.30 1900.00 2000.00 2100.00 2178.18 2200.00 2300.00 2400.00 2499.76 2500.00 2600.00 2800.00 2822.73 2870.83 2900.00 3000.00 3153.64 3200.00 3360.00 3400.00 3600.00 3626.79 3700.00 3800.00 4000.00 4070.00 4170.81 4714.83 4990.91 5000.00 5440.18

0.3175 0.3030 0.2990 0.2800 0.2650 0.2472 0.2471

0.2000

0.3105 0.3076 0.2976 0.2877 0.2800 0.2778 0.2678 0.2580 0.2493 0.2493 0.2405

1.1944

2.1916

4.2056

5.1639

0.3804 0.3667 0.3603 0.3348 0.3305 0.2973 0.2946 0.2858 0.2775 0.2709 0.2691 0.2608 0.2524 0.2446 0.2446 0.2381 0.2251 0.2240 0.2213 0.2198 0.2148 0.2113 0.2103 0.2069 0.2060

0.3451 0.3350 0.3303 0.3116 0.3084 0.2840 0.2817 0.2739 0.2672 0.2619 0.2604 0.2536 0.2468 0.2400 0.2400 0.2358 0.2273 0.2263 0.2242 0.2230 0.2195 0.2141 0.2125 0.2065 0.2050 0.1975 0.1972 0.1964 0.1953 0.1931

0.3150 0.3082 0.3050 0.2920 0.2899 0.2730 0.2714 0.2660 0.2615 0.2580 0.2570 0.2525 0.2480 0.2435 0.2435 0.2397 0.2322 0.2313 0.2295 0.2288 0.2263 0.2224 0.2212 0.2172 0.2162 0.2112 0.2105 0.2086 0.2059 0.2006 0.1988 0.1961 0.1910 0.1904 0.1903

0.3137 0.3073 0.3043 0.2921 0.2901 0.2742 0.2727 0.2676 0.2634 0.2600 0.2591 0.2548 0.2505 0.2463 0.2463 0.2426 0.2354 0.2346 0.2328 0.2321 0.2296 0.2258 0.2246 0.2206 0.2196 0.2148 0.2142 0.2124 0.2099 0.2050 0.2032 0.2008 0.1957 0.1949 0.1949 0.1938

76

Exotic Option Pricing and Advanced L´evy Models 1600

1400

1200

Option price

1000

800

600

400

200

0 1000

Figure 4.1

1500

2000

2500

3000 3500 Strike

4000

4500

5000

5500

Calibration of the NIG-CIR model (Eurostoxx 50 index; October 7th, 2003)

Contrary to the approach described in Hirsa et al. [15], we search for a single global set of parameters per model which we ﬁt (and which captures smile information) across the full range of maturities in the data set. This global parameter set can then be used to price path-dependent derivatives (e.g. payoffs at multiple points during its lifetime or moment derivatives; see Sections 4.6 and 4.7). This is in contrast with the parameter set resulting from a ﬁtting procedure at a single maturity date, which can in principle only be used to price option payoffs occurring at that speciﬁc maturity. The results of the global calibration are visualized in Figures 4.1 and 4.2 for the NIGCIR and the BN–S model, respectively. The other models give rise to completely similar ﬁgures. Here, the circles are the market prices and the plus signs are the analytical prices (calculated via equation (4.8) using the respective characteristic functions and obtained parameters). In Table 4.2, one ﬁnds the risk-neutral parameters for the different models. For comparative purposes, one computes several global measures of ﬁt. We consider the root mean square error (rmse), the average absolute error as a percentage of the mean price (ape), the average absolute error (aae) and the average relative percentage error (arpe): ' ( ( (Market price − Model price)2 rmse = ) number of options options

ape =

|Market price − Model price| 1 mean option price number of options options

Model Risk for Exotic and Moment Derivatives

1600

1400

1200

Option price

1000

800

600

400

200

0 1000

1500

2000

2500

3000 3500 Strike

4000

4500

5000

5500

Figure 4.2 Calibration of the BN–S model (Eurostoxx 50 index; October 7th, 2003)

Table 4.2 Risk-neutral parameters for the different models HEST σ02 = 0.0654, κ = 0.6067, η = 0.0707, θ = 0.2928, ρ = −0.7571 HESJ σ02 = 0.0576, κ = 0.4963, η = 0.0650, θ = 0.2286, ρ = −0.9900, µj = 0.1791, σj = 0.1346, λ = 0.1382 BN–S ρ = −4.6750, λ = 0.5474, b = 18.6075, a = 0.6069, σ02 = 0.0433 VG-CIR C = 18.0968, G = 20.0276, M = 26.3971, κ = 1.2145, η = 0.5501, λ = 1.7913, y0 = 1 VG-OU C = 6.1610, G = 9.6443, M = 16.0260, λ = 1.6790, a = 0.3484, b = 0.7664, y0 = 1 NIG-CIR α = 16.1975, β = −3.1804, δ = 1.0867, κ = 1.2101, η = 0.5507, λ = 1.7864, y0 = 1 NIG-OU α = 8.8914, β = −3.1634, δ = 0.6728, λ = 1.7478, a = 0.3442, b = 0.7628, y0 = 1

77

78

Exotic Option Pricing and Advanced L´evy Models

aae = arpe =

|Market price − Model price| number of options options |Market price − Model price| 1 number of options Market price options

In Table 4.3, an overview of these measures of ﬁt are given. Table 4.3 Global ﬁt error measures for the different models Model HEST HESJ BN–S VG-CIR VG-OU NIG-CIR NIG-OU

rmse

ape

aae

arpe

3.0281 2.8101 3.5156 2.3823 3.4351 2.3485 3.2737

0.0048 0.0045 0.0056 0.0038 0.0056 0.0038 0.0054

2.4264 2.2469 2.8194 1.9337 2.8238 1.9194 2.7385

0.0174 0.0126 0.0221 0.0106 0.0190 0.0099 0.0175

4.4 SIMULATION In this section, we describe in detail how the particular processes presented in Section 4.2, can be implemented in practice in a Monte Carlo simulation pricing framework. For this, we ﬁrst discuss the numerical implementation of the four building block processes which drive them. This will be followed by an explanation of how one assembles a time-changed L´evy process. 4.4.1 NIG L´evy process To simulate a NIG process, we ﬁrst describe how to simulate NIG(α, β, δ) random numbers. The latter can be obtained by mixing Inverse Gaussian (IG) random numbers and standard Normal numbers in the following manner. An IG(a, b) random variable X has a characteristic function given by: E[exp(iuX)] = exp(−a −2ui + b2 − b). First, simulate IG(1, δ α 2 − β 2 ) random numbers ik , for example, by using the Inverse Gaussian generator of Michael, Schucany and Haas (see Devroye [13]). Then sample a sequence of standard Normal random variables uk . NIG random numbers nk are then obtained via: nk = δ 2 β ik + δ ik uk . Finally, the sample paths of a NIG(α, β, δ) process X = {Xt , t ≥ 0} in the time points tn = nt, n = 0, 1, 2, . . . can be generated by using the independent NIG(α, β, δt) random numbers nk as follows: X0 = 0,

Xtk = Xtk−1 + nk ,

k ≥ 1.

Model Risk for Exotic and Moment Derivatives

79

4.4.2 VG L´evy process Since a VG process can be viewed as the difference of two independent Gamma processes, the simulation of a VG process becomes straightforward. A Gamma process with parameters a, b > 0 is a L´evy process with Gamma(a, b) distributed increments, i.e. following a Gamma distribution with mean a/b and variance a/b2 . A VG process X(V G) = {Xt(V G) , t ≥ (2) (1) = 0} with parameters C, G, M > 0 can be decomposed as Xt(V G) = G(1) t − Gt , where G (1) (2) (2) {Gt , t ≥ 0} is a Gamma process with parameters a = C and b = M and G = {Gt , t ≥ 0} is a Gamma process with parameters a = C and b = G. The generation of Gamma numbers is quite standard. Possible generators are Johnk’s gamma generator and Berman’s gamma generator [13]. 4.4.3 CIR stochastic clock The simulation of a CIR process y = {yt , t ≥ 0} is straightforward. Basically, we discretize the SDE: 1/2

dyt = κ(η − yt ) dt + λyt

dWt ,

y0 ≥ 0,

where Wt is a standard Brownian motion. Using a ﬁrst-order accurate explicit differencing scheme in time, the sample path of the CIR process y = {yt , t ≥ 0} in the time points t = nt, n = 0, 1, 2, . . . , is then given by: 1/2 √ ytn = ytn−1 + κ(η − ytn−1 )t + λytn−1 t vn , where {vn , n = 1, 2, . . . } is a series of independent standard Normally distributed random numbers. For other more involved simulation schemes, like the Milstein scheme, resulting in a higher-order discretization in time, we refer to J¨ackel [17]. 4.4.4 Gamma-OU stochastic clock Recall that for the particular choice of an OU-Gamma process, the subordinator z = {zt , t ≥ 0} in (equation (4.3)) is given by the compound Poisson process (equation (4.4)). To simulate a Gamma(a, b)-OU process y = {yt , t ≥ 0} in the time points tn = nt, n = 0, 1, 2, . . . , we ﬁrst simulate in the same time points a Poisson process N = {Nt , t ≥ 0} with intensity parameter aλ. Then (with the convention that an empty sum equals zero) ytn = (1 − λt)ytn−1 +

Ntn

xk exp(−λt u˜ k ),

k=Ntn−1 +1

where u˜ k is a series of independent uniformly distributed random numbers and xk can be obtained from your preferred uniform random number generator via xk = − log(uk )/b. 4.4.5 Path generation for time-changed L´evy process The explanation of the building block processes above allow us next to assemble all the parts of the time-changed L´evy process simulation puzzle. For this one can proceed through the following ﬁve steps [22]:

80

(i) (ii) (iii) (iv) (v)

Exotic Option Pricing and Advanced L´evy Models

simulate the rate of time change process y = {y tt , 0 ≤ t ≤ T }; calculate from (i) the time change Y = {Yt = 0 ys ds, 0 ≤ t ≤ T }; simulate the L´evy process X = {Xt , 0 ≤ t ≤ YT }; calculate the time changed L´evy process XYt , for 0 ≤ t ≤ T ; calculate the stock price process using equation (4.6). The mean correcting factor is calculated as: exp((r − q)t) exp((r − q)t) = . E[exp(XYt )|y0 ] ϕ(−iψX (−i); t, 1)

4.5 PRICING OF EXOTIC OPTIONS As evidenced by the quality of the calibration on a set of European call options in Section 4.3, we can hardly discriminate between the different processes on the basis of their smileconform pricing characteristics. We therefore put the models further to the test by applying them to a range of more exotic options. These range from digital barriers, one-touch barrier options, lookback options and ﬁnally cliquet options with local as well as global parameters. These ﬁrst-generation exotics with path-dependent payoffs were selected since they shed more light on the dynamics of the stock processes. At the same time, the pricings of the cliquet options are highly sensitive to the forward smile characteristics induced by the models. 4.5.1 Exotic options Let us consider contracts of duration T , and denote the maximum and minimum process, respectively, of a process Y = {Yt , 0 ≤ t ≤ T } as MtY = sup{Yu ; 0 ≤ u ≤ t} and mYt = inf{Yu ; 0 ≤ u ≤ t},

0 ≤ t ≤ T.

4.5.1.1 Digital barriers We ﬁrst consider digital barrier options. These options remain worthless unless the stock price hits some predeﬁned barrier level H > S0 , in which case they pay at expiry a ﬁxed amount D, normalized to 1 in the current settings. Using risk-neutral valuation, assuming no dividends and a constant interest rate r, the time t = 0 price is therefore given by: digital = e−rT EQ [1(MTS ≥ H )], where the expectation is taken under the risk-neutral measure Q. Observe that with the current deﬁnition of digital barriers their pricing reﬂects exactly the chance of hitting the barrier prior to expiry. The behaviour of the stock after the barrier has been hit does not inﬂuence the result, in contrast with the classic barrier options deﬁned below. 4.5.1.2 One-touch barrier options For one-touch barrier call options, we focus on the following four types: • The down-and-out barrier call is worthless unless its minimum remains above some ‘low barrier’ H , in which case it retains the structure of a European call with strike K. Its

Model Risk for Exotic and Moment Derivatives

81

initial price is given by: DOB = e−rT EQ [(ST − K)+ 1(mST > H )] • The down-and-in barrier is a normal European call with strike K, if its minimum went below some ‘low barrier’ H . If this barrier was never reached during the lifetime of the option, the option remains worthless. Its initial price is given by: DIB = e−rT EQ [(ST − K)+ 1(mST ≤ H )] • The up-and-in barrier is worthless unless its maximum crossed some ‘high barrier’ H , in which case it obtains the structure of a European call with strike K. Its price is given by: UIB = e−rT EQ [(ST − K)+ 1(MTS ≥ H )] • The up-and-out barrier is worthless unless its maximum remains below some ‘high barrier’ H , in which case it retains the structure of a European call with strike K. Its price is given by: UOB = e−rT EQ [(ST − K)+ 1(MTS < H )] 4.5.1.3 Lookback options The payoff of a lookback call option corresponds to the difference between the stock price level at expiry ST and the lowest level it has reached during its lifetime. The time t = 0 price of a lookback call option is therefore given by: LC = e−rT EQ [ST − mST ]. Clearly, of the three path-dependent options introduced so far, the lookback option depends the most on the precise path dynamics. 4.5.1.4 Cliquet options Finally, we also test the proposed models on the pricing of cliquet options. These still are very popular options in the equity derivatives world which allow the investor to participate (partially) in the performance of an underlying over a series of consecutive time periods [ti , ti+1 ] by ‘clicking in’ the sum of these local performances. The latter are measured relative to the stock level Sti attained at the start of each new subperiod, and each of the local performances is ﬂoored and/or capped to establish whatever desirable mix of positive and/or negative payoff combination. Generally, on the ﬁnal sum an additional global ﬂoor (cap) is applied to guarantee a minimum (maximum) overall payoff. This can all be summarized through the following payoff formula:

min capglob , max ﬂoorglob ,

N i=1

Sti − Sti−1 min caploc , max ﬂoorloc , Sti−1

82

Exotic Option Pricing and Advanced L´evy Models

Observe that the local ﬂoor and cap parameters effectively border the relevant ‘local’ price ranges by centering them around the future, and therefore unknown, spot levels Sti . The pricing will therefore depend in a non-trivial subtle manner on the forward volatility smile dynamics of the respective models, further complicated by the global parameters of the contract. For an in-depth account of the related volatility issues, we refer to Wilmott [24]. 4.5.2 Exotic option prices We price all exotic options through Monte Carlo simulation. We consistently average over 1 000 000 simulated paths. All options have a lifetime of three years. In order to check the accuracy of our simulation algorithm, we simulated option prices for all European calls available in the calibration set. All algorithms gave a very satisfactory result, with pricing differences with respect to their analytic calibration values of less than 0.5 percent. An important issue for the path-dependent lookback, barrier and digital barrier options above, is the frequency at which the stock price is observed for purposes of determining whether the barrier or its minimum level have been reached. In the numerical calculations below, we have assumed a discrete number of observations, namely at the close of each trading day. Moreover, we have assumed that a year consists of 250 trading days. In Figure 4.3, we present simulation results with models for the digital barrier call option as a function of the barrier level (ranging from 1.05S0 to 1.5S0 ). As mentioned before, aside from the discounting factor e−rT , the premiums can be interpreted as the chance of hitting the barrier during the option lifetime. In Figures 4.4–4.6, we show prices for all

0.8

0.7

Option price

0.6

0.5

0.4

0.3

1.05

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR 1.1

1.15

1.2

1.25 1.3 1.35 Digital as percentage of spot

1.4

1.45

Figure 4.3 Digital barrier prices (Eurostoxx 50 index; October 7th, 2003)

1.5

Model Risk for Exotic and Moment Derivatives

500

450

Option price

400

350

300

250

200 0.5

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR 0.55

Figure 4.4

0.6

0.65

0.7 0.75 0.8 0.85 Barrier as percentage of spot

0.9

0.95

DOB prices (Eurostoxx 50 index; October 7th, 2003)

350 NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR

300

Option price

250

200

150

100

50

0 0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

Barrier as percentage of spot

Figure 4.5

DIB prices (Eurostoxx 50 index; October 7th, 2003)

0.95

83

84

Exotic Option Pricing and Advanced L´evy Models

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR

180 160 140

Option price

120 100 80 60 40 20 0 1.05

1.1

1.15

1.2

1.25

1.3

1.35

1.4

1.45

1.5

Barrier as percentage of spot

Figure 4.6 UOB prices (Eurostoxx 50 index; October 7th, 2003) 520 500 480

Option price

460 440 420 400 380 360 340 320 1.05

NIG-OUGamma VG-CIR VG-OUGamma HEST HESJ BN–S NIG-CIR 1.1

1.15

1.2 1.25 1.3 1.35 Barrier as percentage of spot

1.4

1.45

Figure 4.7 UIB prices (Eurostoxx 50 index; October 7th, 2003)

1.5

Model Risk for Exotic and Moment Derivatives

85

Table 4.4 Exotic option prices H/S0

NIG-OUT

VG-CIR

VG-OUT

HEST

HESJ

BN–S

NIG-CIR

724.80 511.80 293.28 391.17 448.10 479.83 496.95 505.24 509.10 510.75 511.40 511.67

713.49 509.33 318.35 402.24 452.97 481.74 496.80 504.05 507.21 508.53 509.06 509.24

844.51 510.88 173.85 280.79 359.05 414.65 452.76 477.37 492.76 501.74 506.46 508.91

845.18 510.89 174.64 282.09 360.99 416.63 454.33 479.12 494.25 502.84 507.41 509.51

771.28 509.89 230.25 352.14 423.21 461.82 481.85 492.62 498.93 503.17 505.93 507.68

730.84 512.21 284.10 387.83 446.52 479.77 496.78 505.38 509.34 511.09 511.80 512.08

LC Call DOB DOB DOB DOB DOB DOB DOB DOB DOB DOB

0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

722.34 509.76 300.25 396.80 451.61 481.65 497.00 504.31 507.53 508.88 509.43 509.64

DIB DIB DIB DIB DIB DIB DIB DIB DIB DIB

0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

209.51 112.95 58.14 28.11 12.76 5.45 2.23 0.88 0.33 0.12

218.51 120.62 63.69 31.96 14.84 6.55 2.70 1.04 0.39 0.13

190.98 107.08 56.35 27.59 12.53 5.28 2.11 0.79 0.26 0.09

337.03 230.09 151.83 96.24 58.13 33.51 18.12 9.14 4.42 1.98

336.25 228.80 149.90 94.26 56.56 31.77 16.64 8.05 3.48 1.38

279.61 157.72 86.65 48.04 28.01 17.24 10.94 6.69 3.94 2.19

228.10 124.37 65.68 32.43 15.42 6.83 2.87 1.11 0.40 0.13

UIB UIB UIB UIB UIB UIB UIB UIB UIB UIB

1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

509.32 506.68 500.33 489.05 472.47 450.54 423.62 393.01 359.77 325.25

511.52 509.80 505.21 496.50 482.84 463.62 439.32 410.46 378.05 343.46

508.84 506.11 499.56 488.30 471.39 449.23 422.32 391.36 357.80 322.79

510.78 500.90 507.08 501.04 490.73 475.30 454.77 428.96 399.24 365.57

510.81 510.00 507.28 501.31 490.93 474.86 452.47 424.09 389.56 350.68

509.73 508.38 504.28 495.95 482.66 464.48 441.48 414.98 385.50 354.90

511.98 510.37 505.93 497.41 483.94 465.16 441.00 412.16 380.04 345.79

UOB UOB UOB UOB UOB UOB UOB UOB UOB UOB

1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

0.44 3.08 9.43 20.71 37.29 59.22 86.14 116.75 149.98 184.50

0.27 2.00 6.59 15.29 28.95 48.17 72.47 101.33 133.74 168.33

0.49 3.22 9.77 21.03 37.94 60.10 87.00 117.96 151.52 186.53

0.103 0.979 3.80 8.96 20.15 35.58 56.10 81.93 111.65 145.31

0.08 0.89 3.61 9.85 19.96 36.03 58.42 86.80 121.33 160.21

0.13 1.48 5.58 13.91 27.20 45.38 68.39 94.88 124.36 154.96

0.23 1.84 6.27 14.80 28.26 47.04 71.21 100.04 132.16 166.41

DIG DIG

1.05 1.1

0.7995 0.7201

0.8064 0.7334

0.7909 0.7120

0.8218 0.7478

0.8189 0.7421

0.8173 0.7360

0.8118 0.7380

(continued overleaf )

86

Exotic Option Pricing and Advanced L´evy Models Table 4.4 (continued ) H/S0

DIG DIG DIG DIG DIG DIG DIG DIG

NIG-OUT

VG-CIR

VG-OUT

HEST

HESJ

BN–S

NIG-CIR

0.6458 0.5744 0.5062 0.4418 0.3816 0.3264 0.2763 0.2321

0.6628 0.5940 0.5273 0.4630 0.4021 0.3456 0.2940 0.2474

0.6382 0.5678 0.5003 0.4363 0.3767 0.3217 0.2722 0.2280

0.6762 0.6069 0.5408 0.4769 0.4169 0.3603 0.3087 0.2610

0.6685 0.5971 0.5290 0.4637 0.4012 0.3426 0.2877 0.2374

0.6580 0.5836 0.5138 0.4493 0.3893 0.3355 0.2870 0.2446

0.6670 0.5977 0.5308 0.4668 0.4059 0.3490 0.2975 0.2510

1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5

Table 4.5 Lookback option prices for the different models HEST

HESJ

BN–S

VG-CIR

VG-OU

NIG-CIR

844.51

845.19

771.28

724.80

713.49

730.84

NIG-OU 722.34

one-touch barrier options (as a percentage of the spot). The strike K was always taken equal to the spot S0 . For reference, we summarize in Table 4.4 all option prices for the above discussed exotics. One can check that the barrier results agree well with the identity DIB + DOB = vanilla call = UIB + UOB, suggesting that the simulation results are well converged. Lookback prices are presented in Table 4.5. Consistently over all of the ﬁgures the Heston prices suggest that this model (for the current calibration) results in paths dynamics that are more volatile, breaching more frequently the imposed barriers. The results for the L´evy models with stochastic time change seem to move in pairs, with the choice of stochastic clock dominating over the details of the L´evy model upon which the stochastic time change is applied. The ﬁrst couple, VG- and NIG- display very similar results, overall showing the least volatile path dynamics, whereas the VG-CIR and NIG-CIR prices consistently fall midway of the pack. Finally, the OU- results without stochastic clock typically fall between the Heston and the VG-CIR and NIG-CIR prices. Besides these qualitative observations, it is important to note the magnitude of the observed differences. Lookback prices vary over about 15 percent and the one-touch barriers over 200 percent, whereas for the digital barriers we found price differences of over 10 percent. For the cliquet options, the prices are shown in Figures 4.8 and 4.9 for two different combinations. The numerical values can be found in Tables 4.6 and 4.7. These results are in-line with the previous observations. Variations of over 40 percent are noted.

4.6 PRICING OF MOMENT DERIVATIVES These derivatives depend on the realized higher moments of the underlying. More precisely, their payoff is a function of powers of the (daily) log-returns and allows to cover different kinds of market shocks. Variance swaps were already created to cover changes in the volatility regime. Besides the latter, skewness and kurtosis also play an important role. To protect against a wrongly estimated skewness or kurtosis, moment derivatives of higher order can

Model Risk for Exotic and Moment Derivatives

87

0.18

Option price

0.16

0.14

0.12

0.1

NIG-CIR NIG-OUGamma VG-OUGamma VG-CIR HESJ HEST BN–S

0.08

0.06

0

0.02

0.04

0.06

0.08 0.1 0.12 Global floor

0.14

0.16

0.18

0.2

Figure 4.8 Cliquet prices: caploc = 0.08; ﬂoloc = −0.08; capglo = +∞; N = 3; t1 = 1; t2 = 3 (Eurostoxx 50 index; October 7th, 2003) 0.16 0.15

Option price

0.14 0.13 0.12 0.11 0.1 NIG-CIR NIG-OUGamma VG-OUGamma VG-CIR HESJ HEST BN–S

0.09 0.08

−0.05

0

0.05 Global floor

0.1

0.15

Figure 4.9 Cliquet Prices: caploc = 0.05; ﬂoloc = −0.03; capglo = +∞; T = 3; N = 6; ti = i/2 (Eurostoxx 50 index; October 7th, 2003)

88

Exotic Option Pricing and Advanced L´evy Models

Table 4.6 Cliquet prices: caploc = 0.08; f lo−loc = 0.08; capglo = +∞; f loglo ∈ [0, 0.20]; N = 3; t1 = 1; t2 = 2; t3 = 3 f loglo

NIG-CIR

NIG-OUT

VG-OU

VG-CIR

HESJ

HEST

BN–S

0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.20

0.0785 0.0817 0.0850 0.0885 0.0922 0.0960 0.1000 0.1042 0.1086 0.1144 0.1203 0.1264 0.1327 0.1391 0.1456 0.1523 0.1591 0.1661 0.1732 0.1805 0.1880

0.0837 0.0866 0.0897 0.0930 0.0964 0.1000 0.1037 0.1076 0.1117 0.1174 0.1232 0.1292 0.1353 0.1415 0.1478 0.1543 0.1610 0.1677 0.1747 0.1817 0.1889

0.0835 0.0865 0.0896 0.0928 0.0963 0.0998 0.1036 0.1075 0.1116 0.1173 0.1231 0.1291 0.1352 0.1414 0.1478 0.1543 0.1610 0.1678 0.1747 0.1818 0.1890

0.0785 0.0817 0.0850 0.0885 0.0921 0.0960 0.1000 0.1042 0.1085 0.1144 0.1203 0.1264 0.1327 0.1391 0.1456 0.1523 0.1591 0.1661 0.1733 0.1806 0.1880

0.0667 0.0704 0.0743 0.0783 0.0825 0.0868 0.0913 0.0959 0.1008 0.1072 0.1137 0.1204 0.1272 0.1342 0.1412 0.1485 0.1558 0.1633 0.1709 0.1787 0.1866

0.0683 0.0719 0.0757 0.0796 0.0837 0.0879 0.0923 0.0969 0.1017 0.1080 0.1145 0.1211 0.1279 0.1348 0.1418 0.1489 0.1562 0.1637 0.1712 0.1789 0.1868

0.0696 0.0731 0.0767 0.0805 0.0845 0.0887 0.0930 0.0976 0.1024 0.1085 0.1149 0.1214 0.1280 0.1348 0.1418 0.1489 0.1561 0.1635 0.1711 0.1788 0.1867

Table 4.7 Cliquet prices: ﬂoloc = −0.03; caploc = 0.05; capglo = +∞; T = 3; N = 6; ti = i/2 f loglo

NIG-CIR

NIG-OU

VG-OU

VG-CIR

HESJ

HEST

BN–S

−0.05 −0.04 −0.03 −0.02 −0.01 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10 0.11 0.12 0.13 0.14 0.15

0.0990 0.0997 0.1005 0.1015 0.1028 0.1044 0.1060 0.1079 0.1099 0.1121 0.1145 0.1171 0.1204 0.1239 0.1277 0.1317 0.1361 0.1406 0.1456 0.1508 0.1567

0.1092 0.1098 0.1104 0.1112 0.1124 0.1137 0.1152 0.1168 0.1185 0.1205 0.1226 0.1250 0.1280 0.1312 0.1346 0.1382 0.1421 0.1463 0.1508 0.1556 0.1611

0.1131 0.1137 0.1144 0.1151 0.1162 0.1175 0.1189 0.1204 0.1221 0.1240 0.1260 0.1283 0.1311 0.1342 0.1375 0.1410 0.1448 0.1488 0.1531 0.1576 0.1630

0.1001 0.1008 0.1017 0.1026 0.1039 0.1054 0.1071 0.1089 0.1109 0.1131 0.1154 0.1180 0.1213 0.1248 0.1286 0.1326 0.1368 0.1414 0.1462 0.1514 0.1573

0.0724 0.0734 0.0745 0.0757 0.0776 0.0798 0.0821 0.0847 0.0874 0.0904 0.0937 0.0971 0.1016 0.1063 0.1113 0.1165 0.1220 0.1278 0.1339 0.1403 0.1474

0.0762 0.0771 0.0781 0.0762 0.0811 0.0831 0.0853 0.0877 0.0904 0.0932 0.0963 0.0996 0.1039 0.1084 0.1132 0.1183 0.1237 0.1293 0.1352 0.1415 0.1484

0.0788 0.0796 0.0805 0.0815 0.0831 0.0849 0.0869 0.0891 0.0915 0.0942 0.0972 0.1004 0.1045 0.1088 0.1135 0.1185 0.1238 0.1294 0.1353 0.1415 0.1484

Model Risk for Exotic and Moment Derivatives

89

be useful. Recent studies by Nualart and Schoutens [20] [21] and Corcuera et al. [10] [11] suggest that functionals of powers of returns seem the natural choice to complete the market. It was shown that allowing trade in the power-assets of all orders in an incomplete L´evy market leads to a complete market. Power assets are strongly related to the realized higher moments and they mainly coincide in a discrete time framework [11]. 4.6.1 Moment swaps Consider a ﬁnite set of discrete times {t0 = 0, t1 , . . . , tn = T } at which the path of the underlying is monitored. We denote the price of the underlying at these points, i.e. Sti , by Si for simplicity. Typically, the ti correspond to daily closing times and Si is the closing price at day i. Note that then: log(Si ) − log(Si−1 ),

i = 1, . . . , n,

correspond to the daily log-returns. Next, we deﬁne the moment swaps. The kth-moment swap is a contract where the parties agree to exchange at maturity:

n n

k Si (k) k log , MOMS = N × (log(Si ) − log(Si−1 )) = N × Si−1 i=1

i=1

where N is the nominal amount. A special case of these swaps is the second moment swap, better known as the Variance Swap. The non-centred payoff function in that case is given by: n

2 VS = N × (log(Si ) − log(Si−1 )) . i=1

Basically, this contract swaps ﬁxed (annualized) variance by the realized variance (second moment) and as such provides protection against unexpected or unfavourable changes in volatility. Higher moment swaps provide the same kind of protection. The MOMS(3) is related to realized skewness and provides protection against changes in the symmetry of the underlying distribution. MOMS(4) derivatives are linked to realized kurtosis and provide protection against the unexpected occurrences of very large jumps, or in other words, changes in the tail behaviour of the underlying distribution. 4.6.2 Moment options Related to the above discussed swaps, we deﬁne the associated options on the realized kth moment. More precisely, a moment option of order k, pays out at maturity T :

+ n (log(Si /Si−1 ))k − K . i=1

The price of these options under risk-neutral valuation is given by: % n

+ & (k) k . (log(Si /Si−1 )) − K MOMO (K, T ) = exp(−rT )EQ i=1

90

Exotic Option Pricing and Advanced L´evy Models

Note that since odd moments can be negative, the strike price for these options can range over the whole real line. 4.6.3 Hedging moment swaps We focus on hedging the moment swaps which are written on the future price as underlying. The price process of the future is given by F = {Ft = exp((r − q)(T − t))St }; we write Fi = Fti . In line with the results obtained by Carr and Lewis [6], ﬁrst consider the following (Taylor-like) expansion of the kth power of the logarithmic function: (log(x))2 (log(x))k = k! x − 1 − log(x) − 2!

3 (log(x)) (log(x))k−1 − − ··· − + O((x − 1)k+1 ) . 3! (k − 1)! Substituting x by Fi /Fi−1 leads to (log(Fi /Fi−1 ))k = k!

(log(Fi /Fi−1 ))j Fi − log(Fi /Fi−1 ) − Fi−1 j! k−1

j =2

+ O((Fi /Fi−1 )k+1 ) , where Fi = Fi − Fi−1 . Summing over i gives a decomposition of the MOMS(k) (on a future) payoff: MOMS(k) = N ×

n i=1

= N k!

n i=1

(log(Fi /Fi−1 ))k

Fi − log(Fi /Fi−1 ) − Fi−1

k−1 (log(Fi /Fi−1 ))j j =2

j!

+ O((Fi /Fi−1 )k+1 ) = −N k!(log(FT ) − log(F0 ))

n

n k−1 Fi k! (j ) k+1 MOMS + O +N k! −N (Fi /Fi−1 ) Fi−1 j! i=1

j =2

i=1

(4.11)

Thus, up to (k + 1)th-order terms the sum of the kth powered log-returns decomposes into the payouts from: • −k! log-contracts on the future with payoff log(FT ) − log(F0 ); i ) in futures; • a self-ﬁnancing dynamic strategy (k! ni=1 FF i−1 • a series of moment contracts of order strictly smaller than k.

Model Risk for Exotic and Moment Derivatives

91

The log-contract can be hedged by a dynamic trading strategy in combination with a static position in bonds, European vanilla call and put options maturing at time T. More precisely, ﬁrst note that for any L > 0: log(FT ) − log(F0 ) =

1 (FT − F0 ) − u(FT ) + u(F0 ), L

(4.12)

for u(x) =

x−L − log(x) + log(L) . L

Moreover Carr and Lewis [6] show that:

L

u(FT ) = u(ST ) =

1 (K − ST )+ dK + K2

0

n

Since FT − F0 = implies:

i=1 Fi ,

MOMS

(k)

+∞

L

1 (ST − K)+ dK. K2

(4.13)

substituting equation (4.13) into equations (4.12) and (4.11)

+∞ 1 1 + + ≈ N k! (K − ST ) dK + (ST − K) dK. 2 K2 0 K L

n k−1 1 k! 1 Fi − u(F0 ) − N +N k! MOMS(j ) . − Fi−1 L j! L

j =2

i=1

4.6.4 Pricing of moments swaps We calculate under the different models, the risk-neutral expectation: EQ MOMS(k) . We consistently average over 1 000 000 simulated paths. All options have a lifetime of 1 year. In Table 4.8, we clearly see how the price differences are even more pronounced as compared to the exotic option pricings discussed in Section 4.5.2.

Table 4.8 Moment swaps (N = 10 000) for the different models Order e−rT EQ MOMS(2) e−rT EQ MOMS(3) e−rT EQ MOMS(4)

HEST 623.89

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

804.60

557.55

628.85

557.75

641.71

−0.0807

−312.58

−21.03

−74.91

−21.69

−88.82

0.6366

322.40

7.8698

33.89

8.554

47.99

92

Exotic Option Pricing and Advanced L´evy Models 800 Heston BN–S VG-CIR VG-GAM NIG-CIR NIG-GAM

700

600

Price

500

400

300

200

100

0 100

200

300

400

500 600 K (bp)

700

800

900

1000

Figure 4.10 Moment option of order 2 (N = 10 000) 100 Heston BN–S VG-CIR VG-GAM NIG-CIR NIG-GAM

90 80 70

Price

60 50 40 30 20 10 0 −0.01 −0.008 −0.006 −0.004 −0.002

0 K

0.002

0.004

0.006

Figure 4.11 Moment option of order 3 (N = 10 000)

0.008

0.01

Model Risk for Exotic and Moment Derivatives

93

4.6.5 Pricing of moments options Next, we calculate the prices of moment call option, MOMO(k) , paying out at maturity T : n

+ (log(Si /Si−1 )) − K k

,

i=1

where the price of these call moment options is by the risk-neutral valuation: MOMO (K, T ) = exp(−rT )EQ (k)

% n

& (log(Si /Si−1 )) − K) k

+

.

i=1

We plot in Figures 4.10–4.11 the price for moment options of order 2 and 3; corresponding values for these options and fourth order moment option prices can be found in Tables 4.9–4.11.

The disparity between the models is ampliﬁed. The L´evy models with stochastic timechange seem again to move in the same pairs as in agreement with the results in Section 4.5.2, but now only up to the third-order moment option. The BN–S model has very pronounced second- and fourth-order moment option prices, while HEST drops (in absolute value) to very low values for the fourth-order moment option when compared to the other models.

4.7 CONCLUSIONS We have looked at different models, all reﬂecting non-normal returns and stochastic volatility. Empirical work has generally supported the need for both ingredients. We have demonstrated the clear ability of all proposed processes to produce a very convincing ﬁt to a market-conform volatility surface. At the same time, we have shown that this calibration could be achieved in a timely manner by using a very fast computational Table 4.9

Moment option data of order 2 (N = 10 000) for the different models

K (bp)

HEST

100 200 300 400 500 600 700 800 900 1000

302.3301 219.1410 156.7058 109.5242 75.9747 52.4440 37.0312 25.9978 17.4472 11.2276

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

491.4817 436.4667 394.0581 357.8177 326.6175 300.0791 277.1135 256.4670 238.1357 221.1757

212.0101 152.6484 110.6050 80.4430 58.2142 42.2200 30.0486 21.2445 14.8481 10.4386

183.3647 121.4477 83.4256 58.7503 42.1646 31.0541 23.5572 17.9608 14.1807 11.1815

249.068 186.381 140.753 106.257 80.235 60.917 46.407 36.175 28.402 24.19

161.5099 100.5233 67.6916 47.7753 35.2828 26.9061 20.2450 15.0211 11.2782 8.4718

94

Exotic Option Pricing and Advanced L´evy Models Table 4.10 Moment option data of order 3 (N = 10 000) for the different models

K −0.010 −0.009 −0.008 −0.007 −0.006 −0.005 −0.004 −0.003 −0.002 −0.001 0.000 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.010

HEST

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

98.0459 88.2341 78.4223 68.6105 58.7987 48.9869 39.1751 29.3686 19.5673 9.8393 0.7274 0.0008 0 0 0 0 0 0 0 0 0

79.9869 71.3740 62.8212 54.3490 46.0100 37.7781 29.6576 21.7306 14.0430 6.6657 0.0997 0 0 0 0 0 0 0 0 0 0

89.5862 79.9497 70.3448 60.8159 51.3447 41.9813 32.7639 23.7993 15.3273 7.5947 1.9022 0.8012 0.4866 0.3267 0.2293 0.1614 0.1052 0.0648 0.0416 0.0268 0.0170

84.6786 75.1442 65.6663 56.2683 46.9840 37.8554 28.9710 20.3739 12.3057 5.3112 0.8162 0.2915 0.1559 0.0938 0.0486 0.0322 0.0224 0.0126 0.0028 0 0

87.4962 78.1320 68.8121 59.5498 50.3831 41.3352 32.4824 23.8966 15.6840 8.1994 2.4520 1.1520 0.6438 0.4213 0.2819 0.1998 0.1325 0.0873 0.0578 0.0287 0.0091

82.2679 72.8796 63.5948 54.3998 45.3475 36.4821 27.8108 19.4166 11.4694 4.4442 0.1462 0.0173 0.0074 0 0 0 0 0 0 0 0

Table 4.11 Moment option data of order 4 (N = 10 000) for the different models K

HEST

BN–S

VG-CIR

VG-OU

NIG-CIR

NIG-OU

0.0001 0.0002 0.0003 0.0004 0.0005 0.0006 0.0007 0.0008 0.0009 0.0010

0.0781 0.0259 0.0120 0.0065 0.0033 0.0011 0.0001 0 0 0

35.7465 35.4823 35.2471 35.0274 34.8220 34.6307 34.4525 34.2810 34.1127 33.9479

1.9322 1.5977 1.3603 1.1821 1.0428 0.9281 0.8328 0.7506 0.6790 0.6202

2.3416 2.0015 1.7655 1.5879 1.4542 1.3386 1.2403 1.1552 1.0805 1.0160

5.2360 4.8754 4.6077 4.3995 4.2249 4.0755 3.9403 3.8221 3.7211 3.6303

3.6095 3.3309 3.1274 2.9601 2.8158 2.6878 2.5750 2.4746 2.3819 2.2979

procedure based on FFT. Note that an almost identical calibration means that at the timepoints of the maturities of the calibration data set the marginal distribution is ﬁtted accurately to the risk-neutral distribution implied by the market. If we have different models all leading to such almost perfect calibrations, all models have almost the same marginal distributions. It should, however, be clear that even if at all time-points 0 ≤ t ≤ T marginal distributions among different models coincide, this does not imply that exotic prices should also be

Model Risk for Exotic and Moment Derivatives

95

the same. This can be seen from the following discrete-time example. Let n ≥ 2 and X = {Xi , i = 1, . . . , n} be an iid sequence and let {ui , i = 1, . . . , n} be an independent sequence which randomly varies between ui = 0 and 1. We propose two discrete (be it unrealistic) stock price models, S (1) and S (2) , with the same marginal distributions: Si(1) = ui X1 + (1 − ui )X2 and Si(2) = Xi . The ﬁrst process ﬂips randomly between two states X1 and X2 , both of which follow the distribution of the iid sequence, and so do all of the marginals at the time points i = 1, . . . , n. The second process changes value in all time-points. The values are independent of each other and all follow again the same distribution of the iid sequence. In both cases, all of the marginal distributions (at every i = 1, . . . , n) are the same (as the distribution underlying the sequence X). It is clear, however, that the maximum and minimum of both processes behave completely different. For the ﬁrst process, the maximal maxj ≤i Si(1) = max(X1 , X2 ) and the minimal process minj ≤i Si(1) = min(X1 , X2 ) for i being large enough, whereas for the second process there is much more variation possible and it clearly leads to other distributions. In summary, it should be clear that equal marginal distributions of a process do not at all imply equal marginal distributions of the associated minimal or maximal process. This explains why matching European call prices do not lead necessarily to matching exotic prices. It is the underlying ﬁne-grain structure of the process that will have an important impact on the path-dependent option prices. We have illustrated this by pricing exotics by Monte Carlo simulation, showing that price differences for one-touch barriers of over 200 percent are no exception. For lookback call options, a price range of more than 15 percent among the models was observed. A similar conclusion was valid for the digital barrier premiums. Even for cliquet options, which only depend on the stock realizations over a limited amount of time-points, prices vary substantially among the models. Moment derivatives amplify pricing disparity. At the same time, the presented details of the Monte Carlo implementation should allow the reader to embark on his/her own pricing experiments. The conclusion is that great care should be taken when employing attractive ‘fancy-dancy’ models to price (or even more important, to evaluate hedge parameters for) exotics. As far as we know, no detailed study about the underlying path structure of assets has been carried out yet. Our study motivates such a deeper investigation.

ACKNOWLEDGEMENTS The ﬁrst author is a Postdoctoral Fellow of the Fund for Scientiﬁc Research, Flanders, Belgium (FWO – Vlaanderen). We thank Marc Jeannin for his devoted programming work.

REFERENCES [1] Bakshi, G., Cao, C. and Chen, Z. (1997), “Empirical performance of alternative option pricing models”, The Journal of Finance, LII(5), 2003–2049.

96

Exotic Option Pricing and Advanced L´evy Models

[2] Barndorff-Nielsen, O.E. and Shephard, N. (2001), “Non-Gaussian Ornstein–Uhlenbeck-based models and some of their uses in ﬁnancial economics”, Journal of the Royal Statistical Society, B, 63, 167–241. [3] Barndorff-Nielsen, O.E., Nicolata, E. and Shephard, N. (2002), “Some recent developments in stochastic volatility modelling”, Quantitative Finance, 2, 11–23. [4] Bertoin, J. (1996), L´evy Processes, Cambridge Tracts in Mathematics, 121, Cambridge University Press, Cambridge, UK. [5] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities”, Journal of Political Economy, 81, 637–654. [6] Carr, P. and Lewis, K. (2004), “Corridor variance swaps”, Risk Magazine, 17(2), 67–72. [7] Carr, P. and Madan, D. (1998), “Option valuation using the fast fourier transform”, Journal of Computational Finance, 2, 61–73. [8] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2001), Stochastic Volatility for L´evy Processes, Pr´epublications du Laboratoire de Probabilit´es et Mod`eles Al´eatoires, 645, Universit´es de Paris 6 and Paris 7, Paris, France. [9] Clark, P. (1973), “A subordinated stochastic process model with ﬁnite variance for speculative prices”, Econometrica, 41, 135–156. [10] Corcuera, J.M., Nualart, D. and Schoutens, W. (2005a), “Completion of a L´evy market by powerjump assets”, Finance and Stochastics, 9, 109–127. [11] Corcuera, J.M., Nualart, D. and Schoutens, W. (2005b), “Moment derivatives and L´evy-type market completion”, in A.E. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 169–193. [12] Cox, J., Ingersoll, J. and Ross, S. (1985), “A Theory of the term structure of interest rates”, Econometrica, 53, 385–408. [13] Devroye, L. (1986), Non-Uniform Random Variate Generation, Springer-Verlag, New York. [14] Heston, S. (1993), “A closed-form solution for options with stochastic volatility with applications to bond and currency options”, Review of Financial Studies, 6, 327–343. [15] Hirsa, A., Courtadon, G. and Madan, B.D. (2003), “The effect of model risk on the valuation of barrier options”, Journal of Risk Finance, Winter, 1–8. [16] Hull, J. and Suo, W. (2001), A Methodology for Assessing Model Risk and its Application to the Implied Volatility Function Model, Working Paper, University of Toronto, Toronto, ON, Canada, July. [17] J¨ackel, P. (2002), Monte Carlo Methods in Finance, Wiley, Chichester, UK. [18] Knudsen, Th. and Nguyen-Ngoc, L. (2003), Pricing European Options in a Stochastic VolatilityJump-Diffusion Model, DBQuant Working Paper, Deutsche Bank, London, UK; Journal of Financial and Quantitative Analysis, to be published. [19] Madan, D.B., Carr, P. and Chang, E.C. (1998), “The variance gamma process and option pricing”, European Finance Review, 2, 79–105. [20] Nualart, D. and Schoutens W. (2000), “Chaotic and predictable representations for L´evy processes”, Stochastic Processes and their Applications, 90, 109–122. [21] Nualart, D. and Schoutens W. (2001), “Backwards stochastic differential equations and feynman–Kac formula for L´evy processes with applications in ﬁnance”, Bernoulli, 7, 761–776.

Model Risk for Exotic and Moment Derivatives

97

[22] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [23] Schoutens, W., Simons E. and Tistaert, J. (2004), “A perfect calibration! Now what?”, Wilmott Magazine, March. [24] Wilmott, P. (2002), “Cliquet options and volatility models”, Wilmott Magazine, December.

5 Symmetries and Pricing of Exotic Options in L´evy Models Ernst Eberlein and Antonis Papapantoleon University of Freiburg, Freiburg, Germany Abstract Standard models fail to reproduce observed prices of vanilla options because implied ´ volatilities exhibit a term structure of smiles. We consider time-inhomogeneous Levy processes to overcome these limitations. Then the scope of this paper is two-fold. On the one hand, we apply measure changes in the spirit of Geman et al., to simplify the valuation problem for various options. On the other hand, we discuss a method for the valuation of ´ models. European options and survey valuation methods for exotic options in Levy

5.1

INTRODUCTION

The efforts to calibrate standard Gaussian models to the empirically observed volatility surfaces very often do not produce satisfactory results. This phenomenon is not restricted to data from equity markets, but it is observed in interest rate and foreign exchange markets as well. There are two basic aspects to which the classical models cannot respond appropriately: the underlying distribution is not ﬂexible enough to capture the implied volatilities either across different strikes or across different maturities. The ﬁrst phenomenon is the so-called volatility smile and the second one the term structure of smiles; together they lead to the volatility surface, a typical example of which can be seen in Figure 5.1. One way to improve the calibration results is to use stochastic volatility models; let us just mention Heston (1993) for a very popular model, among the various stochastic volatility approaches. A fundamentally different approach is to replace the driving process. L´evy processes offer a large variety of distributions that are capable of ﬁtting the return distributions in the real world and the volatility smiles in the risk-neutral world. Nevertheless, they cannot capture the term structure of smiles adequately. In order to take care of the change of the smile across maturities, one has to go a step further and consider time-inhomogeneous L´evy processes – also called additive processes – as the driving processes. For term structure models this approach was introduced in Eberlein et al. (2004) and further investigated in Eberlein and Kluge (2004), where cap and swaption volatilities were calibrated quite successfully. As far as plain vanilla options are concerned, a number of explicit pricing formulas is available for L´evy-driven models, one of which is also discussed in this article. The situation is much more difﬁcult in the case of exotic options. The aim of this paper is to derive symmetries and to survey valuation methods for exotic options in L´evy models. By symmetries, we mean a relationship between pricing formulae for options of different Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

100

Exotic Option Pricing and Advanced L´evy Models

14

Implied volatility (%)

13.5 13 12.5 12 11.5 11 10.5 10 10 20 30 40 50 60 70 80 Delta (%) or strike 90

1

2

3

4

5

6

8 7 Maturity

9

10

Figure 5.1 Implied volatilities of vanilla options on the Euro/Dollar rate: spot, 0.93; date, 5 November 2001. Data available at http://www.Mathfinance.de/FF/sampleinputdata.txt

type. Such a relation is of particular interest if it succeeds to derive the value of a complex payoff from that of a simpler one. A typical example is Theorem 5.1 (see below), where a ﬂoating strike Asian or lookback option can be priced via the formula for a ﬁxed strike Asian or lookback option. Moreover, some symmetries are derived in situations where a put-call parity is not available. The discussion here is rather general as far as the class of time-inhomogeneous L´evy processes is concerned. For implementation of these models, a very convenient class are the processes generated by the Generalized Hyperbolic distributions (cf. Eberlein and Prause (2002)). The paper is organized as follows: in the next section, we present time-inhomogeneous L´evy processes, the asset price model and some useful results. In Section 5.3, we describe a method for exploring symmetries in option pricing. The next section contains symmetries and valuation methods for vanilla options while exotic options are tackled in the following section. Finally, in Section 5.6 we present symmetries for options depending on two assets.

5.2

MODEL AND ASSUMPTIONS

Let (, F, F, IP) be a complete stochastic basis in the sense of Jacod and Shiryaev (2003, I.1.3). Let T ∈ R+ be a ﬁxed time horizon and assume that F = FT . We shall consider T ∈ [0, T ]. The class of uniformly integrable martingales is denoted by M; for further notation, we refer the reader to Jacod and Shiryaev (2003). Let D = {x ∈ Rd : |x| > 1}. Following Eberlein et al. (2004), we use as driving process L a timeinhomogeneous L´evy process, more precisely, L = (L1 , . . . , Ld ) is a process with

Exotic Options in L´evy Models

101

independent increments and absolutely continuous characteristics, in the sequel abbreviated PIIAC. The law of Lt is described by the characteristic function t* 1 iu, bs − u, cs u IE eiu,Lt = exp 2 0 + + (eiu,x − 1 − iu, x)λs (dx) ds, (5.2.1) Rd

where bt ∈ Rd , ct is a symmetric non-negative deﬁnite d × d matrix and λt is a L´evy measure on Rd , i.e. it satisﬁes λt ({0}) = 0 and Rd (1 ∧ |x|2 )λt (dx) < ∞ for all t ∈ [0, T ]. The Euclidean scalar product on Rd is denoted by ·, ·, the corresponding norm by | · | while · denotes a norm on the set of d × d matrices. The transpose of a matrix or vector v is denoted by v and 1 denotes the unit vector, i.e. 1 = (1, . . . , 1) . The process L has c`adl`ag paths and F = (Ft )t∈[0,T ] is the ﬁltration generated by L; moreover, L satisﬁes Assumptions (AC) and (EM) given below. Assumption (AC). Assume that the triplets (bt , ct , λt ) satisfy

T

*

|bt | + ct +

Rd

0

+ (1 ∧ |x|2 )λt (dx) dt < ∞.

Assumption (EM). Assume there exists a constant M > 1, such that the L´evy measures λt satisfy

T 0

expu, xλt (dx)dt < ∞,

∀u ∈ [−M, M]d .

D

Under these assumptions, L is a special semimartingale and its triplet of semimartingale characteristics (cf. Jacod and Shiryaev (2003, II.2.6)) is given by t t t Bt = bs ds, Ct = cs ds, ν([0, t] × A) = λs (dx)ds, (5.2.2) 0

0

0

A

where A ∈ B(Rd ). The triplet of semimartingale characteristics (B, C, ν) completely characterizes the distribution of L. Additionally, L is exponentially special (cf. Kallsen and Shiryaev (2002) pp. 2.12–2.13). We model the asset price process as an exponential PIIAC St = S0 exp Lt 1

d

(5.2.3)

with (S 1 , . . . , S d ) = (S01 eL , . . . , S0d eL ), where the superscript i refers to the i -th coordinate, i ≤ d. We assume that IP is a risk neutral measure, i.e. the asset prices have mean rate i of return µi r − δ i and the auxiliary processes Sti = eδ t Sti , once discounted at the rate r, i are IP-martingales. Here, r is the risk-free rate and δ is the dividend yield of the i -th asset. Notice that ﬁniteness of IE[ ST ] is ensured by Assumption (EM).

102

Exotic Option Pricing and Advanced L´evy Models

The driving process L has the canonical decomposition (cf. Jacod and Shiryaev (2003), II.2.38 and Eberlein et al. (2004)) Lt =

t

t

bs ds +

0

0

cs1/2 dWs +

t Rd

0

x(µL − ν)(ds, dx)

(5.2.4)

1/2

where, ct is a measurable version of the square root of ct , W a IP-standard Brownian motion on Rd , µL the random measure of jumps of the process L and ν(dt, dx) = λt (dx)dt is the IP-compensator of the jump measure µL . Because S is modelled under a risk neutral measure, the drift characteristic B is completely determined by the other two characteristics (C, ν) and the rate of return of the asset. Therefore, the i -th component of Bt has the form Bti =

t

(r − δ i )ds −

0

1 2

t

(cs 1)i ds −

0

t 0

i

Rd

(ex − 1 − x i )ν(ds, dx).

(5.2.5)

In a foreign exchange context, δ i can be viewed as the foreign interest rate. In general, markets modelled by exponential time-inhomogeneous L´evy processes are incomplete and there exists a large class of risk neutral (equivalent martingale) measures. An exception occurs in interest rate models driven by L´evy processes, where – in certain cases – there is a unique martingale measure; we refer to Theorem 6.4 in Eberlein et al. (2004). Eberlein and Jacod (1997) provide a characterization of the class of equivalent martingale measures for exponential L´evy models in the time-homogeneous case; this was later extended to general semimartingales in Gushchin and Mordecki (2002). In this article, we do not dive into the theory of choosing a martingale measure; we rather assume that the choice has already taken place. We refer to Eberlein and Keller (1995), Kallsen and Shiryaev (2002) for the Esscher transform, Frittelli (2000), Fujiwara and Miyahara (2003) for the minimal entropy martingale measure and Bellini and Frittelli (2002) for minimax martingale measures, to mention just a small part of the literature on this subject. A unifying exposition – in terms of f-divergences – of the different methods for selecting an equivalent martingale measure can be found in Goll and R¨uschendorf (2001). Alternatively, one can consider the choice of the martingale measure as the result of a calibration to the smile of the vanilla options market. Hakala and Wystup (2002) describe the calibration procedure in detail; we refer to Cont and Tankov (2004) for a numerically stable calibration method for L´evy driven models. Remark 2.1. In the above setting, we can easily incorporate dynamic interest rates and dividend yields (or foreign and domestic rates). Let Dt denote the domestic and Ft the foreign savings account, respectively; then, they can have the form Dt = exp

t

rs ds 0

and

t

Ft = exp

δs ds 0

and equation (5.2.5) has a similar form, taking rs and δs into account. Remark 2.2. The PIIAC L is an additive process, i.e. a process with independent increments, which is stochastically continuous and satisﬁes L0 = 0 a.s. (Sato (1999) Deﬁnition 1.6).

Exotic Options in L´evy Models

103

Remark 2.3. If the triplet (bt , ct , λt ) is not time-dependent, then the PIIAC L becomes a (homogeneous) L´evy process, i.e. a process with independent and stationary increments (PIIS). In that case, the distribution of L is described by the L´evy triplet (b, c, λ), where λ is the L´evy measure and the compensator of µL becomes a product measure of the form ν = λ ⊗ λ\1 , where λ\1 denotes the Lebesgue measure. In that case, equation (5.2.1) takes the form IE[exp(iu, Lt )] = exp[t · ψ(u)] where 1 ψ(u) = iu, b − u, cu + 2

Rd

(eiu,x − 1 − iu, x)λ(dx)

(5.2.6)

which is called the characteristic exponent of L. Lemma 2.4. For ﬁxed t ∈ [0, T ], the distribution of Lt is inﬁnitely divisible with L´evy triplet (b , c , λ ), given by

b :=

t

bs ds, 0

t

c :=

cs ds, 0

λ (dx) :=

t

λs (dx) ds.

(5.2.7)

0

(The integrals should be understood componentwise.) Proof. We refer to the proof of Lemma 1 in Eberlein and Kluge (2004). Remark 2.5. The PIIACs L1 , . . . , Ld are independent, if and only if, the matrices Ct are diagonal and the L´evy measures λt are supported by the union of the coordinate axes; this follows directly from Exercise 12.10 in Sato (1999) or I.5.2 in Bertoin (1996) and Lemma 2.4. Describing the dependence is a more difﬁcult task; we refer to M¨uller and Stoyan (2002) for a comprehensive exposition of various dependence concepts and their applications. We also refer to Kallsen and Tankov (2004), where a L´evy copula is used to describe the dependence of the components of multidimensional L´evy processes. Remark 2.6. Assumption (EM) is sufﬁcient for all our considerations, but is in general too strong. In the sequel, we will replace (EM), on occasion, by the minimal necessary assumptions. From a practical point of view though, it is not too restrictive to assume (EM), since all examples of L´evy models we are interested in, e.g. the Generalized Hyperbolic model (cf. Eberlein and Prause (2002)), the CGMY model (cf. Carr et al. (2002)) or the Meixner model (cf. Schoutens (2002)), possess moments of all order. We can relate the ﬁniteness of the g-moment of Lt for a PIIAC L and a submultiplicative function g, with an integrability property of its compensator measure ν. For the notions of the g-moment and submultiplicative function, we refer to Deﬁnitions 25.1 and 25.2 in Sato (1999). Lemma 2.7. (g-Moment). Let g be a submultiplicative, locally bounded, measurable function on Rd . Then the following statements are equivalent T

g(x)ν(dt, dx) < ∞ (2) IE g(LT ) < ∞.

(1)

0

D

104

Exotic Option Pricing and Advanced L´evy Models

Proof. The result follows from Theorem 25.3 in Sato (1999) combined with Lemma 6 in Eberlein and Kluge (2004). Now, since g(x) = expu, x is a submultiplicative function, we immediately get the following equivalence concerning Assumption (EM). Corollary 2.8. Let M > 1 be a constant. Then the following statements are equivalent T

expu, xν(dt, dx) < ∞, ∀u ∈ [−M, M]d D (2) IE expu, LT < ∞, ∀u ∈ [−M, M]d . (1)

0

We can describe the characteristic triplet of the dual of a one-dimensional PIIAC in terms of the characteristic triplet of the original process. First, we introduce some necessary notation and the next lemma provides the result. Notation. We denote by −λt the L´evy measure deﬁned by −λt ([a, b]) := λt ([−b, −a]) for a, b ∈ R, a < b, t ∈ R+ . Thus, −λt is a non-negative measure and the mirror image of the original measure with respect to the vertical axis. For a compensator of the form ν(dt, dx) = λt (dx)dt, we denote by −ν the (non-negative) measure, deﬁned as −ν(dt, dx) := −λt (dx)dt. Whenever we use the symbol “−” in front of a L´evy measure or a compensator, we will refer to measures deﬁned as above. Lemma 2.9 (dual characteristics). Let L be a PIIAC, as described above, with characteristic triplet (B, C, ν). Then L := −L is again a PIIAC with characteristic triplet (B , C , ν ), where B = −B, C = C and ν = −ν. Proof. From the L´evy-Khintchine representation we have that t* + iuLt cs 2 ϕLt (u) = IE e ibs u − u + (eiux − 1 − iux)λs (dx) ds. = exp 2 R 0 We get immediately ϕ−Lt (u) = ϕLt (−u) t* + cs 2 ibs (−u) − u + (ei(−u)x − 1 − i(−u)x)λs (dx) ds = exp 2 R 0 t* + cs 2 = exp i(−bs )u − u + (eiu(−x) − 1 − iu(−x))λs (dx) ds. 2 R 0 (AC). Hence, we can conThen bt = −bt , ct = ct , and λt = −λt clearly satisfy Assumption t t clude that L is also a PIIAC and has characteristics Bt = 0 bs ds = −Bt , Ct = 0 cs ds = Ct and ν (dt, dx) = λt (dx)dt = −ν(dt, dx).

