Energy Derivatives: Pricing and Risk Management Les Clewlow and Chris Strickland Lacima PUBLICATIONSPublished in 2000 by Lacima Publications London, England Email: contact@lacimagroup.com Website: www lacimagroup.com Copyright © Les Clewlow and Chris Strickland All rights reserved. No part of this publication may be reproduced, copied, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without either prior written permission from the publisher, the copyright owner, or a licence permitting restricted copying in the United Kingdom issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London, UK WIP OLP. ISBN 0-9538896-0-2 Typeset by Alden Bookset, Oxford Front cover photography by Anthony Marsland Printed and bound in Great Britain by Biddles LimitedContents ABOUT THE AUTHORS PREFACE ACKNOWLEDGEMENTS CHAPTER t INTRODUCTION TO ENERGY DERIVATIVES AND FUNDAMENTALS OF MODELLING AND PRICING 1.1 INTRODUCTION TO ENERGY DERIVATIVES 1.2. FUNDAMENTALS OF MODELLING AND PRICING 1.3. NUMERICAL TECHNIQUES 1.3.1 The Trinomial Method 1.3.2. Monte Carlo Simulation 1.4. SUMMARY CHAPTER 2 UNDERSTANDING AND ANALYSING SPOT PRICES 2.1. INTRODUCTION 2.2 MEAN REVERSION 2.3. SIMULATING MEAN REVERSION 2.4 THE HALF-LIFE OF A MEAN REVERTING PROCESS 2.5 STOCHASTIC VOLATILITY, 2.6 SIMULATING STOCHASTIC VOLATILITY 2.7 JUMPS IN THE SPOT PRICE, 2.8 SIMULATING JUMPS IN THE SPOT PRICE. 2.9 ESTIMATION OF THE MEAN REVERSION RATE 2.10 INTUITIVE ESTIMATION OF JUMP PROCESS PARAMETERS 2.11 MAXIMUM LIKELIHOOD ESTIMATION OF JUMP PARAMETERS 2.12. INCORPORATING SEASONAL PATTERNS INTO THE MODELS 2.13. CONCLUSIONS, CHAPTER 3. VOLATILITY ESTIMATION IN ENERGY MARKETS, 3.1. INTRODUCTION 3.2. ESTIMATING VOLATILITY 3.2.1 Estimation of Volatility From Historical Data 3.2.2 Estimation of Volatility for a Mean Reverting Process 3.2.3 Volatility Estimation: Special Issues 3.2.4 Intraday Price Variability xii xv SeoaseContents 3.2.5. Estimation of Volatility for a Basket 43 3.2.6 Implied Volatility 44 3.3 STOCHASTIC VOLATILITY MODELS 46 3.3.1 Estimation and Testing 49 3.3.2 Ordinary Least Squares 50 3.3.3. Maximum Likelihood 52 3.3.4 Testing 54 3.3.5 Examples from the Energy Commodity Markets 56 3.4 SUMMARY 63 CHAPTER 4 ENERGY FORWARD CURVES 64 4.1 INTRODUCTION 64 42 CONSTRUCTING FORWARD CURVES 66 4.2.1. Cost of Carry Relationship 66 4.2.2. Forward Price Bounds for Energies 67 4.2.3. Seasonality in Prices 68 4.3. FORWARD CURVES IN THE ELECTRICITY MARKET 68 4.3.1 Arbitrage Pricing Approach 70 43.2. The Econometric Approach mn 4.3.3. The Spot Price Modelling Approach nm 4.4 SUMMARY 2 CHAPTER 5 ENERGY DERIVATIVES: STRUCTURES AND APPLICATIONS 3 5.1 INTRODUCTION B 5.2. EXCHANGE TRADED INSTRUMENTS B 5.3. SWAPS 4 5.3.1 Vanilla Swap 4 5.3.2 Variable Volume Swap 4 5.3.3. Differential Swap 15 5.3.4 Margin or Crack Swap 5 5.3.5. Participation Swap 16 5.3.6 Double-Up Swap 16 5.3.7 Extendable Swap 7 5.4 CAPS, FLOORS AND COLLARS 1 5.5 SWAPTIONS B 5.6 COMPOUND OPTIONS - CAPTIONS AND FLOPTIONS 9 5.7 SPREAD AND EXCHANGE OPTIONS 80 5.7.1 Calendar Spreads 80 5.7.2. Crack Spreads 80 5.7.3 Exchange Options 81 5.8 PATH DEPENDENT OPTIONS, 81 5.8.1 Asian Options — Average Price and Average Strike 82 5.8.2. Barrier Options 83 5.8.3 Lookback Options — Fixed Strike and Floating Strike 85 5.8.4 Ladder and Cliquet Options 86 5.9 SUMMARY 88 viContents CHAPTER 6 SPOT PRICE MODELS AND PRICING STANDARD INSTRUMENTS 89 6.1 6.2 6.3 64 6.5 6.6 INTRODUCTION SINGLE FACTOR MODELS Futures and Forward Pricing Option Pricing The Schwartz Single Factor Model Futures and Forward Pricing Option Pricing TWO FACTOR MODE! Futures and Forward Pricing Option Pricing The Schwartz 1 Factor Approximation Option Pricing in the Long Term Model THREE FACTOR MODELS Futures and Forward Pricing Option Pricing CHOOSING A SPOT PRICE MODEL SUMMARY CHAPTER 7 SPOT PRICE MODELS: PRICING PATH DEPENDENT AND AMERICAN STYLE OPTIONS. 7.1 INTRODUCTION 7.2 SIMULATION METHODS FOR ENERGY SPOT PRICE MODELS 7.2.1. Single Factor Simulation 7.2.2. Example — Valuing Spread Options 7.2.3. Example — Valuing European Enetgy Swaptions 7.2.4 Incorporating Seasonality into Monte Carlo Simulations 7.2.5 Computation of Hedge Sensitivities 7.2.6 Simulation of Multi-Factor Models 7.2.7 Generation of the Random Numbers 7.2.8 Valuing Path Dependent Options Using Simulation 7.3 BUILDING TRINOMIAL TREES FOR THE ENERGY SPOT PROCESS 7.4 IMPLIED TRINOMIAL TREES 7.5 PRICING GENERAL PATH DEPENDENT ENERGY OPTIONS IN SPOT PRICE TREES 7.5.1 Pricing Asian Energy Options in a Trinomial Tree 7.5.2 Pricing Swing Options 7.6 SUMMARY CHAPTER 8 FORWARD CURVE MODELS 8.1 INTRODUCTION 8.2 A SIMPLE MODEL FOR THE FORWARD CURVE 8.3. THE DYNAMICS OF THE FORWARD CURVE 8.4 A GENERAL MULTI-FACTOR MODEL OF THE FORWARD CURVE 8.5 RELATIONSHIP BETWEEN FORWARD CURVE AND SPOT 8.6 PRICE MODELS ESTIMATION OF THE VOLATILITY FUNCTIONS 89 89 90 91 91 93 95 97 98 100 102 103 104 104 106 107 108 109 109 109 110 il4 14 115 7 119 120 123 127 128 129 131 133 134 134 135 137 142 142 143Contents 8.6.1 Historical Estimation of the Forward Curve Volatility Functions 8.6.2 Example Analysis of NYMEX Crude Oil Futures 8.6.3 Incorporating Seasonality into the Volatility Functions 8.7 PRICING STANDARD EUROPEAN OPTIONS, 8.7.1 Options on the Spot 8.7.2. Options on Forward Contracts 8.8 PRICING GENERAL EUROPEAN CONTINGENT CLAIMS Example: Pricing a European Energy Swaption 8.9 PRICING EXOTIC OPTIONS 8.9.1 Calendar Spread Options 8.9.2 Crack Spread Options 8.9.3 Asian Options 8.9.4 Amercian Options in Forward Curve Models 8.10 IMPLIED ESTIMATION OF THE VOLATILITY FUNCTIONS 8.10.1 Handling Volatility Skews and Smiles 8.11 SUMMARY CHAPTER 9 RISK MANAGEMENT OF ENERGY DERIVATIVES 9.1 INTRODUCTION 9.2 DELTA HEDGING OF ENERGY DERIVATIVE POSITIONS 9.3 GAMMA HEDGING 9.4 HEDGING VOLATILITY AND OTHER SOURCES OF MARKET RISK 9.5 FACTOR HEDGING 9.6 SUMMARY CHAPTER 10 RISK MANAGEMENT OF ENERGY DERIVATIVES 10.1. INTRODUCTION 10.2. WHAT IS VALUE AT RISK? 10.3. MODELLING THE MARKET VARIABLES 10.4 FORECASTING STANDARD DEVIATIONS (VOLATILITY) 10.5 FORECASTING COVARIANCES AND CORRELATIONS 10.6 VALUE AT RISK METHODOLOGIES 10.6.1 Variance-Covariance (Delta VaR) 10.6.2 Delta-Gamma VaR 10.6.3. Monte Carlo Simulation 10.6.4 Historical Simulation 10.7. TESTING VALUE-AT-RISK ESTIMATES 10.7.1 Testing the VaR Forecast for Market Variables 10.8 CONCLUSIONS CHAPTER 11 CREDIT RISK IN ENERGY MARKETS 11 112 11.3 viii INTRODUCTION WHAT IS CREDIT RISK? THE KEY COMPONENTS OF CREDIT RISK MODELS 113.1 Credit Ratings and Credit Rating Transition Probabilities 11.3.2. Credit Spreads 11.3.3 Recovery Rates in Default 144 145 146 149 151 152 153 155 156 156 156 187 159 160 162 163 163 164 169 172 173 179 180 180 180 182 183 187 188 191 195 199 202 205 208 208 210 210 210 211 211 214 214Contents 11.3.4 A Summary of Major Credit Risk Models 11.4 CREDITMETRICS 11.4.1 Modelling Credit Rating Changes 11.4.2. Simulation of Counterparty Market Returns 11.4.3. Credit Spreads, Recovery Rates and the Calculation of Derivative Values 11.4.4 Modelling Recovery Rates 11.4.5 Calculation of the Credit Risk for an Energy Derivative Portfolio 11.5 CREDITRISK+ 11.5.1 Modelling the Occurrence of Defaults 11.6 CREDIT RISK MEASURES AND MANAGEMENT 11.6.1 Marginal Risk 11.6.2 Concentration Risk 11.6.3 Managing Credit Risk 11.6.4 Assessing the Models 11.7. CONCLUSIONS REFERENCES: INDEX 215 215 216 219 221 221 222 228 230 233 233 234 235 236 236 238 243About the Authors Les Clewlow is a Principal of Lacima Group which has been supplying training, software and consulting to financial, commodity and energy institutions and software companies worldwide for more than ten years. He is an Associate Research Fellow at the School of Finance and Economics, University of Technology Sydney and the Financial Options Research Centre, University of Warwick in the United Kingdom. He has a First Class Honours degree in Physics with Astrophysics (Birmingham), and a PhD in Computer Science (Warwick). He has recent publications in Risk Magazine, Energy and Power Risk Management and Journal of Economic Dynamics and Control. His current interests include energy and commodity risk management, pricing and hedging exotic options, interest rate models and numerical algorithms. Chris Strickland co-founded and is principal of Lacima Group. He is an Associate Research Fellow at the School of Finance and Economics, University of Technology Sydney and the Financial Options Research Centre, University of Warwick in the United Kingdom, and an associate researcher at the Instituto de Estudios Superiores de Administracién, Caracas, Venezuela. Previously, Chris worked for RBC Gilts Ltd, and Kitcat and Aitken. & Co. in London. He is the co-author, with Les Clewlow, of the book “Implementing Derivatives Models” and co-editor of the book “Exotic Options: The State of the Art”. He has a First Class Honours degree in Pure Mathematics (Liverpool), an MSc in Mathematics (Warwick) and a PhD in Finance (Warwick). His current interests include energy derivative pricing, interest rate models and numerical algo- rithms. Vincent (Vince) Kaminski is Managing Director, Enron North America, Enron Corp. Vince joined Enron in June of 1992. Previously, Vince was a vice president in the research department of Salomon Brothers in New York (Bond Portfolio Analysis Group) and a manager in AT&T Communications (Long Lines) in Bedminster, New Jersey. In his current position, Vince is responsible for development of analytical tools for pricing of commodity options and other commodity transactions, hedging strategies, optimisation of financial and physical transactions, as well as development of value-at- risk systems. Enron Corp. manages the largest portfolio of fixed-price and derivative natural gas contracts in the world and has invested significant time and effort in the development of state-of-the-art risk management systems. Vince is a recipient of the 1999 James H. McGraw Award for Energy Risk Management (Energy Risk Manager ofAbout the Authors the Year), Vince holds an MS degree in international economics, a PhD degree in mathematical economics from the Main School of Planning and Statistics in Warsaw, Poland, and an MBA from Fordham University in New York. Grant Massonis a vice president with the research group of Enron Corp. He oversees the quantitative support for asset and derivative structure valuations and market analysis for both domestic and international electricity trading and origination. He is also responsible for quantitative support for corporate Value- and Credit-at-Risk measure- ment systems. Prior to joining Enron in 1994, Grant spent five years at the University of Basel as a research scientist specialising in experimental nuclear physics. He received a BA from Rice University in Houston and MS and PhD degrees in physics from the University of Wisconsin — Madison. Ronnie Chahal is a manager with the research group at Enron Corp. He has a PhD in Finance from Georgia State University, Atlanta, Georgia. He currently provides analytical support to the risk management group in Enron Energy Services. xiPreface One of our main objectives in writing Energy Derivatives: Pricing and Risk Management has been to bring together as many of the various approaches for the pricing and risk management of energy derivatives as possible, to discuss in-depth the models, and to show how they relate to each other. In this way, we hope to help the reader to analyse the different models, price a wide range of energy derivatives, or to build a risk management system which uses a consistent modelling framework, We believe that for practitioners this last point is very important and we continue to stress in our articles and presenta- tions the dangers of using flawed risk management and pricing systems. Using ad-hoc and inconsistent models for different instruments and markets gives arbitrage opportu- nities to your competitors. However, it is not our wish to concentrate on one particular model or models, to the exclusion of the others because we believe that the choice should rest with the user (although it will probably be clear from our discussions which model(s) we prefer). We therefore try and give as clear account as possible of the advantage and disadvantages of all the models so that the reader can make an informed choice as to the models which best suit their needs. In order to meet our objectives, the book is divided into 11 Chapters. In Chapter | we give an overview of the fundamental principles needed to model and price energy derivatives which will underpin the remainder of the book. In addition to introducing the techniques that underlie the Black-Scholes modelling framework, we outline the numerical techniques of trinomial trees and Monte Carlo simulation for derivative pricing, which are used throughout the book. In Chapter 2 we discuss the analysis of spot energy prices. As well as analysing empirical price movements, we propose a number of processes that can be used to model the prices. We look at the well-know process of Geometric Brownian Motion as well as mean reversion, stochastic volatility and jump processes, discussing each and showing how they can be simulated and their parameters estimated. Chapter 3, written by Vince Kaminski, Grant Masson and Ronnie Chahal of Enron Corp., discusses volatility estimation in energy commodity markets. This chapter builds on the previous one. It examines in detail the methods, merits and pitfalls of the volatility estimation process assuming different pricing models introduced in Chapter 2. Examples from crude, gas, and electricity markets are used to illustrate the technical and interpretative aspects of calculating volatility. Chapter 4 examines forward curves in the energy markets. Although such curves are well understood and straight-forward in many financial and commodity markets, the xiiPreface difficulty of storage in many energy markets leads to less well defined curves. In this chapter we describe forward price bounds for energy prices and the building of forward curves from market instruments. We outline the three main approaches which have been applied to building forward curves in energy markets; the arbitrage approach, the econometric approach, and deriving analytical values by modelling underlying stochas- tic factors. Chapter 5 presents an overview of structures found in the energy derivative markets and discusses their uses. Examples of products analysed in this chapter include a variety of swaps, caps, floors and collars, as well as energy swaptions, compound options, Asian options, barrier options, lookback options, and ladder options. Chapter 6 investigates single and multi-factor models of the energy spot price and the pricing of some standard energy derivatives. Closed form solutions for forward prices, forward volatilities, and European option prices (both on the spot and forwards) are derived and presented for all the models in this chapter including a three factor, stochastic convenience yield and interest rate model. Chapter 7 shows how the prices of path dependent and American style options can be evaluated for the models in Chapter 6. Simulation schemes are developed for the evaluation of European style options and applied to a variety of path dependent options. In order to price options which incorporate early exercise opportunities, a trinomial tree approach is developed. This tree is built to be consistent with the observed forward curve and can be used to price exotic as well as standard European and American style options. Chapter 8 describes a methodology for valuing energy options based on modelling the whole of the market observed forward curve. The approach results in a multi-factor model that is able to capture realistically the evolution of a wide range of energy forward curves, The user defined volatility structures can be of an extremely general form. Closed-form solutions are developed for pricing standard European options, and efficient Monte Carlo schemes are presented for pricing exotic options. The chapter closes with a discussion of the valuation of American style options. Chapter 9 focuses on the risk management of energy derivative positions. In this chapter we discuss the management of price risk for institutions that trade options or other derivatives and who are then faced with the problem of managing the risk through time. We begin with delta hedging a portfolio containing derivatives and look at extensions to gamma hedging ~ illustrating the techniques using both spot and forward curve models. The general model presented in Chapter 8 is ideally suited to multi-factor hedging of a portfolio of energy derivatives which is also discussed. Chapter 10 examines the key risk management concept of Value at Risk (VaR) applied to portfolios containing energy derivative products. After discussing the concept of the measure, we look at how the key inputs (volatilities, covariances, correlations, etc) can be estimated. We then compare the four major methodologies for computing VaR, delta, delta-gamma, historical simulation and Monte-Carlo simulation, applying each to the same portfolio of energy options. In this chapter we also look at testing the VaR estimates for various underlying energy market variables. xiiiPreface Finally, in Chapter 11 we review modelling approaches to credit risk. We look in detail at two quite different approaches, Credit Metrics (J. P. Morgan (1997)) and CreditRisk + (Credit Suisse Financial Products (1997)), for which detailed information is publicly available. Together, these provide an extensive set of tools with which to measure credit risk. We present numerical examples of applying these techniques to energy derivatives. We must stress that the models and methods we present in this book are tools,which should be used with the benefit of an understanding of how both the ‘tool’ and the market works. The techniques we describe are certainly not “magic wands” which can be waved at data and risk management problems to provide instant and perfect solutions. To quote from the RiskMetrics Technical Document “... no amount of sophisticated analytics will replace experience and professional judgement in managing risk”. How- ever, the right tools correctly used make the job a lot easier! We are, as always, pleased to receive any comments on the book — please use the dedicated email address: energyderivatives@lacimagroup.com. Les Clewlow, Lacima Group Chris Strickland, Lacima Group Vince Kaminski, Enron Corp. Research August 2000Acknowledgements It is a pleasure for us to acknowledge numerous colleagues, both academic and practitioner, for the many useful discussions that have contributed to our knowledge over the years, as well as numerous friends, all of whom have helped us in the production of this book. To those we fail to mention by name, please accept our apologies; Peter Ash, Carl Chiarella, Brian Collins, Nigel Collins, Gary Cox, Olivier Croissant, Nadima El-Hassan, Doug Gardner, Tony Hall, Simon Hurst, Chris Leeds, Vincent Kaminski, Adam Kucera, Andrew Jameson, Sandra Jackson, Rob de Rozario, Jamie Rogers, Peter Israel, Glenn Kentwell, John Martin, Grant Masson, Ian Merker, Sudip Mukherjee, Bernard Murphy, Christina Nikitopoulos, Mike Roffey, Mark Schneider, Vito Scoppio, Gary Stanford, Max Stevenson, Steve Thomas, Louise Walker, Amanda Walker and Joe Winsen. Although too numerous to mention individually, participants at Lacima’s energy related courses in Sydney, Melbourne, Houston, New York, London, and Brazil have helped to clarify our ideas and provided us with data and examples. Special thanks go to Michel Booth, Tlia Bouchiev, Alexander Edyeland, Mark Garman, Corwin Joy, and David Shimko who read the original manuscript and gave us extensive feedback which led to significant improvements in the finished version. Julie Brennan and Rubin Rajendram greatly assisted in the production process. The authors would also like to acknowledge the financial support of the School of Finance and Economics, University of Technology Sydney, Boral Energy and Société Generale. LC would like to thank and dedicate this book to his Mum, Dad, brother, sister and grandparents. CS would like to thank his Mum and Dad and to apologise to Lorna for all the disruptions this text has caused. This is for ‘Mugs’. LC cs xv1 Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing 1.1. Introduction to Energy Derivatives Energy markets around the world are under going rapid deregulation, leading to more competition, increased volatility in energy prices, and exposing participants to poten- tially much greater risks. Deregulation impacts both consumers and producers and has lead to a heightened awareness of the need for risk management and the use of derivatives for controlling exposure to energy prices. However, this is not the only source of the development of risk management products. Investment banks are being drawn into the area as they look for new markets in which to operate. There is also an increasing number of power marketers entering the market and companies like Enron! are establishing themselves in a role which might be described as an ‘energy investment bank’. This combination of the two different sides of the market, along with the sheer size of the market at the sales level, has the potential to make energy derivatives one of the fastest growing of all derivatives markets”. For many market participants, energy derivatives appear to be a new phenomenon. Although it is true that traded derivatives are a relatively new concept in the energy markets, the structures have been around for centuries and contracts with derivative characteristics have existed in energy markets for decades. For example, options have been embedded in supply and purchasing agreements which have traditionally offered a high degree of flexibility in terms of price, volume, timing and location of delivery. Although there is now a realisation that these contracts should be priced to reflect the optionality in such agreements, they have been trading for many years. ‘There are many contracts that enable the user to manage their exposure to energy prices, with derivatives often providing the simplest and most flexible solutions for precise risk management. A derivative security can be defined as a security whose payoff depends on the value of ' http://www.enron.com Readers interested in the market growth and the development of competitive electricity markets are referred to Kaminski (1997) and Masson (1999)Energy Derivatives: Pricing and Risk Management other more basic variables*. The simplest types of derivatives are forward and futures contracts. Futures and Forward Contracts A futures contract is an agreement to buy or sell the underlying asset in the spot market (often called the spot asset) at a predetermined time in the future for a certain price, which is agreed today. Futures contracts are standardised, in terms of the future date, amount traded, etc. and can be retraded through time on a futures exchange. Forward contracts are also agreements to transact on fixed terms at a future date, but these are direct agreements between two parties. Although forwards and futures are similar contracts involving an agreement to buy or sell on a certain date for a certain price, important differences exist. Firstly, as we have just seen, futures are exchange standar- dised contracts, whereas forward contracts trade between individual institutions. Secondly, the cash flows of the two contracts occur at different times — futures are daily marked to market with cashflows passing between the long and the short position to reflect the daily futures price change, whereas forwards are settled once at maturity* Despite these differences, if future interest rates are known with certainty then futures and forwards can be treated as the same for pricing purposes and we will, for the most part, use the terms interchangeably. There are two sides to every forward contract. The party who agrees to buy the asset is said to hold a long forward position, whilst the seller is said to hold a short forward position. At the maturity of the contract (the “forward date’) the short position delivers the asset to the long position in return for the cash amount agreed in the contract — which is often called the delivery price. If T represents the contract maturity date, then mathematically this long forward payoff can be expressed as S; — K where S; represents the asset price at time 7, and K represents the agreed delivery price. Figure 1.1 shows the profit and loss profile to the long forward position at the maturity of the contract. The payoff can obviously be positive or negative, depending on the relative values of S; and K. The short position, by definition, has the opposite payoff to the long position (i.e. —S'p + K) as every time the long position makes a profit, a loss is suffered by the short, and vice versa. Since the holder of a long forward contract is guaranteed to pay a known fixed price for the spot asset, futures and forwards can be seen as insurance contracts providing protection against the price uncertainty in the spot markets. A straight-forward arbitrage relationship means that the forward price must be equal to the cost of financing the purchase of the spot asset today and holding it until the forward maturity date®. Let F represent the price of a forward contract on the spot asset > Derivatives are often referred to as contingent claims as the payout to the security (often referred to as the maturity payoff) and hence value, is contingent on other events. * For credit purposes, some forward contracts are also marked to market on a regular basis. 5 See Chapter 4 for more details 2Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Profit Spot Energy Price (S_) At Maturity of Contract (7') aD Loss FIGURE 1.1 Payoff to Long Forward Position that is currently trading at S. If the maturity date of the contract is T, ¢ represents the cost of holding the spot asset (which includes the borrowing costs for the initial purchase and any storage costs), and 6 the continuous dividend yield paid out by the underlying asset, then the price of a forward contract, at current time 1, and the spot instrument on which it is written are related via the ‘cost of carry’ formula®: F = Se) (1.1) The continuous dividend yield can be interpreted as the yield on an index for index futures, as the foreign interest rate in foreign exchange futures contracts, and as the convenience yield for various energy contracts’. Options Contracts Options contracts are the second cornerstone to the derivatives market. There are two basic types of options. A call option gives the holder the right, but not the obligation, to buy the spot asset on or before a predetermined date (the maturity date) at a certain price (the strike price), which is agreed today. Figure 1.2 graphs the payoff to the holder of such an option. Options differ from forward and futures contracts in that a payment, usually at the time the contract is entered into, must be made by the buyer ~ this is the option price or © Tris measured in years. 7 See Chapter 6 for a discussion of convenience yields in energy and commodity markets.Energy Derivatives: Pricing and Risk Management Profit Present Value of Premium Spot Energy Price (5) FIGURE 1.2 Payoff to Call Option premium. At the maturity date, for spot asset prices below the agreed strike price (denoted by ), the holder lets the option expire worthless, forfeits the premium, and buys the asset in the spot market. For asset prices greater than K, the holder exercises the option, buying the asset at K and has the ability to immediately make a profit equal to the difference between the two prices (less the initial premium). Therefore, the holder of the call option essentially has the same positive payoff as the long forward contract, but without the downside, resulting in a so-called ‘dog leg’ payoff profile. The payoff to a call option can be described mathematically as follows: max(0, 5 — K) (1.2) The second basic type of option, a put option, gives the holder the right, but not the obligation, to sell the asset on or before the maturity date at the strike price. Figure 1.3 shows the payoff profile to the holder of a put option. Mathematically, the payoff for a put option can be written: max(0,K — S) (13) One of the difficulties experienced by newcomers to the derivatives market is the amount of terminology involved. As we have already seen, the date specified in the contract is known as the maturity date, but it is also known as the expiration, exercise, or strike date. The strike price is often referred to as the exercise price. Options are also classified 4Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Profit Present ‘Spot Energy Price (S) Value of Premium Loss FIGURE 1.3 Payoff to Put Option with respect to their exercise conventions. European options can only be exercised on the maturity date itself, whereas American style options can be exercised at any time up to and including the expiration date. As with forwards, there are two sides to every option contract. One side has bought the option and has the long position, whilst the other side has sold (or written) the option and has taken a short position. Figure 1.4 shows the four possible combinations of terminal payoffs for long and short positions in European call and put options with maturity date T. The futures and options of this section describe the basic building blocks of all derivative securities and the principals are consistent across all underlying markets. However, derivative structures in energy markets exhibit a number of important differences from other underlying markets. The differences arise because of the complex contract types that exist in the energy industry as well as the complex characteristics of energy prices. Both the type of derivatives that trade, and the modelling needed to capture the evolution of prices, reflect these differences. For example, many contracts in the energy industry are based on averages (often weekly or monthly in the oil industry and hourly or less in the electricity markets) of prices and this has led to the wide acceptance of so-called Asian or average price options®. Basis risk, widely defined to * See Chapter 5 for further details on Asian options.Energy Derivatives: Pricing and Risk Management Option payoff Option payoff Long Call Short Call 1max(0, S,~ K) max(0, $,~ K) Option payoff Option payoff Long Put Short Put max(0, $,— K) max(0, Sp-K) FIGURE 1.4 Terminal Payoffs tor European Options mean the difference between two different prices, is very important in energy markets as production processes often involve the conversion of one energy (say natural gas) into another (electricity) thus exposing the company to the price differential. This has lead to the development of a wide variety of spread options. The complications extend to modelling of the price dynamics. For example, large variations in the cost of electricity generation and high demand variability contribute to high price volatility and jumps in prices. High levels of seasonality exhibited by energies are also important to capture”. 1.2 Fundamentals of Modelling and Pricing The modern theory of option pricing is possibly one of the most important contributions to the whole area of financial economics. The breakthrough came in the early 1970s with work by Fisher Black, Myron Scholes and Robert Merton (Black and Scholes (1973), ° See Chapter 4 for a description of types of contract provision that cause embedded options to take on complicating characteristics and their impact on pricing. See also Kaminski (1997).Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Merton (1973)). The Black-Scholes-Merton (BSM) modelling approach proved not only important for providing a computationally efficient and relatively easy way of pricing the then recently developed exchange traded equity options in Chicago, but also for demonstrating the principal of no-arbitrage, risk neutral, valuation. Their analysis showed that the payoff to an option could be perfectly replicated with a continuously adjusted holding in the underlying asset and the risk free bond. Since the risk of writing an option can be completely eliminated the risk appetites or preferences of market participants are irrelevant to the valuation problem, and we can assume they are risk neutral. In such a world, all assets earn the riskless rate of interest, thus the actual expected return on the asset does not appear in the Black-Scholes formula. In many energy markets the concept of being able to perfectly replicate options by continuously trading the underlying asset is unrealistic. For example. spot electricity cannot be easily stored'® and therefore a continuously adjusted position is not possible. Similar arguments can be applied, albeit in a less extreme sense to many other spot energies. However, many energy derivatives actually depend on futures prices rather than the spot price and futures can be used to replicate options positions allowing the application of the risk neutral pricing approach. In cases where it is not reasonable to apply risk neutral pricing we can argue that the risk neutral price provides a good reference with which to compare other pricing methods. An alternative and useful ‘insurance’ based approach is to calculate the expected payoff of the option under a model for the actual behaviour or real-world measure (as opposed to the risk neutral measure) of the market prices. The methods we describe in this book can also be used in this way. Finally, as a writer of options we may wish to price them on the basis of our expected cost of hedging the option given our access to hedging markets and manage- ment systems. This can also be done using the models and methods which we describe. In this section we look at a number of different methodologies that have been developed for pricing options. We start with approaches that have been developed under the BSM assumptions of costless trading in continuous time, infinite divisibility of the underlying asset, a non-dividend paying asset, constant interest rates and constant volatility. From the perspective of this chapter, however, the most important assumption in the BSM model is the mathematical description of how asset prices evolve through time. This is the well-known Geometric Brownian Motion (GBM) assumption where proportional changes in the asset price, denoted by S, are assumed to have constant instantaneous drift, 1, and volatility, o. The mathematical description of this property is given by the following stochastic differential equation!!: dS = pSdt + oSdz (14) 1 Electricity can be stored by hydroelectric schemes by using it to pump water into the reservoir, the electricity can then be recovered by releasing the water through the turbines. Electricity can also be indirectly stored by generators in the form of the fuel used to generate it. '! Most models of asset price behaviour for pricing derivatives are formulated in a continuous time framework by assuming a stochastic differential equation describing the stochastic process followed by the asset price. 7Energy Derivatives: Pricing and Risk Management Here dS represents the increment in the asset price process during a (infinitesimally) small interval of time dr, and dz is the underlying uncertainty driving the model and represents an increment in a Weiner process during dt. The risk-neutral assumption implies that the drift can be replaced by the riskless rate of interest (i.e. 42 =r). Any process describing the stochastic behaviour of the asset price will lead to a characterisa- tion of the distribution of future asset values and the assumption in equation (1.4) implies that future asset prices are lognormally distributed, or alternatively, that returns are to the asset are normally distributed. Let C represent the value of any derivative security (a call or a put, a forward, or any of the other more complicated derivatives we look at throughout this book). The arguments of BSM allow for the derivation of the following partial differential equation describing the evolution of the derivative price through time, oc Ol 1 eC ort ast 2°* as The value of any derivative whose payoff is contingent on the level of the asset price following equation (1.4), and time, must satisfy this equation. In order to evaluate the prices of specific options (for example, European call and put options) this equation must be solved with the appropriate boundary conditions — given as the option maturity payoff (i.e. Cp = max(0, $7 ~ K) for a European call and Cy = max(0,K — Sp) for a put). For a European option equation (1.5) can be solved in a variety of ways, yielding the familiar Black-Scholes formula (here for a call evaluated at time 1), C(t) = SN(d,) — Ke-9N(ds) (1.6) where n(S/K) + (r+30°)(T — oVT a, =d,-oVT-1 and where the parameters 8, K, r, ‘, 7, and o have been previously defined, In(.) is the natural logarithm and N(.) is the standard cumulative normal distribution function. The use of the risk free rate for discounting is based on the notion of risk neutral valuation. One of the qualities that has led to the enduring success of the Black-Scholes model is its simplicity. The inputs of the model are defined by the contract being priced or are directly observable from the market. The only exception to this is the volatility parameter and there is now a vast amount of published material in the finance literature for deriving estimates of this figure either from historical data or as implied by the market prices of options. Although the pricing formula (1.6) was originally applied to equity markets, some of the rigid assumptions have been relaxed by later authors, extending the model to other markets. For example, Merton (1973) extended the model by firstly showing that the 1 aSIntroduction to Energy Derivatives and Fundamentals of Modelling and Pricing discounting in the model could be done in terms of a pure discount bond of the same maturity as the option, thus taking into account non-constant interest rates. A pure discount bond is defined as a bond which pays one unit of cash at its maturity date only. If we represent with, P(t,7), the price at time ¢ of a pure discount bond with maturity date 7, the BSM formula can be written: C(t) = SN(d,) — KP(t, T)N(d2) (1.7) Merton also showed how a non-constant, but deterministic, volatility can be handled by using the average volatility over the life of the option. Another, widely used, relaxation of the original formula takes into account assets that pay a constant proportional dividend. Assets of this kind are handled by reducing the expected growth rate of the asset by the amount of the dividend yield. If the asset pays a constant proportional dividend at a rate 6, over the life of the option, then we can use the original Black-Scholes call formula (1.6) with the adjustment that the parameter S is replaced by the term Se~S7-9, This adjustment has been applied to value options on broad-based equity indices as well as options on foreign exchange rates — see Garman and Kohlhagen (1983) for the latter’, In practice it is now well established that the Black-Scholes-Merton model is used not with the constant parameter volatility assumption of equation (1.6) but in conjunction with what is termed ‘implied volatility smiles’. This is the practice of adjusting the volatility which is entered into the Black-Scholes formula for options which are away from the money!?. As we shall see in section 3.3, volatility smiles correspond to the probability distribution implied in option prices differing from the lognormal distribution implied by the GBM assumption of equation (1.4), resulting in options being priced at different levels of volatility. In Chapter 2 we show that smiles can be introduced into models by incorporating stochastic volatility and jumps. In addition to varying volatility dependent on strike price, traders frequently adjust volatility dependent on the maturity of the option often resulting in the smile becoming less pronounced as the option maturity increases. Although it is possible to obtain closed form solutions such as equation (1.6) for certain derivative pricing problems there are many situations when analytical solutions are not obtainable and numerical techniques need to be applied. Examples that we will see in this book include American options, and other options where there are early exercise opportunities, ‘path dependent’ options with discrete observation frequencies, models that incorporate jumps, and models dependent on multiple random factors. The description of two of these techniques is the subject of the next section. 1.3 Numerical Techniques In this section we describe two numerical techniques which are most commonly used by practitioners to value derivatives in the absence of closed-form solutions. Although we restrict our attention to (trinomial) tree building and Monte Carlo simulation, other !? The proportional dividend is interpreted as the risk free rate on the foreign currency in options of this type. '3 See section 3.3 for a discussion of volatility smiles in energy markets.Energy Derivatives: Pricing and Risk Management techniques such as finite difference schemes (see Clewlow and Strickland (1998)), numerical integration, finite element methods, and others, are also sometimes used by practitioners. However, these methods require more advanced expertise in numerical techniques. For both of the techniques we outline it is possible to price not only derivatives with complicated payoff functions dependent on the final energy price, but also derivatives whose payoff is determined also by the path the underlying price follows during its life. Monte Carlo simulation provides a simple and flexible method for valuing complex derivatives for which analytical formulae are not possible. The method can easily deal with multiple random factors; for example, options on multiple energy prices or models with random volatility, convenience yield, or interest rates. Monte Carlo simulation can also be used to value complex path dependent options, such as average rate, barrier, and lookback options, and also allows the incorporation of more realistic energy price processes, such as jumps in prices and more realistic market conditions such as the discrete fixing of exotic path dependent options. For many American style options early exercise can be optimal depending on the level of the underlying energy price. It is rare to find closed-form solutions for prices and risk parameters of these options, so numerical procedures must again be applied. However, using Monte Carlo simulation for pricing American style options is difficult. The problem arises because simulation methods generate trajectories of state variables forward in time, whereas a backward dynamic programming approach is required to efficiently determine optimal exercise decisions for pricing American options. Therefore, binomial and trinomial trees are usually used by practitioners for pricing American options'*. 