Exotic Options in L´evy Models

5.3

105

GENERAL DESCRIPTION OF THE METHOD

In this section, we give a brief and general description of the method we shall use to explore symmetries in option pricing. The method is based on the choice of a suitable num´eraire and a subsequent change of the underlying probability measure; we refer to Geman et al. (1995) who pioneered this method. The discounted asset price process, corrected for dividends, serves as the num´eraire for a number of cases, in case the option payoff is homogeneous of degree one. Using the num´eraire, evaluated at the time of maturity, as the Radon–Nikodym derivative, we form a new measure. Under this new measure, the num´eraire asset is riskless while all other assets, including the savings account, are now risky. In case the payoff is homogeneous of higher degree, say α ≥ 1, we have to modify the asset price process so that it serves as the num´eraire. As a result, the asset dynamics under the new measure will depend on α as well. We consider three cases for the driving process L and the asset price process(es): (P1): L = L1 is a (1-d) PIIAC, L2 = k is constant and S 1 = S01 exp L1 , S 2 = exp L2 = K; (P2): L = L1 is a (1-d) PIIAC, S 1 = S01 exp L1 and S 2 = h(S 1 ) is a functional of S 1 ; (P3): L = (L1 , L2 ) is a 2-dimensional PIIAC and S i = S0i exp Li , i = 1, 2. Consider a payoff function f : R+ × R+ → R+

(5.3.1)

which is homogeneous of degree α ≥ 1, that is, for κ, x, y ∈ R∗+ f (κx, κy) = κ α f (x, y); for simplicity we assume that α = 1 and later – in the case of power options – we will treat the case of a more general α. According to the general arbitrage pricing theory (Delbaen and Schachermayer (1994, 1998)), the value V of an option on assets S 1 , S 2 with payoff f is equal to its discounted expected payoff under an equivalent martingale measure. Throughout this paper, we will assume that options start at time 0 and mature at T ; therefore we have V = e−rT IE f ST1 , ST2 .

(5.3.2)

We choose asset S 1 as the num´eraire and express the value of the option in terms of this num´eraire, which yields % & f ST1 , ST2 V −rT $= = e IE V S01 S01 % & 2

−rT S 1 e S 1 T = e−δ T IE f 1, T1 . ST e−δ 1 T S01

(5.3.3)

106

Exotic Option Pricing and Advanced L´evy Models

Deﬁne a new measure $ IP via the Radon–Nikodym derivative e−rT ST1 d$ IP = = ηT . dIP e−δ 1 T S01

(5.3.4)

After the change of measure, the valuation problem, under the measure $ IP, becomes + * $ = e−δ 1 T $ (5.3.5) IE f 1, ST1,2 V where we deﬁne the process S 1,2 :=

S2 . S1

loc IP ∼ IP The measures IP and $ IP are related via the density process ηt = IE[ηT |Ft ]; therefore $ and we can apply Girsanov’s theorem for semimartingales (cf. Jacod and Shiryaev (2003, IP. III.3.24)); this will allow us to determine the dynamics of S 1,2 under $ After some calculations, which depend on the particular choice of L2 or S 2 , we can transform the original valuation problem into a simpler one.

5.4 VANILLA OPTIONS These results are motivated by Carr (1994), where a symmetry relationship between European call and put options in the Black and Scholes (1973) and Merton (1973) model was derived. This result was later extended by Carr and Chesney (1996) to American options for the Black–Scholes case and for general diffusion models; see also McDonald and Schroder (1998) and Detemple (2001). This relationship has an intuitive interpretation in foreign exchange markets (cf. Wystup (2002)). Consider the Euro/Dollar market; then a call option on the Euro/Dollar exchange rate St with payoff (ST − K)+ has time-t value Vc (St , K; rd , re ) in dollars and Vc (St , K; rd , re )/St in euros. This euro-call option can also be viewed as a dollar-put option on the Dollar/Euro rate with payoff K(K −1 − ST−1 )+ and time-t value KVp (K −1 , ST−1 ; re , rd ) in euros. Since the processes S and S −1 have the same (Black–Scholes) volatility, by the absence of arbitrage opportunities, their prices must be equal. 5.4.1 Symmetry For vanilla options, the setting is that of (P1): L1 = L is the driving R-valued PIIAC with triplet (B, C, ν), S 1 = exp L1 = S and L2 = k, such that S 2 = ek = K, the strike price of the option. In accordance with the standard notation, we will use σs2 instead of cs , which corresponds to the volatility in the Black–Scholes model. Therefore, the characteristic C in equation t (5.2.2) has the form Ct = 0 σs2 ds. We will prove a more general version of Carr’s symmetry, namely a symmetry relating power options; the payoff of the power call and put option, respectively, is α α and (K − ST )+ (ST − K)+ where α ∈ N (more generally α ∈ R). We introduce the following notation for the value of a power call option with strike K and power index α α Vc (S0 , K, α; r, δ, C, ν) = e−rT IE (ST − K)+

Exotic Options in L´evy Models

107

where the asset is modelled as an exponential PIIAC according to equations (5.2.3)–(5.2.5) and x + = max{x, 0}. Similarly, for a power put option we set α Vp (S0 , K, α; r, δ, C, ν) = e−rT IE (K − ST )+ . Of course, for α = 1 we recover the European plain vanilla option and the power index α will be omitted from the notation. Assumption (EM) can be replaced by the following weaker assumption, which is the minimal condition necessary for the symmetry results to hold. Let D+ = D ∩ R+ and D− = D ∩ R− . Assumption (M). The L´evy measures λt of the distribution of Lt satisfy 0

T

|x|λt (dx)dt < ∞

T

xeαx λt (dx)dt < ∞.

and

D−

0

D+

Theorem 4.1. Assume that (M) is in force and the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the power call and put option via the following symmetry: ∗

Vc (S0 , K, α; r, δ, C, ν) = K α S0α CT eαCT Vp (S0−1 , K, α; δ, r, C, −f ν)

(5.4.1)

where the constants C and C∗ are given by equations (5.4.3) and (5.4.10), respectively (see ∗ below), K = K −1 e−CT and f (x) = eαx . Proof. First, we note that [e(δ−r)t St ]α = S0α exp(α(δ − r)t + αLt ) is not a IP-martingale; we denote by Lα the martingale part of the exponent; hence Lαt =

t

ασs dWs +

0

t 0

R

αx(µL − ν)(ds, dx).

Since Lα is exponentially special, with Theorem 2.18 in Kallsen and Shiryaev (2002) we have that its exponential compensator, denoted CLα , has the form CLαt =

1 2

0

t

α 2 σs2 ds +

t 0

R

(eαx − 1 − αx)ν(ds, dx)

and exp(Lα − CLα ) ∈ M. The price of the power call option expressed in units of the num´eraire yields −rT + α $c := Vc = e V α α IE (ST − K) S0 S0 −rT α α * +α e ST K −1 + −1 (K = e−δT IE − S ) T e−δT S0α

108

Exotic Option Pricing and Advanced L´evy Models

= e−δT K α IE exp (δ − r)T + α * α

× exp LαT − CLT =e

−δT

K CT IE exp α

T

0

(K −1 − ST−1 )+

LαT

−

CLαT

*

bs ds + CLαT

+α

(K

−1

−

ST−1 )+

+α (5.4.2)

where, by using equations (5.2.5) and (5.2.2), we have that log CT = (δ − r)T + αBT + CLαT = (α − 1)(r − δ)T + + 0

T

R

α(α − 1) 2

T 0

σs2 ds

(eαx − αex + α − 1)ν(ds, dx).

(5.4.3)

Deﬁne a new measure $ IP via its Radon–Nikodym derivative d$ IP = exp LαT − CLαT = ηT dIP and the valuation problem (equation (5.4.2)) becomes α $c = e−δT K α CT I$ $−$ V E (K ST )+

(5.4.4)

(5.4.5)

$ = K −1 and $ St := St−1 . where K Since the measures IP and $ IP are related via the density process (ηt ), which is a positive loc martingale with η0 = 1, we immediately deduce that $ IP ∼ IP and we can apply Girsanov’s theorem for semimartingales (cf. Jacod and Shiryaev (2003) III.3.24). The density process can be represented in the usual form $ dIP Ft = exp Lαt − CLαt ηt = IE dIP t t = exp ασs dWs + αx(µL − ν)(ds, dx) 0

1 − 2

0

0 t

α 2 σs2 ds −

R

t 0

R

(eαx − 1 − αx)ν(ds, dx) .

(5.4.6)

Consequently, we can identify the tuple (β, Y ) of predictable processes β(t) = α and Y (t, x) = exp(αx) which characterizes the change of measure. From Girsanov’s theorem, combined with Theorem II.4.15 in Jacod and Shiryaev (2003), we deduce that a PIIAC remains a PIIAC under the measure $ IP, because the processes β and Y are deterministic and the resulting characteristics under $ IP satisfy Assumption (AC).

Exotic Options in L´evy Models

109

$ t As a consequence of Girsanov’s theorem for semimartingales, we infer that Wt = Wt − $ $ ν = Y ν is the IP compensator of the jumps of L. 0 ασs ds is a IP-Brownian motion and $ Furthermore, as a corollary of Girsanov’s theorem, we can calculate the canonical decomposition of L under $ IP; t t t $ $ σs dWs + x(µL − $ ν)(ds, dx) (5.4.7) bs ds + Lt = 0

0

0

R

where $t = B

1 t 2 $ σ ds bs ds = (r − δ)t + α − 2 0 s 0 t (e−αx − e(1−α)x + x)$ ν(ds, dx) +

t

R

0

(5.4.8)

$ C,$ and hence, its triplet of characteristics is (B, ν). Deﬁne its dual process, L := −L $ C, −$ ν). The canonical and by Lemma 2.9, we get that its triplet is (B , C , ν ) = (−B, decomposition of L is t t t $ σs dWs + x(µL − ν )(ds, dx) (5.4.9) bs ds + Lt = − 0

0

0

R

IP-martingale for α = 1. and we can easily deduce that e(r−δ)t St is not a $ Adding the appropriate terms, we can re-write L as L := C∗ + L, where · · σs2 ds − (e−αx − e(1−α)x + 1 − e−x )$ ν(ds, dx) C∗ = (1 − α) 0

0

R

(5.4.10)

and L is such that e(r−δ)t S t is a $ IP-martingale. The characteristic triplet of L is (B − −1 ∗ C , C, ν ) and S t = S0 exp Lt . Therefore, we can conclude the proof α $−$ $c = e−δT K α CT $ IE (K ST )+ V +α * $ − eC∗T S T )+ = e−δT K α CT $ IE (K α ∗ = e−δT K α CT eαCT $ IE (K − S T )+ ∗

∗

$ −CT = K −1 e−CT . where K = Ke Setting α = 1 in the previous theorem, we immediately get a symmetry between European plain vanilla call and put options. Corollary 4.2. Assuming that (M) is in force and the asset price evolves as an exponential PIIAC, we can relate the European call and put option via the following symmetry: (5.4.11) Vc (S0 , K; r, δ, C, ν) = KS0 Vp S0−1 , K −1 ; δ, r, C, −f ν where f (x) = ex .

110

Exotic Option Pricing and Advanced L´evy Models

This symmetry relating European and also American plain vanilla call and put options, in exponential L´evy models, was proved independently in Fajardo and Mordecki (2003). Schroder (1999) proved similar results in a general semimartingale model; however, using a L´evy or PIIAC as the driving motion allows for the explicit calculation of the distribution under the new measure. A different symmetry, again relating European and American call and put options, in the Black–Scholes model, was derived by Peskir and Shiryaev (2002), where they use the mathematical concept of negative volatility; their main result states that Vc (ST , K; σ ) = Vp (−ST , −K; −σ ).

(5.4.12)

See also the discussion – and the corresponding cartoon – in Haug (2002). In this framework, one can derive symmetry relationships between self-quanto and European plain vanilla options. This result is, of course, a special case of Theorem 6.4; nevertheless, we give a short proof since it simpliﬁes considerably because the driving process is one-dimensional. The payoff of the self-quanto call and put option is ST (ST − K)+

ST (K − ST )+ ,

and

respectively. Introduce the following notation for the value of the self-quanto call option Vqc (S0 , K; r, δ, C, ν) = e−rT IE ST (ST − K)+ and similarly, for the self-quanto put option we set Vqp (S0 , K; r, δ, C, ν) = e−rT IE ST (K − ST )+ . Assumption (EM) can be replaced by the following weaker assumption, which is the minimal condition necessary for the symmetry results to hold. Assumption (M ). The L´evy measures λt of the distribution of Lt satisfy

T 0

|x|λt (dx)dt < ∞

T

e2x λt (dx)dt < ∞.

and

D−

0

D+

Theorem 4.3. Assume that the asset price evolves as an exponential PIIAC and (M ) is in force. We can relate the self-quanto and European plain vanilla call and put options via the following symmetry: ∗

Vqc (S0 , K; r, δ, C, ν) = S0 eCT Vc (S0 , K ∗ ; δ, r, C, f ν) C∗T

Vqp (S0 , K; r, δ, C, ν) = S0 e Vp (S0 , K ∗ ; δ, r, C, f ν)

(5.4.13) (5.4.14)

∗

where C ∗ is given by equation (5.4.16) (see below), K ∗ = Ke−CT and f (x) = ex . Proof. Expressing the value of the self-quanto call option in units of the num´eraire as described in Section 5.3, we deﬁne a new measure $ IP via its Radon–Nikodym derivative

Exotic Options in L´evy Models

111

given by equation (5.3.4) and the original valuation problem becomes $qc = e−δT $ IE (ST − K)+ . V

(5.4.15)

Now it sufﬁces to calculate the characteristic triplet of L under $ IP. Arguing as in the proof of Theorem 4.1, the density process η has the form of equation (5.4.6) for α = 1; hence, the tuple (β, Y ) of predictable processes that describes the change of measure is β(t) = 1 and Y (t, x) = exp(x). Therefore, L has the canonical decomposition under $ IP

t

Lt =

$ bs ds +

0

t

$s + σs dW

t

0

R

0

x(µL − $ ν)(ds, dx)

where σ2 $ bt = r − δ + t + 2

R

(e−x − 1 + x)ex λt (dx).

IP-martingale, but if we deﬁne L∗ as Notice that e(r−δ)t eLt is not a $ L∗t := (δ − r)t +

t

− 0

t 0

σt2 ds − 2

$s + σs dW

R

0

t 0

t

R

x(µL − $ ν)(ds, dx)

(ex − 1 − x)ex ν(ds, dx)

∗

then e(r−δ)t eLt ∈ M. Next, we re-express L as L = L∗ + C ∗ , where CT∗

= exp 2(r − δ)T +

T 0

σs2 ds

T

+ 0

(e + e x

R

−x

− 2)e ν(ds, dx) . x

(5.4.16)

By re-arranging the terms in equation (5.4.15), the result follows. 5.4.2 Valuation of European options We outline a method for the valuation of vanilla options, based on bilateral Laplace transforms, that was developed in the PhD thesis of Sebastian Raible (see Raible (2000) Chap. 3). The method is extremely fast and allows for the valuation not only of plain vanilla European derivatives, but also of more complex payoffs, such as digital, self-quanto and power options; in principle, every European payoff can be priced by using this method. Moreover, a large variety of driving processes can be handled, including L´evy and additive processes. The main idea of Raible’s method is to represent the option price as a convolution of two functions and consider its bilateral Laplace transform; then, by using the property that the Laplace transform of a convolution equals the product of the Laplace transforms of the factors, we arrive at two Laplace transforms that are easier to calculate analytically than the original one. Inverting this Laplace transform yields the option price.

112

Exotic Option Pricing and Advanced L´evy Models

A similar method, in Fourier space, can be found in Lewis (2001). See also Carr and Madan (1999) for some preliminary results that motivated this research. Lee (2004) uniﬁes and generalizes the existing Fourier-space methods and develops error bounds for the discretized inverse transforms. We ﬁrst state the necessary assumptions regarding the distribution of the asset price process and the option payoff respectively. (L1): Assume that ϕLT (z), the extended characteristic function of LT , exists for all z ∈ C with z ∈ I1 ⊃ [0, 1]. (L2): Assume that IPLT , the distribution of LT , is absolutely continuous w.r.t. the Lebesgue measure λ\1 with density ρ. (L3): Consider a European-style payoff function f (ST ) that is integrable. (L4): Assume that x → e−Rx |f (e−x )| is bounded and integrable for all R ∈ I2 ⊂ R. In order to price a European option with payoff function f (ST ), we proceed as follows. −rT −rT V = e IE[f (ST )] = e f (ST )dIP = e−rT = e−rT

R

R

f (S0 ex )dIPLT (x) f (S0 ex )ρ(x)dx

(5.4.17)

because of absolute continuity. Deﬁne ζ = − log S0 and g(x) = f (e−x ), and then −rT g(ζ − x)ρ(x)dx = e−rT (g ∗ ρ)(ζ ) (5.4.18) V =e R

which is a convolution at point ζ . Applying bilateral Laplace transforms on both sides of equation (5.4.18) and using Theorem B.2 in Raible (2000), we get LV (z) = e−rT e−zx (g ∗ ρ)(x)dx R

= e−rT

R

e−zx g(x)dx

= e−rT Lg (z)Lρ (z)

R

e−zx ρ(x)dx (5.4.19)

Lh (z) denotes the bilateral Laplace transform of a function h at z ∈ C, i.e. Lh (z) := where −zx h(x)dx. The Laplace transform of g is very easy to compute analytically and the e R Laplace transform of ρ can be expressed as the extended characteristic function ϕLT of LT . By numerically inverting this Laplace transform, we recover the option price. The next theorem gives us an explicit expression for the price of an option with payoff function f and driving PIIAC L. Theorem 4.4. Assume that (L1)–(L4) are in force and let g(x) := f (e−x ) denote the modiﬁed payoff function of an option with payoff f (x) at time T . Assume that I1 ∩ I2 = ∅

Exotic Options in L´evy Models

113

and choose an R ∈ I1 ∩ I2 . Letting V (ζ ) denote the price of this option, as a function of ζ := − log S0 , we have eζ R−rT V (ζ ) = 2π

R

eiuζ Lg (R + iu)ϕLT (iR − u)du,

(5.4.20)

whenever the integral on the right-hand side exists. Proof. The claim can be proved by using the arguments of the proof of Theorem 3.2 in Raible (2000); there, no explicit statement is made about the driving process L; hence, it directly transfers to the case of a time-inhomogeneous L´evy process. Remark 4.5. In order to apply this method, validity of the necessary assumptions has to be veriﬁed. (L1), (L3) and (L4) are easy to certify, while (L2) is the most demanding one. Let us mention that the distributions underlying the most popular L´evy processes, such as the Generalized Hyperbolic L´evy motion (cf. Eberlein and Prause (2002)), possess a known Lebesgue density. Remark 4.6. The method of Raible for the valuation of European options can be applied to general driving processes that satisfy Assumptions (L1)–(L4). Therefore, it can also be applied to stochastic volatility models based on L´evy processes that have attracted much interest lately; we refer to Barndorff-Nielsen and Shephard (2001), Eberlein et al. (2003) and Carr et al. (2003) for an account of different models. 5.4.3 Valuation of American options The method of Raible, presented in the previous section, can be used for pricing several types of European derivatives, but not path-dependent ones. The valuation of American options in L´evy-driven models is quite a hard task and no analytical solution exists for the ﬁnite horizon case. For perpetual American options, i.e. options with inﬁnite time horizon, Mordecki (2002) derived formulae in the general case in terms of the law of the extrema of the L´evy process, using a random walk approximation to the process. He also provides explicit solutions for the case of a jump-diffusion with exponential jumps. Alili and Kyprianou (2005) recapture the results of Mordecki by making use of excursion theory. Boyarchenko and Levendorskiˇı (2002c) obtained formulae for the price of the American put option in terms of the Wiener–Hopf factors and derive some more explicit formulae for these factors. Asmussen et al. (2004) ﬁnd explicit expressions for the price of American put options for L´evy processes with two-sided phase-type jumps; the solution uses the Wiener–Hopf factorization and can also be applied to regime-switching L´evy processes with phase-type jumps. For the valuation of ﬁnite time horizon American options, one has to resort to numerical methods. Denote by x = ln S the log price, τ = T − t the time to maturity and v(τ, x) = f (ex , T − τ ) the time-t value of an option with payoff function g(ex ) = φ(x). One approach is to use numerical schemes for solving the corresponding partial integro-differential inequality (PIDI), ∂v − Av + rv ≥ 0 ∂τ

in (0, T ) × R

(5.4.21)

114

Exotic Option Pricing and Advanced L´evy Models

subject to the conditions a.e. in [0, T ] × R v(τ, x) ≥ φ(x), ∂v − Av + rv = 0, in (0, T ) × R (v(τ, x) − φ(x)) ∂τ v(0, x) = φ(x)

(5.4.22)

where σ 2 d2 v σ 2 dv + Av(x) = r − δ − 2 dx 2 dx 2 dv v(x + y) − v(x) − (ey − 1) (x) λ(dy) + dx R

(5.4.23)

is the inﬁnitesimal generator of the transition semigroup of L; see Matache et al. (2003, 2005) for all of the details and numerical solution of the problem using wavelets. Almendral (2004) solves the problem numerically by using implicit–explicit methods in case the CGMY is the driving process. Equation (5.4.21) is a backward PIDE in spot and time to maturity; Carr and Hirsa (2003) develop a forward PIDE in strike and time of maturity and solve it by using ﬁnite-difference methods. Another alternative is to employ Monte Carlo methods adapted for optimal stopping problems, such as the American option; we refer here to Rogers (2002) or Glasserman (2003). K¨ellezi and Webber (2004) constructed a lattice for L´evy-driven assets and applied it to the valuation of Bermudan options. Levendorskiˇı (2004) develops a non-Gaussian analog of the method of lines and uses Carr’s randomization method in order to formulate an approximate algorithm for the valuation of American options. Chesney and Jeanblanc (2004) revisit the perpetual American problem and obtain formulae for the optimal boundary when jumps are either only positive or only negative. Using these results, they approximate the ﬁnite horizon problem in a fashion similar to Barone-Adesi and Whaley (1987). Empirical tests show that this approximation provides good results only when the process is continuous at the exercise boundary.

5.5 EXOTIC OPTIONS The work on this topic follows along the lines of Henderson and Wojakowski (2002); they proved an equivalence between the price of ﬂoating and ﬁxed strike Asian options in the Black–Scholes model. We also refer to Vanmaele et al. (2002) for a generalization of these results to forward-start options and discrete averaging in the Black–Scholes model. 5.5.1 Symmetry For exotic options, the setting is that of (P2): L1 = L is the driving R-valued PIIAC with triplet (B, C, ν), S 1 = S01 exp L1 = S and S 2 = h(S) is a functional of S. The most prominent candidates for functionals are the maximum, the minimum and the (arithmetic) average; let 0 = t1 < t2 < · · · < tn = T be equidistant time points, and then the resulting processes, in case of discrete monitoring, are 1 Sti . n n

MT = max Sti , 0≤ti ≤T

NT = min Sti 0≤ti ≤T

and "T =

i=1

Exotic Options in L´evy Models

115

Table 5.1 Types of payoffs for Asian and lookback options Option type

Asian payoff

Lookback payoff

Fixed strike call Fixed strike put Floating strike call Floating strike put

("T − K)+ (K − "T )+ (ST − "T )+ ("T − ST )+

(MT − K)+ (K − NT )+ (ST − NT )+ (MT − ST )+

Therefore, we can exploit symmetries between ﬂoating and ﬁxed strike Asian and lookback options in this framework; the different types of payoffs of the Asian and lookback option are summarized in Table 5.1. We introduce the following notation for the value of the ﬂoating strike call option, be it Asian or lookback Vc (ST , h(S); r, δ, C, ν) = e−rT IE (ST − h(S)T )+ and similarly, for the ﬁxed strike put option we set Vp (K, h(S); r, δ, C, ν) = e−rT IE (K − h(S)T )+ ; a similar notation will be used for the other two cases. Now we can state a result that relates the value of ﬂoating and ﬁxed strike options. Notice that because stationarity of the increments plays an important role in the proof, the result is valid only for L´evy processes. Theorem 5.1. Assuming that the asset price evolves as an exponential L´evy process, we can relate the ﬂoating and ﬁxed strike Asian or lookback option via the following symmetry: Vc ST , h(S); r, δ, σ 2 , λ = Vp S0 , h(S); δ, r, σ 2 , −f λ Vp h(S), ST ; r, δ, σ 2 , λ = Vc h(S), S0 ; δ, r, σ 2 , −f λ

(5.5.1) (5.5.2)

where f (x) = ex . Proof. We refer to the proof of Theorems 3.1 and 4.1 in Eberlein and Papapantoleon (2005). The minimal assumptions necessary for the results to hold are also stated there. Remark 5.2. These results also hold for forward-start Asian and lookback options, for continuously monitored options, for partial options and for Asian options on the geometric and harmonic average; see Eberlein and Papapantoleon (2005) for all of the details. Note that the equivalence result is not valid for in-progress Asian options. 5.5.2 Valuation of barrier and lookback options The valuation of barrier and lookback options for assets driven by general L´evy processes is another hard mathematical problem. The difﬁculty stems from the fact that (a) the distribution of the supremum or inﬁmum of a L´evy process is not known explicitly, and (b) the

116

Exotic Option Pricing and Advanced L´evy Models

overshoot distribution associated with the passage of a L´evy process across a barrier is also not known explicitly. Various authors have treated the problem in the case where the driving process is a spectrally positive/negative L´evy process; see, for example, Rogers (2000), Sch¨urger (2002) and Avram et al. (2004). Kou and Wang (2003, 2004) have derived explicit formulas for the values of barrier and lookback options in a jump diffusion model where the jumps are double-exponentially distributed; they make use of a special property of the exponential distribution, namely the memoryless property, which allows them to explicitly calculate the overshoot distribution. Lipton (2002) derives similar formulas for the same model, making use of ﬂuctuation theory. Fluctuation theory and the Wiener–Hopf factorization of L´evy processes play a crucial role in every attempt to derive closed form solutions for the value of barrier and lookback options in L´evy-driven models. Introduce the notation Mt = sup Ls

Nt = inf Ls

and

0≤s≤t

0≤s≤t

and let θ denote a random variable exponentially distributed with parameter q, independent of L. Then, the celebrated Wiener–Hopf factorization of the L´evy process L states that IE[exp(izLθ )] = IE[exp(izMθ )] · IE[exp(izNθ )]

(5.5.3)

or equivalently q(q − ψ(z))−1 = ϕq+ (z) · ϕq− (z),

z ∈ R,

(5.5.4)

where ψ denotes the characteristic exponent of L. The functions ϕq+ and ϕq− have the following representations ϕq+ (z) ϕq− (z)

= exp = exp

*

∞

t *

−1 −qt

e

0 ∞

t 0

−1 −qt

e

∞

dt dt

(eizx − 1)µt (dx)

+ (5.5.5)

0 0 −∞

(eizx − 1)µt (dx)

+ (5.5.6)

where µt (dx) = IPLt (dx) is the probability measure of Lt . These results were ﬁrst proved for L´evy processes in Bingham (1975) – where an approximation of L´evy processes by random walks is employed – and subsequently by Greenwood and Pitman (1980) – where excursion theory is applied. See also the recent books by Sato (1999, Chapter 9) and Bertoin (1996, Chapter VI) respectively, for an account of these two methods. Building upon these results, various authors have derived formulae for the valuation of barrier and lookback options; Boyarchenko and Levendorskiˇı (2002a) apply methods from potential theory and pseudodifferential operators to derive formulae for barrier and touch options, while Nguyen-Ngoc and Yor (2005) use a probabilistic approach based on excursion theory. Recently, Nguyen-Ngoc (2003) takes a similar probabilistic approach, motivated from Carr and Madan (1999) and derives quite simple formulae for the value of barrier and lookback options, which can be numerically evaluated with the use of Fourier inversion algorithms in two and three dimensions.

Exotic Options in L´evy Models

117

More speciﬁcally, let us denote by Vc (MT , K; T ) the price of a ﬁxed strike lookback option with payoff (MT − K)+ , where MT = max0≤t≤T St and S is an exponential L´evy α,γ process. Choose γ > 1 and α > 0 such that IE[e2L1 ] < er+α and set Vc (MT , K; T ) = −αT −γ k Vc (MT , K; T ) where k = log(K/S0 ). Then, we have the following result. e Proposition 5.3. If k > 0, then for all q, u > 0 we have: ∞ ∞ e−qT dT e−uk Vcα,γ (MT , S0 ek ; T )dk 0

0

1 1 + [ϕ + (i(z − 1)) + (z − 1)ϕq+r+α (−i) − z] = S0 q + r + α z(z − 1) q+r+α

(5.5.7)

where z = u + γ . Proof. We refer to the proof of Proposition 3.9 in Nguyen-Ngoc (2003). The formula for the value of the ﬂoating strike lookback option is – as one could easily foresee – a lot more complicated than equation (5.5.7). Using the symmetry result of Theorem 5.1, this case can be dealt with via a change of the L´evy triplet and strike in the previous proposition. The Wiener–Hopf factors are not known explicitly in the general case and numerical computation could be extremely time-consuming. Boyarchenko and Levendorskiˇı (2002b) provide some more efﬁcient formulas for the Wiener–Hopf factors of – what they call – regular L´evy processes of exponential type (RLPE); for the deﬁnition refer to Section 1.2.2 in the above-mentioned reference. Given that L is an RLPE, ϕq+ (z) has an analytic continuation on the half plane z > ω and * z +∞+iω ln(q + ψ(u)) + du . (5.5.8) ϕq+ (z) = exp 2π i −∞+iω u(z − u) The family of RLPEs contains many popular – in mathematical ﬁnance – L´evy motions such as the Generalized Hyperbolic and Variance Gamma models (see Boyarchenko and Levendorskiˇı (2002b)). Discretely monitored options have received much less attention in the literature than their continuous time counterparts. Borovkov and Novikov (2002) use Fourier methods and Spitzer’s identity to derive formulae for ﬁxed strike lookback options. Various numerical methods have been applied for the valuation of barrier and lookback options in L´evy-driven models. Cont and Voltchkova (2005a, 2005b) study ﬁnite-difference methods for the solution of the corresponding PIDE (see also Matache et al. (2004)). Ribeiro and Webber (2003, 2004) have developed fast Monte Carlo methods for the valuation of exotic options in models driven by the Variance Gamma (VG) and Normal Inverse Gaussian (NIG) L´evy motions; their method is based on the construction of Gamma and Inverse Gaussian bridges, respectively, to speed up the Monte Carlo simulation. The recent book of Schoutens (2003) contains a detailed account of Monte Carlo methods for L´evy processes, also allowing for stochastic volatility. 5.5.3 Valuation of Asian and basket options An explicit solution for the value of the arithmetic Asian or basket option is not known in the Black–Scholes model and, of course, the situation is similar for L´evy models. The difﬁculty

118

Exotic Option Pricing and Advanced L´evy Models

is that the distribution of the arithmetic sum of log-normal random variables – more generally, random variables drawn from some log r inﬁnitely divisible distribution – is not known in closed-form. Veˇceˇr and Xu (2004) formulated a PIDE for all types of Asian options – including inprogress options – in a model driven by a process with independent increments (PII) or, more generally, a special semimartingale. Their derivation is based on the construction of a suitable self-ﬁnancing trading strategy to replicate the average and then a change of num´eraire – which is essentially the one we use – in order to reduce the number of variables in the equation. Their PIDE is relatively simple and can be solved by using numerical techniques such as ﬁnite-differences. Albrecher and Predota (2002, 2004) use moment-matching methods to derive approximate formulae for the value of Asian options in some popular L´evy models such as the NIG and VG models; they also derive bounds for the option price in these models. See also the survey paper by Albrecher (2004) for a detailed account of the above mentioned results. Hartinger and Predota (2002) apply Quasi Monte-Carlo methods for the valuation of Asian options in the Hyperbolic model. Their method can be extended to the class of Generalized Hyperbolic L´evy motions, which contains the VG motion as a special case; see Eberlein and von Hammerstein (2004). Benhamou (2002), building upon the work of Carverhill and Clewlow (1992), uses the Fast Fourier transform and a transformation of dependent variables into independent ones, in order to value discretely monitored ﬁxed strike Asian options. As he points out, this method can be applied when the return distribution is fat-tailed, with L´evy processes being prominent candidates. Henderson et al. (2004) derive an upper bound for in-progress ﬂoating strike Asian options in the Black–Scholes model, using the symmetry result of Henderson and Wojakowski (2002) and valuation methods for ﬁxed strike ones. Their pricing bound relies on a modeldependent symmetry result and a model-independent decomposition of the ﬂoating-strike Asian option into a ﬁxed-strike one and a vanilla option. Therefore, given the symmetry result of Theorem 5.1, their general methodology can also be applied to L´evy models. Albrecher et al. (2004) derive static super-hedging strategies for ﬁxed strike Asian options in L´evy models; these results were extended to L´evy models with stochastic volatility in Albrecher and Schoutens (2005). The method is based on super-replicating the Asian payoff with a portfolio of plain vanilla calls, using the following upper bound + n n Stj − nK ≤ (Stj − nKj )+ (5.5.9) j =1

j =1

and then optimizing the hedge, i.e. the choice of Kj s, using results from co-monotonicity theory. Similar ideas appear in Hobson et al. (2004) for the static super-hedging of basket options. The payoff of the basket option is super-replicated by a portfolio of plain vanilla calls on each individual asset, using the upper bound

+ n n + i wi ST − K ≤ (5.5.10) wi STi − li K n

i =1

i =1

where li ≥ 0 and i =1 li = 1; subsequently, the portfolio is optimized using co-monotonicity theory. Moreover, no distribution is assumed about the asset dynamics, since all of the

Exotic Options in L´evy Models

119

information needed is the marginal distributions which can be deduced from the volatility smile; we refer to Breeden and Litzenberger (1978). This is also observed by Albrecher and Schoutens (2005).

5.6 MARGRABE-TYPE OPTIONS In this section, we derive symmetry results between options involving two assets – such as Margrabe or Quanto options – and European plain vanilla options; therefore, we generalize results by Margrabe (1978) and Fajardo and Mordecki (2003) to the case of timeinhomogeneous L´evy processes. Schroder (1999) provides similar results for semimartingale models; the advantage of using a L´evy process or a PIIAC instead of a semimartingale as the driving motion, is that the distribution of the asset returns under the new measure can be deduced from the distribution of the returns of each individual asset under the risk-neutral measure. For Margrabe-type options, the setting is that of (P3): L = (L1 , L2 ) is the driving 2 R -valued PIIAC with triplet (B, C, ν) and S = (S 1 , S 2 ) is the asset price process. For convenience, we set i = 1, 2, (5.6.1) Sti = S0i exp (r − δ i )t + Lit , modifying the characteristic triplet (B, C, ν) accordingly. With Theorem 25.17 in Sato (1999) and Lemma 2.4, Assumption (EM) guarantees the existence of the moment generating function MLt of Lt for u ∈ Cd such that u ∈ [−M, M]d . Furthermore, for u ∈ Cd with u ∈ [−M, M]d , we have that MLt (u) = ϕLt (−iu) = IE eu,Lt t* 1 u, bs + u, cs u = exp 2 0 + + (eu,x − 1 − u, x)λs (dx) ds. (5.6.2) Rd

The next result will allow us to calculate the characteristic triplet of a one-dimensional process, deﬁned as a scalar product of a vector with the d-dimensional process L, from the characteristics of L under an equivalent change of probability measure. Proposition 6.1. Let L be a d-dimensional PIIAC with triplet (B, C, ν) under IP, let u, v be loc IP ∼ IP, with density vectors in Rd and v ∈ [−M, M]d . Moreover, let $ d$ IP ev,LT = . dIP IE[ev,LT ] := u, L is a $ Then, the one-dimensional process L IP-PIIAC and its characteristic triplet is (B, C, ν) with 1 u, x ev,x − 1 λs (dx) bs = u, bs + u, cs v + v, cs u + 2 Rd cs = u, cs u λs = T (κs )

120

Exotic Option Pricing and Advanced L´evy Models

where T is a mapping T : Rd → R, such that x → T (x) = u, x, and κs is a measure deﬁned by ev,x λs (dx). κs (A) = A

Proof. Because the density process (ηt ) is given by ηt = ev,Lt IE[ev,Lt ]−1 , by using equation (5.6.2) we get $ IE ezu,Lt = IE ezu,Lt ηt * −1 + = IE ezu,Lt ev,Lt IE ev,Lt −1 = IE ezu+v,Lt IE ev,Lt t* 1 zu + v, bs + zu + v, cs (zu + v) = exp 2 0 + + (ezu+v,x − 1 − zu + v, x)λs (dx) ds Rd

t

× exp 0

* 1 − v, bs + v, cs v 2 + + (ev,x − 1 − v, x)λs (dx) ds Rd

t* , 1 = exp z u, bs + u, cs v + v, cs u 2 0 v,x - 1 u, x e − 1 λs ds + z2 u, cs u + 2 Rd + zu,x e − 1 − zu, x ev,x λs (dx) ds. + Rd

If we write κs for the measure on Rd given by ev,x λs (dx), κs (A) =

(5.6.3)

(5.6.4)

A

A ∈ B(Rd ) and T for the linear mapping T : Rd → R given by T (x) = u, x, then we get for the last term in the exponent of equation (5.6.3) v,x zu,x zy − 1 − zu, x e λs (dx) = e e − 1 − zy T (κs )(dy) Rd

R

by the change-of-variable formula. The resulting characteristics satisfy Assumption (AC), and thus the result follows. The valuation of options depending on two assets modelled by a two-dimensional PIIAC can now be simpliﬁed – using the technique described in Section 5.3 and Proposition 6.1 – to

Exotic Options in L´evy Models

121

the valuation of an option on a one-dimensional asset. Subsequently, this option can be priced by using bilateral Laplace transforms, as described in Section 5.4.2. The payoff of a Margrabe option, or option to exchange one asset for another, is + 1 ST − ST2 and we denote its value by Vm (S01 , S02 ; r, δ, C, ν) = e−rT IE

*

ST1 − ST2

+ +

where δ = (δ 1 , δ 2 ). The payoff of the Quanto call and put option is + ST1 ST2 − K

and

+ ST1 K − ST2 ,

respectively, and we will use the following notation for the value of the Quanto call option * + + Vqc (S01 , S02 , K; r, δ, C, ν) = e−rT IE ST1 ST2 − K and similarly for the Quanto put option * + + . Vqp (S01 , S02 , K; r, δ, C, ν) = e−rT IE ST1 K − ST2 The different variants of the Quanto option traded in Foreign Exchange markets are explained in Musiela and Rutkowski (1997). The payoff of a cash-or-nothing and a two-dimensional asset-or-nothing option is 1l{ST >K}

and

ST1 1l{S 2 >K} . T

The holder of a two-dimensional asset-or-nothing option receives one unit of asset S 1 at expiration, if asset S 2 ends up in the money; of course, this is a generalization of the (standard) asset-or-nothing option, where the holder receives one unit of the asset if it ends up in the money. We denote the value of the cash-or-nothing option by Vcn (S0 , K; r, δ, C, ν) = e−rT IE 1l{ST >K} and the value of the two-dimensional asset-or-nothing option by + * Van (S01 , S02 , K; r, δ, C, ν) = e−rT IE ST1 1l{S 2 >K} . T

Notice that in the ﬁrst case, r, δ, C and ν correspond to a one-dimensional driving process, while in the second case to a two-dimensional one. Theorem 6.2. Let Assumption (EM) be in force and assume that the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the value of a Margrabe and a European plain vanilla option via the following symmetry: ν Vm (S01 , S02 ; r, δ, C, ν) = IE[ST1 ]eCT Vp S02 /S01 , K; δ 1 , r, C,

(5.6.5)

122

Exotic Option Pricing and Advanced L´evy Models

ν) where K = e−CT , C is given by equation (5.6.9) (see below) and the characteristics (C, are given by Proposition 6.1 for v = (1, 0) and u = (−1, 1).

Proof. Expressing the value of the Margrabe option in units of the num´eraire, we get * −rT + 1 2 + $ := Vm = e S I E − S V T T S01 S01 %

+ & ST2 e−rT ST1 ηT1 −δ 1 T 1− 1 IE =e ST e−δ 1 T S01 ηT1 where η1 = IE[exp(L1 )] = IE[expv, L], for v = (1, 0), and by using equation (5.6.1) we get % =

1 e−δ T ηT1 IE

1

eLT ηT1

+ & ST2 1− 1 . ST

(5.6.6)

Deﬁne a new measure $ IP via its Radon–Nikodym derivative 1 eLT d$ IP = 1 dIP IE[eLT ]

and the valuation problem takes the form $ = e−δ 1 T ηT1 $ IE V

*

1− ST

+ +

where, by using equation (5.6.1), we get S2 1 2 S2 2 1 St := t1 = 01 e(δ −δ )t+Lt −Lt =: S0 exp (δ 1 − δ 2 )t + Lt St S0

(5.6.7)

:= L2 − L1 = u, L for u = (−1, 1). The characteristic triplet of L, (B, C, ν) under and L $ IP, is given by Proposition 6.1 for v = (1, 0) and u = (−1, 1). 1 IP-martingale. However, if we deﬁne St is not a $ Observe that e(r−δ )t t 1 t 1 cs ds − Lt := (δ − r)t − (ex − 1 − x) ν(ds, dx) 2 0 R 0 t t $s + cs1/2 dW x(µL − ν)(ds, dx) (5.6.8) + 0

0

R

$ is a $ where W IP-standard Brownian motion and µL is the random measure of jumps of L, 1 t + then e(r−δ )t eLt ∈ M. Therefore, we re-express the exponent of equation (5.6.7) as L Ct where (δ 1 − δ 2 )t = Lt + t t 1 t 2 cs ds + (ex − 1 − x) ν(ds, dx) (5.6.9) bs ds + Ct = (r − δ )t + 2 0 R 0 0

S0 exp Lt . and deﬁne S t :=

Exotic Options in L´evy Models

123

Now the result follows, because

* + + 1− ST 1 + = e−δ T ηT1 I$ E 1 − S T eCT

$ = e−δ 1 T ηT1 I$ E V

IE = e−δ T ηT1 eCT $ 1

+ . e−CT − S T

Theorem 6.3. Let Assumption (EM) be in force and assume that the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the value of a Quanto and a European plain vanilla call option via the following symmetry: ν Vqc (S01 , S02 , K; r, δ, C, ν) = IE[ST1 ]eCT Vp S02 , K; δ 1 , r, C, (5.6.10)

C is given by where K = e−CT , the constant t t 1 t 1 2 cs ds + Ct = (2r − δ − δ )t + (ex − 1 − x) ν(ds, dx) bs ds + 2 0 R 0 0 ν) are given by Proposition 6.1 for v = (1, 0) and u = (0, 1). A and the characteristics (C, similar relationship holds for the Quanto and European plain vanilla put options. Proof. The proof follows along the lines of that of Theorem 6.2. Theorem 6.4. Let Assumption (EM) be in force and assume that the asset price evolves as an exponential PIIAC according to equations (5.2.3)–(5.2.5). We can relate the value of a two-dimensional asset-or-nothing and a cash-or-nothing option via the following symmetry: ν (5.6.11) Van (S01 , S02 , K; r, δ, C, ν) = IE[ST1 ]Vcn S02 , K; δ 1 , r, C, C is given by where K = Ke−CT , the constant t t 1 t cs ds + (ex − 1 − x) ν(ds, dx) bs ds + Ct = (2r − δ 1 − δ 2 )t + 2 R 0 0 0

ν) are given by Proposition 6.1 for v = (1, 0) and u = (0, 1). A and the characteristics (C, similar relationship holds for the corresponding put options. Proof. The proof follows along the lines of that of Theorem 6.2. Remark 6.5. Notice that the factor IE[ST1 ] is the forward price of the asset S 1 , the num´eraire asset.

ACKNOWLEDGEMENTS We thank Wolfgang Kluge for helpful discussions during the work on these topics. The second named author acknowledges the ﬁnancial support provided through the European Community’s Human Potential Programme under Contract HPRN-CT-2000-00100 DYNSTOCH.

124

Exotic Option Pricing and Advanced L´evy Models

REFERENCES [1] Albrecher, H. (2004), “The valuation of Asian options for market models of exponential L´evy type”, in M. Van Maele, A. De Schepper, J. Dhaene, H. Reynaerts, W. Schoutens and P. Van Goethem (Eds), Handelingen Contact forum – 2nd Actuarial and Financial Mathematics Day, Royal Flemish Academy of Belgium for Science and Arts, Brussels, Belgium, pp. 11–20. [2] Albrecher, H., Dhaene, J., Goovaerts, M. and Schoutens, W. (2005), “Static hedging of Asian options under L´evy models: the comonotonicity approach”, The Journal of Derivatives, 12(3), 63–72. [3] Albrecher, H. and Predota, M. (2002), “Bounds and approximations for discrete Asian options in a variance-gamma model”, Grazer Mathematische Berichte, 345, 35–57. [4] Albrecher, H. and Predota, M. (2004), “On Asian option pricing for NIG L´evy processes”, Journal of Computational and Applied Mathematics 172, 153–168. [5] Albrecher, H. and Schoutens, W. (2005), “Static hedging of Asian options under stochastic volatility models using Fast Fourier Transform”, in A. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 129–147. [6] Kyprianou A.E. and Alili, L. (2005), “Some remarks on ﬁrst passage of L´evy process, the American put and pasting principles”, The Annals of Applied Probability, to be published. [7] Almendral, A. (2005), “Numerical valuation of American options under the CGMY process”, in A. Kyprianou, W. Schoutens, and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 259–276. [8] Asmussen, S., Avram, F. and Pistorius, M.R. (2004), “Russian and American put options under exponential phase-type L´evy models”, Stochastic Processes and their Applications, 109, 79–111. [9] Avram, F., Kyprianou, A. and Pistorius, M.R. (2004), “Exit problems for spectrally negative L´evy processes and applications to (Canadized) Russian options”, Annals of Applied Probability, 14, 215–238. [10] Barndorff-Nielsen, O.E. and Shephard, N. (2001), “Non-Gaussian Ornstein–Uhlenbeck-based models and some of their uses in ﬁnancial economics”, Journal of the Royal Statistical Society, Series B , 63, 167–241. [11] Barone-Adesi, G. and Whaley, R.E. (1987), “Efﬁcient analytic approximation of American option values”, Journal of Finance, 42, 301–320. [12] Bellini, F. and Frittelli, M. (2002), “On the existence of minimax martingale measures”, Mathematical Finance, 12, 1–21. [13] Benhamou, E. (2002), “Fast Fourier Transform for discrete Asian options”, Journal of Computational Finance, 6, 49–68. [14] Bertoin, J. (1996), L´evy Processes, Cambridge University Press, Cambridge, UK. [15] Bingham, N.H. (1975), “Fluctuation theory in continuous time”, Advances in Applied Probability, 7, 705–766. [16] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities”, Journal of Political Economy, 81, 637–654. [17] Borovkov, K. and Novikov, A. (2002), “On a new approach to calculating expectations for option pricing”, Journal of Applied Probability, 39, 889–895. [18] Boyarchenko, S.I. and Levendorskiˇı, S.Z. (2002a), “Barrier options and touch-and-out options under regular L´evy processes of exponential type”, Annals of Applied Probability, 12, 1261–1298. [19] Boyarchenko, S.I. and Levendorskiˇı, S.Z. (2002b), Non-Gaussian Merton–Black–Scholes Theory, World Scientiﬁc, Singapore. [20] Boyarchenko, S.I. and Levendorskiˇı, S.Z. (2002c), “Perpetual American options under L´evy processes”, SIAM Journal on Control and Optimization, 40, 1663–1696.

Exotic Options in L´evy Models

125

[21] Breeden, D. and Litzenberger, R. (1978), “Prices of state-contingent claims implicit in option prices”, Journal of Business, 51, 621–651. [22] Carr, P. (1994), “European put call symmetry”, Working Paper, Cornell University, Ithaca, New York, NY, USA. ´ [23] Carr, P. and Chesney, M. (1996), “American put call symmetry”, Working Paper, Ecole des hautes e´ tudes commerciales de Montreal (HEC), Montreal, QU, Canada. [24] Carr, P., Geman, H. Madan, D.B. and Yor, M. (2002), “The ﬁne structure of asset returns: an empirical investigation”, Journal of Business, 75, 305–332. [25] Carr, P., Geman, H. Madan, D.B. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [26] Carr, P. and Hirsa, A. (2003), “Why be backward? Forward equations for American options”, Risk , 16(1), 103–107. Reprinted in A. Kyprianou, W. Schoutens and P. Wilmott (Eds) (2005), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 237–257. [27] Carr, P. and Madan, D.B. (1999), “Option valuation using the fast Fourier transform”, Journal of Computational Finance, 2, 61–73. [28] Carverhill, A. and Clewlow, L. (1992), “Flexible convolution”, in R. Tompkins (Ed.), From Black–Scholes to Black Holes, Risk Publications, London, UK, pp. 165–171. [29] Chesney, M. and Jeanblanc, M. (2004), “Pricing American currency options in an exponential L´evy model”, Applied Mathematical Finance, 11, 207–225. [30] Cont, R. and Tankov, P. (2004), “Nonparametric calibration of jump-diffusion option pricing models”, Journal of Computational Finance, 7(3), 1–49. [31] Cont, R. and Voltchkova, E. (2005), “Finite difference methods for option pricing in jumpdiffusion and exponential L´evy models”, SIAM Journal on Numerical Analysis, to be published. [32] Cont, R. and Voltchkova, E. (2005), “Integro-differential equations for option prices in exponential L´evy models”, Finance and Stochastics, 9, 299–325. [33] Delbaen, F. and Schachermayer, W. (1994), “A general version of the fundamental theorem of asset pricing”, Mathematische Annalen, 300, 463–520. [34] Delbaen, F. and Schachermayer, W. (1998), “The fundamental theorem of asset pricing for unbounded stochastic processes”, Mathematische Annalen, 312, 215–250. [35] Detemple, J. (2001), “American options: symmetry properties”, in J. Cvitanic, E. Jouini and M. Musiela (Eds), Option Pricing, Interest Rates and Risk Management, Cambridge University Press, Cambridge, UK, pp. 67–104. [36] Eberlein, E. and Jacod, J. (1997), “On the range of options prices”, Finance and Stochastics, 1, 131–140. [37] Eberlein, E., Jacod, J. and Raible, S. (2005), “L´evy term structure models: no-arbitrage and completeness”, Finance and Stochastics, 9, 67–88. [38] Eberlein, E., Kallsen, J. and Kristen, J. (2003), “Risk management based on stochastic volatility”, Journal of Risk , 5(2), 19–44. [39] Eberlein, E. and Keller, U. (1995), “Hyperbolic distributions in ﬁnance”, Bernoulli , 1, 281–299. [40] Eberlein, E. and Kluge, W. (2004), “Exact pricing formulae for caps and swaptions in a L´evy term structure model”, FDM Preprint 86, University of Freiburg, Freiburg, Germany. [41] Eberlein, E. and Papapantoleon, A. (2005), “Equivalence of ﬂoating and ﬁxed strike Asian and lookback options”, Stochastic Processes and their Applications, 105, 31–40. [42] Eberlein, E. and Prause, K. (2002), “The generalized hyperbolic model: ﬁnancial derivatives and risk measures”, in H. Geman, D. Madan, S. Pliska and T. Vorst (Eds), Mathematical FinanceBachelier Congress 2000, Springer-Verlag, Berlin, Germany, pp. 245–267. [43] Eberlein, E. and von Hammerstein, E.-A. (2004), “Generalized hyperbolic and inverse Gaussian distributions: limiting cases and approximation of processes”, in R. Dalang, M. Dozzi and

126

[44]

[45] [46] [47] [48] [49] [50] [51] [52]

[53] [54] [55]

[56] [57] [58] [59] [60] [61] [62] [63] [64]

Exotic Option Pricing and Advanced L´evy Models F. Russo (Eds), Seminar on Stochastic Analysis, Random Fields and Applications IV, Progress in Probability, Vol. 58, Birkh¨auser Verlag, Basel, Switzerland, pp. 221–264. Fajardo, J. and Mordecki, E. (2003), “Duality and derivative pricing with L´evy processes”, Preprint No. 76, Centro de Matem´atica, Universidad de la Rep´ublica del Uruguay, Montevideo, Uruguay. Frittelli, M. (2000), “The minimal entropy martingale measure and the valuation problem in incomplete markets”, Mathematical Finance, 10, 39–52. Fujiwara, T. and Miyahara, Y. (2003), “The minimal entropy martingale measures for geometric L´evy processes”, Finance and Stochastics, 7, 509–531. Geman, H., El Karoui, N. and Rochet, J.-C. (1995), “Changes of num´eraire, changes of probability measures and option pricing”, Journal of Applied Probability, 32, 443–458. Glasserman, P. (2003), Monte Carlo Methods in Financial Engineering, Springer-Verlag, New York, NY, USA. Goll, T. and R¨uschendorf, L. (2001), “Minimax and minimal distance martingale measures and their relationship to portfolio optimization”, Finance and Stochastics, 5, 557–581. Greenwood, P. and Pitman, J. (1980), “Fluctuation identities for L´evy processes and splitting at the maximum”, Advances in Applied Probability, 12, 893–902. Gushchin, A.A. and Mordecki, E. (2002), “Bounds on option prices for semimartingale market models”, Proceedings of the Steklov Mathematical Institute 237, 73–113. Hakala, J. and Wystup, U. (2002), “Heston’s stochastic volatility model applied to foreign exchange options”, in J. Hakala and U. Wystup (Eds), Foreign Exchange Risk, Risk Publications, London, UK, pp. 267–282. Hartinger, J. and Predota, M. (2002), “Pricing Asian options in the hyperbolic model: a fast quasi Monte Carlo approach”, Grazer Mathematische Berichte, 345, 1–33. Haug, E.G. (2002), “A look in the antimatter mirror”, Wilmott Magazine, September, 38–42. Henderson, V., Hobson, D., Shaw, W. and Wojakowski, R. (2004), “Bounds for in-progress ﬂoating-strike Asian options using symmetry”, Annals of Operational Research (Special Issue on Financial Modelling), (to appear). Henderson, V. and Wojakowski, R. (2002), “On the equivalence of ﬂoating and ﬁxed-strike Asian options”, Journal of Applied Probability, 39, 391–394. Heston, S.L. (1993), “A closed-form solution for options with stochastic volatility with applications to bond and currency options”, Review of Financial Studies, 6, 327–343. Hobson, D., Laurence, P. and Wang, T.-H. (2005), “Static-arbitrage upper bounds for the prices of basket options”, Quantitative Finance, to be published. Jacod, J. and Shiryaev, A.N. (2003), Limit Theorems for Stochastic Processes, 2nd Edn, SpringerVerlag, Berlin, Germany. Kallsen, J. and Shiryaev, A.N. (2002), “The cumulant process and Esscher’s change of measure”, Finance and Stochastics, 6, 397–428. Kallsen, J. and Tankov, P. (2004), “Characterization of dependence of multidimensional L´evy ´ processes using L´evy copulas”, Working Paper, Ecole Polytechnique, Paris, France. K¨ellezi, E. and Webber, N. (2004), “Valuing Bermudan options when asset returns are L´evy processes”, Quantitative Finance, 4, 87–100. Kou, S.G. and Wang, H. (2003), “First passage times of a jump diffusion process”, Advances in Applied Probability, 35, 504–531. Kou, S.G. and Wang, H. (2005), “Option pricing under a double exponential jump diffusion model”, Management Science, 50, 1178–1192.

Exotic Options in L´evy Models

127

[65] Lee, R.W. (2004), “Option pricing by transform methods: extensions, uniﬁcation, and error control”, Journal of Computational Finance, 7(3), 50–86. [66] Levendorskiˇı, S.Z. (2004), “Pricing of the American put under L´evy processes”, International Journal of Theoretical and Applied Finance, 7, 303–335. [67] Lewis, A. (2001), “A simple option formula for general jump-diffusion and other exponential L´evy processes”, Working Paper, Optioncity.net. [68] Lipton, A. (2002), “Assets with jumps”, Risk , 15(9), 149–153. [69] Margrabe, W. (1978), “The value of an option to exchange one asset for another”, Journal of Finance, 33, 177–186. [70] Matache, A.-M., Nitsche, P.-A. and Schwab, C. (2003), “Wavelet Galerkin pricing of American options on L´evy driven assets”, SAM Research Report 2003-06, ETH, Zurich, Switzerland. [71] Matache, A.-M., Schwab, C. and Wihler, T.P. (2005), “Fast numerical solution of parabolic integro-differential equations with applications in ﬁnance”, SIAM Journal on Scientiﬁc Computing, to be published. [72] Matache, A.-M., von Petersdorff, T. and Schwab, C. (2004), “Fast deterministic pricing of options on L´evy driven assets”, ESAIM Mathematical Modelling and Numerical Analysis, 38, 37–72. [73] McDonald, R.L. and Schroder, M.D. (1998), “A parity result for American options”, Journal of Computational Finance, 1, 5–13. [74] Merton, R.C. (1973), “Theory of rational option pricing”, Bell Journal of Economics and Management Science, 4, 141–183. [75] Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493. [76] M¨uller, A. and Stoyan, D. (2002), Comparison Methods for Stochastic Models and Risks, Wiley, Chichester, UK. [77] Musiela, M. and Rutkowski, M. (1997), Martingale Methods in Financial Modelling, SpringerVerlag, Berlin, Germany. [78] Nguyen-Ngoc, L. (2003), “Exotic options in general exponential L´evy models”, Pr´epublication No 850, Universit´es Paris 6, Paris, France. [79] Nguyen-Ngoc, L. and Yor, M. (2005), “Lookback and barrier options under general L´evy processes”, in Y. Ait-Sahalia and L.-P. Hansen (Eds), Handbook of Financial Econometrics, North-Holland, Amsterdam, The Netherlands, (to appear). [80] Peskir, G. and Shiryaev, A.N. (2002), “A note on the call-put parity and a call-put duality”, Theory of Probability and its Applications, 46, 167–170. [81] Raible, S. (2000), “L´evy processes in ﬁnance: theory, numerics, and empirical facts”, Ph.D. Thesis, University of Freiburg, Freiburg, Germany. [82] Ribeiro, C. and Webber, N. (2003), “A Monte Carlo method for the normal inverse Gaussian option valuation model using an inverse Gaussian bridge”, Preprint, City University, London, UK. [83] Ribeiro, C. and Webber, N. (2004), “Valuing path-dependent options in the variance-gamma model by Monte Carlo with a gamma bridge”, Journal of Computational Finance, 7, 81–100. [84] Rogers, L.C.G. (2000), “Evaluating ﬁrst-passage probabilities for spectrally one-sided L´evy processes”, Journal of Applied Probability, 37, 1173–1180. [85] Rogers, L.C.G. (2002), “Monte Carlo valuation of American options”, Mathematical Finance, 12, 271–286. [86] Sato, K.-I. (1999), L´evy Processes and Inﬁnitely Divisible Distributions, Cambridge University Press, Cambridge, UK.

128

Exotic Option Pricing and Advanced L´evy Models

[87] Schoutens, W. (2002), “The Meixner process: theory and applications in ﬁnance”, in O.E. Barndorff-Nielsen (Ed.), Mini-Proceedings of the 2nd MaPhySto Conference on L´evy processes, Aarhus, Denmark, pp. 237–241. [88] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [89] Schroder, M. (1999), “Changes of numeraire for pricing futures, forwards and options”, Reviews of Financial Studies, 12, 1143–1163. [90] Sch¨urger, K. (2002), “Laplace transforms and suprema of stochastic processes”, in K. Sandmann and P.J. Sch¨onbucher (Eds), Advances in Finance and Stochastics, Springer-Verlag, Berlin, Germany, pp. 285–294. [91] Vanmaele, M., Dhaene, J., Deelstra, G., Liinev, J. and Goovaerts, M.J. (2005), “Bounds for the price of discrete arithmetic Asian options”, Journal of Computational and Applied Mathematics, to be published. [92] Veˇceˇr, J. and Xu, M. (2004), “Pricing Asian options in a semimartingale model”, Quantitative Finance, 4, 170–175. [93] Wystup, U. (2002), “Vanilla options”, in J. Hakala and U. Wystup (Eds), Foreign Exchange Risk, Risk Publications, London, UK, pp. 3–14.

6 Static Hedging of Asian Options under Stochastic Volatility Models using Fast Fourier Transform ¨ Albrecher Hansjorg Graz University of Technology, Graz, Austria and

Wim Schoutens Katholieke Universiteit Leuven – UCS, Leuven, Belgium Abstract We present a simple static super-hedging strategy for the payoff of an arithmetic Asian option in terms of a portfolio of European options under various stochastic volatility models. Moreover, it is shown that the obtained hedge is optimal in some sense. The strategy is based on stop-loss transforms and comonotonicity theory. The numerical implementation is based on the Fast Fourier transform. We illustrate the hedging performance for several models calibrated to market data and compare the results with other (trivial) static super-hedging strategies.