1.3.1. The Trinomial Method The binomial model of Cox, Ross and Rubinstein (1979) is a well-known alternative discrete time representation of the behaviour of asset prices to GBM. This model is important in several ways. Firstly, the continuous time limit of the proportional binomial process is exactly the GBM process. Second, and perhaps most importantly, the binomial model is the basis of the dynamic programming solution to the valuation of American options'». Although binomial trees are used by many practitioners for pricing American style options, we and many other practitioners prefer trinomial trees. The trinomial tree has a number of advantages over the binomial tree. Because there are three possible future movements of the asset price over each time period, rather than the two of the binomial approach, the trinomial tree provides a better approximation to a continuous price process than the binomial tree for the same number of time steps. Also, * Tilley (1993), Li and Zhang (1996), Broadie and Glasserman (1997a, 1997b) and Clewlow and Strickland (1998), amongst others have described methods of extending Monte Carlo simulation to the valuation of options with early exercise opportunities. In Chapter 8 we show how Monte Carlo simulation can be applied to American options in conjunction with a tree based approach 'S See Chapter 2 of Clewlow and Strickland (1998) for an in-depth discussion of implementing the binomial method, 10Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing the trinomial tree is easier to work with because of its more regular grid and is more flexible, allowing it to be fitted more easily to market prices of forwards and standard options, an important practical consideration. For an asset paying a continuous dividend yield, the stochastic differential equation for the risk neutral GBM model is given by equation (1.4) where the drift is replaced by the difference between the riskless rate and the continuous yield: dS = (r ~ 6)Sdt + oSdz (1.8) In the following we will find it more convenient to work in terms of the natural logarithm of the spot price, x = In(S), and under this transformation we have the following process for x: dx = vdt + odz (1.9) where v = r — 6 — $0?. Consider a trinomial model of the asset price in which, over a small time interval Az, the asset price can increase by Ax (the space step), stay the same or decrease by Ax, with probabilities p,, p,,, and p, respectively. This is depicted in terms of x in figure 1.5, The drift and volatility parameters of the asset price are now captured in this simplified discrete process by Ax, p,, Pm, and pg. It can been shown that the space step cannot be chosen independently of the time step, and that a good choice is Ax = oV3At. The relationship between the parameters of the continuous time process and the trinomial process is obtained by equating the mean and variance over the time interval Az and requiring that the probabilities sum to one, i-e.: E(Ax] = py(Ax) + Pn(0) + pa(—Ax) = vAt (1.10) tar R Poy x x & x AX iS eeareereee Ae eeeeeaeee FIGURE 1.5 Trinomia! Model of an Asset PriceEnergy Derivatives: Pricing and Risk Management E[AX)] = py(Ax?) + pn (0) + py(Ax) = oP? A+ PAP (1.11) PutPm+Pu= I (1.12) Solving equations (1.10) to (1.12) yields the following explicit expressions for the transitional probabilities: oP At+ vat p=4(=eeeae +e) (1.13) PAtr+ AP i Pm = 1S (1.14) (1.15) The single period trinomial process in figure 1.5 can be extended to form a trinomial tree. Figure 1.6 depicts such a tree. N,-N+2 N,-N4I FIGURE 1.6 A Trinomial Tree Model of an Asset Price 12Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Let / denote the number of the time step and j the level of the asset price relative to the initial asset price in the tree. If 5;,; denotes the level of the asset price at node (i, j) then we have ¢ = ¢;=iAt, and an asset price level of Sexp(jAx). Once the tree has been constructed we know the spot price at every time and every state of the world consistent with our original assumptions about its behaviour process, and we can use the tree to derive prices for a wide range of derivatives. We will illustrate the procedure with reference to pricing a European and American call option with strike price K on the spot price. We represent the value of an option at node (i, j) by C;,. In order to value an option we construct the tree representing the evolution of the spot price from the current date out to the maturity date of the option — let time step N correspond to the maturity date in terms of the number of time steps in the tree, i.e. 7 = NAr. The values of the option at maturity are determined by the values of the spot price in the tree at time step N and the strike price of the option: Cy,; = max(0, Sy; — K); f= N (1.16) It can be shown that we can compute option values as discounted expectations in a risk neutral world!®, and therefore the values of the option at earlier nodes can be computed as discounted expectations of the values at the following three nodes to which the asset price could jump: Cig = PuCin jr + PmCinay + PaCinayy-1) (1.17) where e~! is the single period discount factor. This procedure is often referred to as “backwards induction’ as it links the option value at time i to known values at time i + 1. The attraction of this method is the ease with which American option values can be evaluated. During the inductive stage we simply compare the immediate exercise value of the option with the value if not exercised, computed from equation (1.17). If the immediate exercise value is greater, then we store this value at the node, i.e.: CG. = max fe" PiCirayer +PmCi1y +PaCinipt)Siy—K} (1-18) This method also gives us the optimal exercise strategy for the American option since for every possible future state of the world, i.e. every node in the tree, we know whether we should exercise the option or not. The value of the option today is given by the value in the tree at node (0,0), Co. Although we have so far discussed only the valuation of a simple derivative using the tree structure, it is also possible to price many path dependent options. In section 7.5 we explain how path dependent options can be priced in an energy spot price tree fitted to the observed forward curve. The standard option hedge sensitivities: delta, gamma and theta, can be calculated straightforwardly using the tree since they can be approximated by finite difference ratios. Vega and rho can be computed by re-evaluation of the price for small changes in 1 See Clewlow and Strickland (1998) for an in-depth discussion of the implementation of trinomial trees for derivative pricing 13Energy Derivatives: Pricing and Risk Management the volatility and interest rate respectively (see section 7.2.5 below and section 3.8 of Clewlow and Strickland (1998). 1.3.2 Monte Carlo Simulation’” Monte Carlo simulation is easy to implement, works for a wide range of path dependent options, and is suitable for handling multiple stochastic factors. This last property implies that it is straightforward to add multiple sources of uncertainty such as stochastic volatility or random jumps to the basic model as well as valuing derivatives whose payoff depends on some function of two or more energy prices!®. As we have already seen, in general, the present value of an option is the expectation of its discounted payoff. We can obtain an estimate of this expectation by computing the average of a large number of discounted payoffs computed via Monte Carlo simulation. Originally applied to the pricing of financial instruments by Boyle (1977), the Monte Carlo technique involves simulating the possible paths that the asset price can take from today until the maturity of the option. We can discretise the transformed GBM process represented in equation (1.8) in the following way: Xray =X + (VAL + O(2:,.a1 — 21) (1.19) Alternatively, in terms of the original asset price we have the discrete form Strar = SrexpVAt + o(Z.ar ~ 21)) (1.20) Equations (1.19) or (1.20) can be used to simulate the evolution of the spot price through time. The change in the random Brownian motion, z,,.4, ~ 2;, has a mean of zero and a variance of Ar. It can therefore be simulated using random samples from a standard normal multiplied by VA?, ie. VAve where ¢ ~ N(0,1). In order to simulate the spot price we divide the time period over which we wish to simulate, [0,7], into N intervals such that At = T/N, t; = iAt,i = 1,..., N. Using, for example, equation (1.20) we have _ exp(vAt + ov Ate;) (1.21) It is important to note that, since the drift and volatility terms do not depend on the variables S and t, the discretisation is correct for any time step we choose. This enables us {o jump straight to the maturity date of the option in a single time step, if the payoff to the derivative is only a function of the terminal asset value, and does not depend on the path taken by the asset during the life of the option. Repeating this process N times, choosing ¢; randomly each time, leads to one possible path for the spot price. Figure 1.7 illustrates the result of repeating this single path simulation one thousand times: S = 100,r — 6 = 0.05, ¢ = 0.30, Ar = 1/(365 x 24) " This section is based on Chapter 4 of Clewlow and Strickland (1998). '® For example a crack spread that pays the difference between gas and electricity prices. 14Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Asset Price FIGURE 1.7 Illustration of 1000 Simulated GBM Paths At the end of each simulated path the terminal value of the option (Cy) is evaluated. Let Cr,; represent the payoff to the contingent claim under the j” simulation. For example, for a standard European call option the terminal value is given by: Cr = max(0, Sy, — K) (1.22) Each payoff is discounted using the simulated short-term interest rate sequence: r Cie = exn(- f rats) Cry (1.23) In the case of constant or deterministic interest rates equation (1.23) simplifies to: Co,j = P(0, T)Cr, This value represents the value of the option along one possible path that the asset can follow. The simulations are repeated many (say M) times and the average of all the outcomes is taken to compute the expectation, and hence option price: eee Co My Co.)Energy Derivatives: Pricing and Risk Management Therefore Cy is an estimate of the true value of the option, Cy, but with an error due to the fact that it is an average of randomly generated samples and so is itself random. In order to obtain a measure of the error we estimate the standard error SE(.) as the sample standard deviation, SD(.), of Co; divided by the square root of the number of samples: SE(C) = 52 (1.25) where SD(Cy ;) is the standard deviation of Cy: The biggest criticism of Monte Carlo methods concerns the speed with which derivative values can be evaluated. It is not unknown for the technique to take many hours to return a price that is sufficiently accurate, due to the number of simulations that have to be performed. However, a number of authors, including Kemna and Vorst (1990), Clewlow and Carverhill (1994) and Clewlow and Strickland (1997, 1998), have proposed methods to speed up the process. These techniques are known as variance reduction techniques, as their aim is to reduce the variance of the estimate obtained via the simulations. Further variance reduction techniques involve the implementation of quasi- random sequences (see for example Joy, Boyle, and Tan (1996). Another criticism of Monte Carlo is its perceived inability to handle American options, However, we discuss in Chapter 8 how American options can be priced using a combination of tree and simulation techniques. 1.4 Summary In this chapter we have introduced energy derivatives, describing some simple structures, and outlined the differences between energy and other underlying markets. We have presented an overview of the fundamental pricing principals that are applied to derivative valuation in a Black-Scholes-Merton world, and described two numerical procedures often implemented by practitioners to evaluate derivative prices and risk sensitivities. In the following chapter we look at the applicability of GBM for modelling energy price processes and describe other price processes often applied to energy price movements.2 Understanding and Analysing Spot Prices 2.1 Introduction In this chapter we describe how the Black-Scholes geometric Brownian motion model, introduced in chapter one, can be generalised to a more realistic model for spot energy prices. Other useful discussions of modelling energy prices can be found in Humphreys and Shimko (1997), Pilipovic (1997) and Eydeland and Geman (1998). We discuss mean reversion — the tendency of spot prices to move back towards their long term level, stochastic volatility — the unpredictability of spot price volatility and jumps — sudden large changes in the spot price. The Monte Carlo simulation of these processes is also described, both as an aid to understanding and to begin to develop the numerical techniques needed to price energy derivatives. Simple yet robust estimation methods for the key parameters of the mean reversion and jump models are presented with illustrative numerical examples. The estimation of volatility models is discussed in depth in Chapter 3 We begin discussing the various extensions to the BSM model by assuming constant parameters for ease of explanation. In reality the volatility, mean reversion rate, long term level and jump behaviour will at the very least vary through time with reasonably predictable “seasonal” patterns. Extending the models to handle these seasonal patterns is generally straightforward and we discuss this in more detail in the penultimate section of this chapter. In particular, time varying parameters are very easy to incorporate into the Monte Carlo simulations which we discuss in this chapter and Chapters 7 and 8. Another important energy derivative pricing issue, introduced in Chapter 1, is the skew and smile observed in Black-Scholes implied volatilities which is directly related to skews and fat-tails in the distributions of energy price returns. Fat-tails can be accounted for by incorporating random volatility and more importantly jumps into the models which we discuss in detail in this chapter, Skews are more difficult to incorporate but we mention some approaches in the sections on modelling volatility and jumps. Finally, an issue that occurs in electricity markets is that of negative prices (see section 3.1 for a discussion of this). The models that we discuss are generally designed to exclude the possibility of such prices. If the frequency and range of the negative prices is small relative to the positive prices then they can be excluded without affecting the results discussed in this book. Alternatively, a lower bound can usually be determined for the prices and the actual 17Energy Derivatives: Pricing and Risk Management prices can then be shifted so as to be positive. Although the analytical results we discuss will no longer hold, the numerical methods can still be used to price derivatives. 2.2 Mean Reversion It has been well documented that an important property of energy spot prices is mean- reversion (Gibson and Schwartz (1990), Brennan (1991), Cortazar and Schwartz (1994), Bessembinder, Coughenour, Seguin and Smoller (1995), Ross (1995)). Mean reversion can be understood by looking at a simple model of a mean reverting spot price (Schwartz (1997)) represented by the following equation: dS = aj — In S)Sdt + oSdz (2.1) In this model the spot price mean reverts to the long-term level § = e# ata speed given by the mean reversion rate, a which is taken to be strictly positive. If the spot price is above the long-term level S then the drift of the spot price will be negative and the price will tend to revert back towards the long-term level. Similarly, if the spot price is below the long-term level then the drift will be positive and the price will tend to move back towards S. Note that at any point in time the spot price will not necessarily move back towards the long term level as the random change in the spot price may be of the opposite sign and greater in magnitude than the drift component. Figure 2.1 shows the Henry Hub natural gas spot price for the period from | April 1999 to 31 March 2000. 35 3.3 3.4 2.9 27 25 23 24 1.5 91/04/98 01/05/99 01/0698 01/07/93 01/08/99 31/0BR9 91/1099 31/1099 01/129 91/12/99 91/01100 O1/03/00 01/0400 Date Spot Price (USD/MMBTU) FIGURE 2.1. Henry Hub Natural Gas Spot Price for the Period from 1 April 1999 to 31 March 2000 (USD/MMBTU) 18Understanding and Analysing Spot Prices The mean reversion rate of the spot gas price over this period is approximately eleven (estimation of the mean reversion rate is discussed in section 2.9). This implies that the spot price tends to be pulled back to its long-term level over a period of roughly a month (this is related to the half-life which is discussed in detail in section 2.4). In reality the spot price does not mean revert to a constant long term level, and we show in chapters 6, 7 and 8 how information on the level to which the spot price mean reverts is contained in the forward curve prices and volatilities. 2.3 Simulating Mean Reversion The simple model of a mean reverting spot price represented by equation (2.1) can be written in terms of the natural logarithm of the spot price, x = In S, as follows: dx = [ay — x) — $0" |dt + oz (2.2) Equation (2.2) can be discretised as Ax; = [a(u—x;) — fo? Ar + oV Ate; (2.3) Note that in this case, unlike for GBM (see chapter one), since the drift term depends on the variable x the discretisation is only correct in the limit of the time step tending to zero. We must therefore choose time steps which are small relative to the speed of mean reversion - that is a small fraction of the half-life. In order to obtain a simulated path of the spot price we proceed in exactly the same way as for GBM in chapter one. Firstly, we choose or estimate the parameters a, 1, ¢ and Ar’. Then we repeatedly generate normally distributed random numbers ¢; and from these calculate each new value of x, and thus the new spot price at each time step. Figures 2.2 and 2.3 illustrate a simulated GBM path and mean reverting path for the energy spot price respectively. In both figures exactly the same random shocks have been applied in the spot price simulations. We have set the mean reversion rate in figure 2.3 relatively high, a = 100, in order that the mean reverting behaviour can be easily seen relative to figure 2.2. Comparing figures 2.2 and 2.3, it should be clear that the same random shocks have occurred to the spot price. However, the range of the spot price in figure 2.3 over time is reduced by the effect of the mean reversion continuously driving the spot price towards the long-term mean of 100. 2.4 The Half-Life of a Mean Reverting Process A key property of a mean reverting process is the half-life. This is the time taken for the price to revert half way back to its long-term level from its current level if no more ' We discuss estimation of the mean reversion rate in the following section. In Chapters 6, 7 and 8 we show the parameter 1 can be obtained as a time dependent function from the forward price and volatility curve. Estimation of volatility is discussed in Chapter 3 19Energy Derivatives: Pricing and Risk Management 120 115. 110: 105: 100: Spot Price 95- 90 85. 89 0.00 0.05 0.10 0.15 0.20 0.25 Time (years) FIGURE 2.2 Simulation of a Geometric Brownian Motion Spot Price Path S = 100,r — 5 = 0.05, 0 = 0.30, At = 1/(365 x 24) 120 115 90. 85. 8.0 0.05 0.10 0.15 0.20 0.25 Time (years) FIGURE 2.3 Simulation of a Mean Reverting Path for the Spot Energy Price S = 100, a = 100.0, 5 = 100, 0 = 0.30, At = 1/(365 x 24) 20Understanding and Analysing Spot Prices random shocks arrive. Ignoring the randomness allows us to focus on the mean reverting behaviour alone. Consider the simple mean reverting process, in terms of x = In S, reverting to a level of x. In the absence of randomness we have dx = (8 — x)dt (2.4) Equation (2.4) can be solved by integrating from xo at time 0 to x, at time f to obtain (1 — 8) = (% — xe (2.5) To obtain the half-life, which we denote by 1,2, at which the distance of x from its long term level is half its initial distance we have ) $0 — ¥) = (x0 — Be? (26) Solving for 4)2 we obtain np =In2)/a (2.7) Table 2.1 gives the half-life for a range of values of the mean reversion rate. It is important to realise that these times are averages. Only on average over a long time period, will shocks to the spot price take the time given by the half-life to decay to half their deviation from the long term level 2.5 Stochastic Volatility Even a cursory examination of most energy prices suggests that the volatility of the price is changing through time. Figure 2.4 shows the Brent crude oil spot price for the period from 1 April 1999 to 31 March 2000. The volatility of the spot oil price over the first part of the data is clearly less than that over the second half of the data. Many models have been proposed for the behaviour of volatility (for example — Hull and White (1987), Johnson and Shanno (1987), Wiggins (1987), Hull and White (1988), Madan and Senata (1990), Stein and Stein (1991), Heston (1993), Derman and Kani (1994), Dupire (1994), Duan (1995), Bates (1996), and Scott (1997), Ritchken and Trevor (1999)). Here we focus on two particular types of model. The first model we discuss was introduced independently by Derman and Kani (1994) TABLE 2.1 Mean Reversion Rates and the Corresponding Half-Lives a fe i 8 months 10 25 days 100 2.53 days 1000 6 hours 21Energy Derivatives: Pricing and Risk Management 38> x 8 Spot Price (USD/Barell) ® 8 3 8 0VO489 0170589 01EGD OTA7II9 01MAE SOB OV” >/IOE9 o17IZRG 3171209 3101100 01/0300 01/0800 Date FIGURE 2.4 Brent Crude Oil Spot Price for the Period from 1 April 1999 to 31 March 2000 (USD/Barrel) and Dupire (1994). The idea behind this model was that it could be fitted to or implied from the market prices of standard European options via a trinomial tree, and it is therefore often referred to as the implied tree method”. The second model was first proposed by Hull and White (1988) and subsequently used and extended by Heston (1993), Bates (1996) and Scott (1997). The modelling approach of Derman er a/ and Dupire can be represented by the following process for the spot price: dS = pSdt + o(t, S)dz (2.8) In equation (2.8) the volatility is now a general function of the spot price and time. If this function is specified to depend on a non-linear function of the spot price then the model allows us to obtain spot price returns distributions with skews, A simple example would be: o(t,S) = 3,S* (2.9) where 3, and 8, > 1 are constants — this is referred to as the constant elasticity of variance model (see Cox and Ross (1976)). The general model given by equation (2.8) ? This modelling approach is also discussed in Chapter 7. For a detailed discussion of the implementation of implied trinomial trees see Clewlow and Strickland (1998), Chapter 5. 