6.1

INTRODUCTION

The design of efﬁcient hedging strategies for exotic options is a challenging problem that has received a lot of interest during the last few years. In order to serve the needs of investors, increasingly complex ﬁnancial products have been introduced in the market and the pricing and (in particular) hedging of these products is of great importance for assessing the involved risk when trading these instruments. However, many of the hedging techniques currently used in practice rely on market model assumptions that are clearly not sufﬁciently realistic (such as the Black–Scholes model). A common practice in the hedging of exotics is to calibrate the model to vanilla options traded in the market and then derive the corresponding hedging positions for the exotic option. If the model is then recalibrated to the market on the next day, say, then, in order to make the hedging strategy meaningful, the obtained parameter-set should be rather close to the one from the previous day so that only minor adjustments of the hedging portfolio are needed. That is, in addition to a good ﬁt to historical market data, one crucial requirement for a sound market model is its stability in terms of hedging strategies. Empirical studies in that direction indicate that stochastic volatility models outperform classical models like Black-Scholes by far (see e.g. Bakshi et al. (1997) [7]). Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

130

Exotic Option Pricing and Advanced L´evy Models

Apart from that, proposed dynamic hedging strategies with continuously changing positions in the asset (such as delta-hedging) have various deﬁciencies (see e.g. Allen and Padovani (2002) [6]). These are typically based on assumptions like no limit on the frequency of rebalancing, zero transaction costs and full liquidity of the market. However, in practice these assumptions are usually not fulﬁlled and alternatives are asked for. The most favourable situation is the availability of a static hedging strategy for the exotic option, that is an initial hedge portfolio (in terms of the underlying and vanilla options), which will perfectly replicate the payoff at maturity without any portfolio adjustments during the lifetime of the option. For some exotic options (such as barrier, lookback and cliquet options), it is possible to derive semi-static hedging strategies, where portfolio adjustments are only needed at a ﬁnite (and typically small) number of times before maturity (see, for instance, Allen and Padovani (2002) [6] and Carr et al. (1998) [16]). Another alternative is to look for a static super-hedging strategy, which is a portfolio of the underlying and vanilla options that will dominate the payoff of the exotic option without any adjustments during its lifetime. Such a strategy puts a ﬂoor on the maximum loss whatever the subsequent price path will look like and provides a simple way to hedge the product at the expense of a calculable additional cost (namely the difference of the cost of the hedge portfolio and the actual price of the option). At the same time, this strategy enjoys all the advantages of a static hedge: it is less sensitive to the assumption of zero transaction costs (both commissions and the cost of paying individuals to monitor the positions) and does not face the risk of dried-up liquidity when the market makes large moves (opposed to dynamic hedging (see e.g. Carr and Picron (1999) [19] and Carr and Wu (2002) [20]). Semi-static super-hedging strategies for barrier options are discussed in Brown et al. (2001) [14] and Neuberger and Hodges (2002) [32]). This paper focuses on Asian options. Already, the pricing of these products is far from trivial, especially when leaving the Black–Scholes framework (see e.g. Albrecher and Predota (2002, 2004) [5, 4] and Ve˘ce˘r and Xu (2004) [41] or the recent survey by Albrecher (2004) [3]). Moreover, many of the available pricing techniques do not lead to an effective hedging strategy. A delta-hedging strategy for Asian options in a Black–Scholes model based on approximations was discussed in Jacques (1996) [28]. In Albrecher et al. (2003) [2], a simple static super-hedging strategy for arithmetic Asian call options consisting of a portfolio of European options has been derived and optimized using comonotonicity theory. The performance of the resulting strategy has been studied for models for asset price processes following an exponential L´evy model. In the present paper, we extend this approach to stochastic volatility models and investigate the performance of the resulting hedging strategy. As will be illustrated, the hedging error of this simple super-hedging strategy is very small if the option is in the money. For options at and out of the money, this strategy can be quite conservative, but the static nature of the hedge may compensate for parts of the gap. The paper is structured as follows. In Section 6.2, several stochastic volatility models for the asset price process are introduced. In Section 6.3, we present the static super-hedging strategy in detail and illustrate how it can be optimized by comonotonicity techniques. The numerical implementation of the strategy for the various models on the basis of Fast Fourier transforms is discussed in Section 6.4. In Section 6.5, all of the models are calibrated to market data, namely to the same set of vanilla options on the S&P 500, and the performance

Static Hedging of Asian Options

131

of the corresponding hedging strategies is illustrated. Moreover, the issue of model risk is discussed. Since it turns out that the developed hedging strategy only depends on the marginal risk-neutral densities of the asset price process at each averaging day of the Asian option, it can actually be implemented in a completely model-independent setup by estimating the marginal risk-neutral densities directly from the call option surface. This extension is discussed in Section 6.6.

6.2

STOCHASTIC VOLATILITY MODELS

In the sequel, we will brieﬂy introduce various stochastic volatility models, all of which proved their smile-conform pricing abilities, and consider their risk-neutral dynamics. Let S = {St , 0 ≤ t ≤ T } denote the stock price process and φ(u, t) the characteristic function of the random variable log St , i.e. φ(u, t) = E[exp(iu log(St ))]. We assume the stock pays out a continuous dividend; the dividend yield is assumed to be constant and denoted by q. We also have at our disposal a risk-free bank account, paying out a continuously compounded interest rate, which we assume to be constant and denote by r. The price process for the bank-account (bond) is thus given by B = {Bt = exp(rt), t ≥ 0}. The stochastic dynamics of our stock price process will be driven by L´evy processes. A L´evy process X = {Xt , t ≥ 0} is a stochastic process which starts at zero and has independent and stationary increments such that the distribution of the increment is an inﬁnitely divisible distribution (i.e. a distribution for which the characteristic function is also the nth power of another characteristic function, for every integer n). There is a one-to-one correspondence between L´evy processes and inﬁnitely divisible distributions. A subordinator is a nonnegative nondecreasing L´evy process. A general reference for L´evy processes is Bertoin (1996) [12], while for applications in ﬁnance see Schoutens (2003) [38]. 6.2.1 The Heston stochastic volatility model In the Heston Stochastic Volatility model (HEST), the stock price process follows a Black– Scholes stochastic differential equation, in which the volatility behaves stochastically over time: dSt = (r − q)dt + σt dWt , St

S0 ≥ 0,

where the (squared) volatility follows the classical Cox-Ingersoll-Ross (CIR) process: dσt2 = κ(η − σt2 )dt + θ σt dW˜ t ,

σ0 ≥ 0.

Here, W = {Wt , t ≥ 0} and W˜ = {W˜ t , t ≥ 0} are two correlated standard Brownian motions such that Cov[dWt , dW˜ t ] = ρ dt.

132

Exotic Option Pricing and Advanced L´evy Models

The characteristic function φ(u, t) is in this case given by (see Bakshi et al. [7] or Heston (1993) [27]): φ(u, t) = E[exp(iu log(St ))|S0 , σ0 ] = exp(iu(log S0 + (r − q)t)) × exp(ηκθ −2 ((κ − ρθ ui − d)t − 2 log((1 − ge−dt )/(1 − g)))) × exp(σ02 θ −2 (κ − ρθ iu − d)(1 − e−dt )/(1 − ge−dt )), where d = ((ρθ ui − κ)2 − θ 2 (−iu − u2 ))1/2 ,

(6.1)

g = (κ − ρθ ui − d)/(κ − ρθ ui + d).

(6.2)

An extension of HEST introduces jumps in the asset price (Bakshi et al. [7]), while other extensions also allow for jumps in the volatility (see e.g. Knudsen and Nguyen-Ngoc (2003) [29]). Since for these extensions the characteristic function of the log stock price is also available, one can straightforwardly apply the methods described below for these models too. 6.2.2 The Barndorff-Nielsen–Shephard model This class of models, denoted by BN–S, was introduced in Barndorff-Nielsen and Shephard (2000) [10] and has a structure similar to the Heston model. The difference is basically that here the volatility is modelled by an Ornstein–Uhlenbeck (OU) process driven by a subordinator. In this way, jumps are introduced into the volatility process. Volatility can only jump upwards and then will decay exponentially. A co-movement effect between up-jumps in volatility and (down)-jumps in the stock price is also incorporated. The squared volatility now follows a SDE of the form: dσt2 = −λσt2 dt + dzλt ,

(6.3)

where λ > 0 and z = {zt , t ≥ 0} is a subordinator. The risk-neutral dynamics of the log-price Zt = log St are given by dZt = (r − q − λk(−ρ) − σt2 /2)dt + σt dWt + ρdzλt ,

Z0 = log S0 ,

where W = {Wt , t ≥ 0} is a Brownian motion independent of z = {zt , t ≥ 0} and where k(u) = log E[exp(−uz1 )] is the cumulant function of z1 . Note that the parameter ρ introduces the co-movement effect between the volatility and the asset price process. We use the classical and tractable example of the Gamma-OU process (other choices for OU-processes include the inverse Gaussian-OU process, which also leads to a tractable model [Schoutens (2003) [38], Section 7.2.1]). For a Gamma-OU process, z = {zt , t ≥ 0} is a compound-Poisson process: zt =

Nt n=1

xn ,

(6.4)

Static Hedging of Asian Options

133

where N = {Nt , t ≥ 0} is a Poisson process with intensity parameter a, i.e. E[Nt ] = at and {xn , n = 1, 2, . . . } is an independent and identically distributed sequence of exponential random variables with mean 1/b. One then has log E[exp(−uz1 )] = −au(b + u)−1 , and it can be shown that σ 2 has a stationary marginal law that follows a Gamma distribution. The characteristic function of the log price can, in this case, be written in the form (cf. Barndorff-Nielson et al. (2002) [11]) φ(u, t) = E[exp(iu log St )|S0 , σ0 ] = exp iu(log(S0 ) + (r − q − aλρ(b − ρ)−1 )t) × exp −λ−1 (u2 + iu)(1 − exp(−λt))σ02 /2

b − f1 + f2 λt , × exp a(b − f2 )−1 b log b − iuρ where f1 = f1 (u) = iuρ − λ−1 (u2 + iu)(1 − exp(−λt))/2, f2 = f2 (u) = iuρ − λ−1 (u2 + iu)/2. 6.2.3 L´evy models with stochastic time Another way to incorporate stochastic volatility effects into the price process is by making time stochastic. Periods with high volatility can be interpreted as if time runs faster than in periods with low volatility. Applications of stochastic time change to asset pricing go back to Mandelbrot and Taylor (1967) [31] (see also Clark (1973) [21]). We consider the models introduced by Carr et al. (2003) [18]. The L´evy models with stochastic time considered in this paper are built out of two independent stochastic processes. The ﬁrst process is a L´evy process. The behavior of the asset price will then be modelled by the exponential of the L´evy process, suitably timechanged. Typical examples for the generator of the L´evy process are the normal distribution (leading to Brownian motion), the Normal Inverse Gaussian (NIG) distribution (BarndorffNielsen (1995) [8] and Rydberg (1997) [34], the Variance Gamma (VG) distribution (Madan et al. (1998) [30]), the generalized hyperbolic distribution (Eberlein (1999) [25] and Rydberg (1999) [35], the Meixner distribution (Grigelionis (1999) [26], Schoutens and Teugels (1998) [36] and Schoutens (2002) [37]) and the CGMY distribution (Carr et al. (2002) [17]) (see Schoutens (2003) [38] for an overview). We opt to work with the VG and NIG processes for which simulation issues become quite standard. The second process is a stochastic clock that builds in a stochastic volatility effect. The above mentioned (ﬁrst) L´evy process will be subordinated (i.e. time-changed) by this stochastic clock. By deﬁnition of a subordinator, the time needs to increase and the process modelling the rate of time change y = {yt , t ≥ 0} also needs to be positive. The economic time elapsed in t units of calendar time is then given by the integrated process Y = {Yt , t ≥ 0} with t ys ds. (6.5) Yt = 0

134

Exotic Option Pricing and Advanced L´evy Models

Since y is a positive process, Y is an increasing process. We will consider two processes for the rate of time change y: the CIR process (which is continuous) and the Gamma-OU process (which is a jump process). We will ﬁrst discuss NIG and VG and subsequently introduce the stochastic clocks CIR and Gamma-OU. 6.2.3.1 NIG and VG processes The NIG(α, β, δ) distribution with parameters α > 0, |β| < α and δ > 0 has a characteristic function given by φNI G (u; α, β, δ) = exp −δ α 2 − (β + iu)2 − α 2 − β 2 and the VG(C, G, M) distribution with parameters C > 0, G > 0 and M > 0 has a characteristic function given by φV G (u; C, G, M) =

GM GM + (M − G)iu + u2

C .

Since both distributions are inﬁnitely divisible, each of them generates a L´evy process X = {Xt , t ≥ 0} where the increment X1 follows a NIG(α, β, δ) law (VG(C, G, M) law, respectively). The resulting process is called a NIG process (VG process, respectively). Due to convolution properties of these two distributions, increments of arbitrary length again follow the same law with just a change in parameters: An increment of the NIGprocess over the time interval [s, s + t] follows a NIG(α, β, δt) law and the increment of a VG-process over [s, s + t] is VG(Ct, G, M)-distributed (see also Barndorff-Nielsen (1997) [9]). 6.2.3.2 Stochastic clocks CIR Stochastic Clock:. Carr et al. (2003) [18] use as the rate of time change the CIR process that solves the SDE: 1/2

dyt = κ(η − yt )dt + λyt dWt , where W = {Wt , t ≥ 0} is a standard Brownian motion. The characteristic function of Yt (given y0 ) is explicitly known (see Cox et al. (1985) [22]): ϕCI R (u, t; κ, η, λ, y0 ) = E[exp(iuYt )|y0 ] =

exp(κ 2 ηt/λ2 ) exp(2y0 iu/(κ + γ coth(γ t/2))) (cosh(γ t/2) + κ sinh(γ t/2)/γ )2κη/λ2

where γ =

κ 2 − 2λ2 iu.

,

Static Hedging of Asian Options

Gamma-OU Stochastic Clock:. of the SDE:

135

Another choice for the rate of time change is the solution

dyt = −λyt dt + dzλt ,

(6.6)

where the process z = {zt , t ≥ 0} is, as in equation (6.4), a compound Poisson process. In the Gamma-OU case, there is an explicit expression for the characteristic function of Yt (given y0 ): ϕ −OU (u; t, λ, a, b, y0 ) = E[exp(iuYt )|y0 ] = exp iuy0 λ−1 (1 − e−λt ) +

b λa b log − iut . iu − λb b − iuλ−1 (1 − e−λt )

6.2.3.3 Time-changed L´evy process Let Y = {Yt , t ≥ 0} as deﬁned in equation (6.5), be the process modelling our business time. The (risk-neutral) price process S = {St , t ≥ 0} is now modelled as follows: St = S0

exp((r − q)t) exp(XYt ), E[exp(XYt )|y0 ]

(6.7)

where X = {Xt , t ≥ 0} is a L´evy process. The factor exp((r − q)t)/E[exp(XYt )|y0 ] puts us immediately into the risk-neutral world by a mean-correcting argument. Essentially, the stock price process is modelled as the ordinary exponential of a time-changed L´evy process. The process incorporates jumps (through the L´evy process Xt ) and stochastic volatility (through the time change Yt ). The characteristic function φ(u, t) for the logarithm of our stock price is given by: φ(u, t) = E[exp(iu log(St ))|S0 , y0 ] = exp(iu((r − q)t + log S0 ))

ϕ(−iψX (u); t, y0 ) , ϕ(−iψX (−i); t, y0 )iu

(6.8)

where ψX (u) = log E[exp(iuX1 )] and ϕ(u; t, y0 ) denotes the characteristic function of Yt given y0 . Since we consider two L´evy processes (VG and NIG) and two stochastic clocks (CIR and Gamma-OU), we will ﬁnally end up with four resulting models abbreviated as VG-CIR, VG-OU , NIG-CIR and NIG-OU . Due to (time-)scaling effects, one can without loss of generality scale the present rate of time change to 1 (y0 = 1). For more details, see Carr et al. (2003) [18] or Schoutens (2003) [38].

136

Exotic Option Pricing and Advanced L´evy Models

6.3

STATIC HEDGING OF ASIAN OPTIONS

Consider now a European-style arithmetic average call option with strike price K, maturity T and n averaging days 0 ≤ t1 < · · · < tn ≤ T . Then, its price according to a risk-neutral pricing measure Q at time t is given by

+ % n & exp(−r(T − t)) AAt = Stk − nK EQ Ft , n k=1

where {Ft , 0 ≤ t ≤ T } denotes the natural ﬁltration of S. In general, the distribution of the dependent sum nk=1 Stk is not available, which makes pricing and hedging of these products difﬁcult. However, for our super-hedging purposes it sufﬁces to look for an upper bound of the above payoff. Assume for simplicity that t = 0 and that the averaging has not yet started. First note, that for any K1 , . . . , Kn ≥ 0 with K = nk=1 Kk , we have a.s.

+ n n + + Stk − nK = (St1 − nK1 ) + · · · + (Stn − nKn ) ≤ Stk − nKk . k=1

k=1

Hence exp(−rT ) EQ AA0 (K, T ) = n ≤

% n

+ Stk − nK

F0

&

k=1

n * + + exp(−rT ) EQ Stk − nKk F0 n k=1

exp(−rT ) exp(rtk )EC0 (κk , tk ), n n

=

(6.9)

k=1

where EC0 (κk , tk ) denotes the price of a European call option at time 0 with strike κk = nKk and maturity tk . In terms of hedging, this means that we have the following static super-hedging strategy: for each averaging day tk , buy exp(−r(T − tk ))/n European call options at time t = 0 with strike κk and maturity tk and hold these until their expiry. Then put their payoff on the bank account. the upper bound (equation (6.9)) holds for all combinations of κk ≥ 0 that satisfy Since n κ k=1 k = nK, we still have the freedom to choose strike values that ﬁt best to our purposes. The simplest choice is κk = K (k = 1, . . ., n). If q ≤ r, we have EC0 (K, t) ≤ EC0 (K, T ) for every K ≥ 0 and 0 ≤ t ≤ T , and thus this trivial choice shows that the Asian option price is dominated by the price of a European option with the same strike and maturity, i.e. AA0 (K, T ) ≤ EC0 (K, T ) (this trivial hedging strategy of an Asian option in terms of the corresponding European option was already observed in Nielsen and Sandmann (2003) [33]). However, for our super-hedging purposes, we naturally look for that combination of κk s which minimizes the right-hand side of equation (6.9). In the Black–Scholes setting, this optimization problem

Static Hedging of Asian Options

137

was solved in Nielsen and Sandmann (2003) [33] by using Lagrange multipliers. In the general case of arbitrary arbitrage-free market models, this optimal combination can be determined by using stop-loss transforms and the theory of comonotonic risks (for a general introduction to comonotonicity techniques, see Dhaene et al. (2002a, b) [23]). Let F (x) be a distribution function of a non-negative random variable X; then its stop-loss transform F (m) is deﬁned by +∞ F (m) = (x − m)dF (x) = E[(X − m)+ ], m ≥ 0. m

If we write An =

n

Stk

k=1

and FAt n (x) = PQ (An ≤ x|Ft ) for the distribution function under Q of An given the information Ft , then we have AAt =

exp(−r(T − t)) F t (nK). An n

(6.10)

In this way the problem of pricing an arithmetic average option is transformed to calculating the stop-loss transform of a sum of dependent risks. Concretely, we will look at bounds for stop-loss transforms based on comonotonic risks. A positive random vector (X1 , . . . , Xn ) with marginal distribution functions F1 (x1 ), . . . , Fn (xn ) is called comonotone, if for the joint distribution function FX1 ,...,Xn (x1 , . . . , xn ) = min{F1 (x1 ), . . . , Fn (xn )} holds for every x1 , . . . , xn ≥ 0. It immediately follows that the distribution of a comonotone random vector (X1 , . . . , Xn ) with given marginal distributions F1 (x1 ), . . . , Fn (xn ) is uniquely determined. In Simon et al. (2000) [40], it was shown that an upper bound for the stop-loss transform of the sum of arbitrary dependent positive random variables nk=1 Xk with marginal distributions F1 (x1 ), . . . , Fn (xn ) is given by the stop-loss transform of the sum S c = nk=1 Yk , where (Y1 , . . . , Yk ) is the comonotone random vector with marginal distributions F1 (x1 ), . . . , Fn (xn ). Let FS c (x) denote the distribution function of nk=1 Yk ; then it follows from general comonotonicity results (see e.g. Dhaene et al. (2002a, b) [23]) that its inverse is given by FS−1 c (x)

=

n

FX−1 (x), k

x ≥ 0.

(6.11)

k=1

The crucial result for our purposes is now that the stop-loss transform of a sum of comonotonic random variables can be obtained as a sum of the stop-loss transforms of the marginals evaluated at speciﬁed points (cf. Proposition 2 in Simon et al. (2000) [40]). More precisely, FS c (m) =

n

c (m)) , FXk FX−1 (F S k

m ≥ 0,

(6.12)

k=1

given that the marginal distribution functions involved are strictly increasing (which is always the case in our applications). At the same time, n

+ n n ≤ Yk − m E((Yk − mk )+ ) = FXk (mk ) (6.13) FS c (m) = E k=1

k=1

k=1

138

Exotic Option Pricing and Advanced L´evy Models

whenever nk=1 mk = m. Thus, the stop-loss transform of the comonotonic sum given by equation (6.12) represents the lowest possible bound in terms of a sum of stop-loss transforms of the marginal distributions. This fact immediately translates to our setting of an arithmetic Asian option. Let F (xk ; tk ) (k = 1, . . . , n) denote the conditional distribution of Stk under the risk-neutral measure Q (given the information available at time t = 0), i.e. for xk , tk > 0, F (xk ; tk ) = PQ Stk ≤ xk |F0 .

(6.14)

Combining equations (6.9), (6.10), (6.12) and (6.13), the optimal combination of strike prices κk is given by κk = F −1 (FS c (n K); tk ) ,

k = 1, . . . , n.

(6.15)

In this way, we have obtained the optimal static super-hedge in terms of European call options with maturity dates equal to the averaging dates. For the practical determination of the strike prices κk , the distribution function of the comonotone sum FS c (x), as given by equation (6.11) has to be calculated and evaluated at nK. For this purpose, we need to approximate the risk-neutral marginal densities of the stock price at the averaging dates, which can be carried out efﬁciently by using Fast Fourier transforms (cf. Section 6.4.1 below). The κk s are then obtained by evaluating the inverse distribution function of F (x; tk ).

6.4 NUMERICAL IMPLEMENTATION 6.4.1 Characteristic function inversion using FFT For all of the above mentioned models, we have the characteristic function of the log-price process at our disposal. However, in order to determine the strike prices of our optimal hedge portfolio as described in Section 6.3, we need the corresponding density functions. Recall that the characteristic function, φ(u), is the Fourier-transform of the corresponding density function f (x): φ(u) =

+∞ −∞

exp(iux)f (x)dx.

So, we need to apply an inverse Fourier-transformation. Next, we illustrate how this can be done fast and accurately by using the Fast Fourier Transform (FFT). The latter is an efﬁcient algorithm for computing the following transformation of a vector (αk , k = 1, . . . , m) into a vector (βk , k = 1, . . . , m): βk =

m

exp(−i2π(j − 1)(k − 1)/N )αj .

j =1

Typically, m is a power of 2. The number of multiplications of the FFT algorithm is of the order O(m log m) and this is in contrast to the straightforward evaluation of the above sums which give rise to O(m2 ) multiplications.

Static Hedging of Asian Options

139

We follow closely a technique described in Carr and Madan (1998) [15] in the context of option pricing. The classical inverse Fourier transform reads: +∞ 1 exp(−iux)φ(u)du. f (x) = 2π −∞ Since f is real, we can write 1 f (x) = π

+∞

exp(−iux)φ(u)du. 0

Next, we are going to discretize the above integral and apply the trapezoid rule. We take a grid on the real line with grid-length u > 0: uj = (j − 1)u,

j = 1, . . . , N.

One approximately has N−1 1 u 1 exp(−iuj x)φ(uj ) + exp(−iuN x)φ(uN ) f (x) ≈ exp(−iu1 x)φ(u1 ) + π 2 2 j =2

u wj exp(−ix(j − 1)u)φ((j − 1)u) π N

=

j =1

where the weights wj are given by 1 1 , w2 = 1, w3 = 1, . . . , wN−1 = 1, wN = . 2 2 We will calculate the value of the density function f in the points w1 =

xk = −b + x(k − 1),

k = 1, . . . , N

where x = 2b/(N − 1), thus covering the interval [−b, b] with an equally spaced grid. In these points we have u wj exp(i(j − 1)bu) exp(−i(j − 1)(k − 1)ux)φ((j − 1)u). π N

f (xk ) ≈

j =1

If we choose the grid sizes such that ux =

2π , N

then u wj exp(i(j − 1)ub) exp(−i(j − 1)(k − 1)2π/N )φ((j − 1)u). π N

f (xk ) ≈

j =1

This sum can be easily computed by FFT: (f (xk ), k = 1, . . . , N ) is the FFT of the vector wj φ(uj ) exp(iuj b), j = 1, . . . , N . Choosing N as a power of 2 allows very fast computation of the FFT.

140

Exotic Option Pricing and Advanced L´evy Models

6.4.2 Static hedging algorithm In order to set up our hedge portfolio, we have to determine the inverse distribution function of the asset price at each averaging day (cf. equation (6.14)). This is carried out by numerically building up the distribution function from the approximated density function obtained in Section 6.4.1. The inverse is then determined by a bisection method from the corresponding table and linear interpolation between grid points is employed. In our implementation, we used 214 points in the grid for both the densities and the inverse distribution functions, which turns out to be sufﬁcient (in the sense that a further increase does not change the signiﬁcant digits of the results). Next, the inverse of the distribution of the comonotone sum is built up according to equation (6.11) and then itself inverted in the above way. Finally, the strike prices κk of the European options are obtained by evaluating the inverse distribution functions of the marginals according to equation (6.15). This numerical procedure to obtain the strike prices for our hedging strategy is both accurate and very quick (the determination of the entire hedge portfolio takes less than a minute on a normal PC for each of the discussed stochastic volatility models).

6.5 NUMERICAL ILLUSTRATION We give numerical results for an arithmetic Asian call option with a maturity of 1 year and averaging every month (i.e. 12 averaging days). First, the model parameters have to be determined from the market prices of vanilla options. 6.5.1 Calibration of the model parameters Carr and Madan (1998) [15] developed pricing methods for the classical vanilla options which can be applied whenever the characteristic function of the risk-neutral stock price process is known. Using Fast Fourier transforms, one can compute within a second the complete option surface on an ordinary computer. In Schoutens (2003) [38], this method was used to calibrate the models (minimizing the difference between market prices and model prices in a least-squares sense) on a dataset of 77 option on the S&P 500 Index [Schoutens (2003) [38], Appendix C]. The results of the calibration are visualized in Figures 6.1 and 6.2 for the NIG-CIR and the Heston model, respectively. Here, the circles are the market prices and the plus signs are the model prices (calculated through the Carr–Madan formula by using the respective characteristic functions and obtained parameters). For details of the ﬁt, see Schoutens (2003) [38]. The Heston model, which is not covered in Schoutens (2003) [38], gives rise to the following calibration errors: ape = 1.31%,

rmse = 1.0530,

aae = 0.8095,

arpe = 1.90%.

Table 6.1 depicts the calibrated parameters for each of the six discussed stochastic volatility models, while Figure 6.3 shows the corresponding marginal density functions of log(St ) for t ranging from 1 month up to 1 year for all six models obtained by Fast Fourier transform, as described in Section 6.4.1.

6.5.2 Performance of the hedging strategy After the strike prices of the hedge portfolio are determined according to equation (6.15), the price of the hedging strategy is easily determined by using the Carr–Madan call option

Static Hedging of Asian Options

200 180 160 140 120 100 80 60 40 20 0 900

1000

1100

1200

1300

1400

1500

1400

1500

Figure 6.1 Calibration of the NIG-CIR Model

200 180 160 140 120 100 80 60 40 20 0 900

1000

1100

1200

1300

Figure 6.2 Calibration of Heston’s Model

141

142

Exotic Option Pricing and Advanced L´evy Models

14 1 2 3 4 5 6 7 8 9 10 11 12

(a) 12 10 1 8 2 6

3 4 5

4

7

6

8

14

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

10 8

2 3

4

4 5 6

9 10

2

8 9 10 11

7 11

2

12

12

0

0 4

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8 1 2 3 4 5 6 7 8 9 10 11 12

(c)

16 14 12 10

1

8

2

6

4.9

5

7

6

4.2

4.3

4.4

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

4.6

4.7

(e) 12 10 1 8 2 3

4

4 5 6

4.4

4.5

7

10

4.9 T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

8

14 12 10

1

5

6

2

0

2 3 4 5

4.9

4

4.1

4.2

4.3

4.4

4.5

1 2 3 4 5 6 7 8 9 10 11 12

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

9

10

4.6

4.7

(f) 12 10

1

6

2

4

3 4 5

12

4.8

4.9 T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

7

5

8 9 10 11

2 12

0

11

1 2 3 4 5 6 7 8 9 10 11 12

14

8

5

8

6 11

4.8

7

2

9 10

4.7

16

11 12

4.8

4.6

18

4

1 2 3 4 5 6 7 8 9 10 11 12

14

6

4.3

(d)

20

8 9

4.5

4.2

6

2 4.1

4.1

8

4 5

4

4 22

3

4

0

1

6

T=1m T=2m T=3m T=4m T=5m T=6m T=7m T=8m T=9m T = 10 m T = 11 m T = 12 m

1 2 3 4 5 6 7 8 9 10 11 12

(b) 12

12

0 4

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5

4

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5

Figure 6.3 Marginal density functions for the various stochastic volatility models: (a) NIG-OU ; (b) NIG-CIR; (c) VG-OU ; (d) NIG-CIR; (e) BN–S; (f) Heston

pricing formula for European options and equation (6.9). Tables 6.2–6.7 compare the Monte Carlo simulated price of the Asian option AAMC and the comonotonic super-hedge price AAc , with the prices of two trivial super-hedging strategies, namely the trivial super-hedge using the European option price EC with identical strike and maturity (note that q ≤ r) and the super-hedge equation (6.9) with all κi = K with price AAtr . The strike price is given as a percentage of the spot. For the Monte Carlo price, we used 1 million sample paths. The VG process was simulated as a difference of two Gamma processes (cf. Schoutens (2003)

Static Hedging of Asian Options Table 6.1 Risk-neutral parameters obtained by calibration to vanilla calls on S&P 500 HEST σ02 = 0.0224, κ = 0.5144, η = 0.1094, θ = 0.3354, ρ = −0.7392 BN–S ρ = −1.2606, λ = 0.5783, b = 11.6641, a = 1.4338, σ02 = 0.0145 VG-CIR C = 11.9896, G = 25.8523, M = 35.5344, κ = 0.6020, η = 1.5560, λ = 1.9992, y0 = 1 VG-OU C = 11.4838, G = 23.2880, M = 40.1291, λ = 1.2517, a = 0.5841, b = 0.6282, y0 = 1 NIG-CIR α = 18.4815, β = −4.8412, δ = 0.4685, κ = 0.5391, η = 1.5746, λ = 1.8772, y0 = 1 NIG-OU α = 29.4722, β = −15.9048, δ = 0.5071, λ = 0.6252, a = 0.4239, b = 0.5962, y0 = 1

Table 6.2 Hedging performance in the BN–S model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.6065 11.7478 4.5265 0.9431 0.1385

20.9648 12.3153 5.2411 1.4128 0.2972

21.1889 12.4876 5.2415 1.6417 0.5002

22.8511 14.9462 8.3470 3.8643 1.5736

Table 6.3 Hedging performance in Heston’s model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.2896 11.3823 4.3056 0.6939 0.0368

20.5088 11.8872 5.0132 1.1328 0.1193

20.7022 12.0223 5.0137 1.3568 0.2807

22.0898 14.1997 7.7280 3.2476 0.9834

Table 6.4 Hedging performance in the NIG-OU model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.3713 11.4467 4.4063 0.8751 0.0738

20.6307 11.8830 4.9562 1.2170 0.1566

20.7753 11.9975 4.9566 1.4321 0.3277

22.2822 14.1826 7.6203 3.2497 1.0465

143

144

Exotic Option Pricing and Advanced L´evy Models Table 6.5 Hedging performance in the NIG-CIR model 100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.2817 11.4069 4.4121 0.9102 0.1506

20.4979 11.8418 4.9588 1.2704 0.2864

20.6808 11.9845 4.9598 1.4781 0.4152

22.0975 14.1909 7.6878 3.2162 1.0910

Table 6.6 Hedging performance in the VG-OU model 100K/S0 80 90 100 110 120

Table 6.7

AAMC

AAc

AAtr

20.3528 11.4380 4.4083 0.9070 0.1061

20.5773 11.8695 4.9561 1.2391 0.1988

20.7447 11.9896 4.9567 1.4559 0.3506

EC 22.2073 14.1938 7.6454 3.2408 1.0433

Hedging performance in the VG-CIR model

100K/S0

AAMC

AAc

AAtr

EC

80 90 100 110 120

20.3256 11.4374 4.4383 0.9294 0.1615

20.4907 11.8395 4.9605 1.2723 0.2883

20.6766 11.9758 4.9613 1.4793 0.4152

22.1156 14.2022 7.6906 3.2159 1.0898

[38], Section 8.4.2) while the NIG paths were obtained as described in Schoutens (2003) [38] (Section 8.4.5). From Tables 6.2–6.7, we observe that the more in the money the Asian option is, the less is the difference between the option price and the comonotonic hedge. For an option with moneyness of 80% the difference is typically around 1.5%, whereas the classical hedge with the European call leads to a difference of almost 10%. For options out of the money, the difference increases, but is then substantially smaller than the differences for the other two trivial hedges. In view of the easy and cheap way in which this hedge can be implemented in practice, this static super-hedge approach seems to be competitive also in these cases. As a by-product, we observe from the Monte Carlo estimates in Tables 6.2–6.7 that the model risk for Asian option prices can be quite substantial (note that all of the models are calibrated to the same set of vanilla option prices with a quite acceptable ﬁt (the average percentage error of the ﬁt is less than 2% for all the models (cf. Schoutens (2003) [38])), but the resulting marginal densities differ considerably (cf. Figure 6.3) and consequently the Asian option prices can differ quite a lot, especially if the option is out of the money). The issue of model risk for other exotic options has recently been discussed in Schoutens et al. (2004) [39].

Static Hedging of Asian Options

6.6

145

A MODEL-INDEPENDENT STATIC SUPER-HEDGE

Since the hedging strategy introduced in this paper only depends on the risk-neutral marginal distribution functions on each averaging day of the Asian option, it can also be applied in a model-independent framework, if for all of the needed maturities tk the European call prices C(K, tk ) are available for every strike value K. In this case, the risk-neutral density function fStk is given by the second derivative of C(K, tk ) with respect to K (see e.g. Breeden and Litzenberger (1978) [13]): fStk (K) = er tk

∂ 2 C(K, tk ) . ∂K 2

In practice, call prices are available for a limited number of strike values K only, so that one has to use sophisticated statistical techniques to estimate f (Stk ). For a recently developed efﬁcient nonparametric estimation procedure utilizing shape restrictions due to no-arbitrage (such as monotonicity and convexity of the call price as a function of the strike), we refer to A¨it-Sahalia and Duarte (2003) [1]. Once the density f (Stk ) is available for all of the needed maturities tk , the hedge portfolio can be determined in just the same way as described in the above sections.

6.7 CONCLUSIONS We have shown that staticly hedging an Asian option in terms of a portfolio of European options is a simple and quick alternative to other strategies. Moreover, in contrast to most of the existing techniques, this approach is applicable in general market models whenever the risk-neutral density of the asset price distribution or an approximation of it is available. In particular, there is a fast algorithm to determine the hedge portfolio for various stochastic volatility models. Since the proposed hedging strategy is static, it is much less sensitive to the assumption of zero transaction costs and to the hedging performance in the presence of large market movements; no dynamic rebalancing is required. These advantages may sometimes compensate for the gap of the hedging price and the option price even for Asian options that are out of the money.

ACKNOWLEDGEMENTS The authors would like to thank Peter Carr for fruitful discussions on the topic. H. Albrecher acknowledges support from the K.U. Leuven (Fellowship F/04/009) and the Austrian Science Foundation Project S-8308-MAT. W. Schoutens is a Postdoctoral Fellow of the Fund for Scientiﬁc Research, Flanders, Belgium (FWO – Vlaanderen).

REFERENCES [1] A¨it-Sahalia, Y. and Duarte, J. (2003), “Nonparametric option pricing under shape restrictions”, Journal of Econometrics, 116, 9–47. [2] Albrecher, H., Dhaene, J., Goovaerts, M. and Schoutens, W. (2005), “Static hedging of Asian options under L´evy models”, Journal of Derivatives, 12(3), 63–72. [3] Albrecher, H. (2004), “The valuation of Asian options for market models of exponential L´evy type”, in M. Vanmaele, et al. (Eds), Proceedings of the Second Day of Actuarial and Financial Mathematics, Royal Flemish Academy of Arts and Sciences, Brussels, Belgium, pp. 11–20.

146

Exotic Option Pricing and Advanced L´evy Models

[4] Albrecher, H. and Predota, M. (2004), “On Asian option pricing for NIG L´evy processes”, Journal of Computational and Applied Mathematics, 172, 153–168. [5] Albrecher, H. and Predota, M. (2002), “Bounds and approximations for discrete Asian options in a variance-gamma model”, Grazer Mathematische Berichte, 345, 35–57. [6] Allen, S. and Padovani, O. (2002), “Risk management using quasi-static hedging”, Economic Notes, 31, 277–336. [7] Bakshi, G., Cao, C. and Chen, Z. (1997), “Empirical performance of alternative option pricing models”, The Journal of Finance, LII, 2003–2049. [8] Barndorff-Nielsen, O.E. (1995), “Normal inverse Gaussian distributions and the modeling of stock returns”, Research Report No. 300, Department of Theoretical Statistics, Aarhus University, Denmark. [9] Barndorff-Nielsen, O.E. (1997), “Normal inverse Gaussian distributions and stochastic volatility models”, Scandinavian Journal of Statistics, 24, 1–13. [10] Barndorff-Nielsen, O.E. and Shephard, N. (2000), “Modelling by L´evy processes for ﬁnancial econometrics”, in O.E. Barndorff-Nielsen, T. Mikosch and S. Resnick (Eds), L´evy Processes – Theory and Applications, Birkh¨auser, Boston, MA, USA, pp. 283–318. [11] Barndorff-Nielsen, O.E., Nicolata, E. and Shephard, N. (2002), “Some recent developments in stochastic volatility modelling”, Quantitative Finance, 2, 11–23. [12] Bertoin, J. (1996), L´evy Processes, Cambridge Tracts in Mathematics, Vol. 121, Cambridge University Press, Cambridge, UK. [13] Breeden, D. and Litzenberger, R. (1978), “Prices of state-contingent claims implicit in option prices”, Journal of Business, 51, 621–651. [14] Brown, H., Hobson, D. and Rogers, L.C.G. (2001), “Robust hedging of barrier options”, Mathematical Finance, 11, 285–314. [15] Carr, P. and Madan, D. (1998), “Option valuation using the fast Fourier transform”, Journal of Computational Finance, 2, 61–73. [16] Carr, P., Ellis, K. and Gupta, V. (1998), “Static hedging of exotic options”, The Journal of Finance, 53, 1165–1190. [17] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2002), “The ﬁne structure of asset returns: an empirical investigation”, Journal of Business, 75, 305–332. [18] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [19] Carr, P. and Picron J. (1999), “Static hedging of timing risk”, Journal of Derivatives, 6, 57–70. [20] Carr, P. and Wu, L. (2002), “Static hedging of standard options”, Preprint. [21] Clark, P. (1973), “A subordinated stochastic process model with ﬁnite variance for speculative prices”, Econometrica, 41, 135–156. [22] Cox, J., Ingersoll, J. and Ross, S. (1985), “A theory of the term structure of interest rates”, Econometrica, 53, 385–408. [23] Dhaene, J., Denuit, M., Goovaerts, M.J., Kaas, R. and Vyncke, D. (2002a), “The concept of comonotonicity in actuarial science and ﬁnance: theory”, Insurance: Mathematics and Economics, 31, 3–33. [24] Dhaene, J., Denuit, M., Goovaerts, M.J., Kaas, R. and Vyncke D. (2002b), “The concept of comonotonicity in actuarial science and ﬁnance: applications”, Insurance: Mathematics and Economics, 31, 133–161. [25] Eberlein, E. (1999), “Application of generalized hyperbolic L´evy motions to ﬁnance”, in O.E. Barndorff-Nielsen T. Mikosch and S. Resnick (Eds), L´evy Processes: Theory and Applications, Birkh¨auser, Boston, MA, USA, pp. 319–337.

Static Hedging of Asian Options

147

[26] Grigelionis, B. (1999), “Processes of Meixner type”, Lithuanian Mathematics Journal, 39, 33–41. [27] Heston, S. (1993), “A closed-form solution for options with stochastic volatility with applications to bond and currency options”, Review of Financial Studies, 6, 327–343. [28] Jacques, M. (1996), “On the hedging portfolio of Asian options”, ASTIN Bulletin, 26, 165–183. [29] Knudsen, Th. and Nguyen-Ngoc, L. (2003), “Pricing European options in a stochastic volatilityjump-diffusion model, DBQuant Working Paper, Deutsche Bank, London, UK; Journal of Financial and Quantitative Analysis, to be published. [30] Madan, D.B., Carr, P. and Chang, E.C. (1998), “The variance gamma process and option pricing”, European Finance Review, 2, 79–105. [31] Mandelbrot, B.B. and Taylor, H.M. (1967), “On the distribution of stock price differences”, Operations Research, 15, 1057–1062. [32] Neuberger, A. and Hodges, S. (2002), “Rational bounds on the prices of exotic options”, IFA Working Paper 359, London Business School, London, UK. [33] Nielsen, J.A. and Sandmann, K. (2003), “Pricing bounds on Asian options”, Journal of Financial and Quantitative Analysis, 38, 449–473. [34] Rydberg, T. (1997), “The normal inverse Gaussian L´evy process: simulation and approximation”, Communications in Statistics: Stochastic Models, 13, 887–910. [35] Rydberg, T. (1999), “Generalized hyperbolic diffusions with applications in ﬁnance”, Mathematical Finance, 9, 183–201. [36] Schoutens, W. and Teugels, J.L. (1998), “L´evy processes, polynomials and martingales”, Communications in Statistics: Stochastic Models, 14, 335–349. [37] Schoutens, W. (2002), “The Meixner process: theory and applications in ﬁnance”, EURANDOM Report 2002-004, EURANDOM, Eindhoven, The Netherlands. [38] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [39] Schoutens, W., Simons, E. and Tistaert, J. (2004), “A perfect calibration! Now what?”, Wilmott Magazine, March, 66–78. [40] Simon S., Goovaerts, M. and Dhaene, J. (2000), “An easy computable upper bound for the price of an arithmetic Asian option”, Insurance: Mathematics and Economics, 26, 175–183. [41] Ve˘ce˘r, J. and Xu, M. (2004), “Pricing Asian options in a semimartingale model”, Quantitative Finance, 4, 170–175.

7 Impact of Market Crises on Real Options Pauline Barrieu London School of Economics, London, UK and

Nadine Bellamy ´ Universite d’Evry Val d’Essonne, Evry, France Abstract We study the impact of market crises on investment decisions through real option theory. The framework we consider involves a Brownian motion and a Poisson process, with the jumps characterizing the crisis effects. We ﬁrst analyze the consequences of different modelling choices. We then provide the real option characteristics and establish the existence of an optimal discount rate. We also characterize the optimal time to invest and derive some properties of its Laplace Transform (bounds, monotonicity, robustness). Finally, we specify the consequences of some wrong model speciﬁcations on the investment decision.

7.1

INTRODUCTION

Investment has always been a crucial question for ﬁrms. Should a given project be undertaken? In addition, if so, when is it the best time to invest? In order to answer these questions, the neo-classical criterion of Net Present Value (N.P.V.) is still widely used. It consists in investing, if and only if, the sum of the project discounted beneﬁts is higher than the sum of its discounted costs. Such a criterion does, however, have several weaknesses. Among many others, the following facts are often mentioned: • The N.P.V. method does not take into account potential uncertainty of future cash ﬂows. • It uses an explicit calculation for the cost of the risk. • It focuses on present time: the investment decision can only be taken now or never. However, reality is often more complex and ﬂexible including, for instance, optional components for the project: a ﬁrm may have the opportunity (but not the obligation) to undertake the project, not only at a precise and given time, but during a whole period of time (or even without any time limit). In this sense, these characteristics may be related to that of an American call option, with the underlying asset being, for example, the ratio discounted beneﬁts/discounted costs, and the strike level ‘1’. Therefore, the N.P.V. criterion implies that the American option has to be exercised as soon as it is in the money, which is obviously a sub-optimal strategy. Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

150

Exotic Option Pricing and Advanced L´evy Models

The use of a method based on option theory, such as the real option theory, would improve the optimality of the investment decision. Several articles have appeared as benchmarks in this ﬁeld. The seminal studies of Brennan and Schwartz (1985), McDonald and Siegel (1986), Pindyck (1991) or Trigeorgis (1996) are often quoted as they present the fundamentals of this method, using particularly dynamic programming and arbitrage techniques. The literature on real options has been proliﬁc from very technical papers to case studies and manuals for practitioners (see among many references, the book edited by Brennan and Trigeorgis (2000) or that of Schwartz and Trigeorgis (2001)). Such an approach better suits reality by taking into account project optional characteristics such as withdrawal, sequential investment, delocalization, crisis management, etc. In that sense, real option theory leads to a decision criterion that adapts to each particular project assessment. However, real options have also some speciﬁc characteristics compared to ‘classical’ ﬁnancial options. In particular, the ‘risk-neutral’ logic widely used in option pricing cannot apply here: the real options’ underlying asset corresponds to the investment project ﬂows and is generally not quoted on ﬁnancial markets. Any replicating strategy of the option payoff is then impossible. So, the pricing is made under a prior probability measure (the historical probability measure or another measure chosen according to the investor’s expectations and beliefs). Moreover, a speciﬁc feature of a real option framework is the key points of interest for the investor. More precisely, she is interested in: • The cash ﬂows generated by the project. They are represented by the ‘price’ of the real option. Note that the notion of ‘price’ is not so obvious in this framework. It corresponds rather to the value a particular investor gives to this project. However, for the sake of simplicity in the notations, we will use the terminology ‘price’ in the rest of this paper. • But also, the optimal time to invest. This optimal time corresponds to the exercising time of the real option. Therefore, it is important noticing that real options are above all a management tool for decision taking. Once the investment project has been well-speciﬁed, the major concern for the investor is indeed summarized in the following question: ‘When is it optimal to invest in the project?’. In that sense, knowing the value of the option is less important than knowing its optimal exercising time. For that reason, in this paper, we focus especially on the properties of this optimal time. Moreover, real options studies are usually written in a continuous framework for the underlying dynamics. However, the existence of crises and shocks on investment markets generates discontinuities. The impact of these crises on the decision process is then an important feature to consider. This is especially relevant when some technical innovations may lead to instabilities in production ﬁelds. For all of these reasons, this paper is dedicated to the analysis of the exercising time properties in an unstable framework. The modelling of the underlying dynamics involves a mixed-diffusion (made up of Brownian motion and Poisson process). The jumps are negative so as to represent troubles and difﬁculties occurring in the underlying market. In the second section of this paper, we describe the framework of the study and analyze the consequences of different modelling choices. The crisis effect may be expressed via a Poisson process or the compensated martingale associated with it. Of course, there is an obvious relation between these models and they are equivalent from a static point of view. However, when studying the real option characteristics and their sensitivity towards the jump size, these models lead to various outcomes.

Impact of Market Crises on Real Options

151

After analyzing the real option characteristics in the third section, we focus on the discount rate. We prove the existence of an optimal discount rate, considering the maximization of the Laplace transform of the optimal time to invest as a choice criterion. We also characterize the average waiting time. In the ﬁfth section we study the robustness of the element decision characteristics. We ﬁrst specify the robustness of the optimal time to enter the project with respect to the jump size. We establish, in particular, that its Laplace transform is a decreasing function. Then, assuming that the investor only knows the expected value of the random jump size, we prove that this imperfect knowledge leads him/her to undertake the project too early. In the last section, we focus on the impact of a wrong model speciﬁcation, assuming that the investor believes in continuous underlying dynamics. In such a framework, we specify the error made in the optimal investment time. All proofs are presented in the Appendix.

7.2

THE MODEL

7.2.1 Notation In this paper, we consider a particular investor evolving in a universe, deﬁned as a ﬁltered probability space (, F, (Ft ), P). She has to decide whether to undertake a given investment project and, if so, when it is optimal to invest. We assume that the investor has no time limit to take this decision. Consequently, the time horizon we consider is inﬁnite. The investment opportunity value at time t = 0 is then of the form C0 = sup E exp (−µτ ) (Sτ − 1)+ τ ∈ϒ

where E denotes the expectation with respect to the prior probability measure P, ϒ is the set of the (Ft )- stopping times and (St , t ≥ 0) is the process of the proﬁts/costs ratio. It is worthwhile noticing that the discount rate µ is usually different from the instantaneous risk-free rate. We will come back later to the real meaning of discount rate in such a framework and to the problem related to its choice. The proﬁts/costs ratio related to the investment project is characterized by the following dynamics dSt = St − [αdt + σ dWt + ϕdMt ] (A) S0 = s0 where (Wt , t ≥ 0) is a standard (P, (Ft ))-Brownian motion and (Mt , t ≥ 0) is the compensated martingale associated with a (P, (Ft ))-Poisson process N . The Poisson process is assumed to have a constant intensity λ and the considered ﬁltration is deﬁned by Ft = σ (Ws , Ms , 0 ≤ s ≤ t). Equivalently, the process (St , t ≥ 0) may be written in the form: St = s0 exp(Xt ) where (Xt , t ≥ 0) is a L´evy process with the L´evy exponent E (exp (ξ Xt )) = exp (t (ξ ))

152

Exotic Option Pricing and Advanced L´evy Models

with (ξ ) = ξ 2

σ2 σ2 − λ 1 − (1 + ϕ)ξ + ξ α − λϕ − 2 2

(7.2.1)

Hence, we have

σ2 σ2 − ξ2 + λ eiξ ln(1+ϕ) − 1 E (exp (iX1 )) = exp iξ α − λϕ − 2 2 = exp (− (ξ )) Therefore, the L´evy measure associated with the characteristic exponent is expressed in terms of the Dirac measure δ as: ν (dx) = λδln(1+ϕ) (dx) Assumptions In the rest of the paper, the following hypothesis (H) holds. (i) 0 < s0 < 1, (ii) σ > 0 (iii) 0 > ϕ > −1.

(H)

Assumption (i) states that s0 is (strictly) less than 1: this is not a restrictive hypothesis, since the problem we study is a ‘true’ decision problem. In fact, delaying the project realization is only relevant in the case where the proﬁts/costs ratio is less than one. Assumption (iii) states that the jump size is negative as we study a crisis situation. The jump process allows us to take into account falls in the project business ﬁeld. These negative jumps may be induced, for instance, by a brutal introduction of a direct substitute into the market, leading to a decrease in the potential sales. Moreover, we assume that the jump size is greater than −1. This hypothesis, together with the identity 1

St = s0 (1 + ϕ)Nt × e(α−λϕ)t × eσ Wt − 2 σ

2t

ensure that the process S remains strictly positive. We also impose the integrability condition µ > sup (α; 0)

(7.2.2)

There exists an optimal frontier L∗ϕ such that sup E(e−µτ (Sτ − 1)+ ) = E(e

τ ∈ϒ

−µτL∗

ϕ

(SτL∗ − 1)+ ) ϕ

where τL is the ﬁrst hitting time of the boundary L by the process S, deﬁned as τL = inf {t ≥ 0; St ≥ L}

(7.2.3)

(For the proof, see, for instance, Darling et al. (1972) or Mordecki (1999).) Before the proﬁts/costs ratio S reaches the optimal boundary L∗ϕ , it is optimal for the investor not to undertake the investment project and to wait. However, as soon as S goes beyond this threshold, it is optimal for her to invest.

Impact of Market Crises on Real Options

153

7.2.2 Consequence of the modelling choice In the framework previously described, we may work a priori with either of the two following models: dSt = St − [αdt + σ dWt + ϕdMt ] (A) S0 = s0 (B)

dSt = St − [αdt + σ dWt + ϕdNt ] S0 = s0

In the case where all of the parameters are constant, these models are obviously equivalent and writing α = α + λϕ

(7.2.4)

is sufﬁcient to see why. Note that the integrability condition for model (B) is expressed as µ > max (α + λϕ; 0) However, when studying the sensitivity of the different option characteristics with respect to the jump size, choosing (A) or (B) really matters. Indeed, monotonicity properties are signiﬁcantly different in both frameworks, as underlined below. • Let us ﬁrst focus on the optimal time to enter the project, characterized by its Laplace transform deﬁned as E(exp(−µτL∗ϕ )). Considering model (A), if the initial value of the proﬁts/costs ratio is not ‘too small’, the Laplace transform of the optimal investment time is monotonic (this result is proved in Proposition 7.5.1). However, this monotonicity property does not hold any more for model (B) as is illustrated in Figure 7.1, which is done for the following set of parameters: s0 = 0.8;

λ = 0.1;

α = 0.05;

µ = 0.15;

σ = 0.2. 0.162 0.161 0.16 0.159 0.158 0.157 0.156 0.155 0.154

−1

−0.9

−0.8

−0.7

−0.6

−0.5 j

−0.4

−0.3

−0.2

−0.1

0.153 0

Figure 7.1 Laplace transform of the optimal time to invest (model (B))

154

Exotic Option Pricing and Advanced L´evy Models

• We now focus on the investment opportunity value C0 . Proposition 7.2.1 Let us consider model (B). Then, the investment opportunity value is an increasing function of the jump size. Figure 7.2 illustrates Proposition 7.2.1. It represents the variations of the investment value with respect to the jump size for different values of the jump intensity and for the following set of parameters: s0 = 0.8;

α = 0.05;

µ = 0.15;

σ = 0.2.

However, this property of the investment opportunity value does not hold any more when considering model (A). Intuitively, the studied model leads to a double effect of the jump size on the underlying level: ϕ has a positive effect on the underlying by increasing the drift but it also has a negative effect on the underlying by acting on the Poisson process level: dSt = St − ((α − λϕ) dt + σ dWt + ϕdNt ) This double effect explains the differences between models (A) and (B), and in particular accounts for the following result: in setting (A), the maximum value of C0 is not necessarily obtained for ϕ = 0. As a conclusion, it cannot be said that one of these models is better or more relevant than the other one. From a static point of view (with respect to the parameter ϕ), both are mathematically equivalent. In particular, given the condition shown in equation (7.2.4), they lead to the same ﬁrst and second moments for S. However, from a dynamic point of view with respect to the jump size, they are different. In the setting (B), crisis is only detected as the spread between the level of S before and after a shock while on the other hand, in the setting (A), there is an additional effect of the shocks on the drift term of S. Economically speaking, both have their own interests and motivations. However, once a model is chosen, the consequences of this choice must be 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 −1

−0.9

−0.8

−0.7

−0.6

λ = 0.1

Figure 7.2

−0.5 j

−0.4

λ = 10

−0.3

−0.2

λ=1

Investment opportunity value (model (B))

−0.1

0 0

Impact of Market Crises on Real Options

155

kept in mind, especially the implications for the monotonicity properties of the real option characteristics. In this present study, since we are particularly interested in the optimal time to invest, we choose a martingale representation for the stochastic part of SdS−t ; therefore, the model t deﬁned by (A) prevails in the following.

7.3 THE REAL OPTION CHARACTERISTICS In this section, we ﬁrst recall the classical formulae for the optimal time to invest and for the investment opportunity. We denote by kϕ the unique real number deﬁned in terms of the L´evy exponent deﬁned in equation (7.2.1) since it satisﬁes: kϕ > 1 and kϕ = µ Then the optimal proﬁts/costs ratio L∗ϕ satisﬁes: L∗ϕ =

kϕ kϕ − 1

The investment opportunity value at time 0 is given by: C0 =

s0 kϕ

kϕ

1 kϕ − 1

1−kϕ (7.3.1)

and the optimal investment time is characterized by its Laplace transform: E exp −µτL∗ϕ =

kϕ s0 kϕ − 1 kϕ

(7.3.2)

(For detailed proofs, see, among others, Gerber and Shiu (1994), Bellamy (1999) and Mordecki (1999, 2002).) It can be noticed that kϕ , as well as the optimal proﬁts/costs ratio L∗ϕ , depend on ϕ, λ and µ. Remark 1. In the framework we deal with, the so-called principle of smooth pasting is satisﬁed. Such a principle is always satisﬁed in a continuous framework but if the model is driven by discontinuous L´evy processes, this property can fail. In the model we consider, however, the smooth pasting principle still holds (see, for instance, Chan (2003, 2005), Boyarchenko and Levendorskii (2002), Alili and Kyprianou (2004) or Avram et al. (2004)). It is also easy to check that the optimal proﬁts/costs ratio satisﬁes L∗ϕ > 1. This underlines the interest of waiting before undertaking the project, as well as the gain in optimality obtained from considering a real option approach rather than the standard N.P.V. method (see, for instance, Dixit et al. (1993)).

156

Exotic Option Pricing and Advanced L´evy Models Table 7.1 Values of the optimal beneﬁts/costs ratio L∗ϕ as a function of σ and ϕ σ \ϕ

−0.9

−0.5

−0.3

−0.1

0.1 0.2 0.4 0.6

16.25 17.13 18.24 20.23

6.59 6.93 8.19 10.26

4.39 4.74 6.08 8.19

3.29 3.69 5.12 7.29

The value of the optimal ratio may be much greater than the limit value ‘1’. This fact is at variance with the N.P.V. criterion and perfectly illustrates what McDonald and Siegel (1986) have called ‘The value of waiting to invest’. As an illustration, the optimal ratio L∗ϕ is calculated in Table 7.1 for the following set of parameters: µ = 0.15;

λ = 1;

α = 0.1;

s0 = 0.8.

Note that high values for the volatility coefﬁcient σ are also considered in this study. This is relevant since the underlying market related to the investment project may be more highly volatile than traditional ﬁnancial markets (for instance, markets related to new technology).

7.4 OPTIMAL DISCOUNT RATE AND AVERAGE WAITING TIME 7.4.1 Optimal discount rate We now focus on the discount rate µ and present some general comments about its choice, which is indeed crucial in this study. The rate µ does not correspond to the instantaneous risk-free rate, traditionally used in the pricing of standard ﬁnancial options. In fact, in this real option framework, the rate µ characterizes the preference of the investor for the present or her aversion for the future. Choosing the ‘right’ µ is extremely difﬁcult. Many different authors have been interested in this question (among many others, Weitzman (1998)). Some have also proved the existence of a speciﬁc relationship between discount rate and future growth rate (Gollier (2002), Gollier and Rochet (2002) and Kimball (1990)). The optimal choice criterion for the rate µ depends, however, on the considered framework. We present here a relevant criterion for this particular problem, corresponding to the maximization of the Laplace transform of the optimal investment time. Proposition 7.4.1 (i) There exists a unique real number µ strictly positive such that E(exp(− µτL∗µ )) = max E(exp(−µτL∗µ )) µ

The real number µ agrees with an optimal choice of the discount rate µ. (ii) The optimal discount rate µ increases with the jumps intensity and decreases with the jumps size.

Impact of Market Crises on Real Options

157

This optimal discount rate is increasing with the absolute value of the jump size and with the intensity of the jumps. Such a behaviour seems rather logical as the occurrence and the frequency of negative jumps in the future make the value of the project decrease and represent an additional risk for the investor. The more important the jump intensity and size in absolute values are, the more the investor favours the present. Thus, she will choose a higher discount rate. Figure 7.3 shows the variations of the optimal rate µ with respect to ϕ for different values of λ and for the following set of parameters: s0 = 0.8;

α = 0.1;

σ = 0.2.