22Understanding and Analysing Spot Prices can be simulated in the same way as GBM except for one small complication. Transforming to the natural logarithm of the spot price no longer leads to a constant volatility for the process. In some simple cases, such as equation (2.9), it is possible to find a transformation which leads to a variable with a constant volatility (see Clewlow and Strickland (1998), Chapter 8 for a discussion of this in the context of interest rate models). If a transformation is not possible then small times steps must be used to simulate the process accurately. The second type of model became popular because of its realistic properties and computational tractability. It is described by the following processes for the spot price and the spot price return variance V = 0? dS = uSdt + oSdz (2.10) dV = a(V —V)dt+&VVdw (2.11) Equation (2.10) is the GBM model but with a volatility, 7, which is no longer constant but which changes randomly. The behaviour of the volatility is determined by equation (2.11) which specifies the process followed by the variance — the square of the volatility. The variance mean reverts to a long-term level V at a rate given by a. The absolute volatility of the variance is £/V which is proportional to the square root of the variance ie. the volatility of the spot price. The source of randomness in the variance, dw, is different from the dz driving the spot price although it may be correlated with correlation coefficient p, A more detailed discussion of stochastic volatility and estimation of the parameters of the volatility process is presented in Chapter 3 2.6 Simulating Stochastic Volatility The stochastic volatility model described by equations (2.10) and (2.11) can be discretised in the following way’: Ax, = (w- 40°) Att oV Ate; (2.12) AV, = a(V —Vi)Att+é/ FiBi( pe, +yl- Peas) (2.13) where ¢; and 2 are independent standard normal random variables. These are combined through the correlation coefficient p to obtain a random shock to the variance, which is correlated with the random shock to the spot price (simulating correlated processes is discussed more fully in Chapters 7 and 8). The drift and volatility terms in both equations (2.12) and (2.13) depend on the variance and so the discretisation is only correct in the limit of the time step tending to zero. We must therefore choose time steps which are small relative to the speed of mean reversion and the volatilities. The simulation of a spot price path now requires the joint simulation of both the spot > The spot price process is discretised in terms of its natural logarithm x.Energy Derivatives: Pricing and Risk Management 1207~ Spot Price 8g 0.00 0.05) 0.10 0.15 0.20 0.25 Time (years) FIGURE 2.5(a) A Simulated Stochastic Volatility Path for the Spot Energy Price S = 100, js = 0.05, 7 = 0.30, a = 10, V = 0.09, € = 0.5, p = 0.0, At = 1/(365 x 24) price and the variance of the spot price. We first obtain or choose values for js, 0, a, € p, and Av. Each time step of the simulation then involves generating two independent normally distributed random numbers ¢; and ¢>, from which the variance and the spot price at the end of the time step can then be calculated. Repeating this process yields a simulated path for both the spot price and its variance. Figures 2.5 (a) and 2.5 (b) illustrate simulated paths for the spot price and the variance of the spot price in which the random shocks to the spot price are the same as in figures 2.2 and 2.3. From a comparison of figures 2.2 and 2.5 it can be seen that although the random movements are in the same directions, the movements are larger in the later part of figure 2.5 where the variance has increased 2.7 Jumps in the Spot Price Energy prices often exhibit sudden, unexpected and discontinuous changes. Particularly good examples of this occur in electricity markets. Figure 2.6 is a plot of the Australian New South Wales (NSW) electricity pool price for the period from 13 December 1998 to 8 March 1999. Figure 2.6 shows that the presence of jumps is a significant component of the behaviour of the NSW electricity pool price’. However, it should also be noted that the price does not stay at the level to which it jumps, but, after a jump, rapidly reverts to * See also Joy (1998) for a discussion of the application of jumps to the Australian electricity markets. 24Understanding and Analysing Spot Prices 0.25 0.20 0.15 Variance 0.10 0.05 98 00 0.08) 0.10 0.45) 0.20 0.28 Time (years) Path for the Spot Energy Price FIGURE 2.5 (b) A Simulated Stochastic Volati 0.5, p = 0.0, At = 1/(365 x 24) S = 100, , = 0.05, ¢ = 0.30,a = 10, V = 0.09, 350. - — 300 2vo2ee 202M 07/0399 seca asnente i ta/i29e 201298 27/2Ne OOIMe IAOIAD ITIVE 2WOIRe aWOTHS OTIC2RD 14/00/99 Date FIGURE 2.6 Australian New South Wales Electricity Pool Prices for the Period from 13 December 1998 to 8 March 1999 25Energy Derivatives: Pricing and Risk Management its long-term level. This is an important aspect of the electricity price, which we will deal with in the following section. This type of behaviour where the price exhibits sudden, large changes can be modelled by using jump processes. A simple and realistic model for a spot price, which is identical to the Black-Scholes model except for the addition of a jump process, is the jump-diffusion model introduced by Merton (1976)°. This model is described by the following SDE°: dS = pSdt-+ oSdz + %Sdq (2.14) The annualised frequency of jumps is given by ¢, the average number of jumps per year’. The proportional jump size is « which is random and is determined by the natural logarithm of the proportional jumps being normally distributed: In(1 +5) ~ N(In(1 + &) — 47,7) (2.15) where & is the mean jump size and ¥is the standard deviation of the proportional jump size which we call the jump volatility. The jump process (dq) is a discrete time process ~ jumps do not occur continuously but at specific instants of time. Therefore, for typical jump frequencies, most of the time dg = 0 and only takes the value 1 when a randomly timed jump occurs, When no jump is occurring the spot price behaviour is identical to GBM and is only different when a jump occurs ~ this is illustrated in the following section on simulation. For example, if we imagine the jump frequency becoming very small, so that the chances of a jump occurring are close to zero, then we would get a GBM spot price. Similarly, if the jump volatility were very small, so that even if jumps were very frequent their size would be insignificant, then this would again result in a GBM process The proportional jumps (or equivalently jump returns) in equation (2.14) are normally distributed and therefore symmetrical. That is the number of positive and negative jumps and the range of sizes of the proportional jumps will be equal on average. In reality the distribution of jump return sizes of energy spot prices is positively skewed. A simple way to incorporate this property into equation (2.14) is to have the proportional jumps drawn from a normal distribution but with different jump volatilities for the positive and negative jumps. Another simple alternative would be to have the propor- tional jumps drawn from a negatively shifted lognormal distribution — this would give a lower limit on the negative jump returns. These extensions are straightforward to incorporate into Monte Carlo simulation but lead to the loss of the analytical tractability of the Merton (1976) model. 2.8 Simulating Jumps in the Spot Price The jump-diffusion model described by equation (2.14) can be discretised as follows: Ax; = (r= 6 — 40°) At + oVAtey, + (% + YEn1)(us > OAD) (2.16) > See Merton (1990) for a general discussion of jump processes applied to financial markets. © The risk-neutral process is obtained by setting « = (r — ¢&) ” Technically, ¢ is defined by prob(dq = 1) = ¢ dt 26Understanding and Analysing Spot Prices where €; and € are independent standard normal random variables and uv is a uniform (0,1) random sample. The term (x; < ¢ Ar) is taken to be one if the condition is true and zero otherwise - this generates jumps randomly at the correct average frequency. When a jump occurs, its size is the mean jump size plus a normally distributed random amount with standard deviation y determined by €>. In order that the frequency of jumps is correctly simulated, the time step At must be small relative to the jump frequency such that ¢ At << 1. Once the parameters 4, &, , y and Az have been chosen, at each time step, we first generate a uniform random number, w, in order to determine if a jump should occur. If a jump is to occur we generate the normally distributed random number £) which determines the size of the jump. Finally, we generate the normally distributed random number ¢, which determines the usual GBM random change in the spot price. Figure 2.7 illustrates a simulated path for the spot price under the jump-diffusion model in which the random shocks to the spot price are the same as in figures 2.2, 2.3 and 2.5. The jumps are indicated by the bold lines with diamonds at each end. The behaviour of the simulated spot price does not appear to be a good model for the spot electricity price, partly because the jumps are not sufficiently large but also because once a jump has occurred the spot price stays at the new level and does not mean revert. This problem can be corrected by combining mean reversion and jumps into the same model. This can be represented by the following SDE: dS = alu — In $)Sdt + oSdz + KSdq (2.17) 120. 115 Spot Price 2:2 8 8 6 2 & 90 85 80 0.00 0.05 0.10 0.15 0.20 0.25 Time (years) FIGURE 2.7 A Simulated Jump-Diffusion Path for the Spot Energy Price S = 100, r = 0.08, 7 = 0.30, & = 0,4 = 100, 7 = 0.02, At = 1/(365 x 24) 27Energy Derivatives: Pricing and Risk Management 350 — — — 300 9 0.00 0.05 0.10 0.15 0.20 0.25 Time (years) fusion Path for an Electricity Spot FIGURE 2.8 A Simulated Mean Reverting Jump- 6 = 250,7 = 1.2, At = 1/(365 x 24) Energy Price S = 20,0 = 2000,$ = 20,0 = 30.0, = Figure 2.8 shows a simulated path for the model represented by equation (2.17) with very strong mean reversion and a large jump volatility. The behaviour of the spot price in figure 2.8 resembles much more closely that of the NSW electricity price in figure 2.6. It can be seen that the diffusion volatility, jump volatility and mean reversion of the NSW electricity price must be extremely high — we give examples of estimating these parameters in the following sections, 2.9 Estimation of the Mean Reversion Rate The mean reversion rate of the spot energy price can be estimated relatively simply and robustly via linear regression. Consider the simple mean reverting process for the natural logarithm of the energy spot price®: dx = a(%— x)dt + odz (2.18) This can be discretised as follows: Ax, = 9 +.04x, + 98; (2.19) where ay = Ar and a; = —aAt. This implies that observations of the spot price through time can be considered as observations of the linear relationship between Ax, ® This is essentially the same as equation (2.2) but with the —1/2 0” incorporated into 28Understanding and Analysing Spot Prices and x; in the presence of noise (represented by ge,). Therefore, if we regress observations of Ax, against x, we can obtain estimates of ay = zand a, = —aAtras the estimates of the intercept and slope of the linear relationship. Since we know the time interval between observations Ar we can obtain estimates of a and xX. In the following section we present a specific example using Henry Hub natural gas prices Example: Estimation of the Mean Reversion Rate of Natural Gas Spot Prices The data we use in this example consists of the Henry Hub natural gas spot price for the period from | April 1999 to 31 March 2000. This gives 260 data items with Ai = 0.003846. Figure 2.9 illustrates the data in terms of changes Ax plotted against x and also the linear regression estimated straight-line fit to the data. The estimates from the linear regression are (standard errors in parentheses): 9 = 0.0398(0.0152) 0, = 0.0428(0.0168) which implies 2.5 a= 11.1(44) ‘This value of a and standard error implies a half-life for price shocks of between 16 and 38 days. 0.15: 0.05 Ax 0.05 04 . . 05 06 07 08 09 1.0 1 12 1.8 x FIGURE 2.9 Linear Regression Estimation of the Mean Reversion Rate of Henry Hub Natural Gas 1 April 1999 to 31 March 2000 (USD/MMBTU) 29Energy Derivatives: Pricing and Risk Management 2.10 Intuitive Estimation of Jump Process Parameters The estimation of jump parameters for energy prices is complicated by the fact that the jumps can only be observed as part of a time series of prices which includes the “normal” non-jump behaviour of the price. Typically, we will not have information on the exact time the jumps occur. However, it is clear in figure 2.6 that the very large price “spikes” should be attributed to jumps because the probability of the normal Brownian motion type behaviour generating these large price changes is virtually zero. This observation provides a clue to how we might estimate the jump parameters. Consider figure 2.10 which shows the Australian New South Wales electricity pool prices returns (i.e. changes in the natural logarithm of the price shown in figure 2.8) every half-hour from the 13 December 1998 to the 8 March 1999. If we assume that jumps are relatively infrequent and not too large then we can get an estimate of the diffusion volatility in the usual way by calculating the sample standard deviation of returns. Based on this estimate of the volatility we can then look for actual returns that were larger than we would expect (for a chosen probability) in the absence of jumps and identify these extreme returns as jumps. Given that we have identified some returns as jumps we should recompute the estimate of the diffusion volatility by recalculating the sample standard deviation of returns with the jump returns excluded. This will gives us a lower estimate of the diffusion volatility. Consequently, using this new estimate of the diffusion volatility, we can look for further returns which exceed the chosen ° Spot Price Return = 312 as as | 3433] ssi7 oe an oe mss i Half-hour FIGURE 2.10 Australian New South Wales Pool Price Returns for 13 December 1998 to 8 March 1999 30Understanding and Analysing Spot Prices limit. This process can be repeated until the estimates converge and no new jumps are identified (typically well within ten iterations) - we call this approach a Recursive Filter. Example : Recursive Filter Estimation of the Jump-Diffusion Parameters The data we use in this example is the same as in figure 2.10, ie. Australian NSW electricity pool prices every half-hour from the 13 December 1998 to the 8 March 1999. ‘The data set therefore consists of 4128 price observations giving 4127 returns with Ar = 5.70776 x 10°°. The sample standard deviation of the half-hourly returns is 0.1353. The probability of returns greater than 3 x 0.1353 = 0.4061 is less than 0.003 therefore we begin by identifying returns larger in absolute value than 0.4061 as jumps. We find that there are 57 returns that exceed this limit. The number of jumps divided by the total time period over which they occur (in this case 0.2329 years) gives an estimate of the jump frequency. The mean and standard deviation of the jump returns gives us estimates of the mean jump size and the jump volatility. Thus we have the following relationships: @ = number of jump returns / time period of data (2.20) ® = average of jump returns (2.21) a standard deviation of jump returns (2.22) Table 2.2 gives the results of repeating this process for the NSW electricity data until the estimates converge. Note that the mean jump size is usually very difficult to estimate robustly and should therefore usually be set to zero. In the next section we describe a formal statistical method for estimating jump process parameters which has somewhat different properties. TABLE 2.2 Iterative Estimation of the Jump-Diffusion Parameters of NSW Electricity Iteration sD Jumps 6 & 7 1 0.1183 37 241.9 0.000149 0.5718 i 0.1121 102 432.9 0.000107 0.4957 3 0.1088 129 547.5 —0,000197 0.4683 4 0.1076 139 589.9 0.000696 0.4595 5 0.1065 149 6324 0.001022 0.4513 6 0.1058 156 662.1 0.001105 0.4462 7 0.1050. 163 691.8 —0.001348 0.4411 8 0.1043 170 T7215 —0.001271 0.4368 9 0.1041 172 730.0 0.001430 0.4354 10 0.1041 172 730.0 0.001430 0.4354 SD is the standard deviation of the returns after removing the jump returns, Jumps is the number of jump returns identified. 31Energy Derivatives: Pricing and Risk Management 2.11 Maximum Likelihood Estimation of Jump Parameters Maximum likelihood is a popular method of estimating the parameters of stochastic process when the probability density for the stochastic variable can be written down analytically, It was used successfully by Ball and Torous (1983, 1985) to estimate jump parameters for NYSE stock prices when earlier work by Beckers (1981) using the method of moments often gave negative variance estimates. As we will return to maximum likelihood estimation in section 3.4.3 we will only give an overview of the technique in this section. The idea behind maximum likelihood is based on the observation that for a given choice of parameter values, ©, we can compute the probability or likelihood, L(Ax;6), that a given set of log price changes, Ax, would have occurred. The probability is simply the product of the probabilities, p(Ax; ®), for the individual observations in the data set. By taking the natural log of the probability for the whole data set, L(Ax;@), we obtain a sum of the logs of the probabilities for the individual observations. If we maximise L(A.x;9) with respect to the parameter set O we obtain the set of parameters which maximises the likelihood that we would have observed the set of price changes in our data set. The density function of changes in x is given by: eine : ; p(Ax;@) = 7S eR ean, n(Ax;kAt+jIn(l +8),07At +77) (2.19) j= where n(x;a,b) is the normal density with mean a and variance b and the vector 9 = (&, o, @, 7). Assuming we have m spot price returns, Ax),...,Ax,,, the logarithm of the likelihood function is given by: InL(Ax;®) = Soin p(Axi,8) (2.23) i=l Therefore in order to estimate the parameters we simply maximise this function with respect to the parameters © (this can be done relatively easily using a spreadsheet solver). Example: Maximum Likelihood Estimation of the Jump-Diffusion Parameters The data for this example consists of daily Brent crude oil spot prices from the 20 December 1994 to the 18 February 1999 which gives 1080 data items and Ar = 0.00274. Maximum likelihood estimation gives: ® = 0.0000, ¢ = 0.1938, 6 = 307.7, y = 0.02057 Sample Standard Deviation of returns = 0.3451 A potentially undesirable empirical property of this method of estimating jump parameters is that it tends to converge on the smallest and most frequent jump component of the actual data. Energy price return time series can often be characterised has having numerous different jump components typically ranging from very high 32Understanding and Analysing Spot Prices frequency, low volatility to low frequency, high volatility jumps, We would usually want capture the lower frequency, high volatility component with a jump model. Maximum likelihood estimation may therefore not have the properties we are looking for in a jump parameter estimation procedure. Empirical analysis suggests that the Recursive Filter discussed in section 2.10 does pick out the lower frequency higher volatility jump components. 2.12 Incorporating Seasonal Patterns into the Models Probably the most important variable after price with a strong seasonal component is volatility. There are two main ways that seasonal variations in volatility can be incorporated into the models discussed in this chapter — as a time varying pattern or as a random variable with time varying parameters. If we assume that volatility follows a predictable (i.e. non-random) seasonal pattern then this is equivalent to replacing the constant volatility parameter, c, by a function of time, o(/), which repeats each seasonal cycle. We can estimate the function o(f) from historical data by using a rolling 30 day sample standard deviation for example’. Figure 2.1] illustrates the Henry Hub natural gas seasonal spot volatility over the period from 1 April 1999 to 31 March 2000 estimated using this method. Implementing the seasonal volatility pattern into Monte Carlo simulation is straight- forward — at each time step we simply apply the spot volatility appropriate to the current time point in the relevant simulation equation (e.g. equation (2.3) or (2.16)). If we are modelling volatility as random then we can incorporate the seasonal pattern into the volatility of volatility (and also the mean reversion rate and level if required) in equation (2.11). Seasonal patterns in the mean reversion rate can also be included in a similar way. Mean reversion rates for most energy prices, with the obvious exception of electricity, are relatively low. This implies that typically at least a years worth of daily data is required to obtain reasonably accurate estimates of the mean reversion rate. In electricity markets the spot price can often be observed on an hourly or half-hourly basis! which gives between roughly 720 and 1440 data points per month. The mean reversion rates are also much higher in electricity spot markets than other energy markets which makes estimation easier (although the spot price volatility is also higher which tends to increase the estimation error). Therefore it is possible for electricity spot prices to estimate seasonal mean reversion rates on a monthly or even weekly basis. Figure 2.12 shows estimates of the mean reversion rate for UK electricity spot prices on a weekly basis for the period from January 1999 to September 1999 Figure 2.12 indicates that there is no strong seasonal pattern in the mean reversion rate of the UK spot electricity price over the year 1999 although there is a strong downward trend in the mean reversion rate. This can be attributed to the fact that the ° In practice this “aw” estimate of the seasonal volatility function should be smoothed, for example by fitting a smooth functional form such as a polynomial (see Press er al. (1992)). 1° For example in the UK and Australian spot electricity is settled on a half-hourly basis. 33Energy Derivatives: Pricing and Risk Management 80% 70%! 60%! 50%) 40% Volatility 30%: 20%+ 10%: 0% J o1/o4ie9 = 21/05/99 «10/07/99 29/0B99 == 18/10/99 7/12/99 26/01/00 -—«*16/03/00 Date FIGURE 2.11 Henry Hub Natural Gas Seasonal Spot Volatility Over the Period from 1 April 1999 to 31 March 2000 4500 —- — —— — —— 4000 3500 3000 2500. 2000- 1500 Mean Reversion Rate 1000: 500 0 5 10 15 20 25 30 35 40 Week FIGURE 2.12 Mean Reversion Rate of UK Electricity Pool Prices on a Weekly Basis for January 1999 to September 1999 34Understanding and Analysing Spot Prices 2500 2300 Jump Frequency 2100 1900 | 17004 1500+ 0 5 10 15 20 25 30 35 40 Week FIGURE 2.13 (a) Jump Frequency of UK Electricity Pool Prices on a Weekly Basis for January 1999 to September 1999 12. 