Remark 2. Other criteria may have been considered in order to choose an optimal rate. For instance, the maximization of C0 could appear as an alternative. However, it is not a relevant criterion, since the function kϕ

s0 kϕ − 1 kϕ µ → C0 = −1 kϕ − 1 kϕ is strictly decreasing. 7.4.2 Average waiting time Another question relative to the best time to invest is, of course, that of the characterization of an average waiting time. If we denote this by Tc , it is deﬁned as the unique element of R∗+ such that: E(exp(− µτL∗µ )) = exp(− µT c ) Hence, Tc corresponds to the average waiting time. In fact, it is the certainty equivalent µ. As of τL when the utility criterion is exponential and the risk aversion coefﬁcient is previously seen, this rate µ can easily be interpreted as a future aversion coefﬁcient (or a present preference coefﬁcient) and Tc may be explicitly determined as: Tc = −

1 ln E(exp(− µτL∗µ )) µ 2.5 2

^ m

1.5 1 0.5

−0.9

−0.8

−0.7

−0.6 λ = 0.2

−0.5

j

−0.4

λ=1

−0.3

−0.2 λ=2

Figure 7.3 Optimal discount rate, µˆ

−0.1

0 0

158

Exotic Option Pricing and Advanced L´evy Models 7 6 5

Tc

4 3 2 1

−0.9

−0.8

−0.7

−0.6

−0.5 j

λ = 0.2

Figure 7.4

−0.4 λ=1

−0.3

−0.2

−0.1

0 0

λ=2

Average waiting time, Tc

From Proposition 7.4.1, we deduce that the average waiting time decreases with respect to the jump intensity, as well as to the absolute value of the jump size. This mathematical property can be economically understood as previously. In fact, jumps induce additional risks, increasing with previous jump intensity and the jump size absolute value. The average waiting time can be related to an exponential utility criterion. Therefore, the investor we consider appears to be risk averse, with an exponential utility function and a risk aversion coefﬁcient of µ. So, in her decision process, she will take into account the expected proﬁt as well as the associated risk. She will tend to reduce the risk induced by the business ﬁeld by entering earlier in the project. Obviously, the more she waits, the greater the probability of jumps and then the risk are. Figure 7.4 highlights this fact. It represents the variations of the average waiting time with respect to the jump size. The graphs are produced for different values of the jump intensity. All of these curves converge to the same point as the jump size tends to zero: this point corresponds to the average waiting time in the model without jump, or, in other words, in an universe without crisis. The following set of parameters has been used: s0 = 0.8;

α = 0.1;

σ = 0.2.

7.5 ROBUSTNESS OF THE INVESTMENT DECISION CHARACTERISTICS All of the different parameters of the model have to be estimated using historical data or strategic anticipations. Every estimation and calibration may lead to an error on the choice of the input parameters. Some stability (or robustness) of the results is an essential condition for a real practical use of a model.

Impact of Market Crises on Real Options

159

7.5.1 Robustness of the optimal time to invest As it has already been underlined, the optimal time to invest is the major concern of the investor. Hence, the robustness of its Laplace transform appears as a key point to be checked. We particularly focus on the study of the sensitivity of this quantity with respect to the jump size. We study the behaviour of the Laplace transform of the optimal time to invest when the jump size is not perfectly known: the investor only knows that there exists ϕ and ϕ such that −1 < ϕ ≤ ϕ ≤ ϕ < 0 We ﬁrst provide a monotonicity result. Proposition 7.5.1 Let s0 be the level deﬁned as s0 = satisﬁes s0 < s0 < 1

k0 k0 −1

exp(− k01−1 ). We assume that s0 (7.5.1)

Then, the Laplace transform of the optimal time to invest is an increasing function of the jump size. Proposition 7.5.1 can be heuristically interpreted as follows: the more the jump size increases (hence decreases in absolute value), the more the investor delays entering the investment project. The maximum waiting time is attained in the lack of jump. Remark 3. The Assumption s0 < s0 amounts to consider investment project only if the initial value is not ‘too small’. From an economic point of view, such an assumption is not very restrictive. In fact, the investor will stop being interested in the project as soon as s0 is below a given threshold. If, for example, we consider the following standard set of parameters α = 0.10; σ = 0.20; µ = 0.15, then we get s0 = 0.276. Note that this level s0 is far from the strike value 1. Figure 7.5 shows the changes in the Laplace transform of the optimal time to invest with respect to ϕ for different values of λ. The following set of parameters is used: s0 = 0.8;

α = 0.10;

σ = 0.20;

µ = 0.15.

The robustness property of the Laplace transform is a straightforward consequence of Proposition 7.5.1. Corollary 7.5.2 We assume that the condition shown in equation (7.5.1) holds and −1 < ϕ ≤ ϕ ≤ ϕ < 0 Then, we have E(exp(−µτL∗ϕ )) ≤ E(exp(−µτL∗ϕ )) ≤ E(exp(−µτL∗ϕ )).

160

Exotic Option Pricing and Advanced L´evy Models 0.14 0.12 0.1 0.08 0.06 0.04 0.02 −1

−0.9

−0.8

−0.7

−0.6

λ = 0.2

−0.5 j

−0.4

λ=1

−0.3

−0.2

−0.1

0 0

λ = 10

Figure 7.5 Laplace transform of the optimal time to invest for different values of λ

This result underlines the model robustness as far as the Laplace transform of the optimal time to invest is concerned. More precisely, if the investor does not know exactly the size of the jump, in other words the impact of the market crisis on the project, but knows, however, some boundaries for it, then she has an idea of the optimal time to enter the project. More precisely, the Laplace transform boundaries are expressed in terms of the boundaries for the market crisis impact. Equivalently, having some control or knowledge of the crisis impact enables the investor to have some control of her optimal time to invest. 7.5.2 Random jump size We now consider the situation where the jump size is an unknown random variable . We focus on the impact that this additional hazard may have on the investor decision. Assuming that the investor estimates the jump size by its expected value E (), we focus on the impact of such an error on her decision. Will she invest too early or too late? In order to answer this questions, we compare the ‘true’ Laplace transform of the optimal time to invest, with the Laplace transform estimated by means of E (). The dynamics of the process of the project are now: dSt = St− (αdt + σ dWt + dMt ) ; S0 = s0 and the investor builds her strategy from StE() where E() dStE() = St− (αdt + σ dWt + E () dMt ) ; S0E() = s0

We assume that the random variable is independent of the ﬁltration generated by the Brownian motion and the Poisson process. Let L∗ be the true optimal beneﬁt–cost ratio. If the investor only knows E (), she estimates this ratio by L∗E() . The next proposition provides a comparison between these two quantities.

Impact of Market Crises on Real Options

161

Proposition 7.5.3 We assume that the condition shown in equation (7.5.1) holds. Then, the wrong speciﬁcation in the model leads the investor to underestimate the optimal proﬁts/costs ratio. Moreover we can pr´ecis the consequences of this error on the decision taking. We assume that the investor undertakes the project when the observed process of the beneﬁts/costs ratio reaches what she supposes to be the optimal level. Therefore, her strategy is determined by the ﬁrst hitting time of L∗E() , instead of the ﬁrst hitting time of L∗ by process S. This proposition can be interpreted as follows: when the investor only knows E (), she tends to undertake the project too early.

7.6 CONTINUOUS MODEL VERSUS DISCONTINUOUS MODEL In this section, we focus on the impact of a wrong model choice. This part extends the previous study of robustness. We suppose that the investor believes in a continuous underlying dynamics for S, while its true dynamics is given by (A). As a consequence, the investor governs her strategy according to the following process: St ($ α dt + $ σ dWt ) d$ St = $ where

$ (A)

S0 = s0 $ $ α =α 2 $ σ = σ 2 + λϕ 2 .

$ on the same These equalities come directly from the calibration of both model (A) and (A) $ data set, leading to the same ﬁrst and second moments for S and S. The volatility parameter of the model without jump is different from that of the model with jumps: the absence of jump in the dynamics is indeed compensated by a higher volatility. In order to obtain the ‘equivalent’ volatility, the right brackets of S and $ S have to be equal. The process $ S is called ‘equivalent process without jump’. We now focus on the impact of such a wrong speciﬁcation on the investment time. To this end, we ﬁrst consider the error in the optimal proﬁts/costs ratio. 7.6.1 Error in the optimal proﬁt–cost ratio $ More precisely, $∗ϕ the optimal proﬁts/costs ratio in the model deﬁned by (A). We denote by L ∗ using the same arguments as presented in Section 7.3, $ Lϕ is given by the following ratio $ L∗ϕ =

$ kϕ $ kϕ − 1

where $ kϕ is the solution of

σ 2 + λϕ 2 2 σ 2 + λϕ 2 $ ψ (k) = k + α− k=µ 2 2

(7.6.1)

162

Exotic Option Pricing and Advanced L´evy Models

Note that this optimal ratio depends on the volatility parameter of the model, or equivalently, on both jump parameters ϕ and λ. For the sake of simplicity, as we are especially interested $∗ϕ . in the sensitivity with respect to the jump size, we use the notation L Proposition 7.6.1 The previous wrong speciﬁcation of the model leads the investor to underestimate the optimal proﬁts/costs ratio, if and only if, σ 2 + λϕ 2 + 2α ≥ µ

(7.6.2)

Note that for the usual values of the parameters, the inequality shown in equation (7.6.2) often holds. For instance, if we consider λ = 1, α = 0.1, σ = 0.2 and µ = 0.15, then σ 2 + λϕ 2 + 2α ≥ µ is true for all ϕ in ] − 1, 0[. As an illustration, the relative error (expressed in percentage) on the optimal proﬁts/costs ratio

$∗ϕ L∗ϕ − L ∗ RE(L , ϕ) = 100 × L∗ϕ is calculated in Table 7.2 for different values of the jump size ϕ and for the standard set of parameters: s0 = 0.8; α = 0.1; σ = 0.20; µ = 0.15; λ = 1. Very naturally, the relative error becomes negligible as the jump size tends to zero. This error is still manageable when the jump size is not too large (up to −0.5). For larger values, however, the relative error becomes quite important to reach more than a third of the value of the ratio when the jump size is maximal. Using the same argument as in the previous section, we can pr´ecis the consequences that this wrong speciﬁcation has on the investor’s strategy. The investor’s waiting time is $∗ϕ instead of L∗ϕ . So, if the condition shown in equation (7.6.2) holds, we determined by L can assert that the error in the model leads the investor to undertake the project too early. This fact is brought to the fore by Figure 7.6. The optimal time to enter the project for a well-informed investor, as well as that of the previous investor, are respectively characterized by the Laplace transforms E(exp(−µτL∗ϕ )) and E(exp(−µτ$ L∗ϕ )). Figure 7.6 represents the variations of these Laplace transforms with respect to the jump size ϕ. This is, carried out for the following values: s0 = 0.8;

α = 0.1;

σ = 0.20;

µ = 0.15;

λ = 1.

As another illustration, the relative error (expressed in percentage) on the Laplace transform of the optimal time to invest

. LT − LT RE(LT , ϕ) = 100 × LT Table 7.2

Relative error on the optimal proﬁts/costs ratio as a function of ϕ

ϕ

−0.995

−0.7

RE(L∗ , ϕ)

38.30

15.81

−0.5 7.11

−0.3 1.87

−0.1 0.08

−0.01 0.01

Impact of Market Crises on Real Options

163

0.12 0.10 0.08 0.06 0.04 0.02 −1

−0.8

−0.6

−0.4 j

−0.2

0

~ LT of the hitting time of L∗

LT in the model with jumps

Figure 7.6 Laplace transform of the optimal time to invest in the model with jump and Laplace transform of the hitting time of the proﬁt–cost ratio, L˜ ∗ Table 7.3 Relative error on the Laplace transform of the optimal time to invest as a function of ϕ ϕ RE(LT , ϕ)

−0.995 −51.77

−0.7

−0.5

−0.3

−0.1

−0.02

−14.68

−5.58

−1.26

−0.04

−0.01

is calculated in Table 7.3 for different values of the jump size ϕ and for the previous set of parameters. The interpretation of these results is very similar to those associated with the relative error on the optimal proﬁts/costs ratio. It can be noticed, however, that for large values of the jump size, the relative error becomes quite important to reach more than a half of the Laplace transform when the jump size is maximal. Hence, the impact of a wrong model speciﬁcation could be important if the investor focuses on the optimal time to invest in the project. 7.6.2 Error in the investment opportunity value In the ‘true’ model with jumps, the investment opportunity value is C0 . If we assume that the investor becomes involved in the project when the ‘true’ process S reaches the level $∗ϕ , then her investment opportunity value is L $∗ϕ − 1)E(exp(−µτ$∗ )) $0 = (L C Lϕ in which

$∗ϕ − 1) × $0 = (L C

where $ kϕ is the solution of equation (7.6.1).

s0 $ L∗ϕ

$kϕ

164

Exotic Option Pricing and Advanced L´evy Models 0.65 0.6 0.55 0.5 0.45 0.4 0.35 −1

−0.9

Figure 7.7

−0.8

−0.7

−0.6

−0.5 j ~ C0

−0.4

−0.3

−0.2

−0.1

0.3 0

C0

Investment value estimated with L˜ ∗ and the optimal investment value

$0 with respect to the jump size ϕ. Of Figure 7.7 represents the variations of C0 and C $∗ϕ differs from the optimal frontier L∗ , we have for any ϕ, course, since L $0 ≤ C0 C $0 comes from a wrong investment time. This loss tends to zero when and the loss C0 − C $∗ϕ tends to the optimal frontier the jump size tends to zero and this fact was expected as L ∗ L when ϕ tends to 0. The curves shown in Figure 7.7 are produced by using the following values: s0 = 0.8; α = 0.1; σ = 0.20; µ = 0.15; λ = 1. As another illustration, the relative error (expressed in percentage) on the investment opportunity value RE(C, ϕ) = 100 ×

$0 C0 − C C0

is calculated in Table 7.4 for different values of the jump size ϕ and for the previous set of parameters. The relative error remains manageable even for large values of the jump size since it is always less than 10%. Therefore, the impact of a wrong model speciﬁcation is relatively not so important if the investor focuses on the value of the investment opportunity. Table 7.4 Relative error on the investment opportunity value as a function of ϕ ϕ

−0.995

RE(C, ϕ)

9.02

−0.7 5.33

−0.5 3.19

−0.3 1.14

−0.1 0.06

−0.02 0.01

Impact of Market Crises on Real Options

165

7.7 CONCLUSIONS In this paper, we study the impact of market crises on investment decision via real option theory. The investment project, modelled by its proﬁts/costs ratio, is characterized by a mixed diffusion process, whose jumps represent the consequences of crises on the investment ﬁeld. After having analyzed the implications of different model choices, we study the real option associated with this investment project. We establish the existence of an optimal discount rate, given a criterion based on this investment time and we characterize the average waiting time. We study in detail the properties of the optimal investment time, through its Laplace transform, and focus, in particular, on its robustness when the underlying dynamics of the project are not well-known or are wrongly speciﬁed. We interpret the results in terms of the investment decision. More precisely, when the investor bases his/her decision on the expected value of the random jump size, he/she tends to undertake the project too early. The same property holds if he/she believes in a continuous dynamics for the underlying project. In this paper, we focus on a single investor. The complexity of reality suggests, however, that different other aspects, in particular, strategic relationships between the economic agents, may play an important role. Investigating more general models involving strategic dimensions and game theory is a topic for future research.

APPENDIX Proof of Proposition 7.2.1

ϕ Let S be deﬁned by model (B). We deﬁne C (ϕ, L) as C (ϕ, L) = (L − 1) × E exp −µτL ϕ where τL = inf {t ≥ 0; St ≥ L}. Hence C0 (ϕ) = C ϕ, L∗ϕ where L∗ϕ is the optimal frontier, that is to say, the optimal beneﬁt–cost ratio. Let ϕ2 and ϕ1 be such that −1 < ϕ2 < ϕ1 < 0. We have C0 (ϕ1 ) = C ϕ1 , L∗ϕ1 ≥ C ϕ1 , L∗ϕ2 Then inequality ϕ1 > ϕ2 leads to ∀t ≥ 0, St (ϕ1 ) ≥ St (ϕ2 ) and consequently ϕ

ϕ

E(exp(−µτL∗ϕ1 )) ≥ E(exp(−µτL∗ϕ2 )) 2

∗

2

∗

Finally, we get C0 (ϕ1 ) ≥ C(ϕ1 , L (ϕ2 )) ≥ C(ϕ2 , L (ϕ2 )) = C0 (ϕ2 ). Proof of Proposition 7.4.1 (i) The function k ∈ ]1, +∞[ →

s0 (k−1) k

ln s0 + ln

k

admits a maximum for k = k, deﬁned by:

k−1 1 + =0 k k−1

(7.A.1)

166

Exotic Option Pricing and Advanced L´evy Models

The study of the L´evy exponent leads to the existence of a unique value of µ, denoted µ k. Moreover, µ satisﬁes: by µ, such that µ > α and kϕ() = E(exp(− µτL∗µ )) = max E(exp(−µτL∗µ )) µ

Assertion (ii) comes from the deﬁnition of µ and the following properties of the L´evy exponent: ∀k ∈ ]1, k[ ∀ϕ ∈ ] − 1, 0[,

λ → (k)

is increasing and ∀k ∈ ]1, k[ ∀λ > 0,

ϕ → (k)

is decreasing. Proof of Proposition 7.5.1 Let k be deﬁned by equation (7.A.1). We have k ⇐⇒ s0 ≥ s0 k0 ≤ where s0 =

1 k0 exp − k0 − 1 k0 − 1

and where k0 is the limit: k0 = lim kϕ . ϕ→0

In order to get the conclusion, it sufﬁces to prove that kϕ is strictly increasing with respect to the jump size ϕ. Let F : ] − 1; 0[ × ]1; +∞[→ R be the function deﬁned as: F (ϕ, k) = (k) − µ where is given by equation (7.2.1). For any (ϕ, k) ∈ ] − 1; 0[ × ]1; +∞[ such that F (ϕ, k) = 0, we can easily check that (k) > 0. Using the implicit function theorem, we get: ∂F ∂k ∂ϕ (ϕ, k) = − ∂F ∂ϕ ∂k (ϕ, k)

and the inequality

∂F ∂ϕ

(ϕ, k) < 0 implies

increasing. Proof of Proposition 7.5.3

∂k > 0. Hence the function ϕ → kϕ is strictly ∂ϕ

We denote by and E() the L´evy exponents of the processes Xt t≥0 and XtE() , t≥0

E() St S and XtE() = ln . where Xt = ln t s0 s0

Impact of Market Crises on Real Options

167

Let k (resp.kE() ) be the unique real number strictly greater than 1 such that (k ) = µ (resp. E() kE() = µ). We have (k) = f (, k) + g(k) (resp. E() (k) = f (E(), k) + g(k)) where f (, k) = λ(1 + )k − λk and g (k) =

σ2 2 σ2 k + α− k − λ. 2 2

The convexity of the function x → f (x, k) for any k > 1, together with Jensen inequality, implies that ∀k > 1, E() (k) ≤ (k) Hence k ≤ kE() and from this last inequality, we conclude L∗ ≥ L∗E() . Proof of Proposition 7.6.1

$t $ be the L´evy exponent of the process X Let

t≥0

$ $t = ln St where X s0

and $ kϕ be the

unique real number such that $ kϕ > 1

$ $ kϕ = µ.

Then, from the equalities $ $ (0) = (0) = 0 and (2) = (2) = σ 2 + λϕ 2 + 2α, we get

$ kϕ ≥ kϕ ,

if and only if, σ 2 + λϕ 2 + 2α ≥ µ and therefore we have

$∗ϕ L∗ϕ ≥ L

REFERENCES [1] Alili, L. and Kyprianou, A. (2004), “Some remarks on ﬁrst passage of L´evy processes, the American put and pasting principles”, Annals of Applied Probability, to be published. [2] Avram, F., Kyprianou, A. and Pistorius, M. (2004), “Exit problems for spectrally negative L´evy processes and applications to Russian, American and Canadized options, Annals of Applied Probability, 14, 215–238. [3] Bellamy, N. (1999), Hedging and Pricing in a Market Driven by Discontinuous Processes, Ph. D. Thesis, Universit´e d’Evry Val d’Essonne, Evry, France.

168

Exotic Option Pricing and Advanced L´evy Models

[4] Boyarchenko, S.I. and Levendorskii, S.Z. (2002), “Perpetual American options under L´evy processes”, SIAM Journal on Control and Optimization, 40, 1663–1696. [5] Brennan, M.J. and Schwartz, E.S. (1985), “Evaluating natural resource investments”, Journal of Business, 58, 135–157. [6] Brennan, M.J. and Trigeorgis, L. (Eds) (2000), Project Flexibility, Agency, and Product Market Competition – New Developments in the Theory and Applications of Real Options, Oxford University Press, New York, NY, USA. [7] Chan, T. (2003), “Some applications of L´evy Processes in insurance and ﬁnance”, Finance, 25, 71–94. [8] Chan, T. (2005), “Pricing perpetual American options driven by spectrally one-sided L´evy processes”, in A.E. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 195–216. [9] Darling, D.A., Ligget, T. and Taylor, H.M. (1972), “Optimal stopping for partial sums”, Annals of Mathematical Statistics, 43, 1363–1368. [10] Dixit, A.K. and Pindyck, R. S. (1993). Investment under Uncertainty, Princeton University Press, Princeton, NJ, USA. [11] Gerber, H.U. and Shiu, E.S. (1994), “Martingale approach to pricing perpetual American options”, ASTIN Bulletin, 24, 195–220. [12] Gollier, C. (2002), “Time horizon and the discount rate”, Journal of Economic Theory, 107, 463–473. [13] Gollier, C. and Rochet, J.C. (2002), “Discounting an uncertain future”, Journal of Public Economics, 85, 149–166. [14] Kimball, M.S. (1990), “Precautionary saving in the small and in the large”, Econometrica, 58, 53–73. [15] McDonald, R. and Siegel, D. (1986), “The value of waiting to invest”, Quarterly Journal of Economics, 101, 707–727. [16] Mordecki, E. (1999), “Optimal stopping for a diffusion with jumps”, Finance and Stochastics, 3, 227–236. [17] Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493. [18] Pindyck, R.S. (1991), “Irreversibility, uncertainty and investment”, Journal of Economic Literature, 29, 1110–1148. [19] Schwartz, E.S. and Trigeorgis, L. (2001), Real Options and Investment under Uncertainty: Classical Readings and Recent Contributions, MIT Press, Cambridge, MA, USA. [20] Trigeorgis, L. (1996), Real Options – Managerial Flexibility and Strategy in Resource Allocation, MIT Press, Cambridge, MA, USA. [21] Weitzman, M.L. (1998), “Why the far-distant future should be discounted at its lowest possible rate”, Journal of Environmental Economics and Management, 36, 201–208.

8 Moment Derivatives and L´evy-type Market Completion Jose´ Manuel Corcuera and David Nualart University of Barcelona, Barcelona, Spain and

Wim Schoutens Katholieke Universiteit Leuven – UCS, Leuven, Belgium Abstract ´ We show how moment derivatives can complete Levy-type markets in the sense that by allowing trade in these derivatives any contingent claim can be perfectly hedged by a dynamic portfolio in terms of bonds, stocks and moment-derivative-related products. Moment derivatives depend on the sum of the powered returns, i.e. the realized moments. Squared log-returns are the basis of the nowadays popular Variance Swaps. Higher-powered returns assess other kinds of important characteristics of the underlying distribution such as skewness and kurtosis. We ﬁrst work under a discrete time setting under which we assume that the returns of the stock price process are independent and identically distributed. Out of the Taylor expansion of the payoff function, we extract the positions one has to take in order to perfectly hedge the claim. We illustrate this by some illustrative examples such as the Trinomial tree model. Next, we comment on the continuous time setting. In this case, a Martingale Representation Property lies at the heart of the completion on the market considered. Results in this ´ market were already obtained in previous work of these authors. A survey exponential Levy of the relevant results are given and the relation and similarities with the discrete setting are discussed.

8.1

INTRODUCTION

In this paper, we consider markets where the returns are independent and identically distributed (iid). Typically, these markets are incomplete, and the purpose of this work is to show a systematic way of completing these markets. We shall complete the market by introducing a series of assets related to the powers of the return process. First we present the procedure in a discrete-time setting with discrete returns, while, secondly we consider more general returns, and ﬁnally we consider the continuous-time setting. In fact this latter case has been considered in Corcuera et al. (2004a) [10] and in such a case the new assets are based on the power-jump processes of the underlying L´evy process. In addition, these new assets can be related with options on the stock (see Balland (2002) [2]) Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

170

Exotic Option Pricing and Advanced L´evy Models

and with contracts on realized variance (Carr and Madan (1998) [7] and Demeterﬁ et al. (1999) [15]) that have found their way into OTC markets and are now traded regularly. Higher order power-jump processes have a similar relationship with which one could call realized skewness and realized kurtosis processes. Contracts on these objects, however, are not common. Carr et al. (2002) [8] and Carr and Lewis (2004) [6] have studied contracts on the quadratic variation processes in a model driven by a so-called Sato process. We give an explicit hedging portfolio for claims whose payoff function depends on the prices of the stock and the new assets at maturity. Then, if we introduce utility functions, we can obtain the optimal terminal wealth with respect to these utilities and by the completeness of the enlarged market we can obtain the optimal portfolio by duplicating the optimal wealth. This has been carried out by Corcuera et al. (2004b) [11], where we also analyze the case where the optimal portfolio consists only in stocks and bonds. This corresponds to complete the market with new assets in such a way that they are superﬂuous, that is, we do not improve the terminal expected utility by including these new assets in our portfolio. This is equivalent to choosing an appropriate risk-neutral or martingale measure (see Kallsen (2000) [18] and Schachermayer (2001) [28]). Moreover, this martingale measure is related to the neutral derivative pricing of Davis (1997) [13].

8.2

MARKET COMPLETION IN THE DISCRETE-TIME SETTING

We start by explaining the ideas in the most simple incomplete discrete market setting: the one-step trinomial market model. Next, we will consider a one-step market model, where the stock can attain m different values, then we will consider the same model but with n time-steps, and ﬁnally we will deal with a general multi-step market. 8.2.1 One-step trinomial market In this model, we assume we have a risk-free bond paying out a constant interest rate r > 0, i.e. the bond has a deterministic value process: B0 = 1 and B1 = 1 + r. We have also a risky asset, a stock, which can move from its initial value S0 > 0 to three different values at time 1. More precisely, we have S1 = S0 (1 + X1 ), where X1 can take the values −1 < x1 < x2 < x3 . It is a classical argument, that in order to avoid arbitrage one should have x1 < r < x3 (by investing in stocks you can lose more, but also gain more, than by investing in bonds). This arbitrage-free market is one of the most simple cases of an incomplete market, in the sense that there exist contingent claims which cannot be hedged by positions in bonds and stocks. We will show that by introducing a moment option (a Variance-Swap-like derivative), the model can be completed. Moreover, we show that the position one has to take, in order to hedge any contingent claim, can just be read off from the Taylor expansion of the payoff function of the claim. Indeed, suppose we allow also trade in a contingent claim, paying out X12 at time 1. We will refer to this derivative as the MOM (2) derivative. Let us denote the price of this contingent claim at time zero by z2 . In order to exclude arbitrage, there must be an equivalent martingale measure, making the discounted values of all traded securities martingales. Denoting the risk-neutral probabilities that X1 attains the value xi by qi > 0, i = 1, 2, 3, we must have

Moment Derivatives and L´evy-type Market Completion

171

q1 + q2 + q3 = 1 q1 x1 + q2 x2 + q3 x3 = r q1 x12 + q2 x22 + q3 x32 = z2 (1 + r), where the ﬁrst equation is ensuring that we have a probability measure, the second equation makes the risk-neutral return on the stock equal to r and the third equation ﬁxes the price of the MOM (2) derivative at z2 . This system of equations can be rewritten in matrix form as 1 1 1 q1 1 . $ · q = x1 x2 x3 · q2 = r z2 (1 + r) q3 x12 x22 x32 Since $ is a Vandermonde matrix, det($) = 0 and $ is invertable. So, the system has exactly one solution, namely: 1 . q = $−1 · r z2 (1 + r) If this solution satisﬁes q ∈ (0, 1)3 , i.e. the qi s can be seen as probabilities, we have no arbitrage. Moreover, since the solution is unique, we have also, by the second fundamental theorem of asset pricing, that the market is complete. Note that to choose an arbitrage price z2 is equivalent to choosing a risk-neutral probability q. Consider now a general contingent claim, with payoff function G(X1 ) and develop this function into powers of X1 : G(X1 ) = a0 + a1 X1 + a2 X12 . Since X1 can take only three possible values, the series is cut off after the quadratic term. In order to hedge this claim one needs to carry out the following: • Invest (a0 − a1 )(1 + r)−1 into bond. • Buy a1 /S0 units of stock, for a total price a1 . • Buy a2 units of MOM (2) derivatives, for a total price a2 z2 . At time t = 1, we have the following: • The money invested in bond has grown to a0 − a1 . • We sell the a1 /S0 stocks, giving us a1 (1 + X1 ) of money. • The MOM (2) derivatives each pay out X12 . This leads to a total payout of a2 X12 . In total we thus end up with a0 + a1 X1 + a2 X12 = G(X1 ) of money, exactly the payout of the contingent claim considered. In order to set up this strategy we needed a0 − a1 + a1 + a2 z2 1+r of money, which in order to avoid arbitrage must be the initial price of the contingent claim with payoff function G(X1 ).

172

Exotic Option Pricing and Advanced L´evy Models

8.2.2 One-step ﬁnite markets The above situation can easily be generalized to a (one-step) setting, where the random variable X1 can take a ﬁnite number m of possible values −1 < x1 < · · · < xm , with x1 < r < xm , to avoid arbitrage. For m ≥ 3, the market is an incomplete market; there exist contingent claims which cannot be hedged by holding positions in bonds and stocks alone. Assume trade is allowed into moment derivatives with payoff functions MOM (k) = X1k ,

k = 2, . . . , m − 1.

So, besides investing in bonds and stocks, one can invest also into m − 2 other derivatives, i.e. the MOM (k) ’s moment derivatives. Note that payoff functions and initial prices can be negative. For example, in the case of a negative payoff, the holder must pay the corresponding amount to the issuer. Let zk , k = 2, 3, . . . , m − 1, be the initial price of the MOM (k) derivative. In order to exclude arbitrage, there must be, as above, an equivalent martingale measure, making the discounted values of all traded securities martingales. Denoting the risk-neutral probabilities that X1 attains the value xi by qi > 0, i = 1, . . . , m, we must have q1 + · · · + qm = 1

(8.1)

q1 x1 + · · · + qm xm = r 2 = z2 (1 + r) q1 x12 + · · · + qm xm

.. . m−1 q1 x1m−1 + · · · + qm xm = zm−1 (1 + r),

where the ﬁrst equation is ensuring that we have a probability measure, the second equation makes the risk-neutral return on the stock equal to r and the other equations ﬁx the prices of the MOM (k) derivatives at zk , k = 2, . . . , m − 1. With obvious notation (as above), the system has exactly one solution, namely: 1 r (8.2) q = $−1 · z2 (1 + r) . .. . zm−1 (1 + r) If this solution satisﬁes q ∈ (0, 1)m , i.e. the qi s can be seen as probabilities, we have noarbitrage. Moreover, since the solution is unique, we have also that the market is complete. Since X1 can now take m possible values, the payoff of a contingent claim G(X1 ) can now be written into a Taylor expansion up to degree m − 1: G(X1 ) = a0 + a1 X1 + a2 X12 + · · · + am−1 X1m−1 . Completely analogous as in the trinomial setting, the hedging of this contingent claim can be carried out by performing the following:

Moment Derivatives and L´evy-type Market Completion

173

• Invest (a0 − a1 )(1 + r)−1 into bond. • Buy a1 /S0 units of stock, for a total price a1 . • For each k = 2, 3, . . . , m − 1, buy ak MOM (k) derivatives for a price ak zk . At time t = 1, we have the following: • The money invested in bond has grown to a0 − a1 . • We sell the a1 /S0 stocks, giving us a1 (1 + X1 ) of money. • For each k = 2, 3, . . . , m − 1, each MOM (k) derivatives pays out X1k . This leads to a k total payout of m−1 k=2 ak X1 . k In total, we thus end up with a0 + a1 X1 + m−1 k=2 ak X1 = G(X1 ) of money, exactly the payout of the contingent claim considered. In order to set up this strategy, we needed a0 − a1 ak zk + a1 + 1+r m−1 k=2

of money, which in order to avoid arbitrage must be the initial price of the contingent claim with payoff function G(X1 ). 8.2.3 Multi-step ﬁnite markets In this model, we consider a generalization of the above model, taking into account n timesteps. We assume that we have a risk-free bond paying out an interest rate r, i.e. the bond has a deterministic value process: B0 = 1 and Bi = (1 + r)i , i = 1, . . . , n. We have also a risky asset, a stock, which has the following price process S0 > 0,

Si = Si−1 (1 + Xi ) = S0 (1 + X1 ) · · · (1 + Xi ),

i = 1, . . . , n.

We assume the Xi s are deﬁned on a stochastic basis {, F, P , F}, where F = {Fi }ni=1 is a ﬁltration that describes how the information about the security prices is revealed to the investors. We will suppose that F0 = {∅, }, Fi = σ (S1 , . . . , Si ), i = 1, . . . , n and F = Fn . In addition, we will assume that the Xi s are iid and can attain m possible values −1 < x1 < · · · < xm ,with x1 < r < xm to avoid arbitrage. This arbitrage-free market is again an incomplete market. We will show that by introducing into the market, at each time-step, moment derivatives which mature one time-step later and payoff some power of the return the stock makes over that time-step, the model can be completed. Assume at time t = i − 1, i = 1, . . . , n trade is allowed into, at this time newly introduced, moment derivatives (MOM(k) i ) which mature at time T = i and have a payoff function k MOM(k) i = Xi ,

k = 2, . . . , m − 1.

So, besides investing in bonds and stocks, one can invest at time zero also into the MOM (k) 1 , k = 2, . . . , m − 1 derivatives. These derivatives mature at time T = 1. At this time, a set of m − 2 new derivatives are introduced into the market; these derivatives MOM (k) 2 , k = 2, . . . , m − 1, mature at time T = 2, etc.

174

Exotic Option Pricing and Advanced L´evy Models

The Xi s are iid with respect to P ; in consequence, any possible path of the stock’s prices has non-null P -probability. Then, if the system shown by equation (8.2) satisﬁes q ∈ (0, 1)m , we can ﬁnd a risk-neutral probability Q such that the prices of the kth moment derivatives at their initiation are equal to say zk , independently on the step time. This means that for each time i = 1, . . . , n we take Q(Xi = xj |Fi−1 ) = qj . Note that under Q the Xi s are also iid. Thus, for each i = 1, . . . , n, we have price of MOM (k) i at time i − 1 = zk ,

k = 2, . . . , m − 1.

By the model described in Section 8.2.2, any payoff function G = G(X1, X2, . . . , Xn ) at time t = n can be hedged by a portfolio built at t = n − 1, having ﬁxed the value of (X1 , . . . , Xn−1 ) = (x1 , . . . , xn−1 ). In fact, we can write G(x1 , . . . , xn−1 , Xn ) =

m−1

bn(k) (x1 , . . . , xn−1 )Xnk ,

k=0

and the value of this portfolio at time t = n − 1 will be Vn−1 (x1 , . . . , xn−1 ) =

bn(0) (x1 , . . . , xn−1 ) − bn(1) (x1 , . . . , xn−1 ) 1+r + bn(1) (x1 , . . . , xn−1 ) +

m−1

bn(k) (x1 , . . . , xn−1 )zk .

k=2

Then, we can replicate G(x1 , . . . , xn−2 , Xn−1 , Xn ) by a portfolio built at t = n − 2, by duplicating Vn−1 (x1 , . . . , xn−2 , Xn−1 ). Finally, by backward induction we have that any contingent claim can be hedged by a self-ﬁnancing portfolio. 8.2.4 Multi-step markets with general returns With the same notation as in the previous case, let us assume that the Laplace transform of (X1 , X2 , . . . , Xn ) is deﬁned in an open neighborhood of the origin (under Q); then the polynomials are dense in L2 (Fn , Q). So, for any contingent claim, G = G(X1 , X2 , . . . , Xn ) ∈ L2 (Q), if we are in the trading time n − 1 with (X1 , X2 , . . . , Xn−1 ) = (x1 , x2 , . . . , xn−1 ), we can write L2

G = lim

l→∞

l

bn(k,l) (x1 , x2 , . . . , xn−1 )Xnk .

k=0

and by backward induction we can replicate G by a self-ﬁnancing portfolio (see Corcuera et al. (2005) [10] for more details). 8.2.5 Power-return assets Another way of completing the market, is by allowing trade in the so-called power-return assets. To simplify the exposition, we shall work under the ﬁnite market setting. We thus have a risk-free bond paying out an interest rate r, i.e. the bond has a deterministic value

Moment Derivatives and L´evy-type Market Completion

175

process: B0 = 1 and Bi = (1 + r)i , i = 1, . . . , n. We have also a risky asset, a stock, which is the following price process S0 > 0,

Si = Si−1 (1 + Xi ) = S0 (1 + X1 ) · · · (1 + Xi ),

i = 1, . . . , n,

and where the Xi s are iid (with respect to P ) and can attain m different values −1 < x1 < · · · < xm , and x1 < r < xm . Assume now, that in this market m − 2 new assets are introduced with price process Hi(k)

= (1 + r)i

i

Xjk − µk i ,

k = 2, . . . , m − 1,

j =1

where µk ∈ R. Let us make a few remarks on these assets. The asset with price process Hi(k) will be refereed to as the kth order power-return asset. Remark 1 (Arbitrage) To avoid arbitrage by the introduction of these power-return assets, some conditions are necessary on the constants µk . Classical theory says that to have an arbitrage-free market, there must exist an equivalent martingale measure, under which all of the discounted prices process of the assets are martingales. This condition translates into the existence of probabilities 0 < qi < 1, such that q1 + · · · + qm = 1

(Condition H)

q1 x1 + · · · + qm xm = r 2 = µ2 q1 x12 + · · · + qm xm

.. . m−1 q1 x1m−1 + · · · + qm xm = µm−1 .

Remark 2 The ﬁrst condition forces the qi s to sum up to 1, as probabilities should do. The second condition forces the discounted stock price to be a martingale; the other ones force the discounted power-return asset prices to be martingales. These conditions are almost identical to the conditions in equation (8.1); just replace zk (1 + r) by µk . In fact, if these condition are satisﬁed, it is straightforward to see that the {q1 , . . . , qm } are unique and hence the market is complete. Remark 3 (Relation with MOM (k) derivatives) The two ways of completing the market are very related. To move from the one to the other, one should set zk (1 + r) = µk (as already noted in the previous remark). To exploit the relationship a bit more, we will brieﬂy show how to set up a MOM(k) derivative by an investment strategy in power-return assets. Suppose, that we are at time t = i − 1 and we want to generate at time i a payoff Xik , exactly like the MOM(k) i derivative is doing. In order to achieve this, at time i − 1 one should • invest −(1 + r)−1

i−1 k j =1 Xj

(k) − µk i = −(1 + r)−i Hi−1 + (1 + r)−1 µk in bond;

(k) • buy (1 + r)−i power-return assets of order k, for the total price of (1 + r)−i Hi−1 .

176

Exotic Option Pricing and Advanced L´evy Models

In order to set up this portfolio, an amount (at time i − 1) of (1 + r)−1 µk = zk is needed, exactly the same amount as the time i − 1 price of the MOM(k) i derivative. Next, we will show, under the condition (H), that if trade is allowed in the power-return assets, the market is complete, in the sense that any contingent claim can be perfectly hedged by positions in bond, stock and the power-return assets. Let us consider a general contingent claim which can depend on the complete path followed by the underlying stock, i.e. the claim is characterized by a payoff function: G(X1 , X2 , . . . , Xn ). Write the discounted payoff function in the following form: (1 + r)−n G(X1 , X2 , . . . , Xn ) = M0 +

n

aj (X1 , . . . , Xj −1 )(Xj − r)

(8.3)

j =1

+

n m−1

aj(k) (X1 , . . . , Xj −1 )(Xjk − µk ).

j =1 k=2

Note that the functions aj and aj(k) , k = 2, . . . , m − 1, only depend on X1 , X2 , . . . , Xj −1 and are thus completely known at time t = j − 1; in other words, the aj(k) s are Fj −1 measurable or ‘predictable’. Then, let us consider the martingale Mi = EQ [(1 + r)−n G(X1 , X2 , . . . , Xn )|Fi ],

i = 1, . . . , n,

where Q is the risk-neutral probability deﬁned by Condition (H). Since for j = 1, . . . , n, EQ [Xj |Fj −1 ] = r and EQ [Xjk |Fj −1 ] = µk , k = 2, . . . , m − 1, we have that Mi = M0 +

i

aj (X1 , . . . , Xj −1 )(Xj − r)

j =1

+

i m−1

aj(k) (X1 , . . . , Xj −1 )(Xjk − µk ).

j =1 k=2

We know that the discounted value of any contingent claim is a Q-martingale. Then, EQ [(1 + r)−n G(X1 , X2 , . . . , Xn )] = M0 is the initial price of the claim under consideration and (1 + r)i Mi is the time t = i price of this claim. In order to hedge the claim, one should follow the following self-ﬁnancing strategy. Just before the realization of Si , i = 1, . . . , n take the following positions in, respectively, bonds, stocks, and kth order power-return assets, k = 2, . . . , m − 1: • number of bonds = αi = Mi−1 − (1 + r)−i+1 βi Si−1 − (1 + r)−i+1

m−1

(k) βi(k) Hi−1 ,

k=2

• number of stocks = βi = (1 + r) ai (X1 , . . . , Xi−1 )/Si−1 , i

• number of kth power-jump assets = βi(k) = ai(k) (X1 , . . . , Xi−1 ), k = 2, . . . , m − 1.

Moment Derivatives and L´evy-type Market Completion

177

Note the following • The initial (t = 0) amount needed to set up the initial portfolio is: α 1 B 0 + β 1 S0 +

m−1

β1(k) H0(k) = M0 .

k=2

• Just before the realization of Si , the portfolio value is (1 + r)i−1 Mi−1 . By a straightforward calculation one can see that just after the realization of Si (and before adjusting the portfolio again), the value is given by (1 + r)i Mi . This implies that the portfolio is self-ﬁnancing. Moreover, since the value of the portfolio at time t = n equals (1 + r)n Mn = G(X1 , X2 , . . . , Xn ), the portfolio is replicating the claim. In conclusion, we have that the portfolio (αi , βi , βi(2) , . . . , βi(m−1) ; i = 1, . . . , n) is the selfﬁnancing portfolio which replicates the claim G(X1 , X2 , . . . , Xn ) and has initial value M0 .

8.3

´ THE LEVY MARKET

8.3.1 L´evy processes L´evy processes are the natural continuous time analogs of the sums of iid random variables. Basically, they are processes with the same kind of structure in the increments: stationary and independent. However, not for any general distribution, one can deﬁne such a continuous time stochastic process, where the increments follow the given distribution. We have to restrict ourselves to so-called inﬁnitely divisible distributions (see e.g. Bertoin (1996) [3] or Sato (2000) [27]) Given an inﬁnitely divisible distribution with characteristic function φ(z), one can deﬁne a stochastic process (with c`adl`ag paths), Z = {Zt , t ≥ 0}, called a L´evy process, which starts at zero, has independent and stationary increments and such that the distribution of an increment over [s, s + t], s, t ≥ 0, i.e. Zt+s − Zs , has (φ(z))t as the characteristic function. It is well known that L´evy processes are semimartingales. The function ψ(z) = log φ(z) = log E[exp(izZ1 )] is called the characteristic exponent and satisﬁes the following L´evy–Khintchine formula (see Bertoin (1996) [3]): +∞ c2 2 (exp(izx) − 1 − izx1{|x| 0, and λ > 0 exp(λ|x|)ν(dx) < ∞. (8.5) (−ε,ε)c

This implies that

+∞ −∞

|x|i ν(dx) < ∞,

i ≥ 2,

and that the characteristic function E[exp(iuXt )] is analytic in a neighbourhood of 0 and E[exp(−hZ1 )] < ∞ for all h ∈ (−h1 , h2 ), where 0 < h1 , h2 ≤ ∞. So, all moments of Zt (and Xt ) exist. 8.3.3 Power-jump processes Under our continuous-time setting, the role of the powered returns will be taken by powerjump processes. These are built from the following transformations of Z = {Zt , t ≥ 0}. We set (Zs )i , i ≥ 2, Zt(i) = 0<s≤t

Moment Derivatives and L´evy-type Market Completion

179

where Zs = Zs − Zs− , and for convenience we put Zt(1) = Zt . Note that Zt = 0<s≤t Zs is not necessarily true; it is only true in the bounded variation case (with c necessarily equal to zero). If we deﬁne Xt(i) in an analogous way, we have that Xt(i) = Zt(i) , i ≥ 2. The processes X(i) = {Xt(i) , t ≥ 0}, i ≥ 2, are again L´evy processes and are called the ith-powerjump processes (or the power-jump processes of order i). They jump at the same points as the original L´evy process, but the jumps sizes are the ith power of the jump size of the original L´evy process. We have E[Xt ] = E[Xt(1) ] = tm1 < ∞ and (see Protter (1990), p. 29 [25]) ∞ * + (Xs )i = t E Xt(i) = E x i ν(dx) = mi t < ∞, i ≥ 2. (8.6) 0<s≤t

−∞

We denote by + * Yt(i) = Zt(i) − E Zt(i) = Zt(i) − mi t,

i ≥ 1,

the compensated processes. Using Itˆo’s formula (see Chan (1999) [9] or Protter (1990) [25]) for c`adl`ag semimartingales, one can show that equation (8.4) has an explicit solution

# c2 t (1 + Xs ) exp(−Xs ). St = S0 exp Zt + b − 2 0<s≤t

In order to ensure that St > 0 for all t > 0 almost surely, we need Xt > −1 for all t. We thus need that the L´evy measure ν is supported on a subset of (−1, +∞).

´ 8.4 ENLARGING THE LEVY MARKET MODEL Suppose that we have an equivalent martingale measure Q under which Z remains a L´evy process. Under this measure, the discounted stock price process is a martingale and the ˜ moreprocess Z˜ = {Zt + (b − r)t, t ≥ 0} will be a L´evy process (with L´evy measure ν); over, the process Z˜ is a martingale. Obviously, Z˜ t = Zt and Z˜ t(i) = Zt(i) , i ≥ 2. Let (i) (i) ˜ us consider (based +∞ ion Z) the ith-power-jump processes Y = {Yt , t ≥ 0}. Note that for ˜ and we will require ν˜ to fulﬁl equation (8.5). i ≥ 2, mi = −∞ x ν(dx), We will enlarge the L´evy market with what we will call ith-power-jump assets. More precisely, we will allow trade in assets with price process H (i) = {Ht(i) , t ≥ 0} where Ht(i) = exp(rt)Yt(i) ,

i ≥ 2.

By taking a suitable linear combination of the Y (i) s, one obtains a set of pairwise strongly orthonormal martingales {T (i) , i ≥ 1} (see Protter (1990) [25]). Each T (i) is a linear combination of the Y (j ) , j = 1, 2, . . . , i: T (i) = ci,i Y (i) + ci,i−1 Y (i−1) + · · · + ci,1 Y (1) ,

i ≥ 1.

The constants ci,j can be calculated as described in Nualart and Schoutens (2000) [22]: they correspond to the coefﬁcients of the orthonormalization of the polynomials {x n , n ≥ 0}

180

Exotic Option Pricing and Advanced L´evy Models

with respect to the measure µ(dx) = x 2 ν(dx) + c2 δ0 (dx). The resulting processes T (i) = {Tt(i) , t ≥ 0} are called the orthonormalized ith-power-jump processes. In addition, we will (i) (i) denote their orthonormalized version of Ht(i) by H = {H t , t ≥ 0}, where (i)

H t = exp(rt)Tt(i) ,

i ≥ 2.

Trade in the power-jump assets can be motivated as follows. Consider the 2nd-powerjump asset. This object in some sense measures the volatility of the stock, since it accounts for the squares of the jumps. If one believes that in the future there will be a more volatile environment than the current market’s anticipation, trading the 2nd-power-jump asset can be of interest. In addition, if one would like to cover against periods of high (or low) volatility, they can be useful: Buying 2nd-power-jump assets can cover the possible losses due to such unfavourable periods. The same can be said for the higher order variation assets. Typically, the 3rd-power-jump assets is measuring a kind of asymmetry (cf. skewness) and the 4th-power-jump process is measuring extremal movements (cf. kurtosis). Trade in these assets can be of use if one likes to bet on the realized skewness or realized kurtosis of the stock: one believes that the market is not counting in asymmetry and possible extremal moves correctly. On the other hand, an insurance against a crash can be easily built from the 4th-power-jump (or ith-power jump, i ≥ 4) assets. Note, that clearly the discounted versions of the H (i) are the power-jump processes, and hence martingales: EQ [exp(−rt)Ht(i) |Fs ] = EQ [Yt(i) |Fs ] = Ys(i) ,

0 ≤ s ≤ t.

Hence, the market allowing trade in the bond, the stock and the power-jump assets remains arbitrage-free. 8.4.1 Martingale representation property Our L´evy process Z = {Zt , t ≥ 0} has the Martingale Representation Property (MRP) in terms of the orthonormalized power-jump processes (see also Nualart and Schoutens (2000, 2001) [22] [23]) that is, every square-integrable martingale M = {Mt , t ≥ 0} can be represented as follows: t ∞ t (i) Mt = M0 + hs dZ˜ s + h(i) s dTs , 0

i=2

0

where hs and h(i) s , i ≥ 2 are predictable processes. such that t |hs |2 ds < ∞ E 0

and E

% ∞ t 0 i=2

& 2 |h(i) s |

ds < ∞.

Note the similarity, except for the orthonormalization, between this MRP and equation (8.3).

Moment Derivatives and L´evy-type Market Completion

181

The MRP implies that the market enlarged with the ith-power-jump assets is complete in the sense that for every square-integrable contingent claim X we can set up a sequence of self-ﬁnancing portfolios whose values converge in L2 (Q) to X. These portfolios will consist of ﬁnite number of bonds, stocks and ith-power-jump assets. We will say, for short, that X can be replicated. Note that this notion of completeness is equivalent to the notion of approximately complete of Bj¨ork and co-workers (given in Bj¨ork et al. (1997) [4]). The details of the hedging strategy can be extracted out of the MRP. Consider a squareintegrable contingent claim X ∈ FT with maturity T . Let Mt = EQ [exp(−rT )X|Ft ]. By the MRP given above, if we deﬁne MtN := M0 +

t

hs dZ˜ s +

0

N i=2

t

0

(i) h(i) s dTs .

we have that lim MtN = Mt ,

N→∞

in L2 (Q). Deﬁne the sequence of portfolios (in terms of the orthonormalized ith-powerjump assets) φ N = {φtN = (αtN , βt , βt(2) , βt(3) , . . . , βt(N) ), t ≥ 0}, by N αtN = Mt− − βt St− e−rt − e−rt

N

N ≥2

(i)

βt(i) H t− ,

i=2

βt = e

rt

−1 ht St− ,

βt(i) = h(i) t ,

i = 2, 3, . . . , N.

Here, αtN corresponds to the number of bonds at time t, βt is the number of stocks at that time (i) and βt(i) is the number of assets H , i = 2, 3, . . . , N , one needs to hold at time t. Then, it was shown in Corcuera et al. (2005) [10] that {φ N , N ≥ 2} is the sequence of self-ﬁnancing portfolios which replicates X. In fact, the value VtN of φ N at time t is given by VtN = αtN ert + βt St +

N

(i)

βt(i) H t = ert MtN ,

i=2

and so the sequence of portfolios {φ N , N ≥ 2} is replicating the claim. Moreover, in the case of a contingent claim whose payoff is only a function of the value at maturity of the stock price, i.e. X = f (ST ), one can compute explicitly the sequence of portfolios that replicates the contingent claim. Note that the value of the contingent claim at time t is given by F (t, St ) = exp(−r(T − t))EQ [f (ST )|Ft ]; we call F (t, x) the price function of X.

182

Exotic Option Pricing and Advanced L´evy Models

Denote by D1 the differential operation with respect to the ﬁrst variable, i.e. the time variable, and by D2 the differential operator with respect to the space variable (the second variable – the stock price). Finally, denote by D the following integral operator: DF (t, x) =

+∞

−∞

(F (t, x(1 + y)) − F (t, x) − xyD2 F (t, x)) ν˜ (dy).

If f is Lipschitz and under certain degeneracy conditions (see Chapter 12 in Cont and Tankov (2004) [12]), F ∈ C 1,2 . In this case, we have that, in analogy with the Black–Scholes partial differential equation, in the L´evy market setting F will satisfy a Partial Differential Integral Equation (PDIE). More precisely, the price function (at time t) F (t, x) satisﬁes (see Chan (1999) [9], Nualart and Schoutens (2001) [23] and Raible (2000) [26]): 1 D1 F (t, x) + rxD2 F (t, x) + c2 x 2 D22 F (t, x) + DF (t, x) = rF (t, x). 2

(8.7)

with F (T , ST ) = f (ST ). The sequence of self-ﬁnancing portfolios replicating a contingent claim X, with a payoff only depending on the stock price value at maturity and a price function F (t, x) ∈ C 1,∞ which satisﬁes sup

∞

x 0, t0 > 0, is given at time t by: • number of bonds = αtN = Bt−1 F (t, St− )−St− D2 F (t, St− )−Bt−1

N i D i F (t, S ) St− t− 2

i=2

i!Bt

(i) Ht−

(8.9) • number of stocks = βt = D2 F (t, St− ), • number of ith-power-jump assets = βt(i) =

i D i F (t, S ) St− t− 2 , i!Bt

i = 2, 3, . . . , N .

Remark 4 In the Black–Scholes model, the risk-neutral dynamics of the stock price is given by the stochastic differential equation

1 dSt = r − σ 2 dt + dWt , St 2

S0 > 0,

where W = {Wt , t ≥ 0} is a standard Brownian motion. In this case, all processes H (i) , i ≥ 1 are equal to zero. Hence, it is clear that the market is already complete and that an enlargement is not necessary. Moreover the hedging portfolio is given by F (s,Ss )−SBssD2 F (s,Ss ) number of bonds and D2 F (s, Ss ) number of stocks.

Moment Derivatives and L´evy-type Market Completion

183

8.5 ARBITRAGE We assume our market is already enlarged with the power-jump assets. So, we have chosen constants a (i) , i ≥ 2 and trade is allowed in the bond, the stock and the power-jump assets with price processes Ht(i) = exp(rt)(Xt(i) − a (i) t), i ≥ 2. We investigate whether this enlargement leads to arbitrage or not. For instance, if we choose a (i) and r to equal zero this leads to arbitrage opportunities because all Ht(i) with even i are strictly increasing and starting at zero and trade is allowed in these objects. Actually, the choice of the constants a (i) may prevent arbitrage opportunities. We will discuss below how to make this choice, which is a delicate matter. No arbitrage, in the usual sense and in our portfolios with a ﬁnite number of assets, is implied by the existence of an equivalent martingale measure under which all discounted assets in the market are martingales. This question is related to the moment problem and we will give sufﬁcient conditions to ensure that there exists an equivalent martingale measure (and hence the market is arbitrage free): in continuous time the existence of an equivalent martingale measure is a sufﬁcient but not a necessary condition to ensure no-arbitrage (see Delbaen and Schachermayer (1994) [14]). The problem in its full generality seems to be very hard and challenging. 8.5.1 Equivalent martingale measures In this section, we will describe the measures, equivalent to the canonical (real world) measure under which the discounted stock price process is a martingale and under which Z remains a L´evy process. More precisely, we characterize all structure preserving P equivalent martingale measures Q under which Z remains a L´evy process and the process S˜ = {S˜t = exp(−rt)St , t ≥ 0} is an {Ft }-martingale. Since we are considering a market with ﬁnite horizon T , locally equivalence will be the same as equivalence. We have the following result (see Sato (2000), Theorem 33.1 [27]). Theorem 5 Let Z be a L´evy process with L´evy triplet [α, c2 , ν(dx)] under some probability measure P . Then the following two conditions are equivalent. (a) There is a probability measure Q equivalent to P on Ft for any t ≥ 0, such that Z is a Q-L´evy process with triplet [α, ˜ c˜2 , ν˜ (dx)]. (b) All of the following conditions hold: (i) ν(dx) ˜ = H (x)ν(dx) for some Borel function H : R→ (0, ∞). +∞ (ii) α˜ = α + −∞ x1{|x|≤1} (H (x) − 1)ν(dx) + Gc for some G ∈ R. (iii) c˜ = c. ∞ √ (iv) −∞ (1 − H (x))2 ν(dx) < ∞. The equivalent conditions in the previous theorem imply that the process W˜ = {W˜ t , t ≥ 0} with W˜ t = Wt − Gt is a Brownian motion under Q and also, if ν and ν˜ verify the condition shown in equation (8.5), the process X is a quadratic pure jump L´evy process with Doob–Meyer decomposition

+∞ ˜ Xt = Lt + a + x(H (x) − 1)ν(dx) t, −∞

184

Exotic Option Pricing and Advanced L´evy Models

where L˜ = {L˜ t , t ≥ 0} is a Q-martingale and the L´evy measure is given by ν(dx) ˜ = H (x)ν(dx). We now want to ﬁnd an equivalent martingale measure Q under which the discounted price process S˜ is a martingale. By the above theorem, under such a Q, X has the Doob–Meyer decomposition

+∞ ˜ x(H (x) − 1)ν(dx) t, Xt = Lt + a + −∞

where L˜ = {L˜ t , t ≥ 0} is a Q-martingale. Noting that Lt = L˜ t , we have

c2 S˜t = S0 exp cW˜ t + L˜ t + a + b − r + cG − t 2 +∞

# × exp t x(H (x) − 1)ν(dx) (1 + L˜ s ) exp(−L˜ s ). −∞

0<s≤t

Then, a necessary and sufﬁcient condition for S˜ to be a Q-martingale is the existence of G ∞ √ and H (x), with −∞ (1 − H (x))2 ν(dx) < ∞ such that cG + a + b − r +

+∞ −∞

x(H (x) − 1)ν(dx) = 0.

(8.10)

Remark 6 We remark (see e.g. Eberlein and Jacod (1997) [16]), that if there exists a (nonstructure preserving) locally equivalent martingale measure Q1 under which Z is not a L´evy process, there exists always a (structure preserving) locally equivalent martingale measure Q2 under which Z is a L´evy process. A sufﬁcient condition to guarantee that the enlarged market is free of arbitrage is the existence of an equivalent martingale measure Q making S˜ and the discounted H (i) s martingales. If this measure is structure preserving, the condition that the discounted stock price must be a martingale comes down to the existence of G and H (x) such that equation (8.10) holds. If we also want that the discounted H (i) s, i.e. Xt(i) − a (i) t, be martingales for i ≥ 2, using equation (8.6) together with the fact that the L´evy measure of X under Q is given by H (x)ν(dx), this comes down to +∞ x i H (x)ν(dx) = a (i) , i ≥ 2. (8.11) −∞

The question now is, do there exist G and H (x) such that equations (8.10) and (8.11) hold simultaneously? This question is related to the moment problem: given a series of numbers {µn }, ﬁnd necessary and sufﬁcient conditions for the existence of such a measure with µn as the nth moment. Another point is the uniqueness. A partial result is that if the moment problem has a solution with bounded support, then it will be unique (see Shohat and Tamarkin (1950) [30] or Ahiezer (1965) [1]). We have then the following proposition. Proposition 7 Suppose that ν(dx) has compact support: then, if there is a martingale measure in the market enlarged with the power-jump assets, the martingale measure is unique, structure preserving and the market is complete.