10 os oa 20 o 5 0 8 2 25 30 3s “0 Week FIGURE 2.13 (b) Jump Volatility of UK Electricity Pool Prices on a Weekly Basis for January 1999 to September 1999 35Energy Derivatives: Pricing and Risk Management UK market is not yet fully competitive and the spot price behaviour is dominated by the behaviour of key players in the market. Seasonal patterns in jump frequency and jump volatility are also important, particu- larly in electricity markets where demand is strongly determined by seasonal weather patterns. Regarding the estimation of seasonal jump frequency and jump volatility, similar comments as for the mean reversion rate apply. For energies other than electricity, jump frequencies and volatilities are in general relatively low and therefore a significant period of daily data (at least the order of a year) would be needed to obtain reasonable historical estimates. However, for electricity data it is possible to obtain reliable estimates on a monthly or even weekly basis. Figures 2.13 (a) and 2.13 (b) plot weekly estimates of jump frequency and jump volatility for the same UK data as figure 2.12. The jump frequency does not show a strong seasonal pattern and appears to be roughly constant over the period of the data and should therefore probably be assumed constant. The jump volatility does show a reasonable seasonal pattern together with a similar downward trend as the mean reversion rate. The extension of Monte Carlo simulation to handle the seasonal mean reversion and jump parameters is the same as for volatility. At each time step the relevant value of the parameter for that particular point is time is obtained from the table of seasonal values and used to increment the spot price over the next time step! An alternative method of estimation of seasonal parameter values is to imply them from traded energy option prices. This method is complicated to implement, computa- tionally demanding and it is difficult to obtain stable estimates. We touch on these techniques in Chapter 8. 2.13 Conclusions In this chapter we have discussed how the standard model for financial market prices, the Black-Scholes-Merton GBM model, can be extended to give more realistic behaviour for energy prices. The key extensions are mean reversion, stochastic volatility and jumps. Mean reversion captures the propensity of spot prices to return to a long term level. We showed that the unpredictability of energy spot price volatility can be accounted for using a stochastic volatility model. The combination of the jump model we described together with mean reversion allowed us to reflect the sudden large changes in energy spot prices, particularly electricity. We demonstrated how all of these extensions can be related back to the original BSM model. The Monte Carlo simulation of all of the extensions w: described as an essential technique for understanding and evaluating complex models and as a numerical technique for pricing energy derivatives. Finally, we explained how estimation methods for the parameters of these extended models can be implemented. "| In practice the historically estimated seasonal patterns should be smoothed in the same manner as for seasonal volatility cycles by fitting a smooth function such as a polynomial (see Press e¢ a/. (1992)). 363 Volatility Estimation in Energy Markets Vince Kaminski, Grant Masson, and Ronnie Chahal of Enron Corp. 3.1 Introduction Volatility is one of the critical concepts in option pricing and risk management. Organised exchanges use volatility of the underlying commodities to determine the required level of margins that futures traders are required to post with the clearing house. Volatility has been traditionally defined as the standard deviation of price returns and is routinely estimated from historical price series in countless spreadsheets all over the world. In spite of its popularity, the notion of volatility is often misunderstood, especially when one applies this tool to the energy markets. This concept has been defined and extensively researched in the context of financial instruments and one has to be very careful making generalisations. It is critical that in practical applications in the energy industry one makes an effort to capture the salient features of the energy markets. The points to make are: Volatility can be defined and estimated in a meaningful way only in the context of a specific stochastic process for the prices (price returns). e The volatility definition and measure should capture the key features of energ) markets, such as the seasonal dependence on the price level. In the following we identify a number of practical problems of price process modelling in energy markets, One implication of the assumption of the GBM process of equation (1.4) is the continuity of the line representing the sample trajectory of the price over time. The line could be drawn, in principle, without removing the pencil from the sheet of paper. The assumption of GBM as the process that describes the dynamics of the prices of financial instruments is an approximation of the behaviour observed in real markets and has to be treated as a stylised fact. As a matter of fact, there is growing evidence that the behaviour of market prices did not conform in many past time periods to this standard assumption of financial economics. One especially troubling observation is that the empirical frequency of the occurrence of extreme outcomes is larger than the probability implied by theoretical models. This issue will be revisited below. 37Energy Derivatives: Pricing and Risk Management ‘The assumption of GBM strikes anyone with practical experience in energy commod- ity trading as an unrealistic description of the observed behaviour of energy prices. This has been recognised by a growing number of academics and practitioners who have devoted a lot of attention to developing more realistic models of energy commodity price behaviour, many of which are referred to in the following pages. What follows is a brief review of the most important issues that have been overlooked in early modelling efforts and have been fully addressed only in recent research. One should note that energy commodities are not created equal and many of the observations made below apply only to some subsets of their entire universe. Investment assets vs. consumption goods. The most obvious observation is that energy commodities cannot be treated as purely financial assets, which are treated by owners as an investment. Energy commodities are inputs to production processes and/or con- sumption goods and this explains why many models based on a mechanical extension of the approach developed for financial markets may break down in the case of energy related contingent claims. For example, the GBM assumption does not allow negative prices. This assumption may be violated in practice often enough to require attention, especially in the case of electricity. In some cases, prices of electricity bid into a power pool may drop to zero if some generators want to guarantee that their plants are dispatched for contiguous blocks of time, longer than a single time slot for which separate bids are accepted. In some cases, the price may become negative, as power plants have to get rid of excess output and have no option to store electricity. In other words, an assumption of free disposal, customarily made in theoretical models, does not hold. This problem has been addressed by some recently published papers (see Routledge, Seppi, Spat (1999)). Prices of energy commodities display seasonality. By this we mean recurring regularities in price levels and/or price volatility observed over time. Seasonality may correspond to the time of the year (winter versus summer versus shoulder month), or may be observed in intramonth, intraweek, and in some markets (like power) intraday prices. Seasonality results primarily from regular demand fluctua- tions, driven in most cases, by recurring weather related factors. Fluctuations in demand interact with the supply side factors: increased demand can be satisfied only from more expensive sources or by using more expensive production units. In many cases, increased demand resulting from weather related factors might reach the levels at which supply becomes constrained by the capacity of the existing transportation or the transmission grid. In many markets, seasonality may change over time due to the changes in economic conditions and technology. For example, many natural gas marketers expect a change in seasonal price patterns in the US natural gas markets starting in year 2000, due to increases in the gas-fired generation capacity. It is expected that in addition to the winter peak, one will observe a more pronounced July/August peak, related to air conditioning load. Recognition of the existing and possibly changing patterns of seasonality creates a need for forward-looking modelling. The information about future seasonality is often derived in formal models from the futures/forward prices that summarise all the 38Volatility Estimation in Energy Markets information available to the market about future demand and supply patterns. Some recent papers offer ingenious methods of calibrating prices to forward price curves (see for example Clewlow and Strickland (1999)). Commodity prices often display jump behaviour. Jump behaviour, or ‘gapping’, is driven in many cases by fluctuations in demand and low elasticity of supply, reflecting rigidities in the transportation and transmission system and limited inventories. We saw in section 2.7 a model to incorporate this effect. Prices gravitate to the cost of production. The assumption of GBM allows prices to wander off to unrealistic levels. The same approach, used in the modelling of two related commodities, like natural gas and power, or peak and off-peak electricity prices, may produce unrealistic spreads between them. The departures from the cost of production, or “normal” price spreads is possible in the short run under abnormal market conditions, but in the long-term, the supply will be adjusted and the prices will move to the level dictated by the cost of production. This adjustment can be captured via mean reversion which was introduced in Chapter 2. One could argue that the use of a mean- reversion process represents another case of looking for the car keys under the street light, even if they were lost somewhere else. Vasicek (1977) first used a mean reverting model for modelling interest rate dynamics and subsequently the model was widely adapted. In the case of energy commodities, a pure mean reversion model may not perform well. First of all, the speed of mean reversion may be different below and above the long-term level. Secondly, in many markets, especially in the case of electricity, one can expect more departures to the upside, than to the downside. Thirdly, a price spike in one direction is frequently neutralized by a spike of similar magnitude and opposite sign, occurring shortly after the initial spike. The mean reversion process generally produces a re-adjustment that is less abrupt. Prices of energy commodities behave differently during different periods of their lives. This is especially true of forward prices, According to the so-called Samuelson’s hypothesis, forward price volatility increases as they get closer to their maturity. This can be explained by the fact that more information becomes available as the forward contract gets closer to delivery period, and this results in more trading, which in turn produces more volatility. The authors believe that GBM may represent a reasonable approximation to the reality of forward markets. Once a forward contract reaches maturity and we enter the delivery period, the behaviour of prices becomes more erratic and subject to frequent jumps. This suggests that one can model price behaviour using more traditional apparatus like GBM or mean reversion during the forward stages of their life, switching to more complex processes to describe the dynamics of the spot price during the delivery month. 3.2 Estimating Vola’ ly In this section we look at issues concerning the estimation of volatility including both historical and implied measures. 39Energy Derivatives: Pricing and Risk Management 3.2.1 Estimation of Volatility From Historical Data The first step in the determination of the correct level of volatility is the examination of the historical price data. In the case that the underlying spot price process is assumed to be GBM, volatility can be estimated from the historical price returns. The process can be broken into several steps that can be easily carried out in a spreadsheet. Step 1. Calculate logarithmic price returns. This can be accomplished by forming the price ratios S,/S,_,and taking the natural logarithms of these ratios. Price returns are typically calculated as r = S,/$,) — 1. The logic of the approach described above is that for relatively small x, In(I + x) & x. Taking the natural log of S,/S,_, is equivalent to taking the natural log of | + r, and this in turn is roughly equal to r. The use of natural log returns has also some other additional advantages. If one wants to calculate a log return over a longer time period, say from 1 to ¢-+n, corresponding to the ratio S,.,,/S,, one can convert this into (S:in/Stan-1) X (Seom-t/Sten-2) X (S,11/S,). Given that the log of a product is equal to the sum of the logs, one can easily show that a log return over the longer time period can be calculated as the sum of log returns for the sub periods. Step 2. Calculate standard deviation of the logarithmic price returns. Step 3. Annualise the standard deviation by multiplying it by the correct factor. As a first approximation the annualisation factor depends on the price data frequency. In the case that the data is monthly, the factor is /12; for weekly data it is \/52. For the daily data available for each calendar day one has to use ,/365. If the information is available for trading days only, one should use the relevant number that may vary from jurisdiction to jurisdiction. The standard usage is /250. The logic of annualisation comes from the assumptions regarding additivity of variance. It is implicitly assumed in this approach that each period return is drawn from a certain probability distribution and we are estimating the variance of this unknown distribution from the time series data for the market closing prices, ignoring the information about price behaviour within the period (intraday, intramonth, etc.). If the price returns for each period are iid (independent, and identically Se random variables, the variance of the sum of n random variables, x), %2,...-,Xn5 i equal to the sum of the variances (if price returns are uncorrelated) o 1) Given that all the variances are assumed to be equal, the variance of the sum is equal to the common variance of the underlying variables multiplied by n. Therefore, the standard deviation is equal to the common standard deviation multiplied by the square root of n. Before we address the annualisation problem in more depth we offer a few warnings to try to avoid common pitfalls. Firstly, energy price series tend to be of low quality in daptitty = Oy + OG be +O, ( 40Volatility Estimation in Energy Markets many cases. It is therefore necessary to screen the price history for data errors that may significantly distort the estimates of volatility. Secondly, in many cases the price data is obtained from the organised futures exchanges by stitching together the numbers representing closing prices for the first available contract. One should keep in mind that in calculating price returns, one should discard the observation corresponding to the contract rollover date (the day following the expiration of one contract, when we switch to the price data representing the new prompt contract). The logic of this approach is that one can, hopefully, mitigate the seasonality problem in the price data. Assuming that the prices corresponding to the same contract month are characterised by the same multiplicative seasonality factor, we get: Si Sar Sax Si Sart Saat where s represents the multiplicative seasonality coefficient, common to both prices, S,, corresponds to the deseasonalised price for period t. When we form the price ratio based on different contracts, there is no guarantee that the seasonality coefficients will cancel. Finally, we need to keep in mind that many energy commodities trade for 6 days during a week (for example, an on-peak period in California is defined as Monday through Saturday). 3.2.2 Estimation of Volatility for a Mean Reverting Process For illustration in this section we assume a simple Ornstein-Uhlenbeck process for a particular price (spot or forward) S of the form dS = a(S — S)dt + odz (3.2) If the starting price level is S(0), the expected price at time ¢ will be given by S+(S(0) — S)exp(—ar), As the term exp(—at) goes to 0, as ¢ becomes large, the long-term expectation of price is equal to S. One should observe that the second term in the equation (3.2) is similar to the diffusion part of GBM, but that the interpretation of volatility changes. The change in price, dS, is measured in dollars per physical unit, dz, is unitless, and therefore o in equation (3.2) must be measured in dollars as well, unlike the volatility used in the Black- Scholes-Merton option model. The important lesson here is that we define and estimate volatility in the context of the stochastic process assumption and when this changes, the interpretation of volatility changes as well. Volatility can be equally easily estimated from the historical data, if a mean reversion process is assumed. A spreadsheet-friendly technology that jointly estimates the mean reversion Coefficient (cf. section 2.9) and the volatility parameter, is based on discretisa- tion of the continuous stochastic process given by equation (3.2) into the following equation that represents an autoregressive process of order one: AS, = 0% +S, +e, (3.3) 41Energy Derivatives: Pricing and Risk Management This follows from the observation that equation (3.2) is the limiting case of the following process (as di — 0): SiS; SU -e*)+(e*-DS,+% (3.4) The error term y, in (3.4) is normally distributed with variance given by: 2 ae 20. o=—(l-e 3.5) v = 55 ( ) (3.5) The parameters of the original equation (3.2) can be recovered from the estimated coefficients of the discrete version: S = -a9/o (3.6) a=~—In(1+a1) (3.7) [tnt +) oN ind an) =I 3.2.3 Volatility Estimation: Special Issues This section contains the discussion of several special issues: the choice of the annual- isation factor and the use of intra-period data (typically this means intraday prices). A typical textbook discussion of the annualisation factor typically revolves around the correct number of trading days in a year. However, the issue is more complicated than this. French and Roll (1986) compared the variance of price returns during the week with the variance of price returns calculated for the weekend. One can calculate four daily price returns for weekdays (Monday over Tuesday, Wednesday over Tuesday, etc) and one weekend return (Monday close to Friday close). For the US stock data used in their study the authors determined that the weekend was equivalent to 1.107 trading days. This result can be easily explained. Volatility is related to trading activities that are in turn triggered by the flow of information into the trading rooms. Over the weekend, the information flow slows down when trading is suspended and new information is processed during trading activity on Monday. This explains why the three calendar day period is practically equal to one trading day. Using this finding one can argue for an annualisation factor of 266 [= 52 x (4 + 1.107)]. However, energy markets are different. In many cases the information flow does not slow down significantly on weekends. One can argue that the information regarding political events and weather (which is extremely important for the natural gas, heating oil, propane and electricity markets) keeps flowing roughly at the same rate over the weekend as during the working week. Data analysis confirms our intuition. We have looked at the weekday/weekend volatility for natural gas and crude oil using the prompt month NYMEX contracts yearly from 1995 until the first quarter of 2000. The results are summarised in table 3.1. 42Volatility Estimation in Energy Markets TABLE 3.1. Weekday/Weekend Volatility Ratio (%) Year Natural Gas Crude 1995 103.41 119.81 1996 134,57 138.05 1997 128.72 113.98 1998 125.03 144.51 1999 115.47 88.17 2000 152.78 93.51 Average 125.02% 125.90% As we can see, the ratio of weekend to weekday volatility may be significantly higher than in the case of equities. On average, the volatility ratio implies the annualisation factor of the square root of 273 [= 52 x (4 + 1.25)] What can be done in view of these findings? One natural solution is to agree on the annualisation factor that will be used across the board for all commodities. This guarantees that any volatility quote will be unambiguous and can be restated by another trader based on his personal views of the market. Our recommendation is to use /250. 3.2.4 Intraday Price Variability Volatility is typically estimated from the market closing prices. In many cases, using closing price information for volatility estimation causes the loss of very valuable information as additional information is available about the price changes between the market opening and closing prices. Many markets, and energy markets in particular are characterized by very high volatility of intraday prices. 