Moment Derivatives and L´evy-type Market Completion

185

Proof. If we have a martingale measure in the enlarged market, there exists, using the same arguments as in Eberlein and Jacod (1997) [16], an H (x) verifying equations (8.10) and (8.11) with H (x) > 0. The measure µ(dx) = x 2 H (x)ν(dx) is ﬁnite and has a bounded support. This implies that H (x) is determined by the condition shown in equation (8.11). On the other hand, since the support is bounded, H (x)ν(dx) veriﬁes equation (8.5) and the model enlarged with the power-jump assets is complete. Finally, since the contingent claim BT 1A with A ∈ F is replicable, the uniqueness of its initial arbitrage price, EQ (1A ), implies the uniqueness of the martingale measure. In general, uniqueness of the martingale measure implies completeness. Proposition 8 If the probability measure that makes the discounted stock price and the power-jump assets martingales is unique, that is, the martingale measure is unique, then the market is complete. Proof. Let Q be a martingale measure. We argue by contradiction. If the market is not complete, there exists a contingent claim X ≥ 0, X ∈ L2 (Q), not identically zero, which is orthogonal to any replicable contingent-claim. Deﬁne Q∗ (dω) = (1 + X)Q(dω). Then Q∗ is a martingale measure different from Q. In fact, for any s ≤ t, and A ∈ Fs , we have EQ∗ (1A (Yt(i) − Ys(i) )) = EQ (1A (Yt(i) − Ys(i) )) + EQ (X1A (Yt(i) − Ys(i) )) = 0, and {Yt(i) , t ≥ 0} are Q∗ -martingales for all i ≥ 2. Clearly, S˜t is also a Q∗ -martingale. 8.5.2 Example: a Brownian motion plus a ﬁnite number of Poisson processes Suppose Zt = cWt +

n

cj Nj,t ,

j =1

where c = 0, W = {Wt , t ≥ 0} a standard Brownian Motion and Nj = {Nj,t , t ≥ 0} are j = 1, . . . , n are independent Poisson processes with intensity aj > 0. The constants cj , assumed to be all different from each other and non-zero. Then, Xt = nj=1 cj Nj,t and E[X1 ] = nj=1 cj aj = a, and n Ht(i) = exp(rt) cji Nj,t − a (i) t , i = 2, 3, . . . . j =1

It is not that hard to see that Ht(i) , for i > n + 1 can be written as a linear combination of the Ht(i) , i = 2, . . . , n + 1 (see L´eon et al. (2002) [19]). In this case, we enlarge the market with only n objects, namely the assets following the price processes Ht(i) , i = 2, . . . , n + 1. In order that an equivalent martingale measure Q exists, we must have the existence of a G and H , such that +∞ x(H (x) − 1)ν(dx) = r − cG − a − b −∞

+∞ −∞

x i H (x)ν(dx) = a (i) ,

i = 2, . . . , n + 1.

186

Exotic Option Pricing and Advanced L´evy Models

The support of H will now be the set {c2 , . . . , cn+1 } and the above equations reduce to n

cj H (cj )aj = r − cG − b

j =1 n

cji H (cj )aj = a (i) ,

i = 2, . . . , n + 1.

j =1

There exists an equivalent martingale measure if the following system of equations for H (cj ), j = 1, . . . , n has a positive solution, i.e. H (cj ) > 0, j = 1, . . . , n.

c12 a1 c3 a1 1 ... c1n+1 a1

c22 a2 c23 a2 ... n+1 c2 a2

. . . cn2 an H (c1 ) . . . cn3 an × H (c2 ) ... ... ... n+1 H (cn ) . . . cn an

a (2) a (3) = ... . a (n+1)

The existence (and uniqueness by Proposition 7) of a positive solution H (cj ), j = 1, . . . , n can be translated into the condition C −1 · a > 0, where C −1 is the inverse of the Vandermonde matrix 1 1 ... 1 c1 c . . . c 2 n C= ... ... ... ... c1n−1 c2n−1 . . . cnn−1

(8.12)

and a is the transpose of [a (2) . . . a (n+1) ]. Note that if all of the ci s are different from each other (as we assumed above), that detC = 0. For the calculation of the inverse of Vandermonde matrices, see Graybill (1983) [17] or Macon and Spitzbart (1958) [20], while for other applications of Vandermonde matrices in ﬁnance see Norberg (1999) [21].

8.6 OPTIMAL PORTFOLIOS Deﬁnition 9 A utility function is a mapping U (x) : R → R ∪ {−∞} which is strictly increasing, continuous on {U > −∞}, of class C ∞ , strictly concave on the interior of {U > −∞} and satisﬁes U (∞) := lim U (x) = 0. x→∞

Denoting by dom(U ) the interior of {U > −∞}, we shall consider only two cases: Case 10 dom(U ) = (0, ∞) in which case U satisﬁes U (0) := lim U (x) = ∞. x→0+

Moment Derivatives and L´evy-type Market Completion

187

Case 11 dom(U ) = R in which case U satisﬁes U (−∞) := lim

x→−∞

U (x) = ∞. 1−p

Typical examples for Case 10 are the so-called HARA utilities, U (x) = x1−p for p ∈ R+ \{0, 1}, and the logarithmic utility U (x) = log(x). A typical example for Case 11 is U (x) = − α1 e−αx . 8.6.1 Optimal wealth Given an initial wealth w0 and an utility function U , we want to ﬁnd the optimal terminal wealth WT , that is, the value of WT that maximizes EP (U (WT )). We will consider the optimization problem

5 WT = w0 . max EP (U (WT )) : EQ BT The corresponding Lagrangian is EP (U (WT )) − λEQ

WT − w0 BT

dQT WT − w0 = EP U (WT ) − λT . dPT BT

Then, the optimal wealth is given by −1 WT = U

λT dQT BT dPT

,

where λT is the solution of the equation EQ

1 −1 U BT

λT dQT BT dPT

= w0 .

(8.13)

It is easy to check the existence and uniqueness of the optimal wealth from the conditions on U . From equation (8.5) and under certain conditions on Q (see Corcuera et al. (2004b) [11]), we can write:

G −1 m (T ) STc eVT , WT = U where

λt − Gc c2 1 2 G m (t) := a+b− t S exp − G t − Bt 0 2 c 2

+∞ G G +t (log H (x) − log (1 + x))H (x) − H (x) + 1 + x ν(dx) c c −∞

188

Exotic Option Pricing and Advanced L´evy Models

and Vt =

+∞ −∞

g(x)(Q((0, t], dx) − tH (x)v(dx)),

with g(x) := log H (x) −

G log (1 + x) . c

It can be shown (see Corcuera et al. (2004b) [11]), that if we consider HARA and exponential utilities we have that the price process of the optimal portfolio is given by Bt EQ WT |Ft = F (t, St , Vt ) BT with

F (t, x1 , x2 ) := φ(t, T ) U

−1

G c x2 m (t) x1 e + χ (t, T ).

(8.14)

We know that under an equivalent martingale measure Q, which is structure preserving, T any random variable WT ∈ L2 (, FT , Q) can be replicated and we have w0 = EQ W BT . Now, by a generalization of equation (8.9) (see Theorem 4 in Corcuera et al. (2004b) [11]) we can ﬁnd the composition of the portfolio with this price process. In fact, we have that the number of stocks and new assets are given, respectively, by G −1

βt =

c Gφ(t, T )m (t) St− eVt− G

c Vt− cU ((U )−1 (m (t) St− e ))

=

Gφ(t, T )U (Wt− ) cSt− U (Wt− )

(8.15)

and βt(i) =

G φ(t, T ) ∂ i −1 c Vt− m S U e H , (t) (y) t− i i!Bt ∂y y=0

i = 2, 3, . . .

8.6.2 Examples

(8.16)

−1 Example 12 Consider U (x) = log x. Then U (x) = U (x) = x1 . Therefore, by solving equation (8.13), we have WT = w0 BT

−1 G dPT = m(T )STc eVT . dQT

Therefore, we have that Bt dPT dPt EQ WT |Ft = w0 Bt EQ |Ft = w0 Bt = Wt BT dQT dQt and the price function of WT at time t is

−1 G c Vt Wt = m (t) St e ,

Moment Derivatives and L´evy-type Market Completion

189

that is, the wealth of the optimal portfolio at time t is the optimal terminal wealth for the period [0, t]; in other words, φ(t, T ) = 1 and χ (t, T ) = 0 in equation (8.14). Now, since U (x) = − x12 , if we apply equation (8.15), we have that βt St− G =− , Wt− c that is, the relative wealth invested in stocks is constant. From equation (8.16), the number of new assets is 1 Wt− ∂ i , i = 2, 3, . . . βt(i) = i!Bt ∂y i H (y) y=0 So, maximization with bonds and stocks corresponds to take 1

H (y) =

1−

G cy

.

where G veriﬁes (see equation (8.10)) cG + a + b − r +

G c

+∞ −∞

y2 ν(dy) = 0. 1 − Gc y

−1 1 1−p Example 13 Consider U (x) = x1−p with p ∈ R+ \ {0, 1}. Then, U (x) = x − p and by solving equation (8.13) we have WT = w0 BT

dPT dQT

EQ

p1

dPT dQT

− p1 G c VT . p1 = m(T )ST e

dPt { dQt , 0 ≤ t ≤ T } is a Q-exponential L´evy process (see Corcuera et al. (2005) [11]), and then 1 dPT p EQ dQ |F t T Bt EQ WT |Ft = w0 Bt

1 BT dPT p EQ dQT

EQ = w0 Bt

dPT ,t dQT ,t

EQ

= w0 Bt

dPt dQt

EQ

p1

dPT ,t dQT ,t

dPt dQt

p1

dPt dQt

p1

dPt dQt

p1

p1 = Wt

|Ft 1

p

190

where

Exotic Option Pricing and Advanced L´evy Models dPT ,t dQT ,t

=

dPT dQT dPt dQt

. That is, again the wealth of the optimal portfolio at time t is the optimal

terminal wealth for the period [0, t], and φ(t, T ) = 1 and χ (t, T ) = 0 in equation (8.14). Now, since U (x) = x −p and U (x) = −px −p−1 , if we apply equation (8.15), we have that βt St− G =− ; Wt− cp and by equation (8.16) the number of new assets is given by Wt− ∂ i (i) − p1 H (y) , i = 2, 3, . . . βt = i!Bt ∂y i y=0 So, we will have an optimal portfolio only based in bonds and stocks, if and only if, H (y) =

1 1−

G cp y

p ,

where G veriﬁes cG + a + b − r +

∞ −∞

y

1 p Gy 1− cp

−1

ν(dy) = 0.

Example 14 Consider the exponential utility function 1 U (x) = − e−αx α −1 with α ∈ (0, ∞). Then, U (x) = − α1 log x and by solving equation (8.13) we have WT = w0 BT +

1 α

G dPT 1 dPT log = − log m(T )STc eVT − EQ log dQT dQT α

Note, that in this case, Wt− is not bounded by below and that there arises the problem of the admissibility of this optimal portfolio (see Kallsen (2000) [18]). In addition, we have that

Bt dPT dPT Bt EQ log WT |Ft = w0 Bt + |Ft − EQ log EQ BT αBT dQT dQT

dPT ,t dPT ,t Bt EQ log = w0 Bt + |Ft − EQ log αBT dQT ,t dQT ,t

dPt dPt + log − EQ log dQt dQt

dPt dPt Bt log = w0 Bt + − EQ log αBT dQt dQt

Bt Bt . Wt + w0 Bt 1 − = BT BT

Moment Derivatives and L´evy-type Market Completion

191

Therefore, in this case φ(t, T ) = BBTt and χ (t, T ) = w0 Bt (1 − BBTt ) in equation (8.14). Now, since U (x) = e−αx and U (x) = −αe−αx , if we apply equation (8.15), we have that BT G βt St− = − , Bt cα that is, the forward value of the wealth invested in stocks is constant. From equation (8.16), the number of new assets is constant: βt(i)

BT ∂ i =− log H (y) , i i!α ∂y y=0

i = 2, 3, . . .

and we obtain the optimal portfolio based only in stocks and bonds by taking H (y) = exp

G y , c

with G verifying

G y − 1 ν(dy) = 0. cG + a + b − r + y exp c −∞

∞

The corresponding martingale measure is then the Esscher measure (see Chan (1999) [9]). Example 15 Consider the quadratic utility U (x) = γ x −

x2 , 2

x γ , it is interesting since the solution of the optimal problem with this utility is the same as that of the solution of the mean-variance portfolio problem if we choose γ =

w0 ((1 + ρ)EQ (ξT ) − (1 + r)) EQ (ξT ) − 1

where ρ > r is a speciﬁed return (see Pliska (1997) [24]).

REFERENCES [1] Ahiezer, N.I. (1965), The Classical Moment Problem, Hafner Publishing Company, New York, NY, USA. [2] Balland, P. (2002), “Deterministic implied volatility models”, Quantitative Finance, 2, 31–44. [3] Bertoin, J. (1996), L´evy Processes, Cambridge Tracts in Mathematics, 121, Cambridge University Press, Cambridge, UK. [4] Bj¨ork, T., Di Masi, G., Kabanov, Y. and Runggaldier, W. (1997), “Towards a general theory of bond markets”, Finance and Stochastics 1, 141–174. [5] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities”, Journal of Political Economy, 81, 637–654. [6] Carr, P. and Lewis, K. (2004), “Corridor variance swaps”, Risk, 17(2), 67–72. [7] Carr, P. and Madan, D. (1998), “Towards a theory of volatility trading”, in R. Jarrow (Ed.), Volatility, Risk Publications, London, UK, pp 417–427. Reprinted (2001) in M. Musiela, E. Jouini and J. Cvitanic (Eds), Option Pricing, Interest Rates and Risk Management, Cambridge University Press, Cambridge, UK, pp. 458–476. [8] Carr, P., Geman, H., Madan, D.H. and Yor, M. (2002), Pricing Options on Realized Variance, Pr´epublications du Laboratoire de Probabilit´es et Mod`eles Al´eatoires, 768, Universit´es de Paris 6 and Paris 7, Paris, France. [9] Chan, T. (1999), “Pricing contingent claims on stocks driven by L´evy processes”, Annals of Applied Probability, 9, 504–528 (1999). [10] Corcuera, J.M., Nualart, D. and Schoutens, W. (2005), “Completion of a L´evy market by powerjump assets”, Finance and Stochastics, 9, 109–127.

Moment Derivatives and L´evy-type Market Completion

193

[11] Corcuera, J.M., Guerra J., Nualart, D. and Schoutens, W. (2004), “Optimal investment in a L´evy market”, Preprint Institut de Matem´atica de la Universitat de Barcelona 249, Universitat de Barcelona, Barcelona, Spain. [12] Cont, R. and Tankov, P. (2004), Financial Modelling with Jump Processes, Chapman & Hall/CRC Press, Boca Raton, FL, USA. [13] Davis, M. (1997), “Option pricing in incomplete markets”, in M.A.H. Dempster and S.R. Pliska (Eds), Mathematics of Derivative Securities, Cambridge University Press, Cambridge, UK, pp. 227–254. [14] Delbaen, F. and Schachermayer, W. (1994), “A general version of the fundamental theorem of asset pricing”, Mathematische Annalen, 300, 463–520. [15] Demeterﬁ, K., Derman, E., Kamal, M. and Zhou, J. (1999), “A guide to volatility and variance swaps”, The Journal of Derivatives, 6(4), 9–32. [16] Eberlein, E. and Jacod, J. (1997), “On the range of option prices”, Finance and Stochastics, 1, 131–140. [17] Graybill, A. (1993), Matrices with Applications to Statistics, 2nd Edn, Wadsworth, Belmont, CA, USA. [18] Kallsen, J. (2000), “Optimal portfolios for exponential L´evy processes”, Mathematical Methods of Operations Research, 51, 357–374. [19] L´eon, J.A., Vives, J., Utzet, F. and Sol´e, J.L. (2002), “On L´evy processes, Malliavin calculus and market models with jumps”, Finance and Stochastics, 6, 197–225. [20] Macon, N. and Spitzbart, A. (1958), “Inverses of Vandermonde matrices”, American Mathematics Monthly, 65, 95–100. [21] Norberg, R. (1999), On the Vandermonde matrix and its role in mathematical ﬁnance Working paper, Laboratory of Actuarial Mathematics (working paper), University of Copenhagen, Denmark. [22] Nualart, D. and Schoutens W. (2000), “Chaotic and predictable representations for L´evy processes”, Stochastic Processes and their Applications, 90, 109–122. [23] Nualart, D. and Schoutens W. (2001), “Backwards stochastic differential equations and Feynman–Kac formula for L´evy processes, with applications in ﬁnance”, Bernoulli, 7, 761–776. [24] Pliska, S. (1997), Introduction to Mathematical Finance, Blackwell, Oxford, UK. [25] Protter, Ph. (1990), Stochastic Integration and Differential Equations, Springer-Verlag, Berlin, Germany. [26] Raible, S. (2000), “L´evy process in ﬁnance: theory, numerics, and empirical facts”, PhD Thesis, University of Freiburg, Freiburg, Germany. [27] Sato, K. (2000), L´evy Processes and Inﬁnitely Divisible Distributions, Cambridge Studies in Advanced Mathematics, 68, Cambridge University Press, Cambridge, UK. [28] Schachermayer, W. (2001), “Optimal investment in incomplete ﬁnancial markets”, in H. Geman, D. Madan, S. Pliska and T. Vorst (Eds), Mathematical Finance – Bachelier Congress 2000, SpringerVerlag, Berlin, Germany, pp. 427–462. [29] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [30] Shohat, J.A and Tamarkin, J.D. (1950), The Problem of Moments, Mathematical Surveys No. 1, American Mathematical Society, New York, NY, USA.

9 Pricing Perpetual American Options Driven by Spectrally One-sided L´evy Processes† Terence Chan Heriot-Watt University, Edinburgh, UK Abstract This paper considers the problem of pricing perpetual American put options on stocks ´ whose price process is the exponential of a Levy process (i.e. a process with stationary independent increments). When the price process has no negative jumps, the problem reduces to one of ﬁnding the law of a relevant ﬁrst-passage time of the process, a problem which has already been well-studied. However, if the price process does have negative jumps, the problem is much more delicate as it involves ﬁnding the joint law of a ﬁrstpassage time and position of the process at that time. This problem is only mathematically tractable under the assumption that the price process has no positive jumps. A renewal ´ process equation for the price is obtained in the case where the jump component of the Levy has ﬁnite variation. In the general case, a simple explicit formula is obtained for the optimal exercise boundary and a formula amenable to efﬁcient numerical computation is obtained for the price of a perpetual put.

9.1

INTRODUCTION

Consider a (non dividend-paying) stock whose price at time t, St , is modelled as St = S0 exp{−Yt }, where Yt is a L´evy process (process with independent stationary increments) of the form Yt = σ Bt + Xt + ct,

Y0 = 0,

(9.1.1)

where Bt is a standard Brownian motion and Xt is a jump process with stationary independent increments. We shall suppose that the market is already risk-neutral, so that for some discount factor δ > 0, e−δt St is a martingale. Of course, such a model is incomplete: there are many equivalent martingale measures and contingent claims cannot be hedged perfectly. However, the purpose of this paper is not to address the problems associated with incompleteness of the market in this model; in particular, it does not deal with the question of how to choose a suitable equivalent martingale measure from the inﬁnitely many available – this problem has been studied in, for example, Chan (1999) and the references cited there. Instead, the present article is concerned with the next step in pricing a contingent claim, namely, once an equivalent martingale measure has been chosen, how to calculate the expected payoff with respect to the chosen martingale measure. It is shown in Chan †

This paper was submitted at the special invitation of the editors. Please see Epilogue on page 215 for details.

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

196

Exotic Option Pricing and Advanced L´evy Models

(1999) that for this model, a L´evy process under the original measure remains a L´evy process under any equivalent martingale measure. We may therefore assume that e−δt St is already a martingale. (However, in the context of perpetual options considered here, it is perhaps worth bearing in mind that while an equivalent martingale measure is equivalent to the actual underlying probability measure over every ﬁnite time interval [0, T ], these two measures are mutually singular over [0, ∞).) In this paper, we consider the problem of pricing perpetual American options. These were ﬁrst studied by Samuelson (1965), in relation to call options. In recent years, there has been renewed interest in this problem: some recent work includes Gerber and Shiu (1994, 1998) and Gerber and Landry (1998). While the implication of Samuelson (1965) seems to be that a few perpetual warrants (i.e. call options) once did exist, the main interest in perpetual put options in recent years arises from their relative mathematical tractability compared with ﬁnitely dated puts and the usefulness of the former in pricing the latter. The problem of pricing an American put option with a ﬁnite maturity presents many difﬁcult mathematical problems because the optimal exercise level depends on the time to maturity and there are no reasonable models for which an explicit pricing formula is known. The article by Gerber and Shiu (1994) contains a long list of references to the literature on American options with ﬁnite maturity. The problem of pricing perpetual American puts is a much more mathematically tractable problem because the optimal exercise level is constant in time. The price of a perpetual American option with payoff function (s) (under the equivalent martingale measure) is simply supτ E[e−δτ (Sτ )] (in the case of payoffs depending only on the stock price at the time of exercise), where the supremum is taken over all stopping times τ . The motivation for studying perpetual American options is three-fold. First, at the most obvious level, perpetual options can be thought of as approximations to options with a long time until expiry. Secondly, the mathematics associated with perpetual options have some applications to the ruin problem for certain generalizations of the classical model in ruin theory. Some recent work that have explored the connection between perpetual American options and ruin theory are Gerber and Landry (1998) and Gerber and Shiu (1997). The third and arguably most compelling motivation for studying perpetual American (put) options lies in their applications to numerical methods for pricing American puts with ﬁnite maturities (not necessarily long-dated). Various numerical approximation schemes developed in recent years involve evaluating perpetual options. MacMillan (1986) and Zhang (1995) consider the difference between the prices of a ﬁnite-maturity American option and a European option and show that under a suitable discretization, this difference can be interpreted as the price of a perpetual American option. Even more strikingly, Carr (1998) presents a recursive algorithm which involves calculating the prices of a sequence of American puts (with different payoffs) expiring at a sequence of independent exponential times. The key observation behind this idea of Carr is that the memoryless property of the exponential distribution reduces the problem with a random exponential time horizon to one with an inﬁnite time horizon and an adjusted discount factor and payoff. To see this, let T (λ) be an exponential time with rate λ, independent of Y ; then the price of an American (λ) option expiring at T (λ) is supτ E[e−δ(τ ∧T ) (Sτ ∧T (λ) )] (where x ∧ y = min(x, y)). Writing (λ) (St ) = (Yt ) for notational convenience, deﬁne U (y) = Ey [e−δT (YT (λ) )] (where Ey denotes expectation given Y0 = y). The strong Markov property at stopping time τ says that Y˜s = Ys+τ − Yτ is independent of {Yu : u ≤ τ } and τ . Hence we may calculate as follows: Ey [e−δ(τ ∧T

(λ) )

(Yτ ∧T (λ) )]

Pricing Perpetual American Options

197

(λ)

= Ey [e−δτ (Yτ )1{τ

1

x ν(dx) = lim xν(x, 1] + x→0

0

ν(x, 1] dx,

(9.2.19)

0

1 so that 0 ν(x, 1] dx < ∞, which in turn implies that limx→0 xν(x, 1] = 0. Integrating by parts and noting that 1 − e−θx ∼ θ x for small x gives

∞

ψ(θ ) = −

(1 − e−θx ) ν(dx) − cθ

0

= −θ

1

e 0

−θx

∞

ν(x, 1] dx −

(1 − e−θ x ) ν(dx) − cθ.

(9.2.20)

1

1 ∞ Since 0 e−θ x ν(x, 1] dx → 0 as θ → ∞ and 1 (1 − e−θ x ) ν(dx) remains bounded as θ → ∞, the result shown in equation (9.2.16) follows.

202

Exotic Option Pricing and Advanced L´evy Models

On the other hand, if Y has inﬁnite variation, then the integral in equation (9.2.19) 1 diverges, and so either 0 ν(x, 1] dx = ∞, or limx→0 xν(x, 1] = ∞, which also implies 1 that 0 ν(x, 1] dx = ∞. Integrating by parts as before gives σ 2θ 2 ψ(θ ) = − bθ − 2 =

σ 2θ 2 −θ 2

1

1

(1 − e

−θx

− θ x) ν(dx) −

0

∞

(1 − e−θ x ) ν(dx) − cθ

1

(e−θx − 1)ν(x, 1] dx −

0

∞

(1 − e−θ x ) ν(dx) + O(θ ).

(9.2.21)

1

1 The same argument as before, only this time 0 (e−θ x − 1)ν(x, 1] dx → ∞ as θ → ∞, shows that equation (9.2.17) holds. (ii) The same sort of argument as used for equation (9.2.19), but this time integrating by parts twice, shows that

1

∞>

1 1

x 2 ν(dx) = 2

0

ν(y, 1] dy dx. x

0

Hence, integrating by parts once more at equation (9.2.21) gives ψ(θ ) =

σ 2θ 2 −θ 2

σ 2θ 2 = + θ2 2 Since

9.3

1 0

e−θ x

1 x

1

(e−θx − 1)ν(x, 1] dx −

0

∞

(1 − e−θ x ) ν(dx) + O(θ )

1

1

e 0

−θx

1

∞

ν(y, 1] dy dx −

x

(1 − e−θ x ) ν(dx) + O(θ ). (9.2.22)

1

ν(y, 1] dy dx → 0 as θ → ∞, the result (equation (9.2.18)) follows.

DESCRIPTION OF THE MODEL, BASIC DEFINITIONS AND NOTATIONS

The price process of a stock is given by St = S0 exp{−Yt } = exp{y − Yt }, where Yt = σ Bt + Xt + ct,

Y0 = 0,

(9.3.1)

is a L´evy process satisfying the basic Assumption 9.2.1. In addition, we assume that the market is risk-neutral – in other words, e−δt St = ey−δt−Yt is a martingale. Since e−Yt −ψ(1)t is a martingale, this requires ψ(1) = δ,

$⇒ φ(δ) = 1.

(9.3.2)

In order to achieve this, we require the drift c to be c = −δ +

σ2 + ψX (1). 2

Note that if Y has ﬁnite variation, then c < 0.

(9.3.3)

Pricing Perpetual American Options

203

We shall consider only options whose payoff is a bounded function of the stock price at the time exercise only. For notational convenience, we shall write the payoff of the option as a function of the logarithm of the price at the time of exercise: thus, if the price of the stock at time of exercise is s, the payoff is given by (w), where w = log s. The expected value of the discounted payoff obtained from exercising the option at a stopping time τ is given by E[e−δτ (y − Yτ )]. The price of such an option is then maxτ E[e−δτ (y − Yτ )]. We shall consider only those options whose payoff (w) is a decreasing function of w (i.e. options of ‘put’ type). For such options, it is clear from the form of the expected payoff and the fact that Y has stationary independent increments that the optimal time of exercise is a stopping time of the form τL = inf{t ≥ 0 : St ≤ L} for some L. It is equally clear that the optimal value of the exercise level L cannot depend on the initial stock price S0 . Writing L = ea , we see that under our model, τL = Ty−a = inf{t ≥ 0 : Yt ≥ y − a}.

(9.3.4)

For a ﬁxed choice of a, let V (y, a) denote the expected value of the discounted payoff obtained from exercising the option at time Ty−a : V (y, a) = E[e−δTy−a (y − Y (Ty−a ))].

(9.3.5)

The price of such an option is then maxa V (y, a). The problem is then to ﬁnd the optimal exercise level a which maximizes V (y, a). Notice that there is a small but important difference between the deﬁnition of Tx as shown in equation (9.3.4) and that of T˜x as in equation (9.2.13): Y (Tx ) ≥ x whereas Y (T˜x ) > x. It is easy to see that Tx = T˜x almost surely except when x = 0 and Y has ﬁnite variation; in the latter case, T0 = 0 by deﬁnition whereas, because c < 0 when Y has ﬁnite variation, Y will not become positive immediately (nor will it hit 0 again immediately), so that T˜0 > 0 almost surely and by letting x ↓ 0 in equation (9.2.15) we can obtain the joint law of T˜0 and Y (T˜0 ). We have chosen the deﬁnition of Tx so as to ensure that V (a, a) = (a), which is consistent with the fact that if the initial stock price S0 = ey is at (or below) the chosen exercise level L = ea , the option is exercised immediately at time 0, resulting in a payoff (y). However, it must be emphasized that V (a, a) = (a) is purely a consequence of the deﬁnition of Ty−a , which provides a neat way of expressing the expected payoff associated with a chosen exercise strategy – in particular, the statement that V (a, a) = (a) is not the same as the continuous junction condition discussed in Gerber and Shiu (1998), which says that ˜

V (a+, a) = lim V (y, a) = E[e−δ T0 (a − Y (T˜0 ))] = (a), y→a+

(9.3.6)

From the deﬁnition of Ty−a , we already have V (a−, a) = limy→a− V (y, a) = (a), and so if equation (9.3.6) holds, the function y → V (y, a) would be continuous at y = a. If

204

Exotic Option Pricing and Advanced L´evy Models

Y has inﬁnite variation, then since T0 = T˜0 = 0 almost surely, V (y, a) is actually jointly continuous in y and a and the continuous junction condition (equation (9.3.6)) is always satisﬁed; on the other hand, if Y has ﬁnite variation, then in general V (a+, a) = (a). However, as we shall see in Section 9.5, there is a unique a ∗ for which lim V (y, a ∗ ) = (a ∗ )

y→a ∗ +

and this a ∗ turns out to be the optimal exercise level (when Y has ﬁnite variation). The ∗ meaning of this special a ∗ is that, when the initial stock price is at the level L∗ = ea , the ∗ normal rule is to exercise the option immediately, resulting in a payoff (a ); however, we would get the same payoff if we waited until the stock price has actually fallen below L∗ – the extra advantage in having S(T˜0 ) < L∗ is exactly counter-balanced by the discount ˜ factor e−d T .

9.4 A RENEWAL EQUATION APPROACH TO PRICING Throughout this section, in addition to the basic Assumption 9.2.1, we make the following assumption. Assumption 9.4.1 The jump component X of Y has ﬁnite variation: thus, equation (9.2.10) holds. We derive a renewal equation for V (y, a), the expected value of the payoff from exercising a perpetual option at level a. This is essentially the same renewal equation as obtained by Gerber and Landry (1998) in the case where X is a compound Poisson process; we show that there is an easy generalization to any jump process with ﬁnite variation. Theorem 9.4.1 Suppose Assumptions 9.2.1 and 9.4.1 hold. Let β=−

2 (ψX (1) − δ) σ2

and let 2 h(s) = 2 e−βs , σ

γ (s) = e

s

∞

e−x ν(dx),

s

and

z

g(z) = 0

2 h(z − s)γ (s) ds = 2 ez σ

z

e

−(β+1)(z−s)

∞

e−x ν(dx) ds.

(9.4.1)

s

0

Then, V (y, a), for y > a, satisﬁes

y−a

V (y, a) =

V (y − z, a)g(z) dz + e−β(y−a) (a)

0

+

∞ y−a

(y − z)g(z) dz − e−β(y−a)

0

∞

(a − z)g(z) dz.

(9.4.2)

Pricing Perpetual American Options

205

Proof. We shall approximate X by a compound Poisson process and use the corresponding result of Gerber and Landry (1998). Thus, if ν is the L´evy measure of X, let Xn (t) be a compound Poisson process with jump rate λn = ν(n−1 , ∞) and jump size distribution Pn (dx) = λ−1 n 1{x>1/n} ν(dx). Note that λn Pn (dx) = 1{x>1/n} ν(dx). Let Yn (t) = σ Bt + Xn (t) + cn t, where the drift parameter cn satisﬁes equation (9.3.3) with X replaced by Xn . Then, Yn is precisely the process used by Gerber and Landry (1998) to model the logarithm of the stock price St . Let βn = −

2 (ψXn (1) − δ) σ2

and let

2 hn (s) = 2 e−βn s , σ

γn (s) = λn e

∞

s

e−x Pn (dx),

s

and

z

gn (z) = 0

2λn hn (z − s)γn (s) ds = 2 ez σ

z

e

−(βn +1)(z−s)

∞

e−x Pn (dx) ds.

(9.4.3)

s

0

Deﬁne Vn (y, a) exactly as in equation (9.3.5) but with Y replaced by Yn . Then Gerber and Landry (1998) showed that Vn satisﬁes

y−a

Vn (y, a) =

Vn (y − z, a)gn (z) dz + e−βn (y−a) (a)

0

+

∞

(y − z)gn (z) dz − e

y−a

−βn (y−a)

∞

(a − z)gn (z) dz

(9.4.4)

0

for y > a. To ﬁnish the proof, we only have to let n → ∞. First, if ψn denotes the Laplace exponent of Yn , it is easy to see that ψn (θ ) → ψ(θ ) for θ ≥ 0. This is equivalent to the weak convergence of Yn to Y under a suitable topology, the J1 -topology of Skorohod (see Billingsley (1968)). Next, let Tx (Yn ) = inf{t ≥ 0 : Yn (t) ≥ x}. As a functional of Yn , the ﬁrst-passage time functional Tx (·) is continuous in the J1 -topology (see Whitt (1971)) and hence Tx (Yn ) converges weakly under J1 to Tx (Y ) = Tx . These facts together imply that Vn (y, a) → V (y, a). In addition, βn → β and hn (s) → h(s), γn (s) → γ (s) pointwise. From the forms of hn and γn , it is easy to see that hn and γn – and hence gn can be bounded by integrable functions and hence gn → g by dominated convergence theorem. Finally, since gn is bounded by an integrable function, letting n → ∞ in equation (9.4.4) and using the dominated convergence theorem for the integrals on the right-hand side gives equation (9.4.2). Note that if X has inﬁnite variation, the integral in the deﬁnition of g diverges; this can be most readily seen if we interchange the order of integration and write 2 g(z) = 2 e−βz σ

∞

e 0

−x

e(β+1) min(x,z) − 1 β +1

ν(dx).

There is a probabilistic explanation for why it is necessary to assume that X has ﬁnite variation in Theorem 9.4.1, which is related to the probabilistic interpretation of the function

206

Exotic Option Pricing and Advanced L´evy Models

gn deﬁned at equation (9.4.3). Let Jn denote the ﬁrst time when Yn attains a record high via a jump: formally Jn = inf{t ≥ 0 : Yn (t) = Sn (t), Yn (t−) = Sn (t−)}, where Sn (t) = sups≤t Yn (s). Consider the joint density function fn (y, t) of Yn (Jn ) and Jn . Then Gerber and Landry (1998) showed that ∞ e−δt fn (y, t) dt. gn (y) = 0

By conditioning on Yn (Jn ), the level of the ﬁrst record high in Yn caused by a jump, Gerber and Landry (1998) derived the renewal equation shown by equation (9.4.4). The same probabilistic interpretations ∞ hold when we let n → ∞: thus, if J = inf{t ≥ 0 : Yt = St , Yt− = St− }, then g(y) = 0 e−δt f (y, t) dt where f (y, t) is the joint density of Y (J ) and J . However, if X has inﬁnite variation, it makes a jump which causes a new record immediately, and so J = 0 and Y (J ) = 0 almost surely. Thus, there is a fundamental obstruction to using the conditioning argument of Gerber and Landry (1998) when X has inﬁnite variation, rather than it being merely a question of certain expressions not behaving well in the limit. To ﬁnd the optimal exercise level a ∗ , we can use the smooth pasting condition, which says that the function y → V (y, a) has a continuous ﬁrst derivative at the optimal boundary a ∗ : lim∗

y→a

∂V (y, a ∗ ) = (a ∗ ). + ∂y

By differentiating the right-hand side of equation (9.4.2), putting y = a and equating this with (a), we obtain an equation for a which is identical to that obtained by Gerber and Landry (1998). In the case of an American put option, we have (a) = (K − ea )+ and the optimal exercise level a ∗ is given by ∗

ea =

σ2 − c −

Kδ ∞ −x ν(dx) 0 xe

(9.4.5)

(where c is given by equation (9.3.3)), which is just the obvious extension of the formula obtained by Gerber and Landry (1998) to the present situation. If σ = 0 (so Y has ﬁnite variation), then since c < 0, the time of the ﬁrst record high caused by a jump is just T˜0 , the time of the ﬁrst jump by Y above its initial level 0. In this case, the same method of approximation by compound Poisson processes, but this time using the results of Gerber and Shiu (1998), shows that V satisﬁes the following simpler renewal equation ∞ y−a V (y − z, a)g(z) ˜ dz + (y − z)g(z) ˜ dz (9.4.6) V (y, a) = y−a

0

where ez g(z) ˜ = c

∞

e−x ν(dx).

z

The function g˜ has a similar probabilistic interpretation: g(y) ˜ = f˜(y, t) is the joint density of Y (T˜0 ) and T˜0 .

∞ 0

e−δt f˜(y, t) dt where

Pricing Perpetual American Options

207

9.5 EXPLICIT PRICING FORMULAE FOR AMERICAN PUTS In this section, we present explicit formulae for the optimal exercise level and value of a perpetual American put option, assuming only Assumption 9.2.1. Recall that the payoff function considered here is (w) = (K − ew )+ and that rθ(x) denotes the resolvent density ∞ of Y ; in other words, the density function of the measure 0 e−θ t P(Yt ∈ dx) dt. The main result is encapsulated in the following theorem. Theorem 9.5.1 Suppose that Assumption 9.2.1 holds and consider a perpetual American put option with payoff (y − YT ) = (K − ey−YT )+ . Then (i) the optimal exercise level is given by ∗

L∗ = e a =

Kδ ψ (1)

(9.5.1)

and for y > a ∗ , the value of a perpetual put option is given by ∞ V (y, a ∗ ) = Kδ rδ (z) dz;

(9.5.2)

y−a ∗

(ii) if Y has inﬁnite variation, the optimal exercise level a ∗ is uniquely determined by the smooth pasting condition lim∗

y→a

∂V (y, a ∗ ) ∗ = (a ∗ ) = −ea ; + ∂y

(9.5.3)

(iii) if Y has ﬁnite variation, V (y, a ∗ ) does not satisfy equation (9.5.3) – instead, the optimal exercise level a ∗ is uniquely determined by the continuity condition ∗

V (a ∗ +, a ∗ ) = lim∗ V (y, a ∗ ) = (a ∗ ) = K − ea . y→a +

(9.5.4)

Proof. (i) Suppose the option is exercised at level a at time Ty−a as described in Section 9.3 and we may assume that a < log K. Then the value function is given by V (y, a) = E[e−δT (y − YT )] = KE[e−δT ] − ey E[e−δT −YT ], where we have put T = Ty−a . For y > a, the terms on the right-hand side above are given by respectively putting θ = δ, η = 0 and θ = δ, η = 1 in equation (9.2.15) and noting the relationship shown in equation (9.3.2): ∞ rδ (z) dz − ψ (1)ea rδ (y − a). (9.5.5) V (y, a) = Kδ rδ (y − a) + y−a

To ﬁnd the optimal value of a so as to maximize V (y, a), we differentiate equation (9.5.5) with respect to a to ﬁnd ∂V (y, a) = (Kδ − ψ (1)ea )(rδ (y − a) − rδ (y − a)) = 0. ∂a

(9.5.6)

208

Exotic Option Pricing and Advanced L´evy Models

The only solution to the above which does not depend on y is a ∗ as given by equation (9.5.1) and upon substitution into equation (9.5.5), we obtain equation (9.5.2). Finally, because the function a → V (y, a) has a discontinuity at a = y when Y has ﬁnite variation, we need to check that the optimal value of V (y, a) is not in fact (y) when y > a ∗ . To this end, we show that V (y, y−) > (y), which implies that the optimal exercise level must be strictly less than y. Letting a → y− in equation (9.5.5) gives

∞ rδ (z) dz − ψ (1)ey rδ (0+). (9.5.7) V (y, y−) = Kδ rδ (0+) + 0

Putting λ = 0 into equation (9.2.14) and using equation (9.3.2) shows that ∞ 1 1 rδ (z) dz = − . δ ψ (1) 0

(9.5.8)

To ﬁnd rδ (0+), we use equation (9.2.14) together with equations (9.2.16) and (9.3.2) to obtain ∞ λ φ (δ)λ −λx − rδ (0+) = lim λ e rδ (x) dx = lim λ→∞ λ→∞ δ − ψ(λ) φ(δ) − λ 0 =

1 1 1 + φ (δ) = + . c c ψ (1)

(9.5.9)

Substituting equations (9.5.9) and (9.5.8) into equation (9.5.7) gives V (y, y−) =

Kδ − ψ (1)ey + K − ey . c

(9.5.10)

Since y > a ∗ , Kδ − ψ (1)ey < 0 and recall that c < 0 when Y has ﬁnite variation. Hence, equation (9.5.10) shows that V (y, y−) > (y) = (K − ey )+ . (ii) Differentiating equation (9.5.2) with respect to y shows that ∂V (y, a ∗ ) = −Kδrδ (y − a ∗ ), ∂y and hence lim

y↓a ∗

∂V (y, a ∗ ) = −Kδrδ (0+). ∂y

(9.5.11)

To ﬁnd rδ (0+), calculating as in equation (9.5.9) but this time using equation (9.2.17) gives ∞ λ ψ∗ (δ)λ −λz − rδ (0+) = lim λ e rδ (z) dz = lim λ→∞ λ→∞ δ − ψ(λ) ψ∗ (δ) − λ 0 = ψ∗ (δ) =

1 ψ (1)

.

(9.5.12)

Substituting this into equation (9.5.11) and using equation (9.5.1) immediately gives equation (9.5.3).

Pricing Perpetual American Options

209

(iii) That a ∗ is uniquely determined by equation (9.5.4) follows immediately from equation (9.5.10). It is also immediately apparent upon substituting equation (9.5.9) into equation (9.5.11) that equation (9.5.3) does not hold. We leave the reader to check that equation (9.5.1) agrees with equation (9.4.5) for the case where X has ﬁnite variation, and also that V (a+, a) = (a) for all a when Y has inﬁnite variation. Furthermore, note that the formula shown in equation (9.5.5) provides an explicit solution to the renewal equation (equation 9.4.2) when the jump component X has ﬁnite variation. Of course, in order to actually evaluate equation (9.5.2), one still has to compute the resolvent density rδ (x) for x > 0, for which there is rarely an explicit formula – unlike the simple formula rθ (x) = φ (θ )eφ(θ)x for x < 0 (e.g. see Section 6 of Bingham (1975)). However, very often – and certainly for all the examples considered in the next section – rδ (x) for x > 0 can be computed easily by Fourier inversion. First, observe that the Fourier transform of rδ (x) is given by ∞ −izx −izx rˆδ (z) = e rδ (x) dx = e e−δt P(Yt ∈ dx) dt =

R ∞

R

e−δt E[e−izYt ] dt =

0

Therefore, if

0 ∞

˜

e−(δ−ψ(z))t dt =

0

1 . ˜ δ − ψ(z)

1 dz < ∞, δ − ψ(z) ˜ R

rδ (x) can be recovered by using the Fourier inversion formula eixz 1 1 ixz rδ (x) = e rˆδ (z) dz = dz, ˜ 2π R 2π R δ − ψ(z)

(9.5.13)

(9.5.14)

(9.5.15)

which can be readily computed using a fast and efﬁcient numerical algorithm like the Fast Fourier Transform (FFT). In particular, equation (9.5.14) is true if σ = 0; however, note that the latter equation implies that Y has inﬁnite variation, since equation (9.2.16) implies ˜ 1/|δ − ψ(z)| dz = ∞ when Y has ﬁnite variation. Of course, one could also try to invert the Laplace transform at equation (9.2.14) but this is generally much harder computationally than Fourier inversion. In principle, other payoff functions besides (w) = (K − ew )+ can be treated in the same way but in practice, this is much harder at the computational level as it entails the inversion of the Laplace transform (equation (9.2.15)).

9.6 SOME SPECIFIC EXAMPLES We consider the following examples of L´evy processes for the jump component X. Gamma process A process is called a Gamma (α, β) process if its L´evy measure is ν(dx) = αx −1 e−βx dx,

x > 0.

(9.6.1)

210

Exotic Option Pricing and Advanced L´evy Models

Note that it has ﬁnite variation. It is well-known that Xt has the following Gamma distribution P(Xt ∈ dx) =

β αt αt−1 −βt x e dx.

(αt)

The L´evy exponent is therefore given by

iθ ψ˜ X (θ ) = −α log 1 + β

(9.6.2)

while its Laplace exponent is given by

θ ψX (θ ) = −α log 1 + . β

(9.6.3)

Stable process: index α ∈ (0, 1) A (spectrally positive) process is called stable of index α if its L´evy measure is ν(dx) = βx −α−1 dx,

x > 0.

Thus, if α ∈ (0, 1), it has ﬁnite variation. Its Laplace exponent is given by ∞ βθ ∞ −θ x −α β (1 − α) α θ . ψX (θ ) = −β (1 − e−θx )x −α−1 dx = − e x dx = − α α 0 0 (9.6.4) It is more convenient if we write the L´evy measure as ν(dx) =

βα x −α−1 dx.

(1 − α)

(9.6.5)

The Laplace exponent is then ψX (θ ) = −βθ α .

(9.6.6)

To ﬁnd the L´evy exponent, ﬁrst note the following identity (see Erd´elyi (1954), Section 2.3) ∞ πα . (9.6.7) x −α sin(θ x) dx = sgn(θ )|θ |α−1 (1 − α) cos 2 0 Hence, integrating by parts, ∞ x −α (cos(θ x) − 1) dx = 0

θ 1−α

∞

x −(α−1) sin(θ x) dx

0

π(α − 1)

(2 − α) cos = |θ |α−1 1−α 2 πα α−1 = |θ | (1 − α) sin . 2

(9.6.8)

Pricing Perpetual American Options

211

To ﬁnd the L´evy exponent, we have ∞ βα ˜ ψX (θ ) = (e−iθx − 1)x −(α+1) dx

(1 − α) 0 ∞

∞ βα = x −(α+1) (cos(θ x) − 1) dx − i x −(α+1) sin(θ x) dx

(1 − α) 0 0

π(α + 1) π(α + 1) βα α α |θ | (−α) sin − i sgn(θ )|θ | (−α) cos =

(1 − α) 2 2 πα πα . (9.6.9) + i sgn(θ ) sin = −β|θ |α cos 2 2 (We have also used the identity (z + 1) = z (z) in the above calculation.) Stable process: index α ∈ (1, 2) This is arguably the most interesting of the examples considered here, as it is the only example where the jump component X has inﬁnite variation. The L´evy measure here is given by ν(dx) = −

βα x −α−1 dx,

(1 − α)

x > 0.

(9.6.10)

(Note that if α ∈ (1, 2), (1 − α) < 0.) Such stable processes are purely discontinuous martingales, which means that in equations (9.2.4) and (9.2.8) ∞ b=− x ν(dx) 1

and we write the Laplace exponent of X as ∞ βα ψX (θ ) = (1 − e−θ x − θ x)x −(α+1) dx

(1 − α) 0 ∞ βθ = (e−θx − 1)x −α dx = βθ α ,

(1 − α) 0

(9.6.11)

where we have integrated by parts and used equation (9.6.4). Note the change in sign from the stable 0 < α < 1 case. To ﬁnd the L´evy exponent, we integrate by parts and use equation (9.6.9): ∞ βα ˜ ψX (θ ) = (1 − e−iθx − iθ x)x −(α+1) dx

(1 − α) 0 ∞ β(iθ ) = (e−iθx − 1)x −α dx

(1 − α) 0 β(iθ ) α−1 πα πα |θ | (1 − α) sin − i sgn(θ )|θ |α−1 (1 − α) cos =

(1 − α) 2 2 π α π α + i sgn(θ ) sin . (9.6.12) = β|θ |α cos 2 2 Again, note the change in sign from the stable 0 < α < 1 case.

212

Exotic Option Pricing and Advanced L´evy Models

Finally, note that there is no such thing as a spectrally positive stable process with α = 1 (other than a deterministic drift t → βt). Numerical examples The main purpose of this section is to illustrate the kind of calculations which can be carried out with these models, rather than to give a detailed comparison of the effects of using different L´evy processes. However, to achieve at least a certain degree of comparability among the different examples below, the parameters are chosen so that σ2 +

1

x 2 ν(dx) = 4

and

ν[1, ∞) = 1

(9.6.13)

0

in all of the examples. The ﬁrst condition in equation (9.6.13) says that the contribution to the volatility (as measured by quadratic variation) coming from the Brownian ﬂuctuations and the small jumps is the same in all of the examples, while the second condition in this equation (9.6.13) says that the rate at which large jumps occur is the same in all of the following examples. Throughout the following examples, we take S0 = 1 – so y = 0 – and δ = 0.1. We then compute the prices of a perpetual American put with strikes K = 0.75, 1 and 1.25 using Theorem 9.5.1. In each case, the integrability condition (equation (9.5.14)) is satisﬁed and the resolvent density is computed by approximating the Fourier integral (equation (9.5.15)) with a suitable discrete Fourier sum which is then computed by an FFT algorithm. The details of how this is carried out are described in the Appendix. We consider the following models: 1. Brownian motion plus Gamma (α,1/2) process: here, α = 1.7864, σ 2 = 3.3555 and according to equations (9.6.3) and (9.3.3), c = −0.3848 and from equation (9.2.8) a∗ ψ (1) = 2.5494, and so the optimal exercise level is L∗ = √ e =2 0.1K/2.5494. 2. Brownian motion plus stable 1/2 process: we take β = π, σ = 3.6667 and according to equations (9.6.6) and (9.3.3), c = −0.0391 and ψ (1) = 2.8196, and so L∗ = 0.1K/2.8196. √ 3. Brownian motion plus stable 3/2 process: we take β = 2 π, σ 2 = 1 and from equations (9.6.11) and (9.3.3) we have c = 3.9449 and ψ (1) = 2.3725, and so L∗ = 0.1K/2.3725. The results are summarized in the Table 9.1.

Table 9.1 Summary of the results obtained from the various models Model

K = 0.75

K = 1.00

K = 1.25

Ex. Lev. L∗ Price Ex. Lev. L∗ Price Ex. Lev. L∗ Price Brownian Motion + Gamma (α,1/2) Brownian Motion + Stable (1/2) Brownian Motion + Stable (3/2)

0.029 0.027 0.032

0.64 0.30 0.64

0.039 0.035 0.042

0.86 0.40 0.86

0.049 0.044 0.053

1.09 0.51 1.08

Pricing Perpetual American Options

213

APPENDIX: USE OF FAST FOURIER TRANSFORM Let {x(n)}N n=1 be a real sequence of length N . The discrete Fourier transform xˆ of x, and its inverse, are given by the relations x(k) ˆ =

N

e−2πi(k−1)(n−1)/N x(n)

(A.1)

n=1

x(n) =

N 1 2πi(k−1)(n−1)/N e x(k) ˆ N

(A.2)

k=1

The Fast Fourier Transform (FFT) is a fast and efﬁcient numerical algorithm for computing equations (A.1) and (A.2). Although not strictly necessary, the FFT is at its most efﬁcient if N is a power of 2. The purpose of this appendix is not to describe the workings of the FFT – there are numerous texts written on the subject and the FFT is also implemented in many standard software packages – rather, the purpose is to describe how to recast the problem of computing the Fourier inversion formula (equation (9.5.15)) into a form equivalent to equation (A.2). ∞ Let f be an integrable function: −∞ |f (x)| dx < ∞. Its Fourier transform is given by ∞ −izx fˆ(z) = −∞ e f (x) dx, and so fˆ(2π z) =

∞ −∞

e−2πizx f (x) dx.

(A.3)

Fix an integer M (ideally a power of 2) and let N = M 2 and = M/N = 1/M. We partition the interval [−M/2, M/2] into steps of length and approximate the Fourier integral (equation (A.3)) by a truncated discrete sum involving the values of f (x), for x = −M/2, −M/2 + . . . , M/2 − as follows: N fˆ 2π [(k − 1) − M/2] = e−2πi[(k−1)−M/2][(n−1)−M/2] f (n − 1) − M/2 . n=1

(A.4) giving an approximation for fˆ(z) for z = −M/2, −M/2 + , . . . , M/2 − . (For simplicity, we have chosen the crudest form of discrete approximation to equation (A.3); however, more sophisticated quadrature rules for the most part involve using weighted averages of sums of the form shown by equation (A.4) and so the method described below can be adapted to handle these more sophisticated approximations.) Rearranging equation (A.4) gives e−iπ((k−1)−N/4) fˆ 2π [(k − 1)/M − M/2] =

N n=1

e−2πi(k−1)(n−1)/N eiπ((n−1)−N/4) f (n − 1)/M − M/2 /M.

(A.5)

214

Exotic Option Pricing and Advanced L´evy Models

If fˆ is also an integrable function, then the Fourier inversion formula holds: ∞ ∞ 1 izx ˆ e f (z) dz = e2πiyx fˆ(2πy) dy. f (x) = 2π −∞ −∞

(A.6)

Approximating the above by a truncated sum over [−M/2, M/2] as before gives N e2πi[(k−1)−M/2][(n−1)−M/2] fˆ 2π [(k − 1) − M/2] , f (n − 1) − M/2 = k=1

which when rearranged gives eiπ((n−1)−N/4) f (n − 1)/M − M/2 M N 1 2πi(k−1)(n−1)/N −iπ((k−1)−N/4) ˆ e e = f 2π [(k − 1)/M − M/2] . (A.7) N k=1

Comparing equations (A.5) and (A.7) with equations (A.1) and (A.2), we see that deﬁning x(n) = M −1 eiπ((n−1)−N/4) f (n − 1)/M − M/2 x(k) ˆ = e−iπ((k−1)−N/4) fˆ 2π [(k − 1)/M − M/2]

(A.8) (A.9)

makes these equations identical. For the purposes of Section 9.6, we have f (x) = rδ (x), whose Fourier transform rˆδ (z) is given by equation (9.5.13). Deﬁning x(k) ˆ as in equation (A.9), we can recover x(n) from equation (A.2) (evaluated using FFT) and then equation (A.8) gives values of rδ (x) for a discrete set of grid-points x ∈ [−M/2, M/2], spaced = 1/M apart. The values of rδ (x) for the grid-points x ∈ [y − a ∗ , M/2] are then used to approximate the integral in equation (9.5.2). The values shown in Table 9.1 are obtained by using M = 29 = 512.

ACKNOWLEDGEMENTS I would like to thank the many people, including several anonymous referees and some of the editors of this volume, who have read and made suggestions for improving various earlier drafts of this article. I am also grateful to the editors for providing the opportunity for its publication.

REFERENCES [1] Bertoin, J. (1996), L´evy Processes, Cambridge University Press, Cambridge, UK. [2] Billingsley, P. (1968), Convergence of Probability Measures, Wiley, New York, NY, USA. [3] Bingham, N.H. (1975), “Fluctuation theory in continuous time”, Advances in Applied Probability, 7, 705–766. [4] Carr, P. (1998), “Randomization and the American put”, Review of Financial Studies, 11, 597–626.

Pricing Perpetual American Options

215

[5] Chan, T. (1999), “Pricing contingent claims on stocks driven by L´evy processes”, Annals of Applied Probability, 9, 504–528. [6] Erd´elyi, A. (1954), Tables of Integral Transforms, Vol. 1, McGraw-Hill, New York, NY, USA. [7] Gerber, H.U. and Landry, B. (1998), “On the discounted penalty at ruin in a jump-diffusion and the perpetual put option”, Insurance: Mathematics and Economics, 22, 263–276. [8] Gerber, H.U. and Shiu, E.S.W. (1994), “Martingale approach to pricing perpetual American options”, ASTIN Bulletin, 24, 195–220. [9] Gerber, H.U. and Shiu, E.S.W. (1997), “From ruin theory to option pricing”, in Joint Day Proceedings Volume of XXVIII International ASTIN Colloquium/7th International AFIR Colloquium, pp. 157–176. [10] Gerber, H.U. and Shiu, E.S.W. (1998), “Pricing perpetual options for jump processes”, North American Actuarial Journal, 2(3), 101–112. [11] MacMillan, L. (1986), “Analytic approximation for the American put option”, Advance Futures Options Research, 1, 119–139. [12] Samuelson, P.A. (1965), “Rational theory of warrant pricing”, Industrial Management Review, 6, 13–33. [13] Whitt, W. (1971), “Weak convergence of ﬁrst passage times”, Journal of Applied Probability, 8, 417–422. [14] Zhang, X. (1995), “Formules quasi-explicites pour les options Am´ericaines dans un mod`ele de diffusion avec sauts”, Mathematics and Computers in Simulation, 38, 151–161.

EPILOGUE† This article is a revised (and hopefully improved) version of an original preprint ﬁrst completed in early 2000. In the period between then and the publication of this present volume, there has been a number of new developments and papers which are related to the results presented here. The purely Brownian model considered in Carr (1998) has been extended by Avram et al. (2002) to models such as that presented here involving spectrally one-sided L´evy processes. Avram et al. (2004) further extend this method to Russian options. Many other authors have also contributed to the recent renewed interest in perpetual options of various kinds and below is a limited bibliography which also lists further papers whose themes are closely related to the present one. Much of this interest has centred on one-sided L´evy processes because of the relative ease in carrying out explicit computations; nevertheless, the more interesting case of two-sided L´evy processes has not been neglected and several authors (see, for example, Asmussen et al. (2004)) have succeeded in performing similar explicit computations in certain special cases.

FURTHER REFERENCES 1. Asmussen, S., Avram, F. and Pistorius, M.R. (2004), “Pricing American and Russian options under spectrally two-sided exponential L´evy models”, Stochastic Processes and their Applications, 109, 79–111. 2. Avram, F., Chan, T. and Usabel, M. (2002), “On the valuation of constant barrier options under spectrally one-sided exponential L´evy models and Carr’s approximation for American puts”, Stochastic Processes and their Applications, 100, 75–107. †

Added at the request of the editors.

216

Exotic Option Pricing and Advanced L´evy Models

3. Avram, F. Kyprianou, A.E. and Pistorious, M.R. (2004), “Exit problems for spectrally negative L´evy processes and applications to Canadized Russian options”, Annals of Applied Probability, 14, 215–238. 4. Boyarchenko, S.I. and Levendorski˘i, S.Z. (2002a), “Perpetual American options under L´evy processes”, SIAM Journal of Control Optimization, 40, 1663–1696. 5. Boyarchenko, S.I. and Levendorski˘i, S.Z. (2002b), “Barrier options and touch-and-out options under regular L´evy processes of exponential type”, Annals of Applied Probability, 12, 1261–1298. 6. Duistermaat, J.J., Kyprianou, A.E. and van Schaik, K. (2005), “Finite expiry Russian options”, Stochastic Processes and their Applications, 115, 609–638. 7. Mordecki, E. (1999), “Optimal stopping for a diffusion with jumps”, Finance and Stochastics, 3, 227–236. 8. Mordecki, E. (2002), “Optimal stopping and perpetual options for L´evy processes”, Finance and Stochastics, 6, 473–493.

10 On Asian Options of American Type Goran Peskir University of Aarhus, Aarhus, Denmark† and

Nadia Uys University of the Witwatersrand, Witwatersrand, South Africa Abstract We show that the optimal stopping boundary for the early exercise Asian call option with ﬂoating strike can be characterized as the unique solution of a nonlinear integral equation arising from the early exercise premium representation (an explicit formula for the arbitragefree price in terms of the optimal stopping boundary). The key argument in the proof relies upon a local time-space formula.

10.1 INTRODUCTION According to ﬁnancial theory (see, e.g. Karatzas and Shreve (1998) [7] or Shiryaev (1999) [18]), the arbitrage-free price of the early exercise Asian call option with ﬂoating strike is given as V in equation (10.2.1) below where Iτ /τ denotes the arithmetic average of the stock price S up to time τ . The problem was ﬁrst studied by Hansen and Jørgensen (2000) [5] where approximations to the value function V and the optimal boundary b were derived. The main aim of this present paper is to derive exact expressions for V and b. The optimal stopping problem (equation (10.2.1)) is three-dimensional. When a changeof-measure theorem is applied (as in Shepp and Shiryaev (1994) [16] and Kramkov and Mordecky (1994) [10]) the problem reduces to (equation (10.2.9)) and becomes twodimensional. The problem (equation (10.2.9)) is more complicated than the well-known problems (Peskir (2005, 2003) [12] [13]) since the gain function depends on time in a nonlinear way. From the result of Theorem 3.1 below, it follows that the free-boundary problem (equations (10.2.10)–(10.2.14)) characterizes the value function V and the optimal stopping boundary b in a unique manner. Our main aim, however, is to follow the train of thought initiated by Kolodner (1956) [9] where V is initially expressed in terms of b, and b itself is then shown to satisfy a nonlinear integral equation. A particularly simple approach for achieving this goal in the case of the American put option has been suggested in Kim (1990) [8], Jacka (1991) [6] and Carr et al. (1992) [2] and we will take this up †

Centre for Analytical Finance (funded by the Danish Social Science Research Council) and Network in Mathematical Physics and Stochastics (funded by the Danish National Research Foundation).

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

218

Exotic Option Pricing and Advanced L´evy Models

in the present paper. We will moreover see (as in Peskir (2005, 2003) [12] [13]) that the nonlinear equation derived for b cannot have other solutions. The key argument in the proof relies upon a local time-space formula (see Peskir (2002) [11]). The latter fact of uniqueness may be seen as the principal result of the paper. The same method of proof can also be used to show the uniqueness of the optimal stopping boundary solving nonlinear integral equations derived by Hansen and Jørgensen (2000) [5] and Wu et al. (1999) [19] where this question was not explicitly addressed. These equations arise from the early exercise Asian options (call or put) with ﬂoating strike based on geometric averaging. The early exercise Asian put option with ﬂoating strike can be dealt with analogously to the Asian call option treated here. For ﬁnancial interpretations of the early exercise Asian options and other references on the topic, see Hansen and Jørgensen (2000) [5] and Wu et al. (1999) [19].

10.2 FORMULATION OF THE PROBLEM The arbitrage-free price of the early exercise Asian call option with ﬂoating strike is given by the following expression:

1 + (10.2.1) V = sup E e−rτ Sτ − Iτ τ 0 0 and a ≥ 0 are given and ﬁxed. (Throughout, B = (Bt )t≥0 denotes a standard Brownian motion started at zero.) We recall that T > 0 is the expiration date (maturity), r > 0 is the interest rate and σ > 0 is the volatility coefﬁcient. By the change-of-measure theorem, it follows that: +

+

1 1 −rτ $ = s sup E 1 − Xτ (10.2.4) V = sup E e Sτ 1 − Xτ τ τ 0 0 and x, y ≥ 0, where (s, a) → f (t, s, a) is the probability density function of (St , It ) under $ P with S0 = 1 and I0 = 0 given by: √

2 2 2 2 s r/σ (r + σ 2 /2)2 2 2π f (t, s, a) = 3/2 3 2 √ exp − t − + s) (1 π σ a t σ 2t 2σ 2 σ 2a √

∞ 4π z 4 s 2z2 dz (10.3.4) exp − 2 − 2 cosh(z) sinh(z) sin × σ t σ a σ 2t 0 for s > 0 and a > 0. For a derivation of the right-hand side in equation (10.3.4) see the appendix below. The main result of the paper may be stated as follows. Theorem 3.1 The optimal stopping boundary in the Asian call problem (equation (10.2.9)) can be characterized as the unique continuous increasing solution b : [0, T ] → IR of the nonlinear integral equation: 1−

b(t) = F (T − t, b(t)) t

T −t 1 1 − + r G(u, b(t), b(t + u)) − H (u, b(t), b(t + u)) du t +u t +u 0 (10.3.5)

satisfying 0 < b(t) < t/(1 + rt) for all 0 < t < T . [The solution b satisﬁes b(0+) = 0 and b(T −) = T /(1 + rT ), and the stopping time τb from equation (10.2.17) is optimal in equation (10.2.9).] The arbitrage-free price of the Asian call option (equation (10.2.9)) admits the following ‘early exercise premium’ representation: V (t, x) = F (T − t, x)

T −t 1 1 + r G (u, x, b(t + u)) − H (u, x, b(t + u)) du − t +u t +u 0 (10.3.6) for all (t, x) ∈ [0, T ] × [0, ∞. [Further properties of V and b are exhibited in the proof below.]

Asian Options of American Type

221

Proof. The proof will be carried out in several steps. We begin by stating some general remarks which will be freely used below without further mentioning. 1. The reason that we take the supremum in equations (10.2.1) and (10.2.9) over τ > 0 is that the ratio 1/(t + τ ) is not well deﬁned for τ = 0 when t = 0. Note, however, in equation (10.2.1) that Iτ /τ → ∞ as τ ↓ 0 when I0 = a > 0 and that Iτ /τ → s as τ ↓ 0 when I0 = a = 0. Similarly, note in equation (10.2.9) that Xτ /τ → ∞ as τ ↓ 0 when X0 = x > 0 and Xτ /τ → 1 as τ ↓ 0 when X0 = x = 0. Thus, in both cases the gain process (the integrand in equations (10.2.1) and (10.2.9)) tends to 0 as τ ↓ 0. This shows that in either equation it is never optimal to stop at t = 0. To avoid similar (purely technical) complications in the proof to follow we will equivalently consider V (t, x) only for t > 0 with the supremum taken over τ ≥ 0. The case of t = 0 will become evident (by continuity) at the end of the proof. 2. Recall that it is no restriction to assume that s = 1 and a = x so that Xt = (x + It )/St with I0 = 0 and S0 = 1. We will write Xtx instead of Xt to indicate the dependence on x when needed. It follows that V admits the following representation:

x + Iτ + $ (10.3.7) V (t, x) = sup E 1 − (t + τ ) Sτ 0≤τ ≤T −t for (t, x) ∈ 0, T ] × [0, ∞. From equation (10.3.7) we immediately see that: x → V (t, x) is decreasing and convex on [0, ∞

(10.3.8)

for each t > 0 ﬁxed. 3. We show that V : 0, T ] × [0, ∞ → IR is continuous. For this, using sup(f ) − sup(g) ≤ sup(f − g) and (z − x)+ − (z − y)+ ≤ (y − x)+ for x, y, z ∈ IR, we get:

x + Iτ + y + Iτ + $ −$ E 1− E 1− V (t, x) − V (t, y) ≤ sup (t + τ ) Sτ (t + τ ) Sτ 0≤τ ≤T −t

1 1 ≤ (y − x) E (10.3.9) ≤ (y − x) sup $ (t + τ ) Sτ t 0≤τ ≤T −t for 0 ≤ x ≤ y and t > 0, where in the last inequality we used equation (10.2.8) to deduce that t − (r + σ 2 /2)t) ≤ exp(σ B t − (σ 2 /2)t) and the latter is a martingale under 1/St = exp(σ B $ P. From equation (10.3.9) with equation (10.3.8) we see that x → V (t, x) is continuous at x0 uniformly over t ∈ [t0 − δ, t0 + δ] for some δ > 0 (small enough) whenever (t0 , x0 ) ∈ 0, T ] × [0, ∞ is given and ﬁxed. Thus, to prove that V is continuous on 0, T ] × [0, ∞ it is enough to show that t → V (t, x) is continuous on 0, T ] for each x ≥ 0 given and ﬁxed. For this, take any t1 < t2 in 0, T ] and ε > 0, and let τ1ε be a stopping time such that $ E((1 − (Xtx1 +τ ε )/(t1 + τ1ε ))+ ) ≥ V (t1 , x) − ε. Setting τ2ε = τ1ε ∧ (T − t2 ) we see that 1 V (t2 , x) ≥ $ E((1 − (Xt +τ ε )/(t2 + τ ε ))+ ). Hence we get: 2

2

2

Xtx2 +τ ε +

Xtx1 +τ ε +

1 2 $ $ −E 1− +ε V (t1 , x) − V (t2 , x) ≤ E 1 − t1 + τ1ε t2 + τ2ε Xx ε Xtx1 +τ ε +

t2 +τ2 1 ≤$ E + ε. (10.3.10) − t2 + τ2ε t1 + τ1ε

222

Exotic Option Pricing and Advanced L´evy Models

Letting ﬁrst t2 − t1 → 0 using τ1ε − τ2ε → 0 and then ε ↓ 0 we see that lim sup t2 −t1 →0 (V (t1 , x) − V (t2 , x)) ≤ 0 by dominated convergence. On the other hand, E((1 − (Xtx2 +τ ε )/(t2 + τ2ε ))+ ) ≥ V (t2 , x) − ε. Then let τ2ε be a stopping time such that $ 2 we have: Xtx2 +τ ε +

Xtx1 +τ ε +

2 2 $ $ −E 1− − ε. V (t1 , x) − V (t2 , x) ≥ E 1 − t1 + τ2ε t2 + τ2ε

(10.3.11)

Letting ﬁrst t2 − t1 → 0 and then ε ↓ 0 we see that lim inf t2 −t1 →0 (V (t1 , x) − V (t2 , x)) ≥ 0. Combining the two inequalities we ﬁnd that t → V (t, x) is continuous on 0, T ]. This completes the proof of the initial claim. 4. Denote the gain function by G(t, x) = (1 − x/t)+ for (t, x) ∈ 0, T ] × [0, ∞ and introduce the continuation set C = { (t, x) ∈ 0, T × [0, ∞ | V (t, x) > G(t, x) } and the stopping set S = { (t, x) ∈ 0, T × [0, ∞ | V (t, x) = G(t, x) }. Since V and G are continuous, we see that C is open and S is closed in 0, T × [0, ∞. Standard arguments based on the strong Markov property (cf. Shiryaev (1978) [17]) show that the ﬁrst hitting time τS = inf { 0 ≤ s ≤ T − t | (t + s, Xt+s ) ∈ S } is optimal in equation (10.2.9) as well as that V is C 1,2 on C and satisﬁes equation (10.2.10). In order to determine the structure of the optimal stopping time τS (i.e. the shape of the sets C and S), we will ﬁrst examine basic properties of the diffusion process X solving equation (10.2.6) under $ P. 5. The state space of X equals [0, ∞ and it is clear from the representation (equation (10.2.5)) with equation (10.2.8) that 0 is an entrance boundary point. The drift of X is given by µ(x) = 1 − rx and the diffusion coefﬁcient of X is given by σ (x) = σ x for x ≥ 0. Hence, we see that µ(x) is greater/less than 0, if and only if, x is less/greater than 1/r. This shows that there is a permanent push (drift) of X towards the constant level 1/r (when X is above 1/r the push of X is downwards and when X is below 1/r the push of 2 2 x X is upwards). The scale function of X is given by s(x) = 1 y 2r/σ e2/σ y dy for x > 0, 2 2 and the speed measure of X is given by m(dx) = (2/σ 2 ) x −2(1+r/σ ) e−2/σ x dx on the Borel σ -algebra of 0, ∞. Since s(0) = −∞ and s(∞) = +∞, we see that X is recurrent. 2 ∞ Moreover, since 0 m(dx) = (2/σ 2 )−2r/σ (1 + 2r/σ 2 ) is ﬁnite we ﬁnd that X has an invariant probability density function given by: 2

f (x) =

1 (2/σ 2 )1+2r/σ 2 e−2/σ x 2) 2 2(1+r/σ

(1 + 2r/σ ) x

(10.3.12)

for x > 0. In particular, it follows that Xt /t → 0 $ P-a.s. as t → ∞. This fact has an important consequence for the optimal stopping problem (equation (10.2.9)): if the horizon T is inﬁnite, then it is never optimal to stop. Indeed, in this case letting τ ≡ t and passing to the limit for t → ∞ we see that V ≡ 1 on 0, ∞ × [0, ∞. This shows that the inﬁnite horizon formulation of the problem (equation (10.2.9)) provides no useful information to the ﬁnite horizon formulation (such as in Peskir (2005, 2003) [12] [13], for example). To examine the latter beyond the trivial fact that all points (t, x) with x ≥ t belong to C (which is easily seen by considering the hitting times τε = inf { 0 ≤ s ≤ T − t | Xt+s ≤ (t + s) − ε } and noting that $ Pt,x (0 < τε < T − t) > 0 if x ≥ t with 0 < t < T ), we will examine the gain process in the problem (equation (10.2.9)) using stochastic calculus as follows.

Asian Options of American Type

223

6. Setting α(t) = t for 0 ≤ t ≤ T to denote the diagonal in the state space and applying the local time–space formula (cf. Peskir (2002) [11]) under $ Pt,x when (t, x) ∈ 0, T × [0, ∞ is given and ﬁxed, we get:

s

G(t + s, Xt+s ) = G(t, x) +

Gt (t + u, Xt+u ) du

0

1 s + Gx (t + u, Xt+u ) dXt+u + Gxx (t + u, Xt+u ) dX, Xt+u 2 0 0 1 s α Gx (t + u, α(t + u)+) − Gx (t + u, α(t + u)−) d%t+u + (X) 2 0

s Xt+u 1 − rXt+u I Xt+u < α(t + u) du = G(t, x) + − 2 (t + u) (t + u) 0 s s α d%t+u (X) Xt+u u + 1 I Xt+u < α(t + u) d B (10.3.13) −σ 2 0 t +u 0 t +u

s

where %αt+u (X) is the local time of X on the curve α given by: 1 %αt+u (X) = $ P−lim ε↓0 2ε =$ P−lim ε↓0

1 2ε

u

I α(t + v)−ε < Xt+v < α(t + v) + ε dX, Xt+v

u

σ2 2 X dv (10.3.14) I α(t + v)−ε < Xt+v < α(t + v) + ε 2 t+v

0

0

and d%αt+u (X) refers to the integration with respect to the continuous increasing function u → %αt+u (X). From equation (10.3.13) we respectively read: G(t + s, Xt+s ) = G(t, x) + As + Ms + Ls

(10.3.15)

where A and L are processes of bounded variation (L is increasing ) and M is a continuous (local) martingale. We note, moreover, that s → Ls is strictly increasing only when Xs = α(s) for 0 ≤ s ≤ T − t, i.e. when X visits α. On the other hand, when X is below α then the integrand a(t + u, Xt+u ) of As may be either positive or negative. To determine both regions exactly, we need to examine the sign of the expression a(t, x) = x/t 2 − (1 − rx)/t. It follows that a(t, x) is larger/less than 0, if and only if, x is larger/less than γ (t) where γ (t) = t/(1 + rt) for 0 ≤ t ≤ T . By considering the exit times from small balls in 0, T × [0, ∞ with centre at (t, x) and making use of equation (10.3.13) with the optional sampling theorem (to get rid of the martingale part), upon observing that α(t) < γ (t) for all 0 < t ≤ T so that the local time part is zero, we see that all points (t, x) lying above the curve γ (i.e. x > γ (t) for 0 < t < T ) belong to the continuation set C. Exactly the same arguments (based on the fact that the favourable regions above γ and on α are far away from X) show that for each x < γ (T ) = T /(1 + rT ) given and ﬁxed, all points (t, x) belong to the stopping set S when t is close to T . Moreover, recalling equation (10.3.8) and the fact that V (t, x) ≥ G(t, x) for all x ≥ 0 with t ∈ 0, T ﬁxed, we see that for each t ∈ 0, T there is a point b(t) ∈ [0, γ (t)] such that V (t, x) > G(t, x) for x > b(t) and V (t, x) = G(t, x) for x ∈ [0, b(t)]. Combining it with the previous conclusion on S we ﬁnd that b(T −) = γ (T ) = T /(1 + rT ). (Yet another argument for this identity will be given

224

Exotic Option Pricing and Advanced L´evy Models

below. Note that this identity is different from the identity b(T −) = T used in Hansen and Jørgensen (2000) [5, page 1126].) This establishes the existence of the non-trivial (nonzero) optimal stopping boundary b on a left-neighbourhood of T . We will now show that b extends (continuously and decreasingly) from the initial neighbourhood of T backward in time as long as it visits 0 at some time t0 ∈ [0, T , and later in the second part of the proof below we will deduce that this t0 is equal to 0. The key argument in the proof is provided by the following inequality. Notice that this inequality is not obvious a priori (unlike in Peskir (2005, 2003) [12] [13]) since t → G(t, x) is increasing and the supremum in equation (10.2.9) is taken over a smaller class of stopping times τ ∈ [0, T −t] when t is larger. 7. We show that the inequality is satisﬁed: Vt (t, x) ≤ Gt (t, x)

(10.3.16)

for all (t, x) ∈ C. (It may be noted from equation (10.2.10) that Vt = −(1 − rx)Vx − (σ 2 /2)x 2 Vxx ≤ (1 − rx)/t since Vx ≥ −1/t and Vxx ≥ 0 by equation (10.3.8), so that Vt ≤ Gt holds above γ because (1 − rx)/t ≤ x/t 2 , if and only if, x ≥ t/(1 + rt). Hence, the main issue is to show that equation (10.3.16) holds below γ and above b. Any analytic proof of this fact seems difﬁcult and we resort to probabilistic arguments.) To prove equation (10.3.16), ﬁx 0 < t < t + h < T and x ≥ 0 so that x ≤ γ (t). Let τ = τS (t + h, x) be the optimal stopping time for V (t + h, x). Since τ ∈ [0, T −t −h] ⊆ [0, T − t], we see that V (t, x) ≥ $ Et,x ((1 − Xt+τ /(t + τ ))+ ) so that using the inequality stated prior to equation (10.3.9) above (and the convenient reﬁnement by an indicator function), we get: V (t + h, x) − V (t, x) − G(t + h, x) − G(t, x)

+

x + Iτ + x x x + Iτ $ $ −E 1− − − ≤E 1− (t + h + τ ) Sτ (t + τ ) Sτ t t +h

x + Iτ x + Iτ x + Iτ xh I ≤$ E − ≤1 − (t + τ ) Sτ (t + h + τ ) Sτ (t + h + τ ) Sτ t (t + h)

1 1 x + Iτ x + Iτ xh I − =$ E ≤1 − Sτ t +τ t +h+τ (t + h + τ ) Sτ t (t + h)

x + Iτ h x + Iτ xh I =$ E ≤1 − (t + h + τ ) Sτ t + τ (t + h + τ ) Sτ t (t + h)

x + Iτ x + Iτ h xh E ≤ $ I ≤1 − ≤0 (10.3.17) t (t + h + τ ) Sτ (t + h + τ ) Sτ t (t + h) where the ﬁnal inequality follows from the fact that with Z := (x + Iτ )/((t + h + τ )Sτ ) we have V (t + h, x) = $ E((1−Z)+ ) = $ E((1−Z) I (Z ≤ 1)) = $ P(Z ≤ 1) − $ E(Z I (Z ≤ $ $ 1)) ≥ G(t + h, x) = 1 − x/(t + h) so that E(Z I (Z ≤ 1)) ≤ P(Z ≤ 1) − 1 + x/(t + h) ≤ x/(t + h) as claimed. Dividing the initial expression in equation (10.3.17) by h and letting h ↓ 0 we obtain equation (10.3.16) for all (t, x) ∈ C such that x ≤ γ (t). Since Vt ≤ Gt above γ (as stated following equation (10.3.16) above) this completes the proof of equation (10.3.16).

Asian Options of American Type

225

8. We show that t → b(t) is increasing on 0, T . This is an immediate consequence of equation (10.3.17). Indeed, if (t, x) belongs to C and t0 from 0, T satisﬁes t0 < t1 , then by equation (10.3.17) we have that V (t0 , x) − G(t0 , x) ≥ V (t1 , x) − G(t1 , x) > 0 so that (t0 , x) must belong to C. It follows that b cannot be strictly decreasing thus proving the claim. 9. We show that the smooth-ﬁt condition equation (10.2.12) holds, i.e. that x → V (t, x) is C 1 at b(t). For this, ﬁx a point (t, x) ∈ 0, T × 0, ∞ lying at the boundary so that x = b(t). Then x ≤ γ (t) < α(t) and for all ε > 0 such that x + ε < α(t) we have: G(t, x + ε) − G(t, x) 1 V (t, x + ε) − V (t, x) ≥ =− . ε ε t

(10.3.18)

Letting ε ↓ 0 and using that the limit on the left-hand side exists (since x → V (t, x) is convex), we get the inequality: ∂ +V ∂G 1 (t, x) ≥ (t, x) = − . ∂x ∂x t

(10.3.19)

To prove the converse inequality, ﬁx ε > 0 such that x + ε < α(t), and consider the stopping times τε = τS (t, x + ε) being optimal for V (t, x + ε). Then we have:

+

V (t, x + ε) − V (t, x) 1 $ x + ε + Iτε + x + Iτε ≤ E 1− − 1− ε ε (t + τε ) Sτε (t + τε ) Sτε

1 x + Iτε 1 x + ε + Iτε = −$ E . (10.3.20) E ≤ $ − ε (t + τε ) Sτε (t + τε ) Sτε (t + τε ) Sτε Since each point x in 0, ∞ is regular for X, and the boundary b is increasing, it follows that τε ↓ 0 $ P − a.s. as ε ↓ 0. Letting ε ↓ 0 in equation (10.3.20) we get: ∂ +V 1 (t, x) ≤ − ∂x t

(10.3.21)

by dominated convergence. It follows from equation (10.3.19) and (10.3.21) that (∂ + V /∂x)(t, x) = −1/t implying the claim. 10. We show that b is continuous. Note that the same proof also shows that b(T −) = T /(1 + rT ) as already established above by a different method. Let us ﬁrst show that b is right-continuous. For this, ﬁx t ∈ 0, T and consider a sequence tn ↓ t as n → ∞. Since b is increasing, the right-hand limit b(t+) exists. Because (tn , b(tn )) ∈ S for all n ≥ 1, and S is closed, it follows that (t, b(t+)) ∈ S. Hence by equation (10.2.16) we see b(t+) ≤ b(t). Since the reverse inequality follows obviously from the fact that b is increasing, this completes the proof of the ﬁrst claim. Let us next show that b is left-continuous. Suppose that there exists t ∈ 0, T such that b(t−) < b(t). Fix a point x in b(t−), b(t)] and note by equation (10.2.12) that for s < t we have: x y Vxx (s, z) − Gxx (s, z) dz dy (10.3.22) V (s, x) − G(s, x) = b(s)

b(s)

upon recalling that V is C 1,2 on C. Note that Gxx = 0 below α so that if Vxx ≥ c on R = { (u, y) ∈ C | s ≤ u < t and b(u) < y ≤ x } for some c > 0 (for all s < t close enough

226

Exotic Option Pricing and Advanced L´evy Models

to t and some x > b(t−) close enough to b(t−)) then by letting s ↑ t in equation (10.3.22) we get: 2 x − b(t) >0 (10.3.23) V (t, x) − G(t, x) ≥ c 2 contradicting the fact that (t, x) belongs to D and thus is an optimal stopping point. Hence, the proof reduces to showing that Vxx ≥ c on small enough R for some c > 0. To derive the latter fact we may ﬁrst note from equation (10.2.10) upon using equation (10.3.16) that Vxx = (2/(σ 2 x 2 ))(−Vt − (1 − rx)Vx ) ≥ (2/(σ 2 x 2 ))(−x/t 2 − (1 − rx)Vx ). Suppose now that for each δ > 0 there is s < t close enough to t and there is x > b(t−) close enough to b(t−) such that Vx (u, y) ≤ −1/u + δ for all (u, y) ∈ R (where we recall that −1/u = Gx (u, y) for all (u, y) ∈ R). Then from the previous inequality, we ﬁnd that Vxx (u, y) ≥ (2/(σ 2 y 2 ))(−y/u2 + (1 − ry)(1/u − δ)) = (2/(σ 2 y 2 ))((u − y(1 + ru))/u2 − δ(1 − ru)) ≥ c > 0 for δ > 0 small enough since y < u/(1 + ru) = γ (u) and y < 1/r for all (u, y) ∈ R. Hence, the proof reduces to showing that Vx (u, y) ≤ −1/u + δ for all (u, y) ∈ R with R small enough when δ > 0 is given and ﬁxed. To derive the latter inequality we can make use of the estimate (equation (10.3.20)) to conclude that

1 V (u, y + ε) − V (u, y) (10.3.24) ≤ −$ E ε (u + σε ) Mσε y+ε

where σε = inf { 0 ≤ v ≤ T − u | Xu+v = b(u) } and Mt = sup0≤s≤t Ss . A simple comparison argument (based on the fact that b is increasing) shows that the supremum over all (u, y) ∈ R on the right-hand side of equation (10.3.24) is attained at (s, x + ε). Letting ε ↓ 0 in equation (10.3.24), we thus get:

1 $ (10.3.25) Vx (u, y) ≤ − E (u + σ ) Mσ x for all (u, y) ∈ R where σ = inf { 0 ≤ v ≤ T − s | Xs+v = b(s) }. Since by regularity of X we ﬁnd that σ ↓ 0 $ P-a.s. as s ↑ t and x ↓ b(t−), it follows from equation (10.3.25) that:

1 1 $ (u + σ ) Mσ − u ≤− +δ (10.3.26) Vx (u, y) ≤ − + E u u (u + σ ) Mσ u

for all s < t close enough to t and some x > b(t−) close enough to b(t−). This completes the proof of the second claim, and thus the initial claim is proved as well. 11. We show that V is given by the formula shown in equation (10.3.6) and that b solves equation (10.3.5). For this, note that V satisﬁes the following conditions: V is C 1,2 on C ∪ D

(10.3.27)

Vt + LX V is locally bounded

(10.3.28)

x → V (t, x) is convex

(10.3.29)

t → Vx (t, b(t)±) is continuous.

(10.3.30)

Indeed, the conditions (10.3.27) and (10.3.28) follow from the facts that V is C 1,2 on C and V = G on D upon recalling that D lies below γ so that G(t, x) = 1 − x/t for all

Asian Options of American Type

227

(t, x) ∈ D and thus G is C 1,2 on D. [When we say in condition (10.3.28) that Vt + LX V is locally bounded, we mean that Vt + LX V is bounded on K ∩ (C ∪ D) for each compact set K in [0, T ] × IR+ .] The condition (10.3.29) was established in equation (10.3.8) above. The condition (10.3.30) follows from equation (10.2.12) since according to the latter we have Vx (t, b(t)±) = −1/t for t > 0. Since conditions (10.3.27)–(10.3.30) are satisﬁed, we know that the local time-space formula (cf. Theorem 3.1 in Peskir (2002) [11]) can be applied. This gives: s V (t + s, Xt+s ) = V (t, x) + Vt + LX V (t + u, Xt+u ) I Xt+u = b(t + u) du 0

s

+ 0

=

1 + 2 s

σ Xt+u Vx (t + u, Xt+u ) I Xt+u = b(t + u) dBu

s 0

b Vx (t + u, Xt+u +) − Vx (t + u, Xt+u −) I Xt+u = b(t + u) d%t+u (X)

Gt + LX G (t + u, Xt+u ) I Xt+u < b(t + u) du + Ms

(10.3.31)

0

the ﬁnal equality follows by the smooth-ﬁt condition (10.2.12) and Ms = where s σ X Vx (t + u, Xt+u ) I Xt+u = b(t + u) dBu is a continuous martingale for 0 ≤ s ≤ t+u 0 T − t with t > 0. Noting that (Gt + LX G)(t, x) = x/t 2 − (1 − rx)/t for x < t we see that equation (10.3.31) yields:

s Xt+u 1 − rXt+u I Xt+u < b(t + u) du + Ms . V (t + s, Xt+s ) = V (t, x) + − 2 (t + u) (t + u) 0 (10.3.32) Setting s = T − t, using that V (T , x) = G(T , x) for all x ≥ 0, and taking the $ Pt,x expectation in equation (10.3.32), we ﬁnd by the optional sampling theorem that:

XT + 1− T

T −t Xt+u 1 − rXt+u $ I Xt+u < b(t + u) du. (10.3.33) Et,x − = V (t, x) + (t + u)2 (t + u) 0

$ Et,x

Making use of equations (10.3.1)–(10.3.3) we see that equation (10.3.33) is the formula (10.3.6). Moreover, inserting x = b(t) in equation (10.3.33) and using that V (t, b(t)) = G(t, b(t)) = 1 − b(t)/t, we see that b satisﬁes the equation (10.3.5) as claimed. 12. We show that b(t) > 0 for all 0 < t ≤ T and that b(0+) = 0. For this, suppose that b(t0 ) = 0 for some t0 ∈ 0, T and ﬁx t ∈ 0, t0 . Then, (t, x) ∈ C for all x > 0 as small as desired. Taking any such (t, x) ∈ C and denoting by τS = τS (t, x) the ﬁrst hitting time to S under $ Pt,x , we ﬁnd by equation (10.3.32) that:

Xt+τS + V (t + τS , Xt+τS ) = G(t + τS , Xt+τS ) = 1 − = V (t, x) + Mt+τS t + τS x = 1 − + Mt+τS . (10.3.34) t

228

Exotic Option Pricing and Advanced L´evy Models

Taking the $ Pt,x -expectation and letting x ↓ 0 we get:

Xt+τS + $ Et,0 1 − =1 t + τS

(10.3.35)

Pt,0 (Xt+τS ≥ T ) > 0 we see that the left-hand side of where τS = τS (t, 0). As clearly $ equation (10.3.35) is strictly smaller than 1, thus contradicting the identity. This shows that b(t) must be strictly positive for all 0 < t ≤ T . Combining this conclusion with the known inequality b(t) ≤ γ (t), which is valid for all 0 < t ≤ T , we see that b(0+) = 0 as claimed. 13. We show that b is the unique solution of the nonlinear integral equation (10.3.5) in the class of continuous functions c : 0, T → IR satisfying 0 < c(t) < t/(1 + rt) for all 0 < t < T . (Note that this class is larger than the class of functions having the established properties of b which is, moreover, known to be increasing.) The proof of the uniqueness will be presented in the ﬁnal three steps of the main proof as follows. 14. Let c : 0, T ] → IR be a continuous solution of the equation (10.3.5) satisfying 0 < c(t) < t for all 0 < t < T . We want to show that this c must then be equal to the optimal stopping boundary b. Motivated by the derivation (10.3.31)–(10.3.33) which leads to the formula (10.3.6), let us consider the function U c : 0, T ] × [0, ∞ → IR deﬁned as follows:

XT + 1− T

T −t Xt+u 1 − rXt+u $ I X Et,x − < c(t + u) du − t+u (t + u)2 (t + u) 0

Et,x U (t, x) = $ c

(10.3.36)

for (t, x) ∈ 0, T ] × [0, ∞. In terms of equations (10.3.1)–(10.3.3), note that U c is explicitly given by: U c (t, x) = F (T − t, x)

T −t 1 1 + r G u, x, c(t + u) − H u, x, c(t + u) du − t +u t +u 0 (10.3.37) for (t, x) ∈ 0, T ] × [0, ∞. Observe that the fact that c solves equation (10.3.5) on 0, T means exactly that U c (t, c(t)) = G(t, c(t)) for all 0 < t < T . We will now, moreover, show that U c (t, x) = G(t, x) for all x ∈ [0, c(t)] with t ∈ 0, T . This is the key point in the proof (cf. Peskir (2005, 2003) [12] [13]) that can be derived using a martingale argument as follows. If X = (Xt )t≥0 is a Markov process (with values in a general state space) and we set F (t, x) = Ex (G(XT −t )) for a (bounded) measurable function G with Px (X0 = x) = 1, then the Markov property of X implies that F (t, Xt ) is a martingale under Px for 0 ≤ t ≤ T . T −t Similarly, if we set F (t, x) = Ex ( 0 H (Xu ) du) for a (bounded) measurable function H t with Px (X0 = x) = 1, then the Markov property of X implies that F (t, Xt ) + 0 H (Xu ) du is a martingale under Px for 0 ≤ t ≤ T . Combining these two martingale facts applied to

Asian Options of American Type

the time–space Markov process (t + s, Xt+s ) instead of Xs , we ﬁnd that:

s Xt+u 1 − rXt+u I Xt+u < c(t + u) du − U c (t + s, Xt+s ) − 2 (t + u) (t + u) 0

229

(10.3.38)

is a martingale under $ Pt,x for 0 ≤ s ≤ T − t. We may thus write:

s Xt+u 1 − rXt+u I Xt+u < c(t + u) du = U c (t, x) + Ns − U c (t + s, Xt+s ) − 2 (t + u) (t + u) 0 (10.3.39) where (Ns )0≤s≤T −t is a martingale with N0 = 0 under $ Pt,x . On the other hand, we know from equation (10.3.13) that:

s Xt+u 1 − rXt+u I Xt+u < α(t + u) du − G(t + s, Xt+s ) = G(t, x) + 2 (t + u) (t + u) 0 + Ms + Ls

(10.3.40)

s u is a continuous martingale under whereMs = −σ 0 (Xt+u /(t + u)) I (Xt+u < α(t + u)) d B s α $ Pt,x and Ls = (1/2) 0 d%t+u (X)/(t + u) is an increasing process for 0 ≤ s ≤ T − t. For 0 ≤ x ≤ c(t) with t ∈ 0, T given and ﬁxed, consider the stopping time: σc = inf { 0 ≤ s ≤ T − t | Xt+s ≥ c(t + s) }.

(10.3.41)

Using that U c (t, c(t)) = G(t, c(t)) for all 0 < t < T (since c solves equation (10.3.5) as pointed out above) and that U c (T , x) = G(T , x) for all x ≥ 0, we see that U c (t + σc , Xt+σc ) = G(t + σc , Xt+σc ). Hence from equations (10.3.39) and (10.3.40) using the optional sampling theorem we ﬁnd: Et,x U c (t + σc , Xt+σc ) U c (t, x) = $ σc

Xt+u 1 − rXt+u I X − < c(t + u) du −$ Et,x t+u (t + u)2 (t + u) 0 =$ Et,x G(t + σc , Xt+σc )

Xt+u 1 − rXt+u I Xt+u < c(t + u) du − (t + u)2 (t + u) 0

σc

Xt+u 1 − rXt+u $ − I Xt+u < α(t + u) du = G(t, x) + Et,x (t + u)2 (t + u) 0 σc

Xt+u 1 − rXt+u $ I Xt+u < c(t + u) du − Et,x − (t + u)2 (t + u) 0 −$ Et,x

= G(t, x)

σc

(10.3.42)

since Xt+u < α(t + u) and Xt+u < c(t + u) for all 0 ≤ u < σc . This proves that U c (t, x) = G(t, x) for all x ∈ [0, c(t)] with t ∈ 0, T as claimed.

230

Exotic Option Pricing and Advanced L´evy Models

15. We show that U c (t, x) ≤ V (t, x) for all (t, x) ∈ 0, T ] × [0, ∞. For this, consider the stopping time: τc = inf { 0 ≤ s ≤ T − t | Xt+s ≤ c(t + s) }

(10.3.43)

under $ Pt,x with (t, x) ∈ 0, T ] × [0, ∞ given and ﬁxed. The same arguments as those given following equation (10.3.41) above show that U c (t + τc , Xt+τc ) = G(t + τc , Xt+τc ). Inserting τc instead of s in equation (10.3.39) and using the optional sampling theorem we get: Et,x U c (t + τc , Xt+τc ) = $ Et,x G(t + τc , Xt+τc ) ≤ V (t, x) (10.3.44) U c (t, x) = $ where the ﬁnal inequality follows from the deﬁnition of V proving the claim. 16. We show that c ≥ b on [0, T ]. For this, consider the stopping time: σb = inf { 0 ≤ s ≤ T − t | Xt+s ≥ b(t + s) }

(10.3.45)

under $ Pt,x where (t, x) ∈ 0, T × [0, ∞ such that x < b(t) ∧ c(t). Inserting σb in place of s in equations (10.3.32) and (10.3.39) and using the optional sampling theorem we get: σb

Xt+u 1 − rXt+u $ du (10.3.46) Et,x V (t + σb , Xt+σb ) = G(t, x) + $ Et,x − (t + u)2 (t + u) 0 $ Et,x U c (t + σb , Xt+σb ) σb

Xt+u 1 − rXt+u I X − < c(t + u) du (10.3.47) = G(t, x) + $ Et,x t+u (t + u)2 (t + u) 0 where we also use that V (t, x) = U c (t, x) = G(t, x) for x < b(t) ∧ c(t). Since U c ≤ V it follows from equations (10.3.46) and (10.3.47) that: σb

Xt+u 1 − rXt+u $ I X − ≥ c(t + u) du ≥ 0. (10.3.48) Et,x t+u (t + u)2 (t + u) 0 Due to the fact that b(t) < t/(1 + rt) for all 0 < t < T , we see that Xt+u /(t + u)2 − (1 − rXt+u )/(t + u) < 0 in equation (10.3.48), so that by the continuity of b and c it follows that c ≥ b on [0, T ] as claimed. 17. We show that c must be equal to b. For this, let us assume that there is t ∈ 0, T such that c(t) > b(t). Pick x ∈ b(t), c(t) and consider the stopping time τb from equation (10.2.17). Inserting τb instead of s in equations (10.3.32) and (10.3.39) and using the optional sampling theorem, we get: $ Et,x G(t + τb , Xt+τb ) = V (t, x) (10.3.49) $ Et,x G(t + τb , Xt+τb ) τb

Xt+u 1 − rXt+u c $ I Xt+u < c(t + u) du − (10.3.50) = U (t, x) + Et,x (t + u)2 (t + u) 0

Asian Options of American Type

231

where we also use that V (t + τb , Xt+τb ) = U c (t + τb , Xt+τb ) = G(t + τb , Xt+τb ) upon recalling that c ≥ b and U c = G either below c or at T . Since U c ≤ V we see from equations (10.3.49) and (10.3.50) that: τb

Xt+u 1 − rXt+u $ I X Et,x − < c(t + u) du ≥ 0. (10.3.51) t+u (t + u)2 (t + u) 0 Due to the fact that c(t) < t/(1 + rt) for all 0 < t < T by assumption, we see that Xt+u /(t + u)2 − (1 − rXt+u )/(t + u) < 0 in equation (10.3.51) so that by the continuity of b and c it follows that such a point (t, x) cannot exist. Thus c must be equal to b, and the proof is complete.

10.4 REMARKS ON NUMERICS 1. The following method can be used to calculate the optimal stopping boundary b numerically by means of the integral equation (10.3.5). Note that the formula (10.3.6) can be used to calculate the arbitrage-free price V when b is known. Set ti = ih for i = 0, 1, . . . , n where h = T /n and denote: J (t, b(t)) = 1 −

b(t) − F (T −t, b(t)) t

(10.4.1)

K(t, b(t);t + u, b(t + u)) (10.4.2)

1 1 + r G(u, b(t), b(t + u)) − H (u, b(t), b(t + u)) . (10.4.3) = t +u t +u Then, the following discrete approximation of the integral equation (10.3.5) is valid: J (ti , b(ti )) =

n

K(ti , b(ti ); tj , b(tj )) h

(10.4.4)

j =i+1

for i = 0, 1, . . . , n − 1. Letting i = n−1 and b(tn ) = T /(1 + rT ) we can solve equation (10.4.4) numerically and get a number b(tn−1 ). Letting i = n−2 and using the values of b(tn−1 ) and b(tn ) we can solve equation (10.4.4) numerically and get a number b(tn−2 ). Continuing the recursion, we obtain b(tn ), b(tn−1 ), . . . , b(t1 ), b(t0 ) as an approximation of the optimal stopping boundary b at points 0, h, . . . , T −h, T . It is an interesting numerical problem to show that the approximation converges to the true function b on [0, T ] as h ↓ 0. Another interesting problem is to derive the rate of convergence. 2. To perform the previous recursion, we need to compute the functions F, G, H from equations (10.3.1)–(10.3.3) as efﬁciently as possible. Simply by observing the expressions (10.3.1)–(10.3.4) it is apparent that ﬁnding these functions numerically is not trivial. Moreover, the nature of the probability density function f in expression (10.3.4) presents a further numerical challenge. Part of this probability density function is the Hartman–Watson density discussed in Barrieu et al. (2003) [1]. As t tends to zero, the numerical estimate of the Hartman–Watson density oscillates, with the oscillations increasing rapidly in both amplitude and frequency as t gets closer to zero. Barrieu et al. (2003) [1] mention that this may be a consequence of the fact that t → exp(2π 2 /σ 2 t) rapidly increases to inﬁnity while z → sin(4π z/σ 2 t) oscillates more and more frequently. This rapid oscillation makes accurate estimation of f (t, s, a) with t close to zero very difﬁcult.

232

Exotic Option Pricing and Advanced L´evy Models

The problems when dealing with t close to zero are relevant to pricing the early exercise Asian call option. To ﬁnd the optimal stopping boundary b as the solution to the implicit equation (10.4.4), it is necessary to work backward from T to 0. Thus, to get an accurate estimate for b when b(T ) is given, the next estimate of b(u) must be found for some value of u close to T so that t = T −u will be close to zero. Even if we get an accurate estimate for f , to solve equations (10.3.1)–(10.3.3) we need to evaluate two nested integrals. This is slow computationally. A crude attempt has been made at storing values for f and using these to estimate F, G, H in equations (10.3.1)–(10.3.3) but this method has not produced reliable results. 3. Another approach to ﬁnding the functions F, G, H from equations (10.3.1)–(10.3.3) can be based on numerical solutions of partial differential equations. Two distinct methods are available. Consider the transition probability density of the process X given by: p(s, x; t, y) =

d $ P(Xt ≤ y | Xs = x) dy

(10.4.5)

where 0 ≤ s < t and x, y ≥ 0. Since p(s, x; t, y) = p(0, x; t −s, y), we see that there is no restriction to assume that s = 0 in the sequel. 4. The forward equation approach leads to the initial-value problem: pt = −((1−ry) p)y + (D y p)yy p(0, x; 0+, y) = δ(y −x)

(t > 0, y > 0)

(y ≥ 0)

(10.4.6) (10.4.7)

where D = σ 2/2 and x ≥ 0 is given and ﬁxed (recall that δ denotes the Dirac delta function). Standard results (cf. Feller (1952) [4]) imply that there is a unique non-negative solution (t, y) → p(0, x; t, y) of equations (10.4.6) and (10.4.7). The solution p satisﬁes the following boundary conditions: p(0, x; t, 0+) = 0 (0 is entrance)

(10.4.8)

p(0, x; t, ∞−) = 0 (∞ is normal).

(10.4.9)

The solution p satisﬁes the following integrability condition:

∞

p(0, x; t, y) dy = 1

(10.4.10)

0

for all x ≥ 0 and all t ≥ 0. Once the solution (t, y) → p(0, x; t, y) of equations (10.4.6) and (10.4.7) has been found, the functions F, G, H from equations (10.3.1)–(10.3.3) can be computed using the general formula: $ E0,x (g(Xt )) =

∞

g(y) p(0, x; t, y) dy 0

upon choosing the appropriate function g : IR+ → IR+ .

(10.4.11)

Asian Options of American Type

233

5. The backward equation approach leads to the terminal-value problem: qt = (1−rx) qx + D x 2 qxx q(T , x) = h(x)

(t > 0, x > 0)

(x ≥ 0)

(10.4.12) (10.4.13)

where h : IR+ → IR+ is a given function. Standard results (cf. Feller (1952) [4]) imply that there is a unique non-negative solution (t, x) → q(t, x) of equations (10.4.12) and (10.4.13). Taking x → h(x) to be x → (1−x/T )+ ( with T ﬁxed ), x → x I (x ≤ y) ( with y ﬁxed ), x → I (x ≤ y) ( with y ﬁxed ), it follows that the unique non-negative solution q of equations (10.4.12) and (10.4.13) coincides with F, G, H from equations (10.3.1)–(10.3.3), respectively. (For numerical results of a similar approach, see Rogers and Shi (1995) [14].) 6. It is an interesting numerical problem to carry out either of the two methods described above and produce approximations to the optimal stopping boundary b by using equation (10.4.4). Another interesting problem is to derive the rate of convergence.

APPENDIX In this section we derive the explicit expression for the probability density function f of (St , It ) under $ P with S0 = 1 and I0 = 0 given in equation (10.3.4) above. Let B = (Bt )t≥0 be a standard Brownian motion deﬁned on a probability space (, F , P). With t > 0 and ν ∈ IR given and ﬁxed, recall from Yor (1992, p. 527) [20] that the random t 2(Bs +νs) variable A(ν) ds has the conditional distribution: t = 0 e P A(ν) (10.A.1) t ∈ dy Bt + νt = x = a(t, x, y) dy where the density function a for y > 0 is given by: a(t, x, y) =

2

1 x + π2 1 2x 1 + e + x − exp πy 2 2t 2y

2 ∞ πz ex z dz. cosh(z) sinh(z) sin × exp − − 2t y t 0

(10.A.2)

This implies that the random vector 2(Bt + νt), A(ν) has the distribution: t P 2(Bt + νt) ∈ dx, A(ν) t ∈ dy = b(t, x, y) dx dy where the density function b for y > 0 is given by:

x 1 x − 2νt √ b(t, x, y) = a t, , y √ ϕ 2 2 t 2 t 2

ν2 1 π ν + 1 1 x x− 1+e = + t− √ exp 2t 2 2 2y (2π )3/2 y 2 t

2 ∞ πz ex/2 z cosh(z) sinh(z) sin dz exp − − × 2t y t 0

(10.A.3)

(10.A.4)

234

Exotic Option Pricing and Advanced L´evy Models

√ 2 and we set ϕ(z) = (1/ 2π)e−z /2 for z ∈ IR (for related expressions in terms of Hermite functions, see Dufresne (2001) [3] and Schr¨oder (2003) [15]). t Denoting Kt = αBt + βt and Lt = 0 eαBs +βs ds with α = 0 and β ∈ IR given and ﬁxed, and using that the scaling property of B implies:

2 α y e2(Bs +νs) ds ≤ 4 0 0 (10.A.5) with t = α 2 t/4 and ν = 2β/α 2 , it follows by applying equations (10.A.3) and (10.A.4) that the random vector (Kt , Lt ) has the distribution:

t αBs +βs P αBt + βt ≤ x, e ds ≤ y = P 2(Bt + νt ) ≤ x,

t

P Kt ∈ dx, Lt ∈ dy = c(t, x, y) dx dy

(10.A.6)

where the density function c for y > 0 is given by: 2

α α2 α2 c(t, x, y) = b t, x, y 4 4 4 √ 2

2 2 β β2 1 1 2 2π x = 3/2 3 2 √ exp x− 2 t − 2 1+e + 2+ π α y t α2t α 2 2α α y

∞ 2 x/2 4e 4π z 2z dz. (10.A.7) × exp − 2 − 2 cosh(z) sinh(z) sin α t α y α2t 0 From equations (10.2.8) and (10.2.3) we see that f satisﬁes: 2

1 1 α2 α α2 f (t, s, a) = c(t, log(s), a) = b t, log(s), a s s 4 4 4

(10.A.8)

with α = σ and β = r + σ 2 /2. Hence equation (10.3.4) follows by the ﬁnal expression in equation (10.A.4).

REFERENCES [1] Barrieu, P., Rouault, A. and Yor, M. (2003), “A study of the Hartman–Watson distribution motivated by numerical problems related to Asian options pricing”, Pr´epublication PMA 813, Universit´e Pierre et Marie Curie, Paris, France. [2] Carr, P., Jarrow, R. and Myneni, R. (1992), “Alternative characterizations of American put options”, Mathematical Finance, 2, 78–106. [3] Dufresne, D. (2001), “The integral of geometric Brownian motion”, Advances in Applied Probability, 33, 223–241. [4] Feller, W. (1952), “The parabolic differential equations and the associated semi-groups of transformations”, Annals of Mathematics, 55, 468–519. [5] Hansen, A.T. and Jørgensen, P.L. (2000), “Analytical valuation of American-style Asian options”, Management Science, 46, 1116–1136. [6] Jacka, S.D. (1991), “Optimal stopping and the American put”, Mathematical Finance, 1, 1–14.

Asian Options of American Type

235

[7] Karatzas, I. and Shreve, S.E. (1998), Methods of Mathematical Finance, Springer-Verlag, Berlin, Germany. [8] Kim, I.J. (1990), “The analytic valuation of American options”, Reviews in Financial Studies, 3, 547–572. [9] Kolodner, I.I. (1956), “Free boundary problem for the heat equation with applications to problems of change of phase I. General method of solution”, Communications in Pure and Applied Mathematics, 9, 1–31. [10] Kramkov, D.O. and Mordecky, E. (1994), “Integral option”, Theory of Probability and Applications, 39, 162–173. [11] Peskir, G. (2002), “A change-of-variable formula with local time on curves”, Research Report No. 428, Department of Theoretical Statistics, University of Aarhus, Denmark (30 pp.); Journal of Theoretical Probability, to be published. [12] Peskir, G. (2005), “On the American option problem”, Research Report No. 431, Department of Theoretical Statistics, University of Aarhus, Denmark (13 pp.); Mathematical Finance, 15, 169–181. [13] Peskir, G. (2003), “The Russian option: Finite horizon”, Research Report No. 433, Department of Theoretical Statistics, University of Aarhus, Denmark (16 pp.); Finance and Stochastics, to be published. [14] Rogers, L.C.G. and Shi, Z. (1995), “The value of an Asian option”, Journal of Applied Probability, 32, 1077–1088. [15] Schr¨oder, M. (2003), “On the integral of geometric Brownian motion”, Advances in Applied Probability, 35, 159–183. [16] Shepp, L.A. and Shiryaev, A.N. (1994), “A new look at the Russian option”, Theory of Probability and Applications, 39, 103–119. [17] Shiryaev, A.N. (1978), Optimal Stopping Rules, Springer-Verlag, Berlin, Germany. [18] Shiryaev, A.N. (1999), Essentials of Stochastic Finance (Facts, Models, Theory), World Scientiﬁc, Singapore. [19] Wu, L., Kwok, Y.K. and Yu, H. (1999), “Asian options with the American early exercise feature”, International Journal of Theoretical and Applied Finance, 1, 101–111. [20] Yor, M. (1992), “On some exponential functionals of Brownian motion”, Advances in Applied Probability, 24, 509–531.

11 Why be Backward? Forward Equations for American Options† Peter Carr Bloomberg LP, New York, NY, USA and

Ali Hirsa Caspian Capital Management, LLC, New York, NY, USA Abstract The purpose of this paper is to develop forward equations for standard American options. We assume that the returns on the underlying assets have stationary independent increments, ´ or in other words, that the log price is a Levy process. In all of these models, except for Black–Scholes, the existence of a jump component implies that the backward and forward equations contain an integral in addition to the usual partial derivatives. Despite the computational complications introduced by this term, we use ﬁnite differences to solve these fundamental partial integro-differential equations (PIDEs). Our approach to determining the forward equation for American options is to start with the well-known backward equation ´ processes. In the process of and then exploit the symmetries which essentially deﬁne Levy developing the forward equation, we also determine two hybrid equations of independent interest. To illustrate that our forward PIDE is a viable alternative to the traditional backward approach, we calculate American option values in the diffusion extended VG option pricing model.

11.1 INTRODUCTION Valuing and hedging derivatives consistent with the volatility smile has been a major research focus for over a decade. A breakthrough occurred in the mid-1990s with the recognition that in certain models, European option values satisﬁed forward evolution equations in which the independent variables are the options’ strike and maturity. More speciﬁcally, Dupire (1994) showed that under deterministic carrying costs and a diffusion process for the underlying price, no arbitrage implies that European option prices satisfy a certain partial differential equation (PDE), now called the Dupire equation. Assuming that one could observe European option prices of all strikes and maturities, then this forward PDE can be used to explicitly determine the underlying’s instantaneous volatility as a function of the underlying’s price † The authors thank Dilip Madan and participants of the 2002 ICBI Barcelona conference and the 2002 Risk Boston conference. Errors are our own responsibility. The views expressed are those of the authors and do not necessarily reﬂect the position of Morgan Stanley. A shorter version of this paper was previously published in Risk, 16(1), pp. 103–107, 2003. The extended version is included here with permission of Incisive Media plc.

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

238

Exotic Option Pricing and Advanced L´evy Models

and time. Once this volatility function is known, the value function for European, American and many exotic options can be determined by a wide array of standard methods. As this value function relates theoretical prices of these instruments to the underlying’s price and time, it can also be used to determine many greeks of interest as well. Aside from their use in determining the volatility function, forward equations also serve a second useful purpose. Once one knows the volatility function, either by an explicit speciﬁcation or by a prior calibration, the forward PDE can be numerically solved to efﬁciently value a collection of European options of different strikes and maturities, all written on the same underlying asset. Furthermore, as pointed out in Andreasen (1998), all the greeks of interest satisfy the same forward PDE and hence can also be efﬁciently determined in the same way. Since the original development of forward equations for European options in continuous models, several extensions have been proposed. For example, Esser and Schlag (2002) develop forward equations for European options written on the forward price rather than the spot price. Forward equations for European options in jump diffusion models were developed in Andersen and Andreasen (1999) and extended by Andreasen and Carr (2002). It is straightforward to develop the relevant forward equations for European binary options or for European power options by differentiating or integrating the forward equation for standard European options. Buraschi and Dumas (2001) develop forward equations for compound options.∗ In contrast to the PDE’s determined by others, their evolution equation is an ordinary differential equation whose sole independent variable is the intermediate maturity date. Given the close relationship between compound options and American options, it seems plausible that there might be a forward equation for American options. The development of such an equation has important practical implications since all of the listed options on individual stocks are American-style. The Dupire equation cannot be used to infer the volatility function from market prices of American options, nor can it be used to efﬁciently value a collection of American options of differing strikes and maturities. The purpose of this paper is to develop forward equations for standard American options. This problem is addressed for American calls on stocks paying discrete dividends in Buraschi and Dumas (2001) and it is also considered in a lattice setting in Chriss (1996). We direct our attention to the more difﬁcult problem of pricing continuously exercisable American puts in continuous time models. To do so, we depart from the diffusive models which characterize most of the previous research on forward equations in continuous time. To capture the smile, we assume that prices jump rather than assuming that the instantaneous volatility is a function of stock price and time. Dumas et al. (1998) ﬁnd little empirical support for the Dupire model whereas there is a long history of empirical support for jump-diffusion models.† In particular, we assume that the returns on the underlying asset have stationary independent increments, or in other words that the log price is a L´evy process. Besides the Black and Scholes (1973) model, our framework includes as special cases the variance gamma (VG) model of Madan et al. (1998), the CGMY model of Carr et al. (2002), the ﬁnite moment logstable model of Carr and Wu (2002), the Merton (1976) and Kou (2002) jump-diffusion models, and the hyperbolic models of Eberlein et al. (1998). In all of these models, except for Black–Scholes, the existence of a jump component implies that the backward and forward equations contain an integral in addition to the usual partial derivatives. Despite the ∗ However, their deﬁnition of a compound option is nonstandard in that the critical stock price is speciﬁed in the contract. † For example, three recent papers documenting support for such models are Anderson et al. (2002), Carr et al. (2002) and Carr and Wu (2002).

Forward Equations for American Options

239

computational complications introduced by this term, we use ﬁnite differences to solve both of these fundamental partial integro differential equations (PIDEs). To illustrate that our forward PIDE is a viable alternative to the traditional backward approach, we calculate American option values in the diffusion extended VG‡ option pricing model and ﬁnd very close agreement. Our approach to determining the forward equation for American options is to start with the well-known backward equation and then exploit the symmetries which essentially deﬁne L´evy processes. In the process of developing the forward equation, we also determine two hybrid equations of independent interest. The advantage of these hybrid equations over the forward equation is that they hold in greater generality. Depending on the problem at hand, these hybrid equations can also have large computational advantages over the backward or forward equations when the model has already been calibrated. In particular, the advantage of these hybrid equations over the backward equation is that they are more computationally efﬁcient when one is interested in the variation of prices or greeks across strike or maturity at a ﬁxed time, e.g. market close. The ﬁrst of these hybrid equations has the stock price and maturity as independent variables. The numerical solution of this hybrid equation is an alternative to the backward equation in producing a spot slide, which shows how American option prices vary with the initial spot price of the underlying. If one is interested in understanding how this spot slide varies with maturity, then our hybrid equation is much more efﬁcient than the backward equation. Our second hybrid equation has the strike price and calendar time as independent variables. The numerical solution of this hybrid equation is an alternative to the forward equation in producing an implied volatility smile at a ﬁxed maturity. If one is interested in understanding how the model predicts that this smile will change over time, then our hybrid equation is much more computationally efﬁcient than the forward equation. This second hybrid equation also allows parameters to have a term structure, whereas our forward equation does not.§ Hence, if one needs to efﬁciently value a collection of American options of different strikes in the time-dependent Black–Scholes model, then it is far more efﬁcient to solve our hybrid equation than to use the standard backward equation. The remainder of this paper is structured as follows. The next section introduces our setting and reviews the backward PIDE which governs American option values in this setting. The following section develops the ﬁrst hybrid equation, while the subsequent section develops the second one. The penultimate section develops the forward equation for American options, while the ﬁnal section summarizes and suggests further research.

11.2 REVIEW OF THE BACKWARD FREE BOUNDARY PROBLEM Throughout this article, we focus on (standard) American puts on stocks leaving American calls and other underlyings as an exercise for the reader. We assume perfect capital markets, continuous trading, no arbitrage opportunities, continuous dividend payments and Markovian stock price dynamics under all martingale measures. We further assume that ‡

For details on the use of ﬁnite differences for solving the backward PIDE for American options in the VG model, see Hirsa and Madan (2003). § Note, however, that implied volatility can have a term or strike structure in our L´ evy setting.

240

Exotic Option Pricing and Advanced L´evy Models

the spot interest rate and dividend yield are given by deterministic functions r(t) > 0 and q(t) ≥ 0, respectively. Thus, we assume that under a risk-neutral measure Q, the stock price st satisﬁes the following stochastic differential equation: dst = [r(t) − q(t)]st− dt + σ (st− , t)st− dWt ∞ + st− (ex − 1)[µ(dx, dt) − ν(st− , x, t) dx dt], −∞

(11.1)

for all t ∈ [0, T ]. Thus, the change in the stock price decomposes into three parts. The ﬁrst part is the risk-neutral drift, comprised entirely of the dollar carrying cost of the stock. The second part is the diffusion part, expressed in terms of the instantaneous volatility function σ (S, t). As usual, the term dWt denotes increments of a standard Wiener process deﬁned on the time set [0, T ] and on a complete probability space (, F , Q). The third part is the jump part. The random measure µ(dx, dt) counts the number of jumps of size x in the log price at time t. The Hunt density t ∞ {ν(S, x, t), S > 0, x ∈ , t ∈ [0, T ]} is used to compensate the jump process Jt ≡ 0 −∞ st− (ex − 1)µ(dx, ds), so that the last term in equation (11.1) is the increment of a Q jump martingale.¶ The jump martingale is speciﬁed in such a way that jumps to negative prices are impossible. Since the last two parts are both martingales, we have: t E Q [st |s0 ] = s0 e 0 [r(u)−q(u)]du , where the initial stock price s0 is positive. Consider an American put option on the stock with a ﬁxed strike price K0 > 0 and a ﬁxed maturity date T0 ∈ [0, T ]. Let pt denote the value of the American put at time t ∈ [0, T0 ]. In this general setup, it is not yet known whether the American put value is monotone in S. Hence, we further assume whatever sufﬁcient conditions on the coefﬁcients that are needed so that the put value is monotone in S. Then, for each time t ∈ [0, T0 ], there exists a unique critical stock price, s(t), below which the American put should be exercised early, i.e. if st ≤ s t , then pt = max[0, K0 − st ]

(11.2)

and if st > s t , then pt > max[0, K0 − st ].

(11.3)

The exercise boundary is the time path of critical stock prices, s t , t ∈ [0, T0 ]. This boundary is independent of the current stock price s0 and is bounded above by K0 . It is a smooth, nondecreasing function of time t whose terminal limit is: r(T0 ) . lim s t = K0 min 1, t↑T0 q(T0 ) Right at expiration, the critical stock price is the strike price, i.e. s T0 = K0 . Hence, when q(T0 ) > r(T0 ), there is a discontinuity in the exercise boundary. Figure 11.1 plots the ¶

The function ν(S, x, t) must have the following properties: ν(S, 0, t) = 0,

∞

−∞

(x 2 ∧ 1)ν(S, x, t) dx < ∞.

Forward Equations for American Options

241

Exercise boundary 1 0.9 0.8

Calendar time

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 50

60

70

80 Stock price

90

100

110

Figure 11.1 Exercise boundary in the DEVG model. Critical stock prices are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400 with initial price S0 = 100

exercise boundary in the Diffusion Extended Variance Gamma (DEVG) model. This model extends the pure jump Variance Gamma model of Madan et al. (1998), by adding a diffusion component with constant volatility. The American put value is also a function, denoted p(s, t), mapping its domain D ≡ (s, t) ∈ [0, ∞) × [0, T0 ] into the nonnegative real line. The exercise boundary, s t , t ∈ [0, T0 ], divides this domain D into a stopping region S ≡ [0, s t ] × [0, T0 ] and a continuation region C ≡ (s t , ∞) × [0, T0 ]. Equation (11.2) indicates that in the stopping region, the put value function p(s, t) equals its exercise value, max[0, K0 − S]. In contrast, the inequality expressed in equation (11.3) shows that in the continuation region, the put is worth more ‘alive’ than ‘dead’. The transition between boundaries is smooth in the following sense: lim p(s, t) = K0 − s t ,

t ∈ [0, T0 ]

(11.4)

∂p(s,t) ∂s

t ∈ [0, T0 ].

(11.5)

s↓s t

lim

s↓s t

= −1,

The value matching condition (equation (11.4)) and equation (11.2) imply that the put value is continuous across the exercise boundary. Furthermore, the high contact condition (equation (11.5) and equation (11.2)) further imply that the put’s delta is continuous. Equations (11.4) and (11.5) are jointly referred to as the ‘smooth ﬁt’ conditions.