3.2.5 Estimation of Volatility For a Basket Tn many energy markets, both producers and end-users hedge their price exposures with so-called basket options. A basket option is defined in terms of a sum of two or more prices. In some cases, the prices may be weighted using weights corresponding to the option buyer’s economic sensitivity to different component prices, For example, an oil and natural gas producer may be interested in buying an option on a composite commodity that is constructed by weighting oil and gas prices (expressed in terms of a common BTU equivalent) by 0.7 and 0.3, respectively. To be more precise, the call option payoff is defined as max(0, 1S + w2S> — K) (3.9) and the put option payoff is defined as max(0, K — (w)S; +125>)) (3.10) 43Energy Derivatives: Pricing and Risk Management where S; and S> stand for the prices of the oil and gas and the w’s stand for the weights. These options can be seen as options written on a composite energy with the volatility derived from the volatilities of the contributing commodities. The question is how to derive this composite volatility in a formal way. The answer is produced through the application of Ito’s lemma. Assume that the prices of the two commodities follow GBM processes given by Sy = ay S\dt + 0) S\dz and Sy = a7Spdt + oySrdw where dz and dw stand for correlated (with a correlation coefficient of ) Brownian motions. The multivariate extension of Ito’s lemma to some function G = G(S\, S2, f) is given by - PG OS\OSy dS} a 72 oG aG OG 1/PG dG 3s, dS + ( PG oo ag dt Bt as d8,dS, +553 4s) (3.11) Applying Ito’s lemma to S$; + S3 renders iniSze7i' + 25,5, Os 4S) = In[S + Sp] (3.12) 3.2.6 Implied Volatility The previous sections have dealt with estimation of volatility from historical data. The job of an option trader is to guess what is the level of volatility expected in the future. A buyer or a writer of an option that is unhedged effectively makes a bet on the price level of the underlying instrument. An option trader who hedges makes a bet on volatility. Volatility that is used as an input to an option pricing formula that equates the model price with the market price is called an implied volatility. It is implied in the sense that once an option market price is known one can use the option pricing formula to calculate the volatility input, given that other inputs are known (maturity, the current price of the underlying, interest rate, the strike price). Volatility can be ‘backed out’ from market prices using a trial and error method or a program allowing for a solution of a nonlinear equation (an option pricing equation in this case). In practice, implied volatility is usually quoted for at-the-money options and is often calculated based on the average of an at-the-money straddle (a call and a put with the same strike price). It is well-known fact that out-of-the-money options trade at volatility levels that diverge from at-the-money volatility. This is often referred to as a volatility smile (or a frown, smirk, etc. depending on the shape of the curve). The traders try to compensate for the shortcomings of the model by adjusting the inputs to the model. Given that most inputs are well defined and directly observable and cannot be adjusted at will, this leaves 44Volatility Estimation in Energy Markets only one free parameter that can be manipulated. The adjustment would lead typically to a volatility smile — the traders increase the volatility for out-of-the-money options to compensate for flat volatility model derived valuations that are perceived to be too low. There are many competing explanations as to why this is the case, which are not necessarily mutually exclusive. The first explanation is related to well-known shortcomings of the Black-Scholes- Merton model which is based on the assumption of normality of price returns and as such understates the probability of extreme outcomes. One consequence of jumps, in for example electricity prices, which is important from the point of view of management of portfolios of power derivatives, is the presence of so-called fat tails. This term is used to refer to the fact that the probability of extreme outcomes inferred from the empirica data exceeds theoretical probabilities which are based on probability distributions typically assumed in the model. Distributions displaying fat tails are referred to as leptokurtic. This property is measured by the kurtosis, or the fourth moment around the mean, which for the normal distribution is equal to three (in some computer programs, this number is adjusted to be equal to 0). Table 3.2 displays calculated values of excess kurtosis for a number of price indices in the US power markets to illustrate the prevalence of leptokurtic distributions in the electricity markets. The program used for this calculation produces zero kurtosis for the normal distribution. Hence, we use the traditional term “excess kurtosis”. One should notice that there are several competing explanations of the presence of fat tails in empirical price distributions. One source of fat tails is the presence of jumps in the prices which, as we have seen, are a typical feature of power markets. Another explanation is that the parameters of the distribution of returns (which is usually assumed to be normal) vary over time. Empirical observations of returns are drawn in this case from a mixture of distributions. This feature of power markets creates not only a challenge from the point of view of pricing electricity derivatives but also from the point of view of portfolio management and value-at-risk calculations. Value-at-risk is a shorthand term for a system that measures the potential loss in the value of a portfolio of financial instruments over a specified time period, typically one day, with a given probability, and a subject which is covered in some depth in Chapter 10. For portfolios that include electricity derivatives, incorporation of gapping prices is of critical importance to insure a necessary degree of realism. The degree to which empirical distributions may diverge from theoretical distributions can be assessed by comparing the third and fourth moments of the empirical distributions with the theoretical levels (both equal to zero for a normal distribution). TABLE 3.2 Excess Kurtosis of Selected Power Prices, 29/6/1995-2/10/1997 Location PJM Palo Verde COB Kurtosis 487 14.48 10.78 45Energy Derivatives: Pricing and Risk Management An alternative explanation for volatility smiles deals with the microstructure of the markets. In many emerging energy markets, insufficient liquidity or lack of complete markets does not allow for effective delta hedging. This means that the supply of options comes either from the trading institutions with exceptionally efficient trading operations or from the organizations that can mitigate the risk of financial options using positions in physical assets. For example, power generators can sell call options on electricity, treating them as covered calls: if an option is exercised an idle unit may be turned on to deliver into the contract. The buyers see out-of-the money options as disaster insurance. This reduces the supply of options relative to demand and increases their prices. This is reflected in turn in implied volatility higher than the volatility for at the money options. 3.3. Stochastic Vola In section 2.5 we introduced the concept of stochastic volatility models for energy price evolution. Recall that a simple model that yields skewed spot price return distributions has a stochastic term that can be generalised as ¢S°dz, where is between 0 and 1. Any dependence on price is factored out, and the volatility term o is explicitly constant with respect to time. A process with constant volatility is termed homoskedastic, while a process with time dependent volatility, be it deterministic or random, is termed heteroskedastic. Part of the appeal of constant-volatility models is their mathematical tractability and relative ease of estimation. There is no mathematical reason, however, that requires volatility to be a constant, and indeed many commodity markets exhibit time varying volatility. For example, figure 3.1 shows the 30-day rolling volatility for Californian average daily PX prices from | April 1998 until 31 May 2000, and clearly suggests a price volatility that is not constant in time but rather is best described by a process with both periodic and stochastic elements. We have already seen a simple example of stochastic volatility in section 2.5 with the Hull-White mean reverting model represented by equations (2.10) and (2.11), where the variance is itself cast as a diffusion process. Price models with deterministically time dependent volatility are treated in Chapter 8 in the context of forward contract price dynamics. As noted in the previous section, energy commodity price return distributions tend to be leptokurtic or fat tailed. In addition, some commodities, such as crude oil, also exhibit volatility clustering. These behaviours are consistent with a process where the volatility is both stochastic and autoregressive. Models that attempt to incorporate and explain these behaviours go by the generic name of Autoregressive Conditional Heteroskedastic or, less formidably, simply ARCH models. First proposed by Engle (1982) to model price inflation, ARCH models have been successfully applied in many different areas of finance including interest rates, foreign exchange, and equities. As a result, variations on the basic concept are myriad and the literature on the subject is extensive. See Bollerslev (1992) for a review ty Models 46Volatility Estimation in Energy Markets 1200% 7 a 1000% ly - 800%: 600% Hal > A o S 400% — 200% + - 0%. 01/04/98 01/07/98 30/09/98 30/12/98 31/03/99 30/06/99 29/09/99 29/12/99 29/03/00 Data FIGURE 3.1 30-Day Rolling Annualized Volatility. Calitornian Day-Ahead Average PX Price (1 April 1 1998 to 31 May 2000) It is typical in ARCH models of asset prices to write the price returns as: y=in(2) =K+u, (3.13) Sit where « is a constant mean term, and yw, is the “innovation at time t”. The innovations are stochastic and written as u, = 0,¢,, Where ¢, is an independent random error distributed as N(0,1). Now the essential feature of ARCH modelling is to make the innovation variance at time 1 conditional on previously realised innovations. For example in an ARCH(q) model, the variance would be represented as: OF = Oy + OUT y + OnUi_ +... + Ogtitg (3.14) Of course, the order of the equation g is arbitrary. In practical applications however, models with several terms are often required in order to capture the ‘decaying memory’ of volatility. To incorporate this feature in a more parsimonious way, Bollerslev (1986) developed a generalised ARCH model, or GARCH(p,q) model. Here the variance equation includes both past innovations (of order q) and past conditional variances (of 47Energy Derivatives: Pricing and Risk Management order p): p OF = 09+ Dantes +S Boi (3.15) i=t iat Note that because the equation is recursive in variance, a; will effectively depend on all innovations. As a result, a GARCH model with a few terms will often fit a data set as well as an ARCH model of much higher order. In practice, the first order GARCH(1,1) model, shown in equation (3.16), is often quite adequate. oF = ay + Oyu + Bory (3.16) Figures 3.2a and b show an example of a spot price process with a GARCH(1,1) volatility structure. Figure 3.2a shows the price path and associated rolling thirty day volatility, and figure 3.2b shows the frequency distribution of the log price returns together with a superimposed normal distribution, the mean and standard deviation of which have been derived from the data. As expected, prices exhibit periods of high and low volatility, and the distribution of log returns is leptokurtic. 350% Spot Price 30-Day Annualized Volatility 250% 200% Price (USD) 150% 1st 401 181 201°" 251 B01 BBY 401481501 Days FIGURE 3.22 Price Process end 30-day Rolling Annualised Volatility modelled using eq. 3.13 with GARCH(1,1) v ity (as described in eq. 3.16). Model parameters: «=0.0, 0006, ay = 0.30, 3,=0.60Volatility Estimation in Energy Markets 3.3.1. Estimation and Testing Once a model is chosen as a hypothesis for a given price behaviour, the parameters for that model must be estimated. In addition to parameter estimation, an equally important and often times more difficult task is to test how well the proposed model describes the data. The fields of estimation and testing are vast, representing core subjects in statistics and econometrics. By necessity we will be brief in this section, concentrating only on the essential aspects of estimation as it relates to energy commodity price modelling. More extensive discussions of the topics touched on here may be found in Greene (1997) and Davidson and MacKinnon (1993), for example. ‘As we shall see in a moment, there is more to parameter estimation than simply running regressions. There are an infinite number of possible estimators for a given model parameter: an arbitrary number, 42 for example, is a perfectly acceptable estimator of the mean of a price process; it is simply not very accurate, and hence not very useful. Two properties, bias and efficiency define the usefulness of a parameter estimator. An estimator of a parameter is a random number with a mean and variance. Clearly, we would like our measurement to be ‘the best guess’ possible of the actual parameter value. In other words, if @ is the estimate of the true value 0, then it is desirable that Z[@ — 6] = 0. An estimator is said to be unbiased if this property holds. Similarly, it is preferable to have an estimator with the smallest possible variance. For example, let 0, 907— ———- — —— hia = Daily Log Returns — Normal Distribution 70 (St. dev. = 0.073) Frequency a & x é 20 10 oe DVB DP MAP AO. gh 8 0! LPP PP Pog Hoo Myo MoPoPomMoK oo!” 9 ooo? FIGURE 3.2b Histogram of the Log of Daily Price Returns 49Energy Derivatives: Pricing and Risk Management be the average of the sum of every fourth price in a time series. This is an entirely acceptable and unbiased estimate of the mean price. However, the variance of 6; is twice as large as that of the estimate 6 derived from taking the average of the sum of every price. Clearly, 62 is more ‘efficient’ than 6; because, £(6?) > E(63) Both properties are important. An estimator that is biased but efficient may be preferable to an unbiased, inefficient estimator because it will provide tighter bounds on the range of possible parameter values. For the sake of completeness, we mention in passing that the concepts of bias and efficiency are applicable to finite-sample estimates only. It is often the case that we do not know the finite properties of an estimator. Therefore we must instead compare the asymptotic (i.e. in the limit of large sample size) properties of estimators to determine which is more suitable and then assume that the suitability holds in the real world of finite sampling. The asymptotic properties most often described in the literature as desirable are consistency, asymptotic normality, and asymptotic efficiency. While consistency and asymptotic normality are akin to bias in that they measure if an estimate converges in some sense to the true value, they are not the same things. Indeed, an estimator can be biased but consistent, or vice versa. 3.3.2 Ordinary Least Squares Ordinary Least Squared estimation (OLS) is the technique most likely to be familiar to readers because it is at the heart of standard linear regression analysis. The concept is simple: we hypothesise that a set of N observations {,, x,} is related via the stochastic equation y=O-xte (3.17) where @ is the set of unknown parameters. We can estimate these parameters by choosing a set 6 such that the sum of squared observed residuals (SSR) is minimised: (3.18) The OLS estimates 6 are found in the usual way by setting 2588 = 0 and solving for a. For example, suppose that we have a set of N observations {y,} that we hypothesise is the result of the pure diffusion equation represented above in equation (3.13): VSR, &~ N(0,1) (3.19) OF course, one must also examine the second derivative to ensure that the solution represents a maximum. 50Volatility Estimation in Energy Markets The sum of squared residuals of this set is then N N = | =O (bP (3.20) T 1 SSR= Taking the derivative with respect to & and setting the resultant equation equal to zero yields: (3.21) as expected. Because o multiplies the stochastic term, we must resort to a slightly different method to find its estimate. From equations (3.18) and (3.19) we have that oe, =" # e=y—k (3.22) We use these relations to define an average innovation and average return: (3.23) and oe (3.24) This in turn allows us to make the following substitutions: e=yn~k =K+o0e,—k =o(e,-é (3.25) Squaring equation (3.25) gives: ef = (2) = 0*(c? — 2 € +) (3.26) which, when summed over all N observations, yields: N ve (3.27) iI 51Energy Derivatives: Pricing and Risk Management Now the random variables ¢,are distributed normally as N(0,1) and are independent, which implies that £(c,e;) = 6;. Therefore, we arrive finally at: x Se? = 07(W 1) wl (3.28) which, of course, is the common definition for standard deviation. If the stochastic equation (3.19) had p parameters rather than just one, it can be shown through a similar analysis that the OLS estimate for o in this case would be equation (3.28) with the denominator N — 1 replaced by N — p as expected. OLS estimators have the desirable properties of unbiasedness, consistency, and efficiency for a variety of problems. However, in the case of heteroskedastic systems, OLS estimators are not generally efficient, therefore other methods are often employed to provide estimates. 3.3.3 Maximum Likelihood We introduced maximum likelihood in section 2.11 as a powerful and general estimation method. In this section, we elaborate on the principle underlying the method by again using the example of the pure diffusion process represented by equation (3.13): v= K+ Oe Recall that ¢, is distributed as N(0,1), which implies that y, is distributed as N(x, 0). The probability density function for each y; is then: 1 ov 20 POBK 2) = (3.29) Now, every y, corresponds to an independent observation, so we can write the joint distribution for a set of, say N, observations as: aol Py 8.0) =] ra oV2a 4 et (3.30) We can think of this function as representing the probability or likelihood that we would draw exactly this set of {y,}. For this reason, the joint density is typically called the likelihood function and is written as: 18; {y/}) = p({v} ®) (3.31) emphasising that its value, given a set {y,} of observations, is a function of the 52Volatility Estimation in Energy Markets distribution’s parameter set 9. With this definition, we can now turn the process around. We now hypothesise that this set of numbers is drawn from a normal distribution of unknown mean and variance and ask the following question. From what distribution is it most likely that these numbers were drawn, or, put another way, which model parameters (the values of « and o” in the case of equation (3.30)) define the distribution that has the maximum likelihood of producing this set of numbers? The answer is found by maximising the likelihood function with respect to the model parameters. This is accomplished in the usual way by setting the gradient with respect to the regression parameters equal to zero and solving for each parameter value. For equation (3.30) the maximising equations are InL(6, 9 (04}) =~“ inn) — Nino? - aK? (3.32a) Aln(L(1s, 05 {71})) tne) (3.32) Aine: 83) (3326) Oo Note the use of the log likelihood function InZ(r,0; {y,}). This is a common practice because the natural log is typically easier to treat mathematically and computationally, and parameters that maximise the log likelihood are guaranteed to maximise the likelihood function as well. The solutions to equation (3.32) are referred to as the Maximum Likelihood Estimates or MLEs of the true distribution parameters. An obvious question is whether or not MLEs are the “‘best guesses” of the true parameter values. We state without proof that, assuming the likelihood function obeys some specific regularity conditions, ML estima- tors are consistent, asymptotically normal, and asymptotically efficient. In fact, ML estimators are at least as asymptotically efficient as any consistent estimator. Itis interesting to note that although ML estimation is consistent, it is not unbiased in general. For example, the MLE for the variance derived from equation (3.32c) is found to be: (3.33) 53Energy Derivatives: Pricing and Risk Management which is not the same as the OLS unbiased estimate As we mentioned before however, the ML estimate is more efficient than OLS and, in cases of small numbers of observations, may be a better estimate in the sense of having the minimum mean squared error. 3.3.4 Testing Having decided on a hypothesis for the price process under study and estimated the model’s parameters, it is essential to test the validity of that hypothesis. Unfortunately, the importance of this step is sometimes ignored in the rush to produce results. For example, the ubiquity of Black-Scholes-Merton option pricing in equity markets means that it is very common to assume that pure diffusion also describes energy commodity prices. Figure 3.3 shows average daily prices for NSW electricity for the calendar year 1999. 