242

Exotic Option Pricing and Advanced L´evy Models

∂p The partial derivatives, ∂p ∂t , ∂s , and differential equation (PIDE):

∂2p ∂s 2

exist and satisfy the following partial integro

∂p(s, t) σ 2 (s, t)s 2 ∂ 2 p(s, t) ∂p(s, t) + − r(t)p(s, t) + [r(t) − q(t)]s 2 ∂t 2 ∂s ∂s ∞ ∂ p(sex , t) − p(s, t) − p(s, t)s(ex − 1) ν(s, x, t) dx + ∂s −∞

+ 1(s < s t ) r(t)K0 − q(t)s −

∞

6 [p(se , t) − (K0 − se )]ν(s, x, t) dx x

x

= 0.

ln(s t /s)

(11.6) The last term on the left-hand side (LHS) of equation (11.6) is the result of applying the integro-differential operator deﬁned by the ﬁrst two lines to the value p(s, t) = K0 − s holding in the stopping region. The American put value function p(s, t) and the exercise boundary s t jointly solve a backward free boundary problem (FBP), consisting of the backward PIDE (equation (11.6)), the smooth ﬁt conditions (equations (11.4) and (11.5)), and the following boundary conditions: p(s, T0 ) = max[0, K0 − s],

s>0

(11.7)

lim p(s, t) = 0,

t ∈ [0, T0 ]

(11.8)

lim p(s, t) = K0 ,

t ∈ [0, T0 ].

(11.9)

s↑∞ s↓0

These Dirichlet conditions force the American put value to its exercise value along the boundaries. As the efﬁcient implementation of a ﬁnite difference scheme usually requires the use of positive ﬁnite spatial boundaries, our implementation replaces the last two conditions in the target problem by: lim pss (s, t) = 0,

t ∈ [0, T0 ]

(11.10)

lim pss (s, t) = 0,

t ∈ [0, T0 ].

(11.11)

s↑∞ s↓0

Hence, the put gamma is forced to zero along the spatial boundaries. Numerical experimentation suggests that imposition of the zero gamma condition on positive ﬁnite spatial boundaries tends to work better than imposing the Dirichlet conditions. The solution to this alternative speciﬁcation is unique under the further condition that it be continuous along the entire boundary. Figure 11.2 plots American put values in the DEVG model against stock price and time.

11.3 STATIONARITY AND DOMAIN EXTENSION IN THE MATURITY DIRECTION The last section assumed that the strike K and maturity T were ﬁxed at K0 and T0 , respectively. To derive a hybrid FBP for American put values, we ﬁrst extend the domain of the problem to all T ∈ [0, T ], keeping the strike price K ﬁxed at K0 .

Forward Equations for American Options

243

100

American put prices

80 60 40 20 0 1 0.8

400 0.6

Calendar time

300 0.4

200 0.2

100 0

Stock prices

0

Figure 11.2 American put values in the DEVG model against stock price and calendar time. American put values are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400. The value of the American put at the initial stock price of S0 = 100 is $23.9875

Note that the exercise boundary depends on t, r(t), q(t), σ (S, t), ν(S, x, t), T and K0 , but not on s. Suppressing the dependence on r(t), q(t), σ (S, t), ν(S, x, t) and K0 , let s(t; T ) be the function relating the exercise surface to t and T : s t = s(t; T ),

t ∈ [0, T ], T ∈ [0, T ].

The extended continuation region is a three-dimensional region denoted by . This can be pictured as stacking the two-dimensional continuation regions up the Z-axis as T increases from 0. For each T ∈ [0, T ], the union of the two-dimensional continuation region and the two-dimensional stopping region is the plane S > 0, t ∈ [0, T ]. As T increases from zero, the area covered by this plane increases. Thus, the extended domain for the backward PIDE is the wedge S > 0, t ∈ [0, T ], T ∈ [0, T ]. We note that the backward PIDE of the last section holds on this wedge with T0 replaced by T . Let (s, t; T ) be the function solving this backward PIDE: ∂(s, t; T ) σ 2 (s, t)s 2 ∂ 2 (s, t; T ) ∂(s, t; T ) + − r(t)(s, t; T ) + [r(t) − q(t)]s 2 ∂t 2 ∂s ∂s ∞ ∂ (sex , t; T ) − (s, t; T ) − (s, t; T )s(ex − 1) ν(s, x, t) dx + ∂s −∞

244

Exotic Option Pricing and Advanced L´evy Models

+ 1(s < s(t; T )) r(t)K0 − q(t)s − 5 − (K0 − sex )]ν(s, x, t) dx = 0.

∞

[(sex , t; T )

ln(s(t;T )/s)

(11.12)

Now suppose stationarity, i.e. that r(t), q(t), σ (S, t) and ν(S, x, t) are all independent of time t. It follows that the time derivative is just the negative of the maturity derivative: ∂ ∂ (s, t; T ) = − (s, t; T ). ∂t ∂T

(11.13)

Dropping the dependence of r(t), q(t), σ (S, t) and ν(S, x, t) on t and substituting equation (11.13) into equation (11.12) implies that the following relation holds in the extended domain: ∂(s, t; T ) ∂(s, t; T ) σ 2 (s)s 2 ∂ 2 (s, t; T ) + − r(s, t; T ) + (r − q)s ∂T 2 ∂s 2 ∂s ∞ ∂ x x + (se , t; T ) − (s, t; T ) − (s, t; T )s(e − 1) ν(s, x) dx ∂s

−

−∞

+ 1(s < s(t; T )) rK0 − qs −

∞

5 [(sex , t; T ) − (K0 − sex )]ν(s, x) dx = 0.

ln(s(t;T )/s)

(11.14) We note that one can ﬁx t at t0 and just solve the above problem in the s, T plane if desired. In this case, the initial condition is: (s, t0 ; t0 ) = max[0, K0 − s],

s > 0.

(11.15)

The Dirichlet boundary conditions are: lim (s, t0 ; T ) = 0,

s↑∞

lim (s, t0 ; T ) ∼ K0 − s, s↓0

T ∈ [t0 , T ] T ∈ [t0 , T ].

(11.16) (11.17)

Alternatively, these Dirichlet conditions can be replaced by the following zero gamma conditions: lim ss (s, t0 ; T ) = 0,

T ∈ [t0 , T ]

(11.18)

lim ss (s, t0 ; T ) = 0,

T ∈ [t0 , T ].

(11.19)

s↑∞ s↓0

The smooth ﬁt conditions are: lim (s, t0 , T ) = K0 − s(t0 ; T ),

s↓s(t0 ;T )

lim

s↓s(t0 ;T )

∂(s, t0 ; T ) = −1, ∂s

T ∈ [t0 , T ]

T ∈ [t0 , T ].

(11.20) (11.21)

Figure 11.3 plots American put values in the DEVG model against stock price and maturity.

Forward Equations for American Options

245

American put prices

100 80 60 40 20 0 1 0.8

400 0.6

300 0.4

Maturity

200 0.2

100 0

Stock prices

0

Figure 11.3 American put values in the DEVG model against stock price and maturity. American put values are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400

11.4 ADDITIVITY AND DOMAIN EXTENSION IN THE STRIKE DIRECTION The last section assumed that the strike K was ﬁxed at K0 and that r(t), q(t), σ (S, t) and ν(S, x, t) are all independent of time t. To derive a new hybrid PIDE for American put values, we further extend the domain of the problem to all K > 0. We also restore the dependence on t of r(t), q(t), σ (S, t) and ν(S, x, t). On this larger domain, let s(t; T , K) be the function relating the exercise surface to t, T , and K: s t = s(t; T , K),

t ∈ [0, T ],

T ∈ [0, T ],

K > 0.

We note that the backward PIDE (equation (11.12)) holding on the three-dimensional domain of the last section holds on the larger four-dimensional domain with K0 replaced by all K > 0. Let (s, t; K, T ) be the function solving this backward PIDE on the extended four-dimensional domain: ∂(s, t; K, T ) σ 2 (s, t)s 2 ∂ 2 (s, t; K, T ) + [r(t) − q(t)] + ∂t 2 ∂s 2 ∂(s, t; K, T ) − r(t)(s, t; K, T ) ×s ∂s

246

Exotic Option Pricing and Advanced L´evy Models

∞ ∂ x x (se , t; K, T ) − (s, t; K, T ) − (s, t; K, T )s(e − 1) ν(s, x, t) dx + ∂s −∞

+ 1(s < s(t; T , K)) r(t)K − q(t)s −

∞

[(sex , t; K, T )

ln(s(t;T ,K)/s)

5

− (K − se )]ν(s, x, t) dx = 0. x

(11.22)

We now assume that the log price process has independent increments, i.e. is additive or equivalently that σ (S, t) and ν(S, x, t) are both independent of the stock price S. Then, for each ﬁxed t and T , the exercise boundary is a linearly homogeneous function of the strike price: s(t; T , λK) = λs(t; T , K), for all λ ≥ 0. Setting λ =

1 K

implies that: s(t; T , K) = Ks(t; T , 1).

(11.23)

For each ﬁxed s, t and T , the condition s > s(t; T , K) is thus equivalent to the condition sK K < s(t;Ts ,1) = s(t;T ,K) ≡ K(s, t; T ). We refer to the output of this function as the critical strike price. For each ﬁxed s, t and T , the critical strike price is the lowest strike price K at which the put is exercised early. Note that the critical strike price depends on s but is independent of K. For an American put, the critical strike price is bounded above by s. In addition, note that the geometric mean of the two critical prices is just the geometric mean of the stock price and strike price: √ s(t; T , K)K(s, t; T ) = sK. (11.24) The additivity of the log price process implies that the function (s, t; K, T ) is linearly homogeneous in s and K. It follows from Euler’s theorem that: (s, t, K, T ) = s

∂ ∂ (s, t; K, T ) + K (s, t; K, T ). ∂s ∂K

(11.25)

Differentiation with respect to s and K and some obvious algebra establishes that: s2

2 ∂2 2 ∂ (s, t; K, T ) = K (s, t; K, T ). ∂s 2 ∂K 2

(11.26)

Dropping the dependence of σ (S, t) and ν(S, x, t) on S and substituting equations (11.25) and (11.26) into equation (11.22) implies: ∂(s, t; K, T ) σ 2 (t)K 2 ∂ 2 (s, t; K, T ) + ∂t 2 ∂K 2 ∂(s, t; K, T ) − q(t)(s, t; K, T ) − [r(t) − q(t)]K ∂K

Forward Equations for American Options

∞ +

(s, t; Ke

−x

−∞

247

∂ −x (s, t; K, T )K(e − 1) ex ν(x, t) dx , T ) − (s, t; K, T ) − ∂K

+ 1(k > k(s, t; T )) r(t)K − q(t)s −

∞

[(s, t; Ke−x , T )

ln(k(s,t;T )/K)

− (Ke

−x

5

− s)]e ν(x, t) dx = 0. x

(11.27)

We note that one can ﬁx s and T at say s0 and T0 and just solve the above problem in the K, t plane if desired. In this case, the terminal condition is: (s0 , T0 ; K, T0 ) = max[0, K − s0 ],

K > 0.

(11.28)

t ∈ [0, T0 ]

(11.29)

The Dirichlet boundary conditions are: lim (s0 , t; K, T0 ) =∼ K − s0 ,

K↑∞

lim (s0 , t; K, T0 ) = 0,

K↓0

t ∈ [0, T0 ].

(11.30)

Alternatively, these Dirichlet conditions can be replaced by: lim kk (s0 , t; KT0 ) = 0,

t ∈ [0, T0 ]

(11.31)

lim kk (s0 , t; K, T0 ) = 0,

t ∈ [0, T0 ].

(11.32)

K↑∞ K↓0

The smooth ﬁt conditions are: lim K↑K(s,t;T0 )

(s0 , t; K, T0 ) = K(s0 , t; T0 ) − s0 , lim K↑K(s,t;T0 )

∂(s0 ,t;K,T0 ) ∂K

= 1,

t ∈ [0, T0 ]

(11.33)

t ∈ [0, T0 ].

(11.34)

Figure 11.4 plots American put values in the DEVG model against strike price and calendar time. We note that setting jumps to zero reduces the PIDE to a PDE arising in the special case of the time-dependent Black–Scholes model. If one wishes to value American options in this model for multiple strikes and maturities and with ﬁxed time and spot, it is much more efﬁcient to solve the hybrid problem of this section once for each T than it is to solve the usual backward problem once for each K and once for each T , as is usually done.

11.5 THE FORWARD FREE BOUNDARY PROBLEM We now assume that we have both stationarity and additivity. In other words, the log price is a L´evy process and r(t), q(t), σ (S, t) and ν(S, x, t) are all independent of both time t

248

Exotic Option Pricing and Advanced L´evy Models

300

American put prices

250 200 150 100 50 0 1 0.8

400 0.6

300 0.4

Calendar time

200 0.2

100 0

Strike

0

Figure 11.4 American put values in the DEVG model against stock price and calendar time. Critical stock prices are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400. The value of the American put at the initial stock price of S0 = 100 is $23.9875

and the stock price S. Stationarity implies that the function (s, t; K, T ) depends on t and T only through T − t. It thus follows that: ∂ ∂ (s, t; K, T ) = − (s, t; K, T ). ∂t ∂T

(11.35)

Substituting equation (11.35) into equation (11.27) implies:

−

∂(s, t; K, T ) σ 2 K 2 ∂ 2 (s, t; K, T ) ∂(s, t; K, T , ) + − q(s, t; K, T ) − (r − q)K ∂T 2 ∂K 2 ∂K ∞ ∂ −x −x (s, t; Ke , T ) − (s, t; K, T ) − (s, t; K, T )K(e − 1) ex ν(x) dx + ∂K −∞

+ 1(k > k(s, t; T )) rK − qs −

∞

[(s, t; Ke−x , T )

ln(k(s,t;T )/K)

− (Ke

−x

5

− s)]e ν(x) dx = 0. x

(11.36)

Forward Equations for American Options

249

We note that one can ﬁx s and t at say s0 and t0 and just solve the above problem in the K, T plane if desired. In this case, the initial condition is: (s0 , t0 ; K, t0 ) = max[0, K − s0 ],

K > 0.

(11.37)

T ∈ [t0 , T ]

(11.38)

The Dirichlet boundary conditions are: lim (s0 , t0 ; K, T ) ∼ K − S0 ,

K↑∞

lim (s0 , t0 ; K, T ) = 0,

K↓0

T ∈ [t0 , T ].

(11.39)

Alternatively, these Dirichlet conditions can be replaced by: lim kk (s0 , t0 ; K, T ) = 0,

T ∈ [t0 , T ]

(11.40)

lim kk (s0 , t0 ; K, T ) = 0,

T ∈ [t0 , T ].

(11.41)

K↑∞ K↓0

The smooth ﬁt conditions are: lim

K↑K(s,t0 ;T )

(s0 , t0 ; K, T ) = K(s0 , t0 ; T ) − s0 , lim

K↑K(s,t0 ;T )

∂(s0 ,t0 ;K,T ) ∂K

= 1,

T ∈ [t0 , T ]

(11.42)

T ∈ [t0 , T ].

(11.43)

300

American put prices

250 200 150 100 50 0 1 0.8

400 0.6

300 0.4

Maturity

200 0.2

100 0

Strike

0

Figure 11.5 American put values in the DEVG model against stock price and maturity. American put values are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400. The value of the American put at the initial stock price of S0 = 100 is $23.9785

250

Exotic Option Pricing and Advanced L´evy Models Critical strike boundary 1 0.9 0.8 0.7

Maturity

0.6 0.5 0.4 0.3 0.2 0.1 0 100

110

120

130

140

150

160

170

Strike

Figure 11.6 Critical strike boundary in the DEVG model. Critical strike prices are computed from the DEVG model for the following inputs: r = 0.06; q = 0.02; σ = 0.4; s = 0.3; v = 0.25; θ = −0.3; K0 = 110; T0 = 1. The ﬁnite difference scheme uses M = 400 space steps and N = 200 time steps on a domain running from 10 to 400

Figure 11.5 plots American put values in the DEVG model against strike price and maturity. The value of the American put at the initial stock price of S0 = 100 is $23.9875 from the backward problem and $23.9785 from the forward problem. The small difference is due to numerical error since the difference gets even smaller as we increase the number of time and spatial steps. Figure 11.6 plots critical strike prices against maturity using the same inputs.

11.6 SUMMARY AND FUTURE RESEARCH We ﬁrst reviewed the backward PIDE governing the arbitrage-free price of an American put option when the underlying spot price process is Markov in itself. By imposing various restrictions on the process, we then derived three new PIDEs for American put values. In particular, by assuming stationarity, we derived a forward PIDE in maturities with spot price still an independent variable. By alternatively assuming that the evolution coefﬁcients for the proportional process are independent of spot, we derived a backward PIDE with the strike price as an independent variable. Finally, by assuming that the log price of the underlying is a L´evy process, we derived the forward PIDE for arbitrage-free American put values. We numerically solved this forward PIDE for the case of the diffusion extended VG model and found very close agreement to the numerical solution of the backward PIDE. A longer version of this paper, downloadable from

Forward Equations for American Options

251

www.math.nyu.edu/research/carrp/papers/pdf, contains an appendix detailing the ﬁnite difference scheme used to numerically solve the forward PIDE for American put options. It is clear how to apply our analysis to American calls or more generally to payoffs which are both monotone and linearly homogeneous in spot and strike. It should be possible to extend our analysis to barrier options in which the payoff is linearly homogeneous in some subset of spot, strike, barrier, or rebate. An open problem is the forward equation for American options when the evolution parameters depend on stock price and/or time. It would also be interesting to extend our univariate approach to additional state variables besides the stock price. If the extra state variable is another asset price, then bivariate American options could be handled. If the extra state variable is a path statistic, then many path-dependent options could be handled. If the extra state variable is the current level of a randomly evolving volatility process, then our approach would encompass stochastic volatility and GARCH models for which there is considerable empirical support. In the interests of brevity, we defer this research to future work.

APPENDIX: DISCRETIZATION OF FORWARD EQUATION FOR AMERICAN OPTIONS This appendix shows how ﬁnite differences can be used to numerically solve the following forward PIDE governing American put values: ∂P (s, t; K, T ) σ 2 2 ∂ 2 P (s, t; K, T ) ∂P (s, t; K, T ) + (r − q)K − K + qP (s, t; K, T ) (11.44) ∂T 2 ∂K 2 ∂K +∞ ∂P (s, t; K, T ) − K(e−y − 1) ey ν(y) dy (11.45) P (s, t; Ke−y , T ) − P (s, t; K, T ) − ∂K −∞ 5 ∞ P (s, t; Ke−y , T ) − (Ke−y − s) ey ν(y) dy = 0 − 1K>K(s,t;T ) rK − qs − ln(K/K(s,t;T ))

(11.46)

We illustrate the solution in the diffusion extended VG model for which the L´evy density has the form: 2 θ2 2 exp(θy/σ ) ν + σ2 |y| . (11.47) ν(y) dy = exp − ν |y| σ Notice that this L´evy density explodes as y approaches zero from either direction. As a result, special measures will have to be taken when approximating the integral containing this L´evy density. One can show that:

+∞ −∞

where: ω≡

(e−y − 1)ey ν(y) dy = ω

1 ln(1 − θ ν − σ 2 ν/2). ν

(11.48)

252

Exotic Option Pricing and Advanced L´evy Models

Dropping the arguments s and t to simplify notation, we can rewrite equation (11.46) as: ∂P (K, T ) σ 2 2 ∂ 2 P (K, T ) ∂P (K, T ) − K + qP (K, T ) + (r − q + ω)K 2 ∂T 2 ∂K ∂K +∞ − (P (Ke−y , T ) − P (K, T ) ey ν(y) dy −∞

− 1K>K(T ) rK − qs −

∞

P (Ke

−y

, T ) − (Ke

−y

5

− s) e ν(y) dy = 0 y

ln(K/K(T ))

By making the change of variable x = ln K we have p(x, T ) = P (K, T ), ∂p ∂P (x, T ) = K (K, T ), ∂x ∂K 2 ∂ 2p ∂p 2∂ P (x, T ) = K (x, T ) − (K, T ), ∂x 2 ∂x ∂K 2 p(x − y, T ) = P (Ke−y , T ),

and hence we obtain the following PIDE for p(x, T ), ∂p σ2 σ 2 ∂ 2 p(x, T ) ∂p + (r − q + (x, T ) − + ω) (x, T ) + qp(x, T ) 2 ∂T 2 ∂x 2 ∂x +∞ − (p(x − y, T ) − p(x, T )) ν˜ (y) dy −∞

− 1x>x(T ) re − qs − x

∞

p(x − y, T ) − (e

x−y

5

− s) ν(y) ˜ dy = 0,

x−x(T )

where:

˜

˜

e −λ p y e−λn |y| 1y>0 + 1yx(Tj ) rK − qe −

∞

xi

xi −x(Tj )

p(xi − y, Tj ) − (K − exi +y ) ν(y) ˜ dy

6 = 0.

Equivalently, we have: (−B − A)pi−1,j +1 + (1 + 2B + qT )pi,j +1 + (−B + A)pi+1,j +1 = +∞ p(xi − y, Tj ) − pi,j ) ν(y) ˜ dy pi,j + T −∞

+ T × 1xi >x(Tj ) re − qs −

∞

xi

xi −x(Tj )

p(xi − y, Tj ) − (e

xi −y

6

− s) ν(y) ˜ dy , (11.52)

T σ2 +ω , A= r −q + 2 2x

where:

σ 2 T , 2 x 2 = (exi − s)+ ,

B= pi,0

x(T0 ) = ln s, and

x(Tj ) = min{xi : pi,j − (exi − s)+ < 0} for j = 1, . . . , M. xi

254

Exotic Option Pricing and Advanced L´evy Models

For the ﬁrst integral on the right-hand side of equation (11.52), we decompose the range of integration into six parts:

+∞

−∞

p(xi − y, Tj ) − pi,j ν˜ (y) dy =

xi −xN

−∞

+

p(xi − y, Tj ) − pi,j ν˜ (y) dy

−x

p(xi − y, Tj ) − pi,j ν(y) ˜ dy

xi −xN

+

0

−x

+x

+

p(xi − y, Tj ) − pi,j ν˜ (y) dy

p(xi − y, Tj ) − pi,j ν˜ (y) dy

0

+

xi −x0

+x

+

+∞

xi −x0

p(xi − y, Tj ) − pi,j ν(y) ˜ dy

p(xi − y, Tj ) − pi,j ν(y) ˜ dy

The six integrals are evaluated as:

0 −x

x 0

p(xi − y, Tj ) − pi,j ν˜ (y) dy ∼ =

p(xi − y, Tj ) − pi,j ν(y) ˜ dy ∼ = −x

xi −xN

=

1 νx λ˜ n 1 νx λ˜ p

˜

(1 − e−λn x )(pi+1,j − pi,j ), ˜

(1 − e−λp x )(pi−1,j − pi,j ).

p(xi − y, Tj ) − pi,j ν˜ (y) dy

N−i−1 1 pi+k,j − pi,j − k(pi+k+1,j − pi+k,j ) ν k=1 " ! × expint(kx λ˜ n ) − expint((k + 1)x λ˜ n )

+

1

N−i−1

λ˜ n νx

k=1

pi+k+1,j − pi+k,j

˜

˜

e−λn kx − e−λn (k+1)x

where:

∞

expint(x) ≡ x

e−t dt t

(11.53)

Forward Equations for American Options

255

is the exponential integral.

xi −x0

p(xi − y, Tj ) − pi,j ν˜ (y) dy

x

=

i−1 , 1 pi−k,j − pi,j − k(pi−k−1,j − pi−k,j ) expint(kxλp ) − expint((k + 1)xλp ) ν k=1

+

i−1 pi−k−1,j − pi−k,j −λp kx e − e−λp (k+1)x . λp νx k=1

xi −xN

−∞

exi expint((N − i)x(λ˜ n − 1)) p(xi − y, Tj ) − pi,j ν(y) ˜ dy = ν −

∞ xi −x0

s + pi,j expint((N − i)x λ˜ n ). ν

1 p(xi − y, Tj ) − pi,j ν(y) ˜ dy = − pi,j expint(ixλp ). ν

The integral inside the Heaviside term in equation (11.52) is treated in the same manner as the other integral. Therefore, we have:

∞

xi −x(Tj )

=

p(xi − y, Tj ) − (exi −y − s) ν˜ (y) dy

i−1 1 pi−k,j − k(pi−k−1,j − pi−k,j ) expint(kx λ˜ p ) − expint((k + 1)x λ˜ p ) ν k=i−l

+

i−1 pi−k−1,j − pi−k,j −λ˜ p kx ˜ e − e−λp (k+1)x λ˜ p νx

k=i−l

−

" 1 ! xi e expint((i − l)x(λ˜ p + 1)) − s expint((i − l)x λ˜ p ) ν

Difference equation Putting all of the pieces together, we obtain the following difference equation at the point (xi , Tj +1 ) Epi−1,j +1 + Fpi,j +1 + Gpi+1,j +1 = pi,j +

T Ri,j + T 1xi >x(Tj ) Hi,j ν

(11.54)

256

Exotic Option Pricing and Advanced L´evy Models

where E = −A − B − Bp , F = 1 + qT + 2B + Bn + Bp +

T (expint(ix λ˜ p ) + expint((N − i)x λ˜ n )), ν

G = A − B − Bn ,

Ri,j =

N−i−1

pi+k,j − pi,j − k(pi+k+1,j − pi+k,j )

k=1

× {expint(kxλn ) − expint((k + 1)xλn )} +

N−i−1 k=1

+

i−1

pi+k+1,j − pi+k,j −λ˜ n kx ˜ (e − e−λn (k+1)x ) λn x

(pi−k,j − pi,j − k(pi−k−1,j − pi−k,j ))

k=1

, × expint(kxλp ) − expint((k + 1)xλp ) +

i−1 pi−k−1,j − pi−k,j −λ˜ p kx ˜ (e − e−λp (k+1)x ) λp x k=1

+ exi expint((N − i)x(λn − 1)) − s expint((N − i)xλn ), Hi,j = rexi − qs −

−

i−1 1 pi−k,j − k(pi−k−1,j − pi−k,j ) expint(kxλp ) − expint((k + 1)xλp ) ν k=i−l i−1 pi−k−1,j − pi−k,j −λ˜ p kx ˜ (e − e−λp (k+1)x ) λp νx

k=i−l

+

1 , xi e expint((i − l)x(λp + 1)) − s expint((i − l)xλp ) , ν

and Bn =

T ˜ (1 − e−λn x ), ˜ νx λn

Bp =

T ˜ (1 − e−λp x ). νx λ˜ p

The initial condition (equation (11.49)) and boundary conditions (equations (11.50) and (11.51)) are discretized in the usual manner. A standard ﬁnite difference solver can then be used to solve the boundary value problem.

Forward Equations for American Options

257

REFERENCES [1] Andersen, L. and Andreasen, J. (1999), “Jumping smiles,” Risk, 12(11), 65–68. [2] Anderson, T.G. Benzoni, L. and Lund, J. (2002), “An empirical investigation of continuous-time equity return models,” Journal of Finance, 57, 1239–1284. [3] Andreasen, J. (1998), “Implied modelling, stable implementation, hedging, and duality,” Working Paper, University of Aarhus, Aarhus, Denmark. [4] Andreasen, J. and Carr, P. (2002), “Put call reversal,” Working Paper, New York University, New York, USA. [5] Black, F. and Scholes, M. (1973), “The pricing of options and corporate liabilities,” Journal of Political Economy, 81, 637–654. [6] Buraschi, A. and Dumas, B. (2001), “The forward valuation of compound options,” Journal of Derivatives, 9(1), 8–17. [7] Carr, P., Geman, H. Madan, D. and Yor, M. (2002), “The ﬁne structure of asset returns: An empirical investigation, Journal of Business, 75, 305–332. [8] Carr, P.P. and Wu, L. (2002), “The ﬁnite moment logstable process and option pricing,” Working Paper, New York University, New York, USA. [9] Chriss, N. (1996), “Transatlantic trees,” Risk, 9(7), 45–48. [10] Dumas, B., Fleming, J. and Whaley, R. (1998), “Implied volatilities: empirical tests,” Journal of Finance, 53, 2059–2106. [11] Dupire, B. (1994), “Pricing with a smile,” Risk, 7(1), 18–20. [12] Eberlein, E., Keller, U. and Prause, K. (1998), “New insights into smile, mispricing, and value at risk: the hyperbolic model,” Journal of Business, 71, 371–406. [13] Esser, A. and Schlag, C. (2002), “A note on forward and backward partial differential equations for derivative contracts with forwards as underlyings,” in J. Hakala and U. Wystup (Eds), Foreign Exchange Risk, Risk Publications, London, pp. 115–124. [14] Hirsa, A. and Madan, D.B. (2003), “Pricing American options under Variance Gamma,” The Journal of Computational Finance, 7(2). [15] Kou, S.G. (2002), “A jump-diffusion model for option pricing,” Management Science, 48, 1086– 1101. [16] Madan, D.B., Carr, P.P. and Chang, E. (1998), “The variance gamma process and option pricing,” European Financial Review, 2, 79–105. [17] Merton, R.C. (1976), “Option pricing when underlying stock returns are discontinuous,” Journal of Financial Economics, 3, 125–144.

12 Numerical Valuation of American Options Under the CGMY Process Ariel Almendral Norwegian Computing Center, Oslo, Norway Abstract American put options written on an underlying stock following a Carr–Madan–Geman–Yor (CGMY) process are considered. It is known that American option prices satisfy a Partial Integro-Differential Equation (PIDE) on a moving domain. These equations are reformulated as a Linear Complementarity Problem, and solved iteratively by an implicit–explicit type of iteration based on a convenient splitting of the Integro-Differential operator. The solution to the discrete complementarity problems is found by the Brennan–Schwartz algorithm and computations are accelerated by the Fast Fourier Transform. The method is illustrated throughout a series of numerical experiments.

12.1 INTRODUCTION In this paper, we propose a numerical method to compute American put options, when the underlying asset is modeled by the Carr–Madan–Geman–Yor (CGMY) process considered in Carr et al. (2002) [8]. Our contribution is to show experimentally that the implicit– explicit method proposed in Cont and Voltchkova (2003) [12] for European options may be successfully applied to the computation of American options under L´evy models. A similar splitting was already proposed in Hirsa and Madan (2004) [13] for the computation of the American price under the Variance Gamma (VG) process (see also, Anon (2004) [3]). Matache et al. (2003) [17] have previously studied the American pricing problem under the CGMY process. They considered a variational inequality formulation combined with a convenient wavelet basis to compress the stiffness matrix. The approach here is different: we essentially work with a formulation as a Linear Complementarity Problem (LCP), and use standard ﬁnite differences. To deal with the singularity of the jump measure at the origin, we ﬁrst approximate the problem by another problem, where small jumps are substituted by a small Brownian component. Next, we solve the approximated problem iteratively, where for each time step one needs to solve tridiagonal linear complementarity problems. The Fast Fourier Transform (FFT) plays also an important role when computing the convolution integrals fast. The sequence of linear complementarity problems are solved with the help of a simple algorithm proposed by Brennan and Schwartz (1977) [6], that works well for the particular case of a put option. We have also veriﬁed numerically the recent results in Alili and Kyprianou (2004) [1] on the smooth-ﬁt principle for general L´evy processes. Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

260

Exotic Option Pricing and Advanced L´evy Models

A statistical study of ﬁnancial time series in Carr et al. (2002) [8] shows that the diffusion component could in most cases be neglected, provided that the remaining part of the process is of inﬁnite activity and ﬁnite variation. We concentrate precisely on the ﬁnite variation case, but also allow for a diffusion component, that may be safely omitted without affecting the pricing algorithm. In Section 12.2, we brieﬂy introduce the CGMY process, the European and American put option problem, and the related PIDEs. For further information on L´evy processes in ﬁnance, we refer the reader to the books by Cont and Tankov (2004) [11] and Schoutens (2003) [20]. An approximation to the equation with a discretization by ﬁnite differences is exposed in Section 12.3 and numerical results are presented in Section 12.4.

´ 12.2 THE CGMY PROCESS AS A LEVY PROCESS A L´evy process is a stochastic process with stationary, independent increments. The L´evy– Khintchine theorem (see Sato (2001) [19]) provides a characterization of L´evy processes in terms of the characteristic function of the process, namely, there exists a measure ν such that, for all z ∈ R and t ≥ 0, E(eizLt ) = exp(tφ(z)), where φ(z) = iγ z −

σ 2 z2 + 2

R

(eizx − 1 − izx1{|x|≤1} )dν(x).

Here σ ≥ 0, γ ∈ R and ν is a measure on R such that ν({0}) = 0 and ∞. Consider a L´evy process {Lt }t≥0 of the form Lt = (r − q + µ)t + σ Wt + Zt ,

R min(1, x

(12.1) 2 )dν(x)

0,

for constants C > 0, G ≥ 0, M ≥ 0 and Y < 2. The process {Zt }t≥0 is known in the literature as the CGMY process (Carr et al. (2002)) [8]; it generalizes a jump-diffusion model by Kou (2002) [15] (Y = −1) and the VG process (Carr et al. (1998)) [10] (Y = 0). The CGMY process is, in turn, a particular case of the Kobol process studied by Boyarchenko and Levendorskiˇı (2002) [5] and Carr et al. (2003) [7], where the constant C is allowed to take on different values on the positive and negative semiaxes. The characteristic function of the CGMY process may be computed explicitly (see Boyarchenko and Levendorskiˇı (2002) [5] and Carr et al. (2002) [8]). In this paper, we

Numerical Valuation of American Options

261

consider only those processes having inﬁnite activity and ﬁnite variation, excluding the VG process, that is, 0 < Y < 1. In such a situation one has σ2 2 φ(z) = (r − q + µ)iz − z 2 , + C (−Y ) (M − iz)Y − M Y + (G + iz)Y − GY .

(12.4)

A market model Let a market consist of one risky asset {St }t≥0 and one bank account {Bt }t≥0 . Let us assume that the asset process {St }t≥0 evolves according to the geometric law St = S0 exp(Lt ),

(12.5)

where {Lt }t≥0 is the L´evy process deﬁned in equation (12.2), and the bank account follows the law Bt = exp(rt). Assume next the existence of some Equivalent Martingale Measure Q (a measure with the same null sets as the market probability, for which the discounted processes {e−(r−q)t St }t≥0 are martingales). In this paper, one works only with a risk-neutral measure Q, where the drift of the L´evy process has been changed. The EMM-condition EQ [St ] = S0 et (r−q) implies φ(−i) = r − q, and so we get the following risk-neutral form for µ: ω := −

, σ2 − C (−Y ) (M − 1)Y − M Y + (G + 1)Y − GY . 2

(12.6)

We keep the same notation for the risk-neutral parameters G and M. The other parameters σ , C and Y are the same in the risk-neutral world (see, e.g. Cont and Tankov (2004) [11] and Raible (2000) [18]). Note that M must be larger than one for ω to be well deﬁned. 12.2.1 Options in a L´evy market 12.2.1.1 European vanilla options Consider a European put option on the asset {St }t≥0 , with time to expiration T , and strike price K. Let us deﬁne the price of a European put option by the formula: v(τ, s) = e−rτ EQ (K − sHτ )+ ,

0 ≤ s < ∞,

0 ≤ τ ≤ T,

(12.7)

where the process {Hτ }τ ≥0 is the underlying risk-neutral process starting at 1, given by Hτ := exp (r − q + ω)τ + σ Wτ + Zτ .

(12.8)

Note that τ means time to expiration T − t. We will not work directly with the asset price s, but rather with its logarithm. Thus, let x = ln s, and deﬁne the new function u(τ, x) := v(τ, ex ).

(12.9)

262

Exotic Option Pricing and Advanced L´evy Models

From a generalization of Ito’s formula it follows that u satisﬁes the following Cauchy problem: uτ − Lu = 0, τ ∈ (0, T ], x ∈ R, u(0, x) = (K − ex )+ , x ∈ R,

(12.10)

where L is an integro-differential operator of the form

σ2 σ2 Lϕ : = ϕxx + r − q − ϕx − rϕ 2 2 ϕ(τ, x + y) − ϕ(τ, x) − (ey − 1)ϕx (τ, x) k(y) dy. +

(12.11)

R

For a derivation of equation (12.10), see Boyarchenko and Levendorskiˇı (2002) [5] or Raible (2000) [18]. 12.2.1.2 American vanilla options Consider an American put option written on the underlying asset {St }t≥0 . The price may be found by solving an optimal stopping problem of the form: * + v(τ, s) = sup EQ e−rτ (K − sHτ )+ .

S

(12.12)

τ ∈ 0,τ

Here S0,τ denotes the set of stopping times taking values in [0, τ ] and {Hτ }τ ≥0 is the process in equation (12.8). The corresponding function u (cf. equation (12.9)) satisﬁes the free-boundary value problem (Boyarchenko and Levendorskiˇı (2002) [5] and Matache et al. (2003) [17]): uτ − Lu = 0, τ > 0, x > $ c(τ ), x c(τ ), u(τ, x) = K − e , τ > 0, x ≤ $ u(τ, x) ≥ (K − ex )+ , τ > 0, x ∈ R, uτ − Lu ≥ 0, τ > 0, x ∈ R, u(0, x) = (K − ex )+ , x ∈ R,

(12.13)

where the operator L is deﬁned in equation (12.11) and the free-boundary is given by , $ c(τ ) = inf x ∈ R | u(τ, x) > (K − ex )+ ,

τ ∈ (0, T ].

(12.14)

The set {x ∈ R | x ≤ $ c(τ )} is the exercise region for the logarithmic prices. Hence, for asset prices s ≤ exp($ c(τ )), the American put should be exercised.

Numerical Valuation of American Options

263

12.3 NUMERICAL VALUATION OF THE AMERICAN CGMY PRICE The function $ c(τ ) is not known a priori, and needs to be found as part of the solution. Thus, rather than solving equation (12.13) directly, it is more convenient to use another formulation as a so-called Linear Complementarity Problem: uτ − Lu ≥ 0 in (0, T ) × R, u≥ψ in [0, T ] × R, (12.15) (uτ − Lu) (u − ψ) = 0 in (0, T ) × R, u(0, x) = ψ(x), where the initial condition is given by ψ(x) := (K − ex )+ .

(12.16)

Note that the dependency on the free-boundary $ c(τ ) has disappeared, but instead we are left with a set of inequalities. The discretization and numerical solution of equation (12.15) is from now on our main goal. The free-boundary is obtained after computing the solution, by making use of equation (12.14). 12.3.1 Discretization and solution algorithm The main idea of the method is to approximate the operator (equation (12.11)) by truncating the integral term close to zero and inﬁnity. The truncation around inﬁnity is harmless, as long as a sufﬁciently large interval is chosen and the price is substituted by the option’s intrinsic value outside the computational domain. However, the truncation around zero gives rise to an artiﬁcial diffusion that must be taken into account. More precisely, the operator L may be split into the sum of two operators: the ﬁrst one containing the Black and Scholes operator and the second accounting for the jumps, namely, L = LBS + LJ . The jump integral part is in turn split into the sum of one operator P for the integration variable in a neighborhood of the origin, and Q for the complementary domain. For P , we use Taylor’s expansion to write the following approximation: (P ϕ)(τ, x) := ϕ(τ, x + y) − ϕ(τ, x) − (ey − 1)ϕx (τ, x) k(y) dy |y|≤

=

|y|≤

ϕ(τ, x + y) − ϕ(τ, x) − yϕx (τ, x) − (ey − 1 − y)ϕx (τ, x) k(y) dy

2 2 $ ϕ)(τ, x) := σ () ϕxx (τ, x) − σ () ϕx (τ, x), ≈ (P 2 2

with the notation: σ () = 2

|y|≤

y 2 k(y)dy.

(12.17)

264

Exotic Option Pricing and Advanced L´evy Models

$ , with a small That is, P has been approximated by a convection–diffusion operator P 2 diffusion coefﬁcient σ (). The operator Q is simply split into a sum, given that this operation is now allowed away from the origin: ϕ(τ, x + y) − ϕ(τ, x) − (ey − 1)ϕx (τ, x) k(y)dy (Q ϕ)(τ, x) := |y|≥

= (J ϕ)(τ, x) − λ()ϕ(τ, x) + ω()ϕx (τ, x), where we have written J for the convolution term, and k(y)dy, λ() = ω() =

|y|≥

|y|≥

(1 − ey )k(y)dy.

(12.18)

(12.19) (12.20)

Remark 12.3.1 These operations have a probabilistic meaning: the pure-jump process has been approximated by a compound Poisson process plus a small Brownian component. As proved in Asmussen and Rosi´nski (2001) [4], this approximation is valid, if and only if, σ ()/ → ∞, as → 0. Note that this condition implies 0 < Y < 1, excluding therefore the VG process and processes with inﬁnite activity. An approximation result in Cont and Voltchkova (2003) [12] states the following. Let $ + Q and u be the solution of the Cauchy problem L := LBS + P uτ − L u = 0, (12.21) u(0, x) = ψ(x), and then there exists a constant C > 0 such that |u(τ, x) − u (τ, x)| < C, for all τ and x. We use here – without proof – the same approximation to numerically solve an American put option. An indication that this approximation works also for American options is shown in Figure 12.1, where one observes that the exercise boundary tends to the theoretical perpetual exercise price, when the time to expiration τ is taken large. The proof of this fact is thus an open problem. Let us focus now on the problem shown in equation (12.15), but with L instead of L. One possible idea to discretize this new problem is to apply Euler’s scheme in time combined with an implicit–explicit iteration in space. Let the time interval [0, T ] be divided into L equal parts, i.e. τj = j τ (j = 0, 1, . . . , L) with τ = T /L and deﬁne the functions uj ≈ u(τj , x). Let operator L be split as L = A + B. We consider the following sequence of problems: j +1 u uj − Auj +1 ≥ d j := + Buj , τ τ uj +1 ≥ ψ, (12.22) j +1

u j +1 j j +1 − Au − d (u − ψ) = 0, τ 0 u = ψ.

Numerical Valuation of American Options

265

(a) 20 18 16 14

Time

12 10 Y = 0.3

8 6 4 2 0

8

8.2

8.4

8.6

8.8 9 9.2 Asset price

9.4

9.6

9.8

10

(b) 20 18 16 14

Time

12 10 Y = 0.7

8 6 4 2 0

6

6.5

7

7.5 8 Asset price

8.5

9

9.5

Figure 12.1 Exercise boundary and perpetual boundary for two different values of the parameter Y , i.e. 0.3 (a) and 0.7 (b): σ = 0; r = 0.1; q = 0; K = 10; T = 20; C = 1; G = 7; M = 9

266

Exotic Option Pricing and Advanced L´evy Models

That is, given the function uj , we compute uj +1 by solving these integro-differential inequalities. A natural choice for the splitting of L is the following: σ 2 + σ 2 () σ 2 + σ 2 () ϕxx + r − q − + ω() ϕx − rϕ Aϕ := (12.23) 2 2 Bϕ := J ϕ − λ()ϕ.

(12.24)

Observe that the integral term is treated explicitly, whereas the differential part is treated implicitly. This method imposes a stability restriction on the time step; see Cont and Voltchkova (2003) [12] for a discussion of this issue for the European case. 12.3.1.1 Spatial discretization of A Consider a computational domain of the form [0, T ] × [xmin , xmax ]. Let ln K ∈ [xmin , xmax ] and deﬁne the uniform spatial grid xi = xmin + ih (i = 0, . . . , N ) where h = (xmax − xmin )/N . Once we have deﬁned the grid, we can discretize A by standard second-order schemes. For the ﬁrst and second derivatives, the central scheme and the standard 3-point scheme are chosen, respectively. Namely, after introducing the notation δ1 (ϕ) := [ϕi+1 − ϕi−1 ]/2h and δ2 (ϕ) := [ϕi+1 − 2ϕi + ϕi−1 ]/ h2 , where ϕi := ϕ(xi ) (i = 0, 1, . . . , N ), we may write (Aϕ)i = βδ2 (ϕ) + γ δ1 (ϕ) − rϕi ,

(12.25)

with the quantities β and γ deﬁned as β :=

σ 2 + σ 2 () , 2

γ := r − q −

σ 2 + σ 2 () + ω(). 2

(12.26) (12.27)

We obtain the following coefﬁcients for the implicit part γ β , + h2 2h 1 2β b= +r + 2, τ h β γ . c=− 2 − h 2h

a=−

(12.28) (12.29) (12.30)

The tridiagonal matrix T associated to the implicit part has constant diagonals: b is on the main diagonal, a is on the subdiagonal and c is on the superdiagonal. From now on, the parameter is taken as the mesh-size h. The artiﬁcial diffusion σ 2 (h) (cf. Matache et al. (2003) [17]) may be approximated by the composite trapezoidal rule on the intervals [−h, 0] and [0, h]. This gives σ 2 (h) ≈

[k(h) + k(−h)] h3 . 2

The quantities λ(h) and ω(h) are approximated in the next section.

(12.31)

Numerical Valuation of American Options

267

12.3.1.2 Spatial discretization of B The discretization of B involves the discretization of J , since Bϕ = J ϕ − λ()ϕ. The discretization of J is explained in detail in Anon (2004) [3]. Brieﬂy, the idea is to truncate the integral to a ﬁnite domain and then apply the composite trapezoidal rule, i.e., Ji := (J ϕ)i =

ϕ(xi + y)k(y)dy

|y|≥h

≈

ϕ(xi + y)k(y)dy h≤|y|≤Mh

≈h

M

ϕi+m km ρm ,

i = 0, 1, . . . , N,

(12.32)

m=−M

where km = k(mh) for m = 0 and we let k0 = 0. The coefﬁcients obtained from applying the trapezoidal rule are: ρm =

1/2 if m ∈ {−M, −1, 1, M}, 1 otherwise.

It is important to substitute ϕ by the payoff function ψ outside the computational domain. The computation of the numbers Ji constitutes the main burden of the method, but thanks to the FFT algorithm, this may be carried out efﬁciently (see next section). However, N must be an even number, and M = N/2, to be able to express this convolution in matrix–vector notation. Finally, we may use the composite trapezoidal rule to compute an approximation to the numbers λ(h) and ω(h) by simply taking ϕ in equation (12.32) as 1 and ey − 1, respectively. 12.3.1.3 Fast convolution by FFT The Fast Fourier Transform is an algorithm that evaluates the Discrete Fourier Transform (DFT) of a vector f = [f0 , f2 . . . , fR−1 ] in O(R log R) operations. The Discrete Fourier Transform is deﬁned as: Fk =

R−1

fn e−i2πnk/R ,

k = 0, 1, . . . , R.

(12.33)

n=0

One of the multiple applications of the DFT is in computing convolutions. Let us ﬁrst introduce the concept of circulant convolution. Let {xm } and {ym } be two sequences with period R. The convolution sequence z := x ∗ y is deﬁned component-wise as zn =

R−1 m=0

xm−n ym .

(12.34)

268

Exotic Option Pricing and Advanced L´evy Models

We now use the FFT to compute the vector [z0 , . . . , zR−1 ]. The periodic structure of x allows the derivation of the following simple relation: Zk = Xk · Yk ,

(12.35)

where X, Y and Z denote the Discrete Fourier Transform of the sequences x, y and z, respectively. That is, the DFT applied to the convolution sequence is equal to the product of the transforms of the original two sequences. The vector [z0 , . . . , zR−1 ] may be recovered by means of the Inverse Discrete Fourier Transform (IDFT): zn =

R−1 1 Zk ei2πkn/R , R

n = 0, 1, . . . , R.

(12.36)

k=0

In the language of matrices, a circulant convolution may be seen as the product of a circulant matrix times a vector. For example, let R = 3, and use the periodicity xk = xk+R to write equation (12.34) as z0 x0 x1 x2 y0 z1 = x2 x0 x1 y1 . (12.37) z2 x1 x2 x0 y2 A circulant matrix is thus a matrix in which each row is a ‘circular’ shift of the previous row. We are interested in the convolution shown in equation (12.32), where the vector ϕ is not periodic. The associated matrix is a so-called Toeplitz matrix, which by deﬁnition is a matrix that is constant along diagonals. A circulant matrix is hence a particular type of Toeplitz matrix. The next idea is to embed a Toeplitz matrix into a circulant matrix. As an example, let M = 1 and N = 2, so that the matrix-vector notation for equation (12.32) reads k1 /2 ϕ1 ϕ0 ϕ−1 ϕ2 ϕ1 ϕ0 k0 . (12.38) ϕ3 ϕ2 ϕ1 k−1 /2 The matrix above may be embedded in a circulant matrix C of size 5 in the following way. (For computational efﬁciency of the FFT algorithm, it is advisable to use a circulant matrix whose size is a power of 2.): ϕ0 ϕ−1 ϕ3 ϕ2 ϕ1 ϕ2 ϕ1 ϕ0 ϕ−1 ϕ3 ϕ2 ϕ1 ϕ0 ϕ−1 (12.39) C = ϕ3 . ϕ−1 ϕ3 ϕ2 ϕ1 ϕ0 ϕ0 ϕ−1 ϕ3 ϕ2 ϕ1 If we deﬁne the vector η := [k1 /2, k0 , k−1 /2, 0, 0]T , then the product (equation (12.38)) is the vector consisting of the ﬁrst three elements in the product Cη. As explained before, a product of a circulant matrix and a vector may be efﬁciently obtained by applying the FFT algorithm.

Numerical Valuation of American Options

269

As a summary, following the ideas explained above, it is possible to compute the convolution (equation (12.32)), with M = N/2, by ‘embedding’ the resulting matrix into a circulant matrix. The product of a circulant matrix and a vector is carried out in three FFT operations, namely, two DFT and one IDFT. In Almendral and Oosterlee (2003) [2], we applied the FFT algorithm in the computation of European options for Merton’s model and Kou’s model, and in Anon (2004) [3] to ﬁnd the American price under the Variance Gamma process. For further details on the computation of convolutions by FFT we refer the reader to Van Loan (1992) [21]. 12.3.1.4 Boundary conditions We used points on the boundary when discretizing the differential operator A. This means that the vector d j needs to be updated. For a put option, this is done by updating the ﬁrst and the last entries of d j as follows: j

j

j

d1 ← d1 − a(K − exmin ),

dN−1 ← 0.

(12.40)

12.3.1.5 Discrete LCP We are now in position to write the discrete inequalities that correspond to the discretization of equation (12.22): T uj +1 ≥ d j , j +1 u ≥ ψ, (12.41) (T uj +1 − d j , uj +1 − ψ) = 0, 0 u = ψ, j

for j = 0, 1, . . . , L − 1. The matrix T has entries given by equations (12.28)–(12.30), di = j j ui /τ + (J uj )i − λ()ui (i = 1, . . . , N − 1) with the update shown in equation (12.40) and ψ is the vector [ψ1 , ψ2 , . . . , ψN−1 ]T , with ψi = ψ(xi ) (cf. Lewis (2001) [16]). The same letter ψ is used to simplify the notation. We proceed now to explain a simple algorithm to solve equation (12.41). 12.3.1.6 Brennan–Schwartz algorithm for a put option Let a tridiagonal matrix

b1 a2 T =

c1 b2 .. .

c2 .. . an−1

..

. bn−1 an

cn−1 bn

(12.42)

and vectors d = [d1 , . . . , dn ]T and ψ = [ψ1 , . . . , ψn ]T be given. Consider the following problem: ﬁnd a vector u satisfying the system T u ≥ d, u ≥ ψ, (12.43) (T u − d, u − ψ) = 0.

270

Exotic Option Pricing and Advanced L´evy Models

The following algorithm to ﬁnd u in equation (12.43) was proposed by Brennan and Schwartz (1977) [6] (for put options) and discussed in detail by Jaillet et al. (1990) [14]: • Step 1: Compute recursively a vector $ b as $ bn = bn , $ bj , bj −1 = bj −1 − cj −1 aj /$

j = n, . . . , 2.

• Step 2: Compute recursively a vector d$ as d$n = dn , bj , d$j −1 = dj −1 − cj −1 d˜j /$

j = n, . . . , 2.

• Step 3: Compute u forward as follows: u1 = max d$1 /b1 , ψ1 , bj , ψ j , uj = max d$j − aj uj −1 /$

j = 2, . . . , n.

We apply these three steps with ai = a, bi = b and ci = c, with a, b and c as in equations (12.28)–(12.30). The splitting proposed in equations (12.23) and (12.24) does not, in general, guarantee the validity of the Brennan–Schwartz algorithm. However, the convection term may be moved to the explicit part of the splitting, so that the conditions of the Brennan– Schwartz algorithm hold Almendral and Oosterlee (2003) [2]. The solutions obtained in both ways are the same, to within the discretization error.

12.4 NUMERICAL EXPERIMENTS In this section, European and American option prices are computed numerically. In the ﬁrst experiment, we compute an European option (problem (12.21)) and compare it with the solution obtained by the Carr–Madan formula in Carr and Madan (1999) [9]; see also the appendix in this paper, formula (12.9). Both solutions are compared in the %∞ -norm, and the results are shown in Table 12.1. A linear convergence rate is observed, and note that the algorithm computes the European price with an error of one cent in about one second. Table 12.1 Linear convergence to exact solution in %∞ -norm and CPU times on a Pentium IV, 1.7 GHz. The parameters are as follows: r = 0; q = 0; K = 10; T = 1; C = 1; G = 7; M = 9; Y = 0.7 N

L

%∞ -error

50 100 200 400

5 10 20 40

0.2675 0.1281 0.0459 0.0160

CPU-time (s) 0.22 0.31 0.34 1.06

Numerical Valuation of American Options

271

A second experiment concerns the veriﬁcation of the theoretical perpetual exercise price against the asymptotic behavior of the free boundary for some large time to expiry. The asymptotic value s ∗ of the American put was veriﬁed with the aid of a formula in Boyarchenko and Levendorskiˇı (2002) [5], (Theorem 3.2 and Theorem 5.1): 6 ∞+iρ ln r + q + φ0 (z) 1 ∗ ∗ dz , (12.44) s = exp(x ) = K exp − 2π −∞+iρ z2 + iz with ρ a positive number (not arbitrary, see Boyarchenko and Levendorskiˇı [5]) and φ0 (z) is given by equation (12.47) (see below). Figure 12.1 above shows two examples of exercise boundaries and their corresponding theoretical asymptotic values. In these examples, ρ = 1 gives the right value. In the next two experiments, we examine the behavior of the option price and free boundary for different values of Y and M. We conclude from Figures 12.2 and 12.3 that the American option price is an increasing function of Y and a decreasing function of M. We mention that the results shown in Figure 12.2 are in accordance with the numerical tests in Matache et al. (2003) [17] (Figure 6). The last test is designed to verify the smooth-ﬁt principle. According to Alili and Kyprianou (2004) [1], the smooth-ﬁt principle holds for perpetual American put options in the bounded variation case considered here, if and only if, the drift r − q + ω is negative, or an additional condition on the jump measure is satisﬁed for zero drift. In Figure 12.4(a), we show the numerical derivative vs at time T = 1, for a set of parameters giving negative drift. In this case, we have smooth-ﬁt. For a second set of parameters chosen such that the drift is positive, we see a discontinuous derivative in Figure 12.4(b) and so there is no smooth-ﬁt.

APPENDIX: ANALYTIC FORMULA FOR EUROPEAN OPTION PRICES We include here the analytic expression given in Lewis (2001) [16] for European options, adapted to the case of a CGMY process: e−rt u(t, x) = 2π

iα+∞

ˆ exp [−izx + tφ0 (−z)] ψ(z)dz,

(12.45)

iα−∞

ˆ where ψ(z) is the generalized Fourier transform of the payoff ψ, which for a put option is given by K iz+1 ˆ ψ(z) =− 2 , z − iz

(12.46)

and the risk-neutral characteristic function φ0 to be used is obtained by substituting µ by ω from equation (12.6) into equation (12.4), i.e. σ2 2 φ0 (z) = (r − q + ω)iz − z 2 , + C (−Y ) (M − iz)Y − M Y + (G + iz)Y − GY .

(12.47)

272

Exotic Option Pricing and Advanced L´evy Models (a)

5

Y = 0.1 Y = 0.3 Y = 0.5 payoff

4.5 4

Option price

3.5 3 2.5 2 1.5 1 0.5 0

(b)

6

8

10 12 Asset price

14

16

5 4.5 4 3.5

Time

3

Y = 0.5

Y = 0.3

Y = 0.1

2.5 2 1.5 1 0.5 0 7.5

8

8.5

9

9.5

10

Asset price

Figure 12.2 (a) Option prices for different values of the parameter Y , i.e. 0.1, 0.3 and 0.7, and (b) the corresponding exercise boundaries: σ = 0; r = 0.1; q = 0; K = 10; T = 5; C = 1; G = 7.8; M = 8.2

Numerical Valuation of American Options (a)

273

M= 5 M= 7 M= 9 payoff

4.5 4

Option price

3.5 3 2.5 2 1.5 1 0.5 0

6

7

8

9

10

11

12

13

14

15

Asset price (b)

5

M=9

4.5 4 3.5 M=7

Time

3 2.5 2 1.5

M=5

1 0.5 0 7.5

8

8.5

9

9.5

10

Asset price

Figure 12.3 (a) Option prices for different values of the parameter M, i.e. 5, 7 and 9, and (b) the corresponding exercise boundaries: σ = 0; r = 0.1; q = 0; K = 10; T = 5; C = 1; G = 7; Y = 0.2

274

Exotic Option Pricing and Advanced L´evy Models (a) −0.1 −0.2

Option Delta

−0.3 −0.4 −0.5 −0.6 −0.7 −0.8 −0.9 6

8

10

12

14

16

Asset price (b) −0.1 −0.2

Option Delta

−0.3 −0.4 −0.5 −0.6 −0.7 −0.8 −0.9 6

7

8

9

10

11

12

Asset price

Figure 12.4 (a) Continuous option Delta for G = 10 and M = 3, and (b) discontinuous option Delta for G = 7 and M = 9: σ = 0; r = 0.1; q = 0; K = 10; T = 1; C = 1

Numerical Valuation of American Options

275

The constant α in equation (12.9) is determined by the region of validity of equation (12.46) together with the strip of regularity of equation (12.47). In this case, we may pick α ∈ (−G, 0). A method using the FFT algorithm was proposed in Carr and Madan (1999) [9] to evaluate an analogous version of equation (12.45).

REFERENCES [1] Alili, L. and Kyprianou, A.E. (2004) “Some remarks on ﬁrst passage of L´evy processes, the American put and pasting principles”, to be published. [2] Almendral, A. and Oosterlee, C.W. (2003), “Numerical valuation of options with jumps in the underlying”, Technical Report, Delft University of Technology, Delft, The Netherlands; to be published. [3] Almendral, A. (2004), “On American options under the Variance Gamma process”, Technical Report, Delft University of Technology, Delft, The Netherlands: to be published. [4] Asmussen, S. and Rosi´nski, J. (2001), “Approximations of small jumps of L´evy processes with a view towards simulation”, Journal of Applied Probability, 38, 482–493. [5] Boyarchenko, S.I. and Levendorski˘ı, S.Z. (2002), Non-Gaussian Merton–Black–Scholes Theory, Advanced Series on Statistical Science and Applied Probability, vol. 9, World Scientiﬁc, River Edge, NJ, USA. [6] Brennan, M.J. and Schwartz, E.S. (1977), “The valuation of the American put option”, Journal of Finance, 32, 449–462. [7] Carr, P., Geman, H., Madan, D.B. and Yor, M. (2003), “Stochastic volatility for L´evy processes”, Mathematical Finance, 13, 345–382. [8] Carr, P.P., Geman, H., Madan, D.B. and Yor, M. (2002), “The ﬁne structure of asset returns: An empirical investigation”, Journal of Business, 75, 305–332. [9] Carr, P.P. and Madan, D.B. (1999), “Option valuation using the Fast Fourier Transform”, Journal of Computational Finance, 2, 61–73. [10] Carr, P.P., Madan, D.B. and Chang, E.C. (1998), “The Variance Gamma process and option pricing”, European Finance Review, 2, 79–105. [11] Cont, R. and Tankov, P. (2004), Financial Modelling with Jump Processes, Chapman & Hall, Boca Raton, FL, USA. [12] Cont, R. and Voltchkova, E. (2003), “A ﬁnite difference scheme for option pricing in jump diffusion and exponential L´evy models”, Technical Report 513, CMAP, Palaiseau, France. [13] Hirsa, A. and Madan, D.B. (2004), “Pricing American options under Variance Gamma”, Journal of Computational Finance, 7. [14] Jaillet, P., Lamberton, D. and Lapeyre, B. (1990), “Variational inequalities and the pricing of American options”, Acta Applied Mathematica, 21, 263–289. [15] Kou, S.K. (2002), “A jump diffusion model for option pricing”, Management Science, 48, 1086– 1101. [16] Lewis, A.L. (2001), “A simple option formula for general jump-diffusion and other exponential L´evy processes”, in Proceedings of the Annual CAP Workshop on Derivative Securities and Risk Management. [17] Matache, A.M., Nitsche, P.A. and Schwab, C. (2003), “Wavelet Galerkin pricing of American options on L´evy driven assets, Working Paper, ETH, Z¨urich, Switzerland. [18] Raible, S. (2000), “L´evy processes in ﬁnance: theory, numerics and empirical Facts”, Ph.D Thesis, Institute f¨ur Mathematische Stochastik, Albert-Ludwigs-Universit¨at Freiburg, Freiburg, Germany.