100 90 80, 70 50 Price (AUD/MWh) 60 nN H Aah Hi ‘awl 2 2 PP 2 SC FPP oF PP Hoo? PP oF gePh? ow FIGURE 3.3 Daily Average NSW Power Prices (Jan. 1, 1999—Dec. 31, 1999) 54Volatility Estimation in Energy Markets The pure diffusion variance estimate, described earlier in this chapter, yields a daily volatility of 21.4%. Without other test statistics, however, it is impossible to judge how meaningful this number is and how well the diffusion process explains the price behaviour. We can make qualitative judgements by graphing the frequency of daily returns and overlaying a Gaussian curve with zero mean and standard deviation equal to the derived volatility as shown in figure 3.4. This figure demonstrates the poor quality of the fit. The standard deviation seems to be overly influenced by a few large valued returns at the expense of a larger number of small valued returns. Indeed, fitting “by eye”, as suggested by the dashed normal distribution in the figure, might lead us to estimate a rather lower volatility. Calculating the skew and excess kurtosis of the return distribution provides a more quantitative test of normality. If the returns are normally distributed, then the two statistics 1 Apres Taw = se) (=) 3.34 so = Jey dy OG (3.34) and 1 Sfaiceys Tron =>) | (*H#—)'-3 3.35 ‘kurt aN Dy ‘ss (3.35) 60) — — --- 50: | Return distribution with | ——— 21.4% daily volatility 40) —- Return distribution with = 15% daily volatility 5 __ 18% dally volatilt 5 e | 3 30) =] 20 10; 0 Bn 110 090 070 050 030 -a10 a1 0M On OM OM FIGURE 3.4 Daily Averege NSW Power Prices Log Returns (Jan. 1, 1999-Dec. 31, 1999)Energy Derivatives: Pricing and Risk Management are asymptotically distributed as N(0,1). A related statistic, suggested by Jarque and Bera (1980), essentially combines equations (3.34) and (3.35) to form: Skew? Excess Kurtosis? Sve Ppeereesane eee reat reeeedeeeeeeeee ae eeeee eee 3.31 n= N[ 7 (3:36) The Jarque-Bera statistic has as chi-squared distribution of degree 2 under the null hypothesis of normality. In the case of the NSW daily price data, we find for the 364 observations the skew and excess kurtosis to be 0.177 and 6.22 respectively. The statistic Typ is then about 590 implying that we reject the hypothesis of normally distributed log price returns at a level of order 10-78. Tests for the presence of heteroskedasticity rely on the fact that the residuals from OLS regression will retain any non-stationary behaviour inherent in the underlying process. Thus, tests sensitive to changing variance can be performed on the regression residuals to detect heteroskedasticity. One such time-honoured test is the Goldfeld~ Quandt test. It assumes that there exists a variable that is related to a model’s variance. Using this variable, the data is sorted into subsets of high and low expected variance. We then perform OLS regression on the two data sets and form the statistic: _SSRin, — K) SSRoiq, — K) Where SSR stands for Sum of Squared Residuals, n; and n2 are the number of observations in the high and low data sets respectively and K is the number of parameters in the regression. Under the null hypothesis of homoskedacity, this statistic is distributed as F[m—K, ny-K] To test specifically for ARCH and GARCH behaviour, Engle (1982) and Bollerslev (1986) suggest variations on the Breusch—Pagan/Godfrey test, which is a form of Lagrange multiplier test. The details of this test are beyond the scope of this chapter, but the interested reader may refer to standard econometric texts such as Davidson and MacKinnon (1993) for a discussion of the Lagrange multiplier principle. A related test is suggested by Hull (1999), which can be used to measure the success of a GARCH model fit. If the model explains the data well, then the set of {u?/o?} should exhibit no autocorrelation. Here u, and o, are the residuals and volatilities derived from the GARCH model under examination. The Ljung-Box statistic is suggested as a method for testing this series for autocorrelation. To (3.37) 3.3.5 Examples from the Energy Commodity Markets We conclude this chapter by comparing different price models fitted to selected energy markets: the California PX day-ahead peak electricity price for zone NP15 (NP15), the NYMEX prompt month gas futures contract (Gas), and the NYMEX prompt month crude oil futures contract (Crude). Table 3.3 details basic statistics for each data set used in this analysis. 56Volatility Estimation in Energy Markets TABLE 3.3 Data set statistics for daily log price returns of California PX NP-15 peak prices, NYMEX prompt month gas futures contract prices, and NYMEX prompt month crude oil futures contract prices NPIS Gas Crude Span January 15", 1999 — March 29", 2000 Number of Observations* 370 303 303 Mean 0.68 x 10° Lol x 10% 2.57 x 10% Variance 0.23 8.52 x lo 5.29 x 10° Skew 0.46 0.07 -0.51 Excess Kurtosis 11.98 2.00 1.10 * California peak prices are quoted six days per week excluding holidays, while NYMEX commodity contracts are traded only five days per week excluding holidays. TABLE 3.4 Model Specifications Return process Volatitity process Equation Reference Pure Diffusion (PD) yaKtoe, @ constant 3.13 Jump Diffusion (SD)* yy andor, @ constant Diffusion + ARCH() (D+ Arch) y= + 06; io + ate 3.14 Diffusion + GARCH(.1) yaRtoe oto, + per, 3.16 (D+ Gareh) Jump Diffusion + GARCH(1,1) Y= K+ OE, (JD + Garch)* ag + 012, + hot, * In the case of Jump Diffusion, the last term is the product of J, a Bernoulli distributed random variable with a probability of being one and probability (1-4) of being zero, and q, a random number distributed as N(w1,?} 2 We estimated the parameters for five different models, outlined in Table 3.4, following maximum likelihood methods similar to those described by Jorion (1988) and Das (1995). In general, we employed the “BHHH” algorithm developed by Berndt, Hall, Hall, and Hausman (1974) to perform the regressions. The BHHH technique is essentially a ? The use of a Bernoulli random number follows the work by Ball and Torous (1983) and more recently Das (1995). It is well known that the Bernoulli process is a stable approximation of the Poisson process and converges to it in the limit. This methodology differs from that employed by Jorion (1988), which invoives truncation of the infinite summation of the Poisson process. In essence, the likelihood function is cast as a mixture of two normals (for a detailed discussion, see Akgiray and Booth (1987)); namely, L(t, te % Bi Y4) = PU = OVP = 0) + PUT = NPQ = 1) ( a Prey ‘ PL 2gF Pre i 57Energy Derivatives: Pricing and Risk Management gradient search method appropriate for maximising the log likelihood. Although usually robust, we found the BHHH method failed to converge for some data sets and model choices. In some cases, we were able to achieve convergence by first using a standard Simplex search of the parameter space to determine the location of the global maximum. The results of the Simplex search were then used as starting points for the BHHH algorithm The relative goodness of fit for different models can be compared using the likelihood ratio test. Under the null hypothesis that the enlarged parameter space does not improve the fit to the data, the ratio test statistic is defined as -2AL(@; y), where AL(O; y) is the difference between the two log likelihood function values. This test statistic has a chi- square distribution with degrees of freedom equal to the difference in the number of parameters estimated between the two spaces. Another related measurement of fit is the Schwarz criterion (Schwarz (1978)) defined as: SC = -2L(@;y) +kInN (3.38) where L(O; y) represents the likelihood function, & is the number of parameters to be estimated and N is the number of observations. The Schwarz criterion accommodates the trade-off between better fit and more parameters by penalising the model with larger number of parameters. The most probable model is the one with the smallest Schwarz criterion. Table 3.5 shows the Maximum Likelihood results of the fitting procedure together with Schwarz criteria for each model and commodity. A number of interesting conclusions can be inferred from the data in Tables 3.3 and 3.5. The electricity data are highly skewed and leptokurtic. The Jarque-Bera statistic described in equation (3.36) is about 2000 for the electricity data, which overwhelmingly rejects the hypothesis of normality. Also interesting are the results for gas and crude oil. Over this time period, both return series exhibited platykurtic, or thin-tailed behaviour, and therefore, as might be expected, the ARCH type models did not outperform the pure diffusion fit. The electricity data, in contrast, appears better described by jump diffusion rather than simple diffusion as demonstrated by the differences in Schwarz criteria. Furthermore, there is some indication that GARCH(1,1) combined with jump diffusion is marginally more descriptive than jump diffusion alone. Table 3.6 shows detailed TABLE 3.5 Maximum Log likelihood Values and Schwarz Criteria (SC) NPIS Gas Crude L(;y) sc L(8;y) sc Le. PD 60 —109 919 —1827 992 -1973 ID 148 ~267 D+Arch 84 -151 917 1817 991 —1965 D+Garch 121 =219 919 1815 995 —1967 JD+Garch 167 =294 993 —1946 (No value means that the regression did not converge). 58TABLE 3.6 Regression Results for NP15 (t-statistics for each parameter are shown in Volatility Estimation in Energy Markets parentheses) Parameter Pure Diffusion (PD) Jump Diffusion (JD) Jump Diffusion + GARCH(1,1) (ID + Garch) r 0.001 + 0.028 (0.01) 0.012 +: 0.017 (0.67) 0.001 + 0.015 (0.04) 2 0.249 +t 0.008 (32.84) 0.066 + 0.008 (8.77) a 0.033 + 0.008 (4.22) a 0.349 + 0.095 (3.66) By 0.208 + 0.094 (2.21) é 0.131 + 0.033 (3.99) 0.063 + 0.023 (2.76) bh 0.086 + 0.214 (0.40) 0.209 +: 0.400 (0.52) 1.391 + 0.385 (3.61) 1,538 + 0.861 (1.79) regression results for Pure Diffusion, Jump Diffusion, and Jump Diffusion + GARCH(1,1) on NPIS electricity prices: We see that the models are generally consistent with each other. The mean return « and mean jump size jz are small and poorly determined in all models. The diffusion variance 9” in the JD model is roughly consistent with the unconditional variance of the D+Garch model, calculated to be 0,075.> Furthermore, the JD total variance, calcu- lated to be 0.248,* is consistent with the PD model’s variance of 0.249, while the JD +Garch model total variance of 0.172 is somewhat lower than expected. Interest- ingly, the jump frequency ¢ in JD is twice that of the D+ Garch model, while the jump variance 7” is smaller (although the difference is statistically insignificant). It is tempting to suggest that perhaps some of the intermediate valued returns are being accounted for differently in each model. In the JD model, there is little choice but to classify them as jumps. In the D +Garch model however, it may be possible to attribute them to the GARCH process, particularly if they tend to cluster in time. While comparative tests like the likelihood ratio test and the Schwarz criterion offer an indication of the relative quality of each model, they do not yield an absolute assessment. Many more tests, just a few of which have been touched on this section, are required to determine if the hypothesised specification yields a model that statistically identical to the observed data. As we alluded to at the beginning of this section, a useful first test is often a simple qualitative inspection of a sample price process and the associated log return distribu- tion. We conclude this chapter then with figure 3.5, which is divided into panels A, B, and C and show, respectively, comparisons of simulated PD, JD, and JD + Garch prices with the actual NP15 data. In each panel, the left graph shows a simulated price history in grey and the observed NPIS data in black. The panels’ right graphs depict the 6 (=o) =f) + We define the total variance for the jump models to be equal to a? + 6-2, where 0” is the unconditional varianee. > The unconditional variance fora GARCH(1.1) process is ¢? = 59Energy Derivatives: Pricing and Risk Management $500 $450 $400 $350 $300 $250 S/MWh $200 NPIS PD 60 Mean — 0.001 0.008 Variance 0.232 0.256 Skew 0.461 -0.043. Kurtosis 11.98 0.227 8 & Frequency 8 8 floes £3. = au ge oer ra 1.8 15 1.2 zsh a a 27 2.4 2.1 0.9) 0.6 0.3 03 Log Returns FIGURE 3.5A Comparison of Simulated Price Processes with NP15 Data. A: Pure Diffusion 60Volatility Estimation in Energy Markets 70 ee NPIS 60 Mean 0001-0003 Variance 0232 0.250 Skew 0.461 1.068 50 Kurtosis 1198 1041 Frequency FIGURE 3.58 Comparison of Simulated Price Processes with NP15 Data. B: Jump Diffusion 61Energy Derivatives: Pricing and Risk Management 70 NPIS JD+Garch 60 Mean 0,001 0.0004 Variance 0.232 0.154 Skew 0461 0.947 50 Kurtosis 1198 14.85 = £ 40 5 8 z E 30 20 10 0 FIGURE 3.5C Comparison of Simulated Price Processes with NP15 Data. C: Jump Diffusion with GARCH(1,1) 62Volatility Estimation in Energy Markets distribution of NPIS daily log returns as a histogram and daily log returns for the simulated prices as an overlaid black line. Casual inspection shows that while the Jump Diffusion models do yield more realistic price processes in a number of important respects, they nevertheless produce price paths that are clearly distinguishable from the observed data. 3.5 Summary In this chapter we have discussed volatility modelling and estimation in the energy commodity markets, emphasising the difference between this market and financial markets. We discuss the estimation of volatility from both historical and implied data, again from the perspective of the energy user. We further discussed a number of stochastic volatility models and have shown how to estimate the models via ordinary least squares and maximum likelihood. We have tested our estimation techniques on a number of examples drawn from energy markets including electricity, gas and crude oil. 634 Energy Forward Curves 4.1 Introduction In this chapter we discuss approaches that have been put forward by a number of authors for constructing energy forward curves. Further useful discussions are provided in Gabillon (1995), Humphreys and Shimko (1997), and Leong (1997). Forwards and futures markets are often used by risk managers to hedge their risk, with liquid forward prices helping the price discovery mechanism to determine the fair value for future delivery. Prices from these markets are also the key inputs to many derivative pricing models — we will see in Chapter 8 that energy derivative prices can be evaluated given the forward curve and the associated volatilities of the forward prices. As such, forward prices and volatilities play an essential role in energy markets and are as important as the level of spot prices. The forward curve contains information about the prices an investor can lock into today for different times in the future. Forward prices allow users to lock into buying (long forward) or selling (short forward) spot energy at a certain time in the future at the forward price. Figure 4.1 illustrates three market forward energy curves. Each curve is a plot of the first 24 monthly contracts with prices taken from NYMEX. The curve for NYMEX crude oil futures at the end of March 2000 is clearly in backwardation (futures prices lower than the spot price) with short dated futures trading at just under $27 dollars, declining to $21 for the two year contract. A price of $21 implies that the holder of the long position can buy oil at $21 at the maturity date of the contract, irrespective of the spot price for oil on that date. The crude oil futures curve just over a year earlier however, at the beginning of March 1999, is in contango (futures prices higher than the spot price) with short dated futures trading at prices below their longer dated counterparts. The third futures curve shown in figure 4.1 is the NYMEX Henry Hub natural gas futures contract observed on the 31 March 2000 and clearly shows the strong seasonality of energy prices associated with the North American seasons. Later in this chapter we explain how forward curves can take these three shapes dependent on the level of transaction costs, the concept of a “convenience yield”, and seasonal changes in supply and demand. 64Energy Forward Curves 28 38 = O11 31/03/00 = + o101/03199 | a. i > Gas 31/0300) |. 4 a 3.2 2g Ps = 20 30 & & 8 3 = @ 18 28 0 : g 5 16 26 i ee 24 12 22 10 2 7/2345 6 7 8 9 10 11 1213 14 15 16 17 18 19 20 21 22 29 24 Months to Maturity FIGURE 4.1 Market Forward Curves Forward Curves in the Debt Markets Forward curves are well-known and understood in the debt markets. Forward rate agreements and exchange traded futures contracts (for example Eurodollar futures contracts) are heavily traded and allow users to lock-in borrowing and lending rates for future time periods. These forward rates are easily implied from the prices of bonds and Spot interest rates. For example, if the price of a 6 month pure discount bond is trading at 0.9608 and the price of a 9 month pure discount bond is trading at 0.9347 then the implied 3 month forward rate starting in 6 months time, with continuous compounding is 11.02%, given by: [esa] (4.1) a 0.25 [0.9608 The pricing of the forward interest rate instrument is therefore determined from the current term structure of interest rates (described above by the prices of pure discount bonds). Equation (4.1) shows that forward rates are not arbitrarily set but depend on the relationship between traded instruments. The calculation also shows that the forward rate is calculated based on the observed current spot values and is not dependent on the forecasted \evel of future spot values. Leong (1997) makes the point that, for many players in the immature electricity markets, “there is not even a consensus among market 65Energy Derivatives: Pricing and Risk Management sol players as to what is meant by the term ‘forward curve’”’, Many participants, and even some data vendors, mistake schedules of price forecasts for forward prices. Price forecasts are predictions on the likely spot price for periods in the future and can differ widely between market participants. Forward prices, on the other hand lock in tradeable prices today for future spot transactions and as such capture the market reality. As far as possible these prices should be observed directly from the market. 4.2 Constructing Forward Curves In the financial markets, no arbitrage models relating spot and forward prices offer a direct and reliable valuation of forward prices. This relationship is governed by arbitrage between the spot and forward contracts which is facilitated by the ability to easily buy and sell the spot instrument. However, for many energies the relationship cannot be described so simply and this will have an impact on the forward prices observed in these markets, 4.2.1 Cost of Carry Relationship For many markets the price of a futures contract can be derived from the price of the spot instrument and the financing cost. This direct relationship results from the simple fact that the payoff to a forward sale of an asset, for example a stock, can be replicated by borrowing the purchase price in the money market, purchasing the stock, holding (or “carrying™) the stock until maturity, and then delivering the asset into the long position, using the funds received to pay off the loan. This type of arbitrage is often referred to as. “cost of carry arbitrage’ and leads to the following pricing formula for the s-maturity futures price at time f: F(t,s) = S(te"? (4.2) where r is the financing cost, assuming continuous compounding, over the life of the forward position, s — ¢, and S(t) is the asset price at time r. It is easy to see that if the market prices for the stock and the future on the stock do not satisfy equation (4.2) then there are arbitrage profits to be made. For example, if the market forward price is greater than the carry relationship (F(t,s) > S(t)e"*”), investors can borrow an amount equal to the price of the asset, buy the asset and take a short position in the forward contract. At the maturity of the contract the spot instrument is sold at the agreed price and the proceeds from the sale used to pay off the loan which has now grown to S(t)e"-9. The difference between the agreed amount and the borrowing is therefore a guaranteed profit. If, on the other hand, the market forward price is less than the carry relationship (F(t,s) < S(aje"—), investors can sell the asset at the spot price, depositing the * Leong (1997) p. 133. 66Energy Forward Curves proceeds in a bank account yielding r and simultaneously take a long position in a forward contract. At the maturity of the contract the asset is bought for F under the forward agreement using the proceeds from the initial deposit, once again yielding a profit equal to the difference between the market price and the cost of carry level. The cost of carry relationship is only slightly more complicated if the asset pays a continuous yield 6, Recall that the resulting futures pricing formula is given by equation (1.1). The above arbitrage arguments hold here as well once the continuous dividend yield accruing to the physical position is taken into account. 4.2.2 Forward Price Bounds for Energies The cost of carry relationship holds in markets where arbitrageurs are able to easily purchase and short sell assets. Now consider the cost of carry relationship for an energy, say oil. For the short future, long spot arbitrage, the arbitrageur would need to pay to store the commodity and transport the energy to the buyer of the forward at the maturity of the contract. The financing cost therefore now con: of the money market interest rate plus the cost of storage. This suggests a cost of carry relationship of the form F(t,8) = S(Helner9 (4.3) where w represents the cost of storage on a yield basis. Both the interest rate and the cost of the storage are positive and so the futures price is greater than the spot level implying an upward sloping (contango) forward curve, and this indeed corresponds to the observed curve for crude oil at the beginning of March 1999. However, often the forward curve is downward sloping (backwardation) as can be seen by the market curve at the end of March 2000 in figure 4.1. This implies that the lower bound on forward pricing is not being met with the cost of carry described here not the only determinant of the futures price. One argument regularly used is that it is often not possible to sell the spot energy short (for example by leasing) and therefore, since the arbitrage cannot be executed, there is nothing to stop forward curves being in backwardation. The current discussion implies that, in times of backwardation, market participants, instead of buying oil and storing it for future use, should purchase oil for future delivery at a discount to the spot price. However, for some participants holding the spot energy has value relative to having a forward contract. The benefit derived has been termed the convenience yield” or consumption value as companies keep the energy in inventory in order to be able to consume it if they want. In the oil industry inventory holders such as end users have a high convenience yield since they cannot afford to be short of physical oil. We can represent the convenience yield therefore in terms of an effective continuous dividend stream which the holder of the spot will receive and which we again denote by 6. 2 Brennan (1989) defined the convenience yield as the “... flow of services which accrues to the owner of a physical inventory but not to the owner of a contract for future delivery” 67Energy Derivatives: Pricing and Risk Management The full cost of carry for an energy therefore becomes: F(t,s) = S(e™" 9) (4.4) Depending on the relative size of the storage costs and convenience yield the resulting forward curve can be in contango or backwardation. Gabillion (1995) discusses the limit to contango and backwardation in terms of the levels of storage and financing costs and on the convenience yield. 