276

Exotic Option Pricing and Advanced L´evy Models

[19] Sato, K.-I. (2001), “Basic results on L´evy processes”, in L´evy Processes, Birkh¨auser, Boston, MA, USA, pp. 3–37. [20] Schoutens, W. (2003), L´evy Processes in Finance: Pricing Financial Derivatives, Wiley, Chichester, UK. [21] van Loan, C. (1992), Computational Frameworks for the Fast Fourier Transform, Frontiers in Applied Mathematics, vol. 10, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, USA.

13 Convertible Bonds: Financial Derivatives of Game Type Jan Kallsen Munich University of Technology, Garching, Germany and

Christoph Kuhn ¨ ¨ Frankfurt am Main, Germany Johann Wolfgang Goethe-Universitat, Abstract A convertible bond is a security that the holder can convert into a speciﬁed number of underlying shares. In addition, very often the issuer can recall the bond, paying some compensation, or force the holder to convert it immediately. Therefore, the pricing problem has also a game-theoretic aspect. When modelling convertible (callable) bonds within the framework of a ﬁrm value model, they can be considered as an example of a standard game contingent claim as long as no dividends are distributed to the equity holders. This article reviews the classical as well as some recent literature in this ﬁeld. Furthermore, we introduce a mathematically rigorous concept of no-arbitrage price processes for these kinds of derivatives, which explicitly incorporates the feature that the contract can be terminated by both counterparties prematurely. We compare this dynamic conception to price derivatives with the static one by Karatzas and Kou (1998) [18].

13.1 INTRODUCTION A game contingent claim (GCC), as introduced in Kifer (2000) [20], is a contract between a seller A and a buyer B which can be terminated by A and exercised by B at any time t ∈ [0, T ] up to a maturity date T when the contract is terminated anyway. More precisely, the contract may be speciﬁed in terms of two stochastic processes (Lt )t∈[0,T ] , (Ut )t∈[0,T ] with Lt ≤ Ut for t ∈ [0, T ) and LT = UT .

(13.1.1)

If A terminates the contract at time t before it is exercised by B, she has to pay B the amount Ut . If B exercises the option before it is terminated by A, he is paid Lt . An example is a put option of game type with constant penalty δ > 0. If S 1 denotes the price process of the underlying and K the strike price, then Lt = (K − St1 )+ and Ut = (K − St1 )+ + δ1{t 0 we deﬁne the recall times σ ε = inf{t ≥ 0 | V t ≥ Ut − ε}. Analogously to Proposition 3.1 in Karatzas and Zamﬁrescu (2003) [19] and Theorem 11 in ε LM, one can show that for all ε > 0 the stopped process V σ is a Q-supermartingale w.r.t. all Q ∈ Me . Note that in LM the payoff processes L and U are supposed to be bounded, but the results still hold under the weaker condition (13.2.1) (see Theorem 1.1 in K¨uhn and Kyprianou (2003b) [23]. By this supermartingale property and L ≤ V we have for all Q ∈ Me , τ ∈ T0 EQ R(τ, σ ε ) ≤ EQ Lτ 1{τ ≤σ ε } + V σ ε 1{σ ε 0 and σ ∈ T0 be a recalling time with hσ := supτ ∈T0 supQ∈Me EQ (R(τ, σ )) ≤ hup + ε. Deﬁne an American contingent claim (ACC) by the c`adl`ag process X = (Xt )t∈[0,T ] , where Xt = Lt 1{t L on [0, T ) × , and U− > L− on [0, T ] × ,

(13.2.10)

which holds true for the American case (Ut = ∞ for t < T ) or for the callable put with constant penalty (see K¨uhn and Kyprianou 2003a [22]). For other sufﬁcient conditions, see Kallsen and K¨uhn (2004) [17]. Proof. Ad ⇒: Step 1: Let S d+1 be an arbitrage-free price process. By condition (2) and the fundamental theorem of asset pricing there exists a probability measure Q ∈ Me (i.e. Q ∼ P and S 1 , . . . , S d are Q-σ -martingales) such that 1{S d+1 =L− } · S d+1 is a Q-σ -supermartingale, −

1{L− <S d+1 0 is the volatility coefﬁcient, and x > 0 is given and ﬁxed. The main purpose of this present paper is to ﬁnd a solution to the following optimal stopping game for the time-homogeneous (strong) Markov process X having the value function: V∗ (x) = inf sup Ex [e−(λ+r)(σ ∧τ ) (G1 (Xσ ) I (σ < τ ) + G2 (Xτ ) I (τ ≤ σ ))] σ

τ

= sup inf Ex [e−(λ+r)(σ ∧τ ) (G1 (Xσ ) I (σ < τ ) + G2 (Xτ ) I (τ ≤ σ ))] τ

σ

(14.2.3)

where Px is a probability measure under which the process X deﬁned in equations (14.2.1) and (14.2.2) starts at some x > 0, the inﬁmum and supremum are taken over all ﬁnite stopping times σ and τ of the process X (i.e. stopping times with respect to (FtX )t≥0 denoting the natural ﬁltration of X: FtX = σ {Xu | 0 ≤ u ≤ t}, t ≥ 0), λ > 0 is a discounting rate, and the functions Gi (x) are deﬁned by: Gi (x) = (x − Li ) I (Li ≤ x < Ki ) + (Ki − Li ) I (x ≥ Ki )

(14.2.4)

The Spread Option Optimal Stopping Game

295

for all x > 0 with some constants Li and Ki such that 0 < Li < Ki , i = 1, 2, as well as L1 < L2 , K1 < K2 and K1 − L1 = K2 − L2 . We will derive sufﬁcient conditions for the existence of a nontrivial closed form solution to the problem (equation (14.2.3)). Note that the existence of a unique value (equation (14.2.3)) was proved in Lepeltier and Mainguenau (1984) [25] and Kifer (2000) [20]. This fact will be re-proved in Theorem 4.1 below under certain conditions on the parameters of the model. It also follows from equation (14.2.3) that the inequalities G2 (x) ≤ V∗ (x) ≤ G1 (x) hold for all x > 0. We will search for optimal stopping times in the problem (equation (14.2.3)) of the following form: σ∗ = inf{t ≥ 0 | Xt ≤ A∗ }

(14.2.5)

τ∗ = inf{t ≥ 0 | Xt ≥ B∗ }

(14.2.6)

for some numbers A∗ and B∗ such that L1 ≤ A∗ ≤ D1 and D2 ≤ B∗ ≤ K2 hold with Di = Li (λ + r)/λ, i = 1, 2 (for an explanation of the latter inequalities see the text following equation (14.4.5) below). In this connection, the points A∗ and B∗ are called optimal stopping boundaries. Note that in this case, A∗ is the largest number from L1 ≤ x ≤ D1 such that V∗ (x) = G1 (x), and B∗ is the smallest number from D2 ≤ x ≤ K2 such that V∗ (x) = G2 (x). The pair of stopping times (σ∗ , τ∗ ) is usually called a saddle point of the optimal stopping game. On a ﬁnancial market, there are investors speculating for a rise of stock prices (so-called ‘bulls’ playing on the increase) and investors speculating for a fall of stock prices (so-called ‘bears’ playing on the decrease), and their strategies on the market are asymmetric (see, e.g. Shiryaev (1999) [31] Chapter I, Section 1c). In order to restrict their losses and gains simultaneously, the investors playing on the increase may turn to a strategy consisting of buying a call option with a strike price L2 and selling a call option with a higher strike price K2 > L2 , while the investors playing on the decrease may turn to a strategy consisting of selling a call option with a strike price L1 and buying a call option with a higher strike price K1 > L1 . Such combinations are called spread options of ‘bull’ and ‘bear’, respectively, and their payoff functions are given by G2 (x) and −G1 (x) from equation (14.2.4), where x denotes the stock price (see Shiryaev (1999) [31] Chapter VI, Section 4e). In this present paper, we consider a contingent claim with arbitrary (random) times of exercise τ and cancellation σ , where according to the conditions of the claim the buyer can choose the exercise time τ and in case τ ≤ σ gets the value G2 (Xτ ) from the seller, and the seller can choose the cancellation time σ and in case σ < τ gives the value G1 (Xσ ) to the buyer. Then, by virtue of the fact that Px is a martingale measure for the given market model (see, e.g. Shiryaev et al. (1994) [29] Section 1, Shiryaev (1999) [31] Chapter VII, Section 3g, and Kifer (2000) [20] Section 3), the value (equation (14.2.3)) may be interpreted as a rational (fair) price of the mentioned contingent claim in the given model. We also observe that from the structure of the problem (equation (14.2.3)) it is intuitively clear that the buyer wants to stop when the process X comes close to L1 (from above) while the seller wants to stop when the process X comes close to K2 (from below) without waiting too long because of the punishment of discounting. Taking into account the arguments stated above, we will call the presented contingent claim a spread game option. Note that the structure of the given option differs from the structure of the game options considered in Kifer (2000) [20] and Kyprianou (2004) [24].

296

Exotic Option Pricing and Advanced L´evy Models

14.3 SOLUTION OF THE FREE-BOUNDARY PROBLEM By means of standard arguments, it is shown that the inﬁnitesimal operator L of the process X acts on an arbitrary function F from the class C 2 on (0, ∞) according to the rule: (LF )(x) = rx F (x) + (θ 2 x 2 /2) F (x)

(14.3.1)

for all x > 0. In order to ﬁnd explicit expressions for the unknown value function V∗ (x) from equation (14.2.3) and the boundaries A∗ and B∗ from equations (14.2.5) and (14.2.6), using the results of general theory of optimal stopping problems for continuous time Markov processes as well as taking into account the results about the connection between optimal stopping games and free-boundary problems (see, e.g. Grigelionis and Shiryaev (1966) [12] and Shiryaev (1978) [30] Chapter III, Section 8; as well as Bensoussan and Friedman (1974, 1977) [3] [4]), we can formulate the following free-boundary problem: (LV )(x) = (λ + r)V (x) V (A+) = A − L1 , V (x) = G1 (x)

for

A<x B

A<x0 B γ1 +1 (γ1 − 1)(γ2 − 1)(B − D2 ) 0. By virtue of the fact that the time spent by the process X at the points L1 , A∗ , B∗ and K2 is of Lebesgue measure zero, from the expression (14.4.2) it therefore follows that the inequalities: e−(λ+r)(σ∗ ∧τ ) G2 (Xσ∗ ∧τ ) ≤ e−(λ+r)(σ∗ ∧τ ) V (Xσ∗ ∧τ ) ≤ V (x) + Mσ∗ ∧τ

(14.4.6)

e−(λ+r)(σ ∧τ∗ ) G1 (Xσ ∧τ∗ ) ≥ e−(λ+r)(σ ∧τ∗ ) V (Xσ ∧τ∗ ) ≥ V (x) + Mσ ∧τ∗

(14.4.7)

are satisﬁed for any ﬁnite stopping times σ and τ of the process X. Let (τn )n∈N be an arbitrary localizing sequence of stopping times for the process (Mt )t≥0 . Then, by using inequalities (14.4.6) and (14.4.7) and taking the expectations with respect to

The Spread Option Optimal Stopping Game

301

Px , by means of the optional sampling theorem (see, e.g. Jacod and Shiryaev (1987) [14] Chapter I, Theorem 1.39) we get: * + Ex e−(λ+r)(σ∗ ∧τ ∧τn ) G1 (Xσ∗ )I (σ∗ < τ ∧ τn ) + G2 (Xτ ∧τn )I (τ ∧ τn ≤ σ∗ ) * + ≤ Ex e−(λ+r)(σ∗ ∧τ ∧τn ) V (Xσ∗ ∧τ ∧τn ) ≤ V (x) + Ex Mσ∗ ∧τ ∧τn = V (x)

(14.4.8)

* + Ex e−(λ+r)(σ ∧τ∗ ∧τn ) G1 (Xσ ∧τn )I (σ ∧ τn < τ∗ ) + G2 (Xτ∗ )I (τ∗ ≤ σ ∧ τn ) * + (14.4.9) ≥ Ex e−(λ+r)(σ ∧τ∗ ∧τn ) V (Xσ ∧τ∗ ∧τn ) ≥ V (x) + Ex Mσ ∧τ∗ ∧τn = V (x) for all x > 0. Hence, letting n go to inﬁnity and using Fatou’s lemma, we obtain that for any ﬁnite stopping times σ and τ the inequalities: * + Ex e−(λ+r)(σ∗ ∧τ ) G1 (Xσ∗ )I (σ∗ < τ ) + G2 (Xτ )I (τ ≤ σ∗ ) * + ≤ V (x) ≤ Ex e−(λ+r)(σ ∧τ∗ ) G1 (Xσ )I (σ < τ∗ ) + G2 (Xτ∗ )I (τ∗ ≤ σ )

(14.4.10)

hold for all x > 0. In order to show that the equalities in expression (14.4.10) are attained at σ∗ and τ∗ from equations (14.2.5) and (14.2.6), let us use the fact that the function V (x) solves the equation (14.3.2) for all A∗ < x < B∗ . In this case, by the expression (14.4.2) and the structure of the stopping times σ∗ and τ∗ , it follows that the equality: e−(λ+r)(σ∗ ∧τ∗ ∧τn ) V (Xσ∗ ∧τ∗ ∧τn ) = V (x) + Mσ∗ ∧τ∗ ∧τn

(14.4.11)

holds, from where, by using the expressions (14.4.6) and (14.4.7), we may conclude that the inequalities: −(K1 − L1 ) ≤ Mσ∗ ∧τ∗ ∧τn ≤ K2 − L2

(14.4.12)

are satisﬁed for all x > 0, where (τn )n∈N is a localizing sequence for (Mt )t≥0 . Hence, letting n go to inﬁnity in the expression (14.4.11) and using the conditions (14.3.3), as well as the obviously fulﬁlled property Px [σ∗ ∧ τ∗ < ∞] = 1 (see, e.g. Shiryaev et al. (1994) [29] Section 8, or Shiryaev (1999), [31] Chapter VIII, Section 2a), by means of the Lebesgue bounded convergence theorem we obtain the equality: * + Ex e−(λ+r)(σ∗ ∧τ∗ ) G1 (Xσ∗ ) I (σ∗ < τ∗ ) + G2 (Xτ∗ ) I (τ∗ ≤ σ∗ ) = V (x)

(14.4.13)

for all x > 0, which together with (14.4.10) directly implies the desired assertion.

302

Exotic Option Pricing and Advanced L´evy Models

14.5 CONCLUSIONS Recall that throughout the paper and particularly in the proof of Theorem 4.1 we have used the assumption that L2 (λ + r)/λ ≤ K2 among others. When the latter condition fails to hold but L1 (λ + r)/λ ≤ K1 holds, let us set B∗ = K2 in equation (14.2.6) and consider the problem (14.2.3) as an optimal stopping problem for the seller. In this case, we can also formulate the free-boundary problem (equations (14.3.2)–(14.3.5)), where L1 ≤ A ≤ D1 and B = K2 with D1 = L1 (λ + r)/λ, and assume that the following condition holds: V (A+) = 1 (smooth ﬁt).

(14.5.1)

By means of the same arguments as in Section 14.3, by using the assumed smoothﬁt condition (14.5.1), it can be shown that the boundary A should satisfy the following equation: γ1 (A − L1 )(K2 /A)γ2 − (K2 − L2 ) γ2 (K2 − L2 ) − (A − L1 )(K2 /A)γ1 + = 1. (14.5.2) A (K2 /A)γ2 − (K2 /A)γ1 A (K2 /A)γ2 − (K2 /A)γ1 In order to ﬁnd sufﬁcient conditions for the existence and uniqueness of solution of the equation (14.5.2) let us deﬁne the function H (A) by: H (A) = [(γ1 − 1)A − γ1 L1 ](K2 /A)γ2 − [(γ2 − 1)A − γ2 L1 ](K2 /A)γ1 + (γ2 − γ1 )(K2 − L2 )

(14.5.3)

for all A such that L1 ≤ A ≤ D1 . By virtue of the fact that for the derivative of the function (14.5.3) the following expression holds:

(γ1 − 1)(γ2 − 1)(A − D1 ) K2 γ2 K2 γ1 K2 , and the conditions (14.5.5) and (14.5.6) are satisﬁed. Then, the value function of the problem (14.2.3) takes the expression (14.4.1) and the optimal stopping times σ∗ and τ∗ have the structure (14.2.5) and (14.2.6) with B∗ = K2 , where the function V (x; A, B) is explicitly given by

The Spread Option Optimal Stopping Game V

303

K1 − L1 = K2 − L2

V∗ (x)

L1 A∗

D1

K1

L2

B∗ = K2 D2 x

Figure 14.2 A computer drawing of the value function V∗ (x) and the optimal stopping boundaries A∗ and K2 V

K1 − L1 = K2 − L2

V∗ (x)

A∗ = L1

D1

K1

L2

B∗ = K2 D2 x

Figure 14.3 A computer drawing of the value function V∗ (x) and the optimal stopping boundaries L1 and K2

equation (14.3.12), and A∗ satisfying the inequalities L1 ≤ A∗ ≤ L1 (λ + r)/λ is determined as a unique solution of the equation (14.5.2) (see Figure 14.2). The veriﬁcation of this assertion can be carried out by means of a slight modiﬁcation of the arguments from the proof of Theorem 4.1, using also the facts that the condition (14.5.6) implies that V (K2 ; D1 , K2 ) < 1 and the function V (K2 ; A, K2 ) is increasing in A on the interval (L1 , D1 ). It is seen that the smooth-ﬁt condition at the point B∗ breaks down in this case. We also note that when the condition (14.5.5) fails to hold, almost the same arguments show that (even when the condition L1 (λ + r)/λ ≤ K1 fails to hold too) the assertion of Proposition 5.1 remains true with A∗ = L1 , while the smooth-ﬁt condition at A∗ also breaks down (see Figure 14.3). Remark 5.2. We also mention that another interesting but difﬁcult question is to present a complete description of the behavior of the optimal stopping boundaries A∗ and B∗ from equations (14.2.5) and (14.2.6) under the changing of the parameters of the model.

ACKNOWLEDGMENTS The author thanks Andreas E. Kyprianou and two anonymous Referees for valuable comments and suggestions.

304

Exotic Option Pricing and Advanced L´evy Models

REFERENCES [1] Alario-Nazaret, M., Lepeltier, J.P. and Marchal, B. (1982), “Dynkin games”, in Stochastic Differential Games (Bad Honnef), Lecture Notes in Control and Information Sciences, Vol. 43, Springer-Verlag, Berlin, Germany, pp. 23–32. [2] Baurdoux, E. and Kyprianou, A. (2004), “Further calculations for Israeli options”, Stochastics and Stochastic Reports, to be published. [3] Bensoussan, A and Friedman, A. (1974), “Non-linear variational inequalities and differential games with stopping times”, Journal of Functional Analysis, 16, 305–352. [4] Bensoussan, A and Friedman, A. (1977), “Nonzero-sum stochastic differential games with stopping times and free-boundary problems”, Transactions of the American Mathematical Society, 231, 275–327. [5] Bismut, J.M. (1977), “Sur un probl`eme de Dynkin”, Zeitschrift f¨ur Wahrscheinlichkeitstheorie und Verwandte Gebiete, 39, 31–53. [6] Cvitani´c, J. and Karatzas, I. (1996), “Backward stochastic differential equations with reﬂection and Dynkin games”, Annals of Probability, 24, 2024–2056. [7] Dynkin, E.B. (1969), “Game variant of a problem on optimal stopping”, Soviet Mathematical Doklady, 10, 270–274. [8] Elbakidze, N.V. (1976), “Construction of the cost and optimal policies in a game problem of stopping a Markov process”, Theory of Probability and its Applications, 16, 163–168. [9] Frid, E.B. (1969), “Optimal stopping of a Markov chain of two persons with opponent interests”, Theory of Probability and its Applications, 14, 746–749. [10] Friedman, A. (1973), “Stochastic games and variational inequalities”, Archive for Rational Mechanics and Analysis, 51, 321–346. [11] Friedman, A. (1976), Stochastic Differential Equations and Applications II, Academic Press, New York, NY, USA. [12] Grigelionis, B.I. and Shiryaev, A.N. (1966), “On Stefan’s problem and optimal stopping rules for Markov processes”, Theory of Probability and its Applications, 11, 541–558. [13] Jacod, J. (1979), Calcul Stochastique et Probl`emes de Martingales, Lecture Notes in Mathematics, Vol. 714, Springer-Verlag, Berlin, Germany. [14] Jacod, J. and Shiryaev, A.N. (1987), Limit Theorems for Stochastic Processes, Springer-Verlag, Berlin, Germany. [15] Kallsen, J. and K¨uhn, C. (2004), “Pricing derivatives of American and game type in incomplete markets”, Finance and Stochastics, 8, 261–284. [16] Kallsen, J. and K¨uhn, C. (2005), “Convertible bonds: ﬁnancial derivatives of game type”, in A.E. Kyprianou, W. Schoutens and P. Wilmott (Eds), Exotic Option Pricing and Advanced L´evy Models, Wiley, Chichester, UK, pp. 277–291. [17] Karatzas, I. and Wang, H. (2001), “Connection Between Bounded-Variation Control and Dynkin games”, in J.L. Menaldi et al. (Eds), Optimal Control and Partial Differential Equations (Volume in Honour of Alain Bensoussan), IOS Press, Amsterdam, The Netherlands, pp. 363–373. [18] Kifer, Yu.M. (1971a), “Optimal stopped games”, Theory of Probability and its Applications, 16, 183–188. [19] Kifer, Yu.M. (1971b), “Optimal stopping in games with continuous time”, Theory of Probability and its Applications, 16, 545–550. [20] Kifer, Y. (2000), “Game options”, Finance and Stochastics, 4, 443–463. [21] Krylov, N.V. (1971), “Control of Markov processes and W-spaces”, Izvestija, 5, 233–266. [22] K¨uhn, C. and Kyprianou, A. (2003a), “Callable puts as composite exotic options”, to be published.

The Spread Option Optimal Stopping Game

305

[23] K¨uhn, C. and Kyprianou, A. (2003b), “Pricing Israeli options: a pathwise approach”, to be published. [24] Kyprianou, A. (2004), “Some calculations for Israeli options”, Finance and Stochastics, 8, 73–86. [25] Lepeltier, J.P. and Mainguenau, M.A. (1984), “Le jeu de Dynkin en th´eorie g´en´erale sans l’hypoth`ese de Mokobodski”, Stochastics, 13, 25–44. [26] Morimoto, H. (1984), “Dynkin games and martingale methods”, Stochastics, 13, 213–228. [27] Neveu, J. (1975), Discrete-Parameter Martingales, North-Holland, Amsterdam, The Netherlands. [28] Protter, Ph. (1992), Stochastic Integration and Differential Equations: A New Approach, Applications of Mathematics, Vol. 21, Springer-Verlag, Berlin, Germany. [29] Shiryaev, A.N., Kabanov, Y.M., Kramkov, D.O. and Melnikov, A.V. (1994), “On the pricing of options of European and American types, II. Continuous time”, Theory of Probability and its Applications, 39, 61–102. [30] Shiryaev, A.N. (1978), Optimal Stopping Rules, Springer-Verlag, Berlin, Germany. [31] Shiryaev, A.N. (1999), Essentials of Stochastic Finance, World Scientiﬁc, Singapore. [32] Snell, J.L. (1952), “Applications of martingale systems theorems”, TAMS, 73, 293–312. [33] Stettner, L. (1982), “Zero-sum Markov games with stopping and impulsive strategies”, Applied Mathematics and Optimization, 9, 1–24.

Index ACCs see American contingent claims additive processes see time-inhomogeneous L´evy processes Albrecher, Hansj¨org 129–47 Almendral, Ariel 259–75 American contingent claims (ACCs) 282, 293–4 see also game options American options 5–16, 29, 31–6, 106, 113–14, 149–50, 195–215, 237–56, 259–75, 278–9 Asian options 217–34 CGMY process 238–9, 259–75 forward equations 237–56, 259–76 LCP 259–76 perpetual American options 16, 113–14, 195–215, 271 PIDEs 113–14, 237–56, 259–75 pricing 5–6, 10, 13–16, 29, 31–6, 106, 113–14, 149–50, 195–215, 237–56, 259–75 puts 5, 6, 13–14, 15–16, 31–2, 35–6, 195–215, 217–34, 239–56, 259–76 arbitrage 52, 105–6, 137, 145, 150, 170–92, 217–21, 231, 250–1, 277–86 classical theory 175, 217–18 concepts 175, 183–92, 217–21, 231 game options 277–86 market completion 183–92

arithmetic averages, early exercise Asian options 217–18 Arrow Debreu Securities 64 Asian options 10, 100, 114–19, 129–47, 217–34 see also early exercise. . . American type 217–34 concepts 114–19, 129–47, 217–34 optimal stopping problems 217–34 pricing 114–19, 129–47, 218–34 static super-hedging strategy 129–47 valuations 114–19, 129–47, 218–34 asset-or-nothing options 121–3 at-the-money options 130–45 autocorrelation, squared returns 58–9 average rate call options 39–41 average waiting time, investment decisions 156–65 backward equations 237–47, 293–4 backward free boundary problems 239–47 bankruptcies, convertible bonds 287 Barndorff-Nielsen–Shephard model (BN–S) 9, 31, 54, 67, 70–95, 132–43 barrier options 15–16, 29, 35, 40–8, 80–6, 115–17, 130 Barrieu, Pauline 149–68 basket options 117–19 bear markets 295 Bellamy, Nadine 149–68 Bermudan options 10, 31–2, 35–6, 114

Exotic Option Pricing and Advanced L´evy Models. Edited by A. E. Kyprianou, W. Schoutens and P. Wilmott Copyright 2005 John Wiley & Sons, Ltd.

308

Index

Bessel function 9–10, 24 beta distribution 37–42 bias correction algorithms 43–8 simulation methods 29, 42–8 bilateral Laplace transforms 111–13, 121–3 binned data, statistical density 58–9 Black-Scholes pricing model assumptions 4–5, 10, 29, 67–9, 106–7, 129, 178, 182 concepts 4–5, 10–14, 29, 67–9, 74, 106–7, 114–19, 129–30, 136–7, 178, 182, 237–9, 247, 263–4, 277, 283, 293–5 Israeli options 13–14, 293–4 Lagrange multipliers 137 SDE 182, 294–5 stochastic-volatility contrasts 129 ‘suicide’ strategies 283 Blumenthal 0–1 law 16 BN–S see Barndorff-Nielsen–Shephard model bonds 52, 169–92, 277–91 contingent claims 52 convertible bonds 277–91 counterparty default 52 Borel function 183, 222 bounded variation, path properties 12, 14–24, 103–4, 112, 136, 178–9, 271, 293–4 Boyarchenko, S.I. 1, 261–2, 271 Brennan, M.J. 150, 259–60, 269–70, 286–7 Brennan–Schwartz algorithm 259–60, 269–70 bridge algorithms concepts 29, 36–48, 117 Monte Carlo simulation 39–42, 117 stratiﬁed sampling 36–42 subordinator representation 37–48 Brownian motion 4–5, 10, 14–17, 30–48, 69–71, 102, 109–11, 122–3, 131–5, 150–1, 160–1, 177–219, 233–4, 259–64, 293–303 see also normal distributions; Wiener processes

market completion 177–92 stable processes 212 BS see Black-Scholes pricing model bull markets 295 c`adl`ag paths 278, 281–5 calibration, model risk 74–8 call options 39–41, 68–95, 106–14, 121–3, 136–45, 149–50, 196, 197–8, 217–34, 259–75, 295–303 callable put options 13–14 caps 99 Carr, Peter 31, 56–63, 68, 73, 74, 90, 106–7, 114, 116, 133–5, 139–40, 145, 170, 196, 215, 237–57, 259–60, 270–1, 275 Carr–Geman–Madan–Yor (CGMY) process 1–2, 4, 7–9, 21–3, 25, 54–65, 71–95, 133–5, 238–9, 259–75 see also generalized tempered stable processes; ‘inﬁnite activity’ L´evy processes American options 238–9, 259–75 change of measure density 54–65 characteristic function 260–2 concepts 1–2, 4, 7–9, 21–3, 25, 54–65, 71–95, 133–5, 238–9, 259–75 European options 271–5 jump models 54–65, 260–75 numerical valuation of American prices 263–71 path properties 21–3, 25, 271 risk-neutral densities 54–65, 261–2 statistical densities 54–65 cash ﬂows, NPV 149–50, 155–6 cash-or-nothing options 121–3 Cauchy sequence 11, 262, 264 CCGMYY processes 7 CGMY process see Carr–Geman–Madan–Yor. . . Chan, Terence 195–216 change of measure density see also Radon–Nikodym derivative concepts 52–65, 99, 105–6, 108–11, 122–3, 217–19 estimation details 57–63

Index

character function inversion, FFT usage 138–9 chi-squared goodness of ﬁt statistic 59 CIR see Cox–Ingersoll–Ross process circulant convolutions, concepts 267–70 classical theory, arbitrage 175, 217–18 cliquet options 68, 81–8, 95, 130 comonotonicity theory, concepts 129–30, 137–47 complete markets 169–92, 280–1 composite trapezoidal rule, spatial discretizations 266–7 compound options 238 compound Poisson process concepts 4–5, 11–12, 17, 30–3, 55–6, 69–70, 132–5, 150–1, 160–1, 177–8, 185–6, 204–6, 264 path properties 17 computer drawings, optimal stopping problems 299, 303 Cont, R. 1, 259–61, 264, 266 contingent claims 51–65, 169–92, 277–89, 294–303 continuation region, exercise boundary 241–5 continuous barrier options 40–8 continuous junction condition, concepts 203–4 continuous-time setting, market completion 169–70, 177–92 continuous/discontinuous models, investment decisions 161–5 continuously reset path-dependent options, valuations 40–8 convertible bonds 277–91 see also game options bankruptcies 287 concepts 277–91 coupon payments 287–9 deﬁnition 277, 287 ﬁrm value 286–9 literature 286–9 perpetual model 286–9 reduced form models 289 convolutions, circulant convolutions 267–70 Corcuera, Jos´e Manuel 169–93

309

correction algorithms, simulation bias 43–8 counterparty default, bonds 52 coupon payments, convertible bonds 287–9 course path properties, L´evy processes 1–2, 10–28 Cox–Ingersoll–Ross process (CIR) 69–95, 131–44 see also stochastic clocks creeping, path properties 1, 14–24 critical stock prices, DEVG model 240–2, 246–50

DAX 61–4 deﬁnitions, L´evy processes 2–4 Delbaen, F. 279–85 delta 130, 241–2, 274 DEVG see Diffusion Extended Variance Gamma DFT see discrete Fourier transform DIB see down-and-in barrier options Diffusion Extended Variance Gamma (DEVG) 241–56 digital barriers 68, 80–1, 82–6, 111 Dirac measure 152, 232 Dirichlet conditions 242, 244–9 discontinuous L´evy processes, real options 155–6 discontinuous martingales, stable processes 211–14 discontinuous models investment decisions 161–5 real options 155–6 discount rates investment decisions 149–66 perpetual American options 197–8, 203–4 discounted payoff function, moment derivatives 176 discrete Fourier transform (DFT) 267–70 discrete LCP, concepts 269–70 discrete-time setting, market completion 169–92 discretely reset path-dependent options, valuations 39–40

310

Index

discretization ﬁnite differences 251–6, 260, 263–75 forward equations 251–6, 263–75 distributional characteristics concepts 4–5, 29, 54–65, 195, 197–202 normal distributions 29–30, 133–5 skewed distributions 29–30, 53–65, 86–95, 169, 180 dividends, convertible bonds 287–9 DOB see down-and-out barrier options Doob–Meyer decomposition 183–4 down-and-in barrier options (DIB) 81–6 down-and-out barrier options (DOB) 80–6 drift 4–5, 17, 33–6, 43–4, 54–65, 102, 240–2, 260–2 Dupire equation 237–9 dynamic hedging 130 dynamic programming 150 dynamic trading strategies 68, 91–5, 130–47 Dynkin’s games 278, 280–1, 288–9, 293–4 see also optimal stopping problems

early exercise Asian options see also Asian options concepts 217–34 numerics 231–3 optimal stopping boundary 217–34 premiums 220–1 pricing 218–34 probability density function 231–4 problem formulation 218–20 proof 220–31 Eberlein, Ernst 31, 99–128, 184–5, 238 efﬁcient markets 52 EMM see equivalent martingale measures enlargement, L´evy market model 179–82 equity indexes 52, 60–5, 130–1, 140–4 equivalent martingale measures (EMM) 183–92, 195–6, 261–2 Esscher measure 191 estimation details, change of measure density 57–63 Euclidean scalar product 101 Euler’s theorem 246–7

European options 29, 31, 35, 67–95, 99, 106–14, 123, 130, 136–45, 237–9, 259, 261–2, 270–5, 278–9, 288–9 calls 68–95, 106–14, 123, 136–45 CGMY 271–5 forward equations 238–9, 270–5 pricing 29, 31, 35, 67–95, 99, 106–14, 123, 130, 136–45, 237–9, 259, 261–2, 270–5 puts 106–14, 261–2, 270–5 Eurostoxx 50 index 75–8, 82–6 exceedance probabilities, barrier options 44–8 excursion theory, concepts 14–15, 113–14, 116 exotic options see also individual types concepts 1–28, 80–95, 114–19 model risk 67–97, 131 path dependency 10, 29, 31–2, 39–48, 67–95, 113, 197 pricing 74–8, 80–95, 99–123, 129–47, 195–215, 218–34 super-hedging strategy 129–47 symmetries 99–123 types 10, 13–14, 15–16, 80–6, 95, 99–100, 114–19 explicit ﬁnite-difference methods 114 exponential L´evy processes 101–23, 130, 178–9, 189–92, 196–7, 264 exponential PIIAC, time-inhomogeneous L´evy processes 101–23 fast Fourier transform (FFT) 74–5, 129–30, 138–45, 209–14, 259–76 Fatou’s lemma 301 FBPs see free boundary problems FFT see fast Fourier transform ﬁnancial mathematics, objectives 29 ﬁne path properties, L´evy processes 1–2, 10–28 ﬁnite expiry American puts 16 ﬁnite markets market completion 172–92 multi-step markets 173–4 one-step markets 172–3 ﬁnite moment logstable model 238–9

Index

ﬁnite variation, L´evy processes 200–13, 260 ﬁnite-difference methods 113–14, 237–57, 259–76 ﬁrm value, convertible bonds 286–9 ﬁrst-passage distributions 195, 197–202 ﬁxed strike Asian options 100, 114–19 ﬁxed strike lookback options 42–8, 100, 115–19 ﬂoating strike Asian options 100, 114–19, 217–34 see also Asian options ﬂoating strike lookback options 42–8, 100, 115–19 ﬂuctuation theory, concepts 10 foreign exchange 102–23 forward equations American options 237–56, 259–76 concepts 232–3, 237–76 DEVG model 241–56 discretization 251–6, 263–75 European options 238–9 hybrid equations 239–56 initial-value problem 232–3, 263 uses 237–8 forward free boundary problems 247–56, 262–71 forward-start options 114–19 Fourier transform methods 2, 31, 56, 57–60, 74–5, 112, 116–17, 129–30, 138–45, 209–14, 259–76 free boundary problems (FBPs) 239–51, 262–71, 293–303 see also optimal stopping. . . concepts 239–51, 262–71, 293–303 solution 296–302 FTSE 61–4 futures prices 51–5 g-moment, PIIAC 103–4 game contingent claim (GCC), concepts 277–89, 294 game options 10, 13–14, 277–303 see also convertible bonds; Israeli. . . arbitrage 277–86 concepts 10, 13–14, 277–303

311

deﬁnition 277–8, 293–4 NFLVR 278–85 optimal stopping problems 278, 280–1, 288–9, 293–4 pricing 277–89 spread game options 293–303 gamma process 1–2, 4, 8–9, 10, 22–3, 25, 32–48, 56–65, 70–95, 117, 132–5, 142–5, 209–14, 242 Gamma-OU stochastic clock 73–4, 79–80, 132–3, 135–44 see also stochastic clocks Gapeev, Pavel 293–305 GARCH models 251 Gaussian processes see also normal distributions concepts 1–2, 9–10, 15, 17, 21, 23–4, 29, 32–48, 53–4 GCC see game contingent claim Geman, H´elyette 51–66, 99, 105, 259 see also Carr–Geman–Madan–Yor process general diffusion model 106 generalized hyperbolic processes see also normal inverse Gaussian. . . concepts 1–2, 5, 9–10, 23–4, 31–48, 54, 71–95, 100–23, 133–44 path properties 23–4 variance gamma process 10 generalized inverse Gaussian distributions (GIG) 33–6 generalized tempered stable processes see also Carr–Geman–Madan–Yor process; truncated. . . ; variance gamma. . . concepts 1–2, 4, 7–9, 19–23, 25, 32–48, 54–65, 71–95, 260–2 path properties 19–23, 25 geometric averages, early exercise Asian options 217–18 geometric Brownian motion 42–3, 71–4, 218–19, 293–303 geometric L´evy model see exponential L´evy processes Gerber, H.U. 196–8, 203–6 German equity indexes, risk-neutral densities 60–5

312

Index

GIG see generalized inverse Gaussian distributions Girsanov’s theorem 106–11 the greeks 238 half lines 6, 16–24 HARA utilities 187–8 Hartman–Watson density 231–2 Heaviside function 255 hedging 29, 57, 68, 90–5, 118–19, 129–47, 169–92 Asian options 129–47 concepts 90–5, 129–47, 169–92 moment derivatives 90–5, 169–92 moment swaps 90–5 static super-hedging strategy 129–47 strategy performance 140–4 Hellinger distances, densities 59 Heston Stochastic Volatility model (HEST) 67–8, 69–95, 131–43 concepts 67–8, 69–70, 82–6, 131–43 jumps 69–70, 132–3 high contact condition, concepts 241–2 Hilbert space 11 Hirsa, Ali 237–57 hitting points concepts 12–14, 17–24, 222–3 path properties 12–14, 17–24 holders, game options 277–91, 293–4 Hunt density 240–2 hybrid equations, forward equations 239–56 hyperbolic process, concepts 9–10, 118–19, 238–9 IBEX 61–4 IDFT see inverse discrete Fourier transform implicit ﬁnite-difference methods 113–14 implicit function theorem 298–301 implied volatilities 29–30, 67–95, 99–123 in-progress Asian options 115 in-the-money options 144–5, 149–50 incomplete markets 169–92, 294 independent increments, L´evy processes 2–4, 100–23, 237–9, 245–7, 260–1

‘inﬁnite activity’ L´evy processes see also Carr–Geman–Madan–Yor process concepts 54–65, 261–2, 264 inﬁnite variation, L´evy processes 200–13 inﬁnitely divisible distributions 2–4, 9, 54–65, 68–74, 131, 177–92 initial-value problem, forward equations 232–3, 263 inner expectations, lattice methods 35–6 instantaneous returns, jump perspectives 58–9 instantaneous volatility 237–56 insurance premiums 54–65 integral equations early exercise Asian options 217–34 PIDEs 113–19, 237–56, 259–75 interest rates models 29–33 risk-free interest rates 75, 174–5, 260–1 simulation methods 29–30, 32–48 inverse discrete Fourier transform (IDFT) 268–70 Inverse Gaussian (IG) random numbers 78 see also normal inverse. . . inverse transform 36–40, 268–70 investment decisions see also real options average waiting time 156–65 concepts 149–65 continuous/discontinuous models 161–5 discount rates 149–66 opportunity value 154–5, 163–6 optimal discount rates 149–50, 156–66 optimal times 150–65, 166, 197, 203–13 optimization 149–66, 197, 203–6 proﬁts/costs ratio 149–65, 167 random jump sizes 160–1, 166–7 relative errors 162–5 robustness checks 158–65 Israeli options 10, 13–14, 293–4 see also game. . . issuers, game options 277–91, 293–4 Itˆo’s formula 218–19, 262, 300 Itˆo –Tanaka–Meyer formula 300

Index

Jacod, J. 278, 300–1 Japanese equity indexes, risk-neutral densities 60–5 joint returns distributions 29–30, 59 jump perspectives concepts 4, 11–12, 51–65, 68–70, 89, 102, 113–17, 149–65, 177–92, 196–8, 238–56, 260–2 HEST 69–70, 132–3 instantaneous returns 58–9 market crises 149–65 optimal discount rates 156–65 perpetual American options 196–8 risks in returns 51–65 jump-diffusion models, concepts 1–2, 4–5, 11–12, 17, 53–65, 116–17, 149–65, 238–56, 260–2 Kallsen, Jan 277–91, 294 Karatzas, I. 277, 280–1, 293 Kifer, Y. 277, 295 KoBoL processes see generalized tempered stable processes Kolmogorov–Smirnov statistic 59 Kolodner, I.I. 217 Kou model 1–2, 5, 260–1 see also jump-diffusion model Kou, S.G. 1–2, 5, 260–1, 277 K¨uhn, Christoph 277–91, 293–4 kurtosis levels 53–4, 86–95, 169, 180 Kyprianou, Andreas E. 1–28, 259, 277, 293–5 Lagrange multipliers 137, 187 Laplace transforms 111–13, 121–3, 149, 151, 153–65, 174, 197, 199–202, 205, 209–14 investment decisions 149, 151, 153–65 relative errors 162–5 lattice methods, concepts 31–3, 35–6 law of a ﬁrst-passage time of the process 195, 197 LC see lookback options LCPs see linear complementary problems Lebesgue measure 112, 200–2, 278, 300–1 Lepeltier, J. 280, 295

313

Levendorskii, S.Z. 1, 261–2, 271 L´evy exponent, concepts 198–214 L´evy measure, concepts 198–214 L´evy processes see also individual classes; stochastic processes bias 42–8 bridge algorithms 29, 36–48, 117 change of measure density 53–65, 99, 105–6, 108–11, 122–3 classes 1–2, 4–10, 17–24, 48, 54, 71–2, 100, 133–5, 195–215, 259–75 concepts 1–48, 53–65, 71–2, 99–123, 131–5, 169–92, 195–215, 237–56, 259–75 deﬁnitions 2–4, 103, 131, 177, 202–4, 260–1 examples 1, 4–25 exponential L´evy processes 101–23, 130, 178–9, 189–92, 196–7, 264 ﬁnite variation 200–13, 260 ﬂuctuation theory 10 geometric L´evy model 178–9 independent increments 2–4, 100–23, 237–9, 245–7, 260–1 ‘inﬁnite activity’ L´evy processes 54–65, 261–2, 264 introduction 1–48 market completion 169–92 model risk 67–97 modelling 1–2, 4–10, 17–24, 30–48, 53–65, 67–97, 202–4 moment derivatives 67–8, 86–95, 169–92 path properties 1–28, 79–80, 103–4, 112, 136, 178–9, 271 problems 10 random walks 116–17 real options 151–5 risk-neutral densities 53–65, 68–78, 89–90, 93–5, 101–23, 131–45, 170–92, 195–7, 202–4, 240–56 simulation methods 29–30, 31–48, 67–8, 72–95, 114, 117–18, 133–44

314

Index

L´evy processes (Continued) stationary independent increments 2–4, 103, 237, 242–56, 260–1 statistical densities 53–65 stochastic time 71–95, 131, 133–5 symmetries 99–123, 237–56 theorems 2–4 time-changed L´evy process 73–4, 78, 79–80, 93–5, 133–44 time-inhomogeneous L´evy processes 99–123, 245–56 L´evy triple, concepts 3–28, 30–3, 100–11, 119–20, 183–4, 278 L´evy–Itˆo composition, concepts 11–12 L´evy–Khintchine formula, concepts 1, 2–4, 6, 7–8, 10–25, 54–65, 104–5, 177–92, 198–202, 260–1 light tails, distributional characteristics 4–5 linear complementary problems (LCPs) 259–76 Lipschitz constant 286 local time-space calculus 217–34 local volatilities 53–4 Loeffen, Ronnie 1–28 log returns 53–4, 67–8, 90–5, 169–92 lookback options (LC) 10, 40–8, 68, 80–6, 95, 100, 115–19, 130 lower half line regularity 6, 16–24 spectrally one-sided processes 6, 17 Madan, Dilip B. 51–66, 74, 116–17, 140, 170, 237, 238, 241, 259, 270–1, 275 see also Carr–Geman–Madan–Yor process Maingueneau, M. 280, 295 management tools, real options 150 Margrabe-type options 119–23 market completion 169–92, 280–1 continuous-time setting 169–70, 177–92 discrete-time setting 169–92 moment derivatives 169–92 market crises, real options 149–65 Markov process 59, 196–7, 219, 222, 228–9, 239–40, 250, 293–9 martingale representation property (MRP) 180–2

martingales 7, 11–12, 30–63, 100–23, 150–1, 169–92, 195–6, 202–4, 211–14, 221–2, 228–31, 240–2, 278–303 compound Poisson process 11–12, 55–6, 150–1, 185–6 early exercise American options 221–2, 228–31 equivalent martingale measures 183–92, 195–6, 261–2 game options 278–89 Meixner processes 7 moment derivatives 169–92 optimal stopping problems 196, 202–4, 211–14, 221–2, 228–31, 240–2, 278–89, 293–303 semi-martingales 33–6, 101–23, 278–89 Matache, A.M. 259, 262, 271 mean-variance mixtures 31, 32–3, 192 Meixner processes concepts 1–2, 5, 6–7, 17–18, 25, 54, 71–2, 103–4, 133–5 path properties 17–18, 25 memoryless property, exponential L´evy processes 116, 196–7 Merton model 1–2, 5 see also jump-diffusion model Miller–Modigliani hypothesis 287 minimal entropy martingale measure 102 minimal martingale measures 192 minimax martingale measures 102 model risk calibration 74–8 exotic options 67–97, 131 model-independent static super-hedges 145 moment derivatives concepts 67–8, 86–95, 169–92 hedging 90–5, 169–92 market completion 169–92 pricing 86–95 moment options (MOMO) 89–95 moment swaps (MOMS) hedging 90–5 pricing 89–95 Mongolian options 10 Monte Carlo simulation

Index

bridge algorithms 39–42, 117 concepts 31–3, 35–6, 39–40, 67–8, 78–95, 114, 117–18, 142–5 NIG 39–42, 78–95, 117–18, 142–4 problems 36 simulation bias 42–8 stratiﬁed sampling 39–42 VG 39–42, 79–95, 117, 142–4 MRP see martingale representation property multi-step ﬁnite markets, market completion 173–4 net present value (NPV) concepts 149–50, 155–6 weaknesses 149 NFLVR see no free lunch with vanishing risk NIG see normal inverse Gaussian processes NIKKEI 61–4 no free lunch with vanishing risk (NFLVR) 278–85 no-arbitrage pricing, game options 279–86 nonlinear integral equations, early exercise Asian options 217–34 normal distributions 29–30, 133–5 see also Brownian motion; Gaussian. . . normal inverse Gaussian processes (NIG) see also generalized hyperbolic processes concepts 1–2, 9–10, 30, 32–48, 54, 71–95, 117–18, 133–44 Monte Carlo simulation 39–42, 78–95, 117–18, 142–4 simulation methods 32–48, 72–95, 117–18, 133–44 NPV see net present value Nualart, David 169–93 numerical approach, simulation methods 33–6, 113–14 one-dimensional driving processes, symmetries 121–3 one-side L´evy processes, two-sided L´evy processes 215 one-step market models 170–92

315

one-touch barriers 68, 80–6, 95 opportunity value, investment decisions 154–5, 163–6 optimal discount rates 149–50, 156–66 optimal portfolios, concepts 186–92 optimal stopping problems 114, 196–215, 217–34, 237–56, 259–76, 278, 280–1, 286–9, 293–303 see also Dynkin’s games; free boundary. . . computer drawings 299, 303 early exercise Asian options 217–34 forward equations 232–3, 237–76 game options 278, 280–1, 288–9, 293–4 perpetual American options 16, 113–14, 195–215, 271 spread game options 293–303 value function 294–303 optimal times, investment decisions 150–65, 166, 197, 203–13 optimal wealth, concepts 187–92 optional sampling theorem 223–31, 301 options see American. . . ; European. . . ; exotic. . . ; game. . . ; real. . . Ornstein Uhlenbeck process (OU) 70–1, 73–95, 132–44 orthonormal martingales 179–82 OU see Ornstein Uhlenbeck process out-of-the-money options 59–63, 130–45 outer expectations, lattice methods 35–6 Papapantoleon, Antonis 99–128 Parisian options 10 partial differential equations (PDEs) 31, 182, 237–57 partial integro-differential equations (PIDEs) 113–19, 237–56, 259–75 American options 113–14, 237–56, 259–75 concepts 113–19, 237–56, 259–75 forward equations 237–56, 259–75 hybrid equations 242–56 partial integro-differential inequality (PIDI) 113–19 passport options 278

316

Index

path dependency, exotic options 10, 29, 31–2, 39–48, 67–95, 113, 197 path properties bounded/unbounded variation 12, 14–24, 103–4, 112, 136, 178–9, 271, 293–4 concepts 1–2, 10–24 L´evy processes 1–28, 79–80, 103–4, 112, 136, 178–9, 271 types 10–24 path variation, path properties 10–24 PDEs see partial differential equations perpetual American options 16, 113–14, 195–215, 271, 293–4 concepts 16, 113–14, 195–215, 271, 293–4 discount rates 197–8, 203–4 jump perspectives 196–8 pricing 113–14, 195–215, 271 renewal equation pricing approach 204–6 spectrally one-sided processes 195–215 perpetual convertible bonds, concepts 286–9 perpetual spread game options 293–303 perpetual warrants 196 Peskir, Goran 217–35 PIDEs see partial integro-differential equations PIDI see partial integro-differential inequality PIIAC see process with independent increments and absolutely continuous characteristics PIIS see process with independent and stationary increments Poisson process, concepts 4–5, 11–12, 17, 30–3, 55–6, 69–71, 132–5, 150–1, 160–1, 177–8, 185–6, 204–6, 264 portfolios, optimal portfolios 186–92 power options concepts 106–11, 174–92 symmetries 106–11 power-jump processes, concepts 178–9 power-return assets, market completion 174–92

premiums creeping 15–17 early exercise Asian options 220–1 insurance premiums 54–65 risks 51–65 pricing see also valuation methods American options 5–6, 10, 13–16, 29, 31–6, 106, 113–14, 149–50, 195–215, 237–56, 259–75 Asian options 114–19, 129–47, 218–34 Black-Scholes pricing model 4–5, 10, 13–14, 29, 67–9, 74, 293 early exercise Asian options 218–34 European options 29, 31, 35, 67–95, 99, 106–14, 123, 130, 136–45, 237–9, 259, 261–2, 270–5 exotic options 74–8, 80–95, 99–123, 129–47, 195–215, 218–34 forward equations 232–3, 237–56 game options 277–89 GCC 277–89 moment derivatives 86–95 perpetual American options 113–14, 195–215, 271 renewal equation approach 204–6 spread game options 293–303 swaps 89–95 symmetries 21–2, 99–123 vanilla options 106–14, 121–3, 261–2 principle of smooth pasting 155–6, 206 process with independent increments and absolutely continuous characteristics (PIIAC), time-inhomogeneous L´evy processes 100–23 process with independent and stationary increments (PIIS) 103, 237, 242–56, 260–1 see also L´evy processes proﬁts/costs ratio, investment decisions 149–65, 167 put options 5, 6, 13–16, 31–2, 35–6, 100, 106–14, 121–3, 195–234, 239–56, 259–79, 284, 288–9 put–call parity 100 p–value, chi-squared goodness of ﬁt statistic 59

Index

qq–plots 59 quadratic utility 191–2 quanto options 110–11, 119–23 Radon–Nikodym derivative 52, 105–6, 108–11, 122–3 see also change of measure density Raible, Sebastian 111–13, 262 random jump sizes, investment decisions 160–1, 166–7 random numbers, simulations 78–9 random walks, L´evy processes 116–17 real options see also investment decisions characteristics 155–6 concepts 149–65 L´evy processes 151–5 management tools 150 market crises 149–65 models 151–5 optimal discount rates 149–50, 156–66 optimal times 150–65, 166 real valued L´evy processes, deﬁnitions 2–4 reduced form models, convertible bonds 289 regular L´evy processes of exponential type (RLPE) 117 regularity of the half line, path properties 6, 16–24 relative errors, Laplace transforms 162–5 renewal equation approach, pricing 204–6 returns jump models 51–65 models 29–33 risks 51–65 simulation methods 29–30, 32–48, 71–95 Ribeiro, C. 37–48 risk management 29, 57 risk-free interest rates 75, 174–5, 260–1 risk-neutral densities CGMY process 54–65, 261–2 concepts 51–65, 68–78, 89–90, 93–5, 101–2, 131–47, 170–92, 195–7, 202–4, 240–2, 261–2 equity indexes 60–5, 68–78 estimation details 57–63, 68–78, 131

317

L´evy processes 53–65, 68–78, 89–90, 93–5, 101–23, 131–45, 170–92, 195–7, 202–4, 240–56, 261–2 stochastic volatility models 131–47 risks jump models 51–65 model risk 67–97, 131 NFLVR 278–85 premiums 51–65 returns 51–65 two-sided features 65 RLPE see regular L´evy processes of exponential type robustness checks, investment decisions 158–65 ruin theory 196 Russian options 6, 10, 15–16, 215 Rydberg, T.H. 31 S&P 500 130–1, 140–4 saddle point, optimal stopping game 295 Sato process 170 Schachermayer, W. 279–85 Schoutens, Wim 1, 7, 53, 54, 67–97, 129–47, 169–93, 260 Schwartz, E.S. 259, 269–70, 286–7 SDE see stochastic differential equations second moment swaps see variance swaps securities, game options 278–89 self-ﬁnancing trading strategies 118, 182 self-quanto options 110–11 semi-heavy tails, distributional characteristics 4–5 semi-martingales 33–6, 101–23, 278–89 Shiryaev, A. 278, 296–7, 301 Simons, Erwin 67–97 simulation methods bias 29, 42–8 bridge algorithms 29, 36–48, 117 concepts 29–30, 32–48, 67–8, 72–95, 118, 133–44 continuously/discretely reset path-dependent options 39–48 L´evy processes 29–30, 31–48, 67–8, 72–95, 114, 117–18, 133–44 Monte Carlo simulation 31–3, 35–6, 39–40, 67–8, 114, 117–18, 142–4

318

Index

simulation methods (Continued) numerical approach 33–6, 113–14 speed-up methods 36–48 Sirbu, M. 286–9 skewed distributions 29–30, 53–65, 86–95, 169, 180 smiles 29–30, 67–95, 99–128, 131–47, 237–9 ‘smooth ﬁt’ conditions, concepts 241–50, 259–60, 271, 293–303 smooth pasting principle 155–6, 206 Spanish equity indexes, risk-neutral densities 60–5 spatial discretizations, CGMY process 266–70 spectrally one-sided processes concepts 1–2, 5, 6, 17–18, 195–215 ﬁrst-passage distributions 195, 197–202 path properties 17–18 perpetual American options 195–215 stable of index 210–14 speed-up methods, simulation methods 36–48 spread game options concepts 293–303 deﬁnition 295 SPX 61–4 squared returns, autocorrelation 58–9 stable of index, spectrally one-sided processes 210–14 static hedging algorithm 140 Asian options 129–30, 136–47 concepts 129–30, 136–47 model-independent super-hedges 145 performance issues 140–4 static positions 68, 91–5, 129–47 stationary independent increments, L´evy processes 2–4, 103, 237, 242–56, 260–1 statistical densities CGMY process 54–65 concepts 51–65 estimation details 57–63 L´evy processes 53–65 stochastic calculus 30–3, 69–71, 79–95, 178–9, 222–31, 240–2, 293, 294–5

stochastic clocks 72–95, 133–44 see also Cox–Ingersoll–Ross. . . ; Gamma-OU. . . stochastic differential equations (SDEs) 30–3, 69–71, 79–95, 178–9, 240–2, 293, 294–5 stochastic processes 1–48, 102–3, 131–5, 178–9, 260–1, 277–89 see also L´evy processes stochastic time, L´evy processes 71–95, 131, 133–5 stochastic volatility 4, 29–33, 58–9, 67–95, 117, 129–47 Black-Scholes contrasts 129 concepts 4, 29–33, 58–9, 67–95, 117, 129–47 models 129–30, 131–47 numerical implementation 138–44 super-hedging strategy 129–47 stocks 52, 60–5, 130–1, 140–4, 169–92, 202–15, 277–89, 295–303 stop-loss transforms, concepts 129–30, 137–47 stopping region, exercise boundary 241–5 stratiﬁed sampling bridge algorithms 36–42 concepts 36, 39–42 Monte Carlo simulation 39–42 sub-optimal strategies 149–50 submultiplicative function, PIIAC 103–4 subordinated Brownian motion 31–3, 34–48 subordinator representation, bridge algorithms 37–48 subordinators, concepts 31–3, 34–48, 68, 70–2, 131–5, 199–200 ‘suicide’ strategies, Black-Scholes pricing model 283 super-hedging strategy, Asian options 129–47 supermartingales, game options 281–6 surveys, valuation methods 99–123 swaps 68, 89–95, 169–71 moment swaps 89–95 pricing 89–95 variance swaps 89, 169–71 swaptions 99

Index

swing-options 278 symmetries 21–2, 99–123, 237–56 concepts 21–2, 99–123 deﬁnition 99–100 exotic options 99–123 Margrabe-type options 119–23 power options 106–11 vanilla options 106–14, 121–3 tails distributional characteristics 4–5, 29, 54–65 insurance claims 57–63 Tankov, P. 1, 260–1 Taylor expansion 90, 169, 170–3, 263–4 term structure of smiles, concepts 99–123 theorems, L´evy processes 2–4 time-changed L´evy process 73–4, 78, 79–80, 93–5, 133–44 time-inhomogeneous L´evy processes concepts 99–123, 245–56 model 100–4 Tistaert, Jurgen 67–97 Toeplitz matrix 268–9 trading strategies dynamic trading strategies 68, 91–5, 130–47 game options 278–89, 295 self-ﬁnancing trading strategies 118, 182 transaction costs 130 Trigeorgis, L. 150 trinomial market model 169–92 truncated stable processes see also generalized tempered. . . concepts 1–2, 4, 19–23 two-agent models 51–65 two-dimensional asset-or-nothing options 121–3 two-dimensional driving processes, symmetries 121–3 two-sided features, risks 65 two-sided L´evy processes, one-side L´evy processes 215 UIB see up-and-in barrier options UK equity indexes, risk-neutral densities 60–5

319

unbounded variation, path properties 12, 14–24 UOB see up-and-out barrier options up-and-in barrier options (UIB) 29, 42–8, 81, 84–6 up-and-out barrier options (UOB) 29, 42–8, 81, 84–6 upper half line, regularity 6, 16–24 USA equity indexes, risk-neutral densities 60–5 utility theory 51–65, 186–92 Uys, Nadia 217–35 valuation methods see also pricing surveys 99–123 value at risk (VaR) 29, 52 value function, optimal stopping game 294–303 value matching condition, exercise boundary 241–2 Vandermonde matrices 186 vanilla options 10, 67, 74–8, 99–100, 106–14, 121–3, 129–30, 140–4, 261–2 pricing 106–14, 121–3, 261–2 symmetries 106–14, 121–3 VaR see value at risk variance gamma process (VG) see also generalized tempered stable processes change of measure densities 56–65 concepts 1–4, 8–10, 22–5, 32–48, 56–65, 71–95, 117, 133–44, 237–9, 241–50, 259–62 DEVG model 241–56 generalized hyperbolic processes 10, 117 Monte Carlo simulation 39–42, 79–95, 117, 142–4 simulation methods 32–48, 72–95, 133–44 variance swaps 89, 169–71 VG see variance gamma process volatility smiles 29–30, 67–95, 99–128, 131–47, 237–9 volatility surface, concepts 99–100 Voltchkova, E. 259, 264, 266

320

Index

waves, Fourier transform methods 2, 31, 56, 57–60, 74–5, 112, 116–17, 129–30, 138–45, 209–14, 259–76 Webber, N. 37–42 Wiener processes 31, 240–2 see also Brownian motion Wiener–Hopf factorization 14–15, 113–14, 116–17

writers see issuers Yor, M. 259 see also Carr–Geman–Madan–Yor process zero-sum Dynkin stopping game see Dynkin’s games

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close