4.2.3 Seasonality in Prices The forward curve for crude oil is downward sloping most of the time, but as we have seen can move in to contango, sometimes very quickly. Forward curves such as natural gas, however, cannot be so easily described, often exhibiting strong seasonal cycles. An example of this is the NYMEX natural gas curve illustrated in figure 4.1. For this curve, the peaks occur during the northern hemispheres winter when natural gas is in high demand for heating. The size of the difference between the low point on the gas curve ($2.69 for the April 2001 contract) and the high point the following winter ($2.98 for the December 2001 contract) suggests that there is a clear arbitrage opportunity — buy natural gas forward when prices are low and store the gas until the next winter peak. Execution of this arbitrage should continue until the peaks disappear. Leong (1997) suggests that the presence of arbitrage opportunities of this kind continue to exist because of the inefficiency of the gas forward market compared to the financial markets. He also suggests several reasons for this inefficiency; — there is little incentive of many players in the gas market to be cost conscious as increases in gas prices can be passed on to end users by many utilities. — in order to be able to execute arbitrages a player needs a presence in both the spot and derivative markets — for example, the arbitrage described above requires the user to buy and store gas. — this kind of relative value trading is new to many gas market players and, as the market becomes more efficient, this kind of discrepancy will diminish. These arguments suggest that as the players in the market become more familiar with arbitrage between spot and forward markets and more conversant with derivatives the level of seasonality should diminish. 4.3 Forward Curves in the Electricity Market Of all energy commodities electricity exhibits the most complicated seasonal patterns. In many markets there are two peaks in the forward curve as a result of weather. A peak is observed in winter due to the reliance on heating and again in summer because of air conditioning. Additionally, there can exist daily and hourly patterns. Figures 4.2(a) and 4.2(b) show spot price seasonality for the Australian electricity market for weekdays and 68Energy Forward Curves 10.00 8.00 _ — — — : 7 7 @ x 7 =99 own ere er PSaRRSkssseasgsrees Half Hour (Note: Monday and day after public holiday (daph) has been grouped together) FIGURE 4.2(a) Median Price For Weekdays weekends respectively. The figures show the median price per MWh for each of the 48 half hours for the different days for the NSW market. Leong (1997) suggests a number of reasons for this strong level of seasonality. Firstly the more complicated patterns are due to the non-storability of electricity — once generated it has to be consumed immediately. This implies that arbitrage relationships which depend on storing the energy breakdown once the electricity has been generated although importantly electricity can be stored indirectly via the generator fuel’. The difficulty of storing electricity means that there is an absence of a leasing market which affects the short spot energy, long forward arbitrage. A third suggested reason is that electricity is still a very regional market. Differences between regions arise due to differing methods used to produce electricity, weather patterns, demographics, local supply and demand conditions, etc. Links between regions, such as that between New South Wales and Victoria in Australia, help to reduce such regional variations, but constraints on the lines mean that they don’t disappear completely. As with other forward markets the forward curve should be constructed to fit available market prices, Unfortunately the illiquidity of many instruments in the + Blectricity can also be stored directly via hydroelectric schemes. 69Energy Derivatives: Pricing and Risk Management 22.00 2000; SunePh | “= Sat 18.00: 16.00. 8 8 a 14.00 12.00 Af 10.00 SO2HRSrReRSrQOR SRC atCssagssersess Half Hour -OON OTE OND (Note: Sunday and public holiday (ph) has been grouped together) FIGURE 4.2(b) Median Price Weekends electricity area means we often have to turn to other techniques. The current literature is dominated by three main approaches to building forward curves in the electricity market and we briefly describe the approaches in the following sections. The arbitrage approach relates the forward electricity curve to the prices of the fuels used to generate the electricity. The econometric approach uses econometric techniques to derive the forward prices of electricity by taking account of key econometric variables such as fuel cost, weather patterns, etc. Finally what we will call the spot price modelling approach derives forward prices by modelling underlying stochastic factors such as the spot price, convenience yield, long term mean price, etc. and deriving analytical formulas to characterise the curve. 4.3.1 Arbitrage Pricing Approach Although electricity cannot be easily stored the fuels used to generate the electricity can be stored, and the link between the two fuels implies that the forward curve for electricity should be related to the input fuels. The arbitrage pricing approach takes this into account by considering the conversion process. 70Energy Forward Curves One of the key steps in the conversion process is the generation process itself and this depends on the efficiency of generation expressed as the heat rate — the number of BTU’s (British Thermal Unit) required to generate one kWh of electricity. A basic electricity forward curve can be obtained from the fuel forward curve via the following relation- ship: CoStetecirieny = Heat Rate x Pricey: (4.5) A constant value of the heat rate implies that the shape of the electricity forward curve should resemble the forward curve of the input fuel. The calculation of the heat rate, especially for long term contracts is notoriously difficult given the speed of technological innovation in the electricity industry. The cost of electricity can be converted into a forward price after taking into account costs associated with fixed assets, transmission and tolling charges and others such as fuel purchase, fuel storage, and fuel transporta- tion. These costs obviously change through time. 4.3.2 The Econometric Approach Griswold (1997) discusses a simulation model which has inputs such as capacity ratings of transmission lines, fuel costs, total load by sub-region on an hourly basis, and data relevant to adjacent regions to compare incremental power prices. This type of analysis is performed by a number of commercial software packages for forward curve calculation. The output to a model of this type, however, is a prediction or forecast of spot hourly energy prices in the future. It is not a forward price in the sense of our definition in section 4. — i.e. the price which we can lock into today for trading the spot instrument at some time in the future. 4.3.3 The Spot Price Modelling Approach This approach derives forward energy prices from assumptions about the stochastic factors driving the spot energy price and other key variables (for example, the long term price, the convenience yield, or interest rates). Models of this kind are well known in the fixed income market. For example, traditional models of interest rates, such as Vasicek (1977) and Cox, Ingersoll, and Ross (1985) explain the whole term structure of interest rates (discount bond prices, spot and forward rates) in terms of a stochastic description of the short term interest rate’. The approach involves specifying the stochastic factors that drive the evolution of the spot price. Forward prices, and all other derivatives whose value depends on the stochastic factors and time, can then be derived after applying various mathematical techniques. Depending on the assumptions employed about the random driving factors futures prices are often derived analytically. Examples of this * Muiti-factor interest rate models such as Longstaff and Schwartz (1992) and Fong and Vasicek (1992) attempt to explain the term structure by modelling extra stochastic factors such as short rate volatility, the long rate, and other factors. 7Energy Derivatives: Pricing and Risk Management approach include Gabillion (1995) and Pilipovic (1997) who both suggest modeling the spot energy price and a long term price. All the models we study in Chapter 6 can be viewed as being consistent with this approach. 4.4 Summary In this chapter we have discussed the importance of forward curves for risk management and derivative pricing in the energy markets. Forward curve construction is more difficult in the energy markets when compared to the financial markets because of the cost of storing the spot energy, the fact that some energies derive a convenience yield to the holder, and the illiquidity of forward instruments in many markets. ps5 Energy Derivatives: Structures and Applications 5.1 Introduction Both exchange traded standard derivatives and over-the-counter exotic options are increasingly used in energy markets as a means of controlling exposure to energy prices. In this chapter we give an overview of these options, beginning briefly with some exchange traded derivatives including both futures and options, and continue with a discussion of standard over-the-counter instruments including caps, floors and swap derivatives. We conclude with a discussion of exotic options including compound options, Asian options, lookback options, barrier options, and spread options. Exotic options are derivatives with a more complicated payoff structure than standard. derivatives. They are also known as second generation, or sometimes path-dependent options. Almost all exotics trade over-the-counter. The distinctive features of the energy markets have led to a range of derivative products which are used for risk management. For example, the high volatility of energy markets and the price spikes observed in electricity markets have made Asian (or average price) options a popular risk manage- ment tool. Another popular class of instrument is spread options which are used by many participants to hedge against basis risk. In an article describing the use of exotic options in energy markets Kaminski and Gibner (1995) point out that “one of the main reasons that exotic options have been accepted within the energy industry is that options were, in fact, embedded in many energy contracts long before they became fashionable tools of financial engineering”. 5.2 Exchange Traded Instruments Although most energy derivatives are traded over-the-counter a number of exchanges worldwide list energy products. For example, the New York Mercantile Exchange (NYMEX) trades futures and American style futures options on oil, gas, and electricity Chapters 6, 7 and 8 discuss pricing derivatives of this type. NYMEX also lists crack spread options which are options written on inter-market spreads such as the heating oil / crude oil spread-the pricing of these is dealt with in Chapters 7 and 8. See also sections 5.3.4 and 5.7.2 for a discussion of crack swaps and crack spread options. Electricity futures and options trade on the Chicago Board Options Exchange with 73Energy Derivatives: Pricing and Risk Management electricity futures also trading on the Sydney Futures Exchange in Australia and on the Nordic Electricity Exchange. Although there is an increasing amount of trading in exchange traded derivatives the majority of trading is in over-the-counter instruments. We begin a description of these products by looking at energy swaps. 5.3 Swaps The first energy swaps were traded in October 1986 when Chase Manhattan Bank acted as counterparty to Cathay Pacific Airways and Koch Industries in an oil-indexed price swap. Other institutions which lead the way in these early days were Citibank, Phibro Energy and Bankers Trust who were to become one of the leading providers of energy derivatives in those early years. Swaps are also known as Contracts-for-Differences' (CFD’s) or Fixed-for-Floating contracts. They are used to lock in a fixed price for a certain predetermined but not necessarily constant quantity. In the following sections we describe a wide variety of swap structures which are now common; vanilla, differential, participation, double-up, and extendable. 5.3.1. Vanilla Swap The vanilla swap is an agreement in which counterparties exchange a floating energy price for a fixed energy price. That is one counterparty pays a fixed price and receives the floating price either by receiving the cash value of the spot energy or the spot energy itself. This counterparty would normally be considered the buyer or receiver of the swap. The other counterparty, the swap provider, receives the fixed prices and cither supplies the spot energy or its cash equivalent. The agreement or contract defines the fixed volume or quantity over a specified period of time. The contract is settled at prede- termined regular intervals, the settlement or reset dates, over the period of the contract — typically monthly, quarterly, semi-annually or annually. An example of a fixed price receiver might be an oil producer who enters into an oil swap in order to hedge future sales of oil. A producer would be concerned that oil prices could fall in the future and may therefore wish to sell forward its anticipated oil production based on existing oil forward prices. A fixed price payer might be a oil refiner who would be adversely affected by future oil price increases. A swap would allow them to buy forward their anticipated oil consumption based on current oil forward prices. Vanilla swaps can be priced directly off the forward energy curve since the appropriate portfolio of forward deals is an exact hedge for the swap. Therefore the payoff at each reset date is the same as for a forward contract. Let the reset dates at which the cashflows occur be denoted by sy: k =1,..., mand the fixed price payment (strike price) be denoted by K. A vanilla swap is usually viewed as a weighted average of the forwards underlying ‘ Note in some markets caps and floors are also known as Contracts-for-Differences. 74Energy Derivatives: Structures and Applications the swap with weights being the discount rates to each of the swap settlement dates. Therefore, we write the value of a vanilla swap today (¢) as the following’. Vanilla_Swap(1) = sy P(t, 4) (F(t, 5,) — K) (3.1) k=l where P(¢,s) discounts the future cashflows back to today (see Chapter 1). §.3.2 Variable Volume Swap This contract is identical to a vanilla swap except that the underlying quantity or volume is not known in advance. For example, in a “Whole of Meter” swap the volume is exactly the metered amount realised for a customer. This type of unknown volume is generically termed “swing” and in order to price contracts involving swing the variability of the volume must be modelled. We discuss variable volume contracts in greater detail in Chapter 7. 5.3.3 Differential Swap A differential swap is similar to a vanilla swap except that the counterparties exchange the difference between two different floating prices (the differential) for a fixed price differential. Differential swaps are often used by companies to hedge basis risk in their normal hedging activities. For example, a refiner’s profitability depends on the market price differential of the raw commodity and the refined products and a differential swap allows a refiner to lock in their refining margin at the fixed price. A differential swap is essentially a portfolio of forward differential contracts with maturity dates corresponding to the settlement dates underlying the differential swap. As a consequence, the differential swap price is a weighted average of the forward price differences underlying the differential swap, with the weights being the discount rates to each of the differential swap settlement dates. Using the same notation as before but denoting the forward prices on the two different floating prices as F)(¢, s) and F(t, s) we have that the time 7 value of a differential swap is given by: 1 Differential Swap(s) = m 37 PU. 5)(Fil65)— False) -K) (52) 5.3.4 Margin or Crack Swap The crack spread or margin is the price differential between the raw commodity or energy, for example crude oil, and the refined product or products, for example gasoline, ? Throughout this book we define swap prices on a per reset basis. 75Energy Derivatives: Pricing and Risk Management jet fuel, or gasoil. The margin or crack swap is a specific form of differential swap in which the fixed price payer, for example a refiner, receives the difference between the market price differential of the raw commodity and the refined products in appropriate fractions and the fixed price. This type of swap therefore allows a refiner to lock in their refining margin at the fixed price. These types of structures are very popular because the market price of the refining margin can often be negative for sustained periods. Crack spreads are discussed further in section 5.7.2 below. 5.3.5 Participation Swap This is similar to a vanilla swap in that, for example, the fixed price payer is fully protected when prices rise above the agreed fixed price but they participate in a certain percentage of the savings if prices fall. For example, consider a cash settled swap for gasoil at a fixed price of $150 per tonne with a participation of 50%. If the gasoil price is above $150 per tonne, say $160 per tonne, the fixed price payer is compensated by the difference between the market price and the fixed price i.e. receives $10 per tonne. If the market price is below $150 per tonne, say $140 per tonne then the fixed price payer would only pay the provider $5 per tonne rather then the $10 per tonne which they would pay under a vanilla swap. For a fixed payer, this instrument is actually a portfolio of a vanilla swap combined with a long position in a floor (see section 5.4) with a strike equal to the fixed payment and the same settlement dates as the swap. A fixed payer will pay for the floor premium through a higher fixed price on the participation swap than that paid on the correspond- ing vanilla swap. Participation swaps illustrate that, even in such apparently simple structured swaps, option pricing techniques are important in obtaining correct valua- tions. 5.3.6 Double-Up Swap In this type of swap the fixed price payer can achieve a better swap price than the market price but in return the swap provider has the option to double the volume before the pricing period starts. This is achieved by the fixed price payer implicitly selling a put swaption to the swap provider (fixed price receiver). The premium is used to reduce the fixed price of the underlying vanilla swap by splitting it across the swap reset dates and deducting it from the market swap price. A natural fixed price payer would be an oil refiner who would enter into a swap to hedge against future oil price increases. The fixed price payer is exposed to the risk that if swap prices fall the fixed price receiver (swap provider) will exercise the right to double the swap volume. In this case, the fixed payer will now have to pay the relatively high fixed price on twice the notional amount compared to the lower market prices. An oil refiner may be willing to take this risk if they believe oil prices will rise as the swaption will then expire worthless and they achieve a better fixed price on the swap. 76Energy Derivatives: Structures and Applications 5.3.7 Extendable Swap This agreement is similar to the double-up swap except that the swap provider has the option to extend the period of the swap for a predetermined period. This is again achieved by the fixed price payer selling a put swaption to the swap provider. The premium is again used to reduce the fixed price of the underlying vanilla swap. 5.4 Caps, Floors and Collars Caps (also sometimes known, together with floors, as one-way CFD’s in some markets) provide price protection for the buyer above a predetermined level the cap price—for a predetermined period of time. A floor guarantees the minimum price that will be paid or received at a predetermined level —the floor price. A collar (also known as two-way CFD’s in the Australian electricity market) is a combination of a long position in a cap and a short position in a floor. Standard caps, floors and collars are for predetermined quantities or volumes. These instruments are usually cash settled at regular intervals over the period of the contract. A generic cap can be viewed as a portfolio of standard European call options with strike prices equal to the cap level and maturity dates equal to the settlement dates of the cap. Let the dates at which the cashflows occur (often called “fixing” or “reset” dates) be denoted by %; k = 1,...,m and the fixed cap price (strike price) be denoted by K. The time t value of a standard cap is given by: Cap_Price(t) = $7 c(t, F(t, te); Ky teste) (5.3) k=l where c(t, F(s,); K, Ts) is the price of a standard European call option with strike price Kand maturity 7 on a s-maturity forward contract. Results in Chapters 6, 7 and 8 show how these option prices can be evaluated for a wide range of energy models. This type of fixed price cap is commonly found in the electricity markets where they are often fixed on an hourly or half-hourly basis. In other energy markets such as oil or natural gas it is more common for the cap to be fixed on a monthly basis and the individual options to be Asian options fixed on the daily price over the relevant month. We describe Asian options in detail in section 5.8.1. An example of the application of a cap might be an oil refiner who wants to protect themselves against an anticipated rise in the price of crude oil over the next twelve months. Suppose the current price of crude oil is $18 per barrel and a refiner purchases a monthly cap on the average of the daily spot price in the month struck at $20 per barrel. A cap with al year maturity on the average daily price over each month is cheaper than a cap on the daily price for a year and achieves roughly the same level of protection due to the mean reversion in oil prices. If the average price in any month exceeds $20 per barrel then the refiner will still purchase crude oil in the market at the higher price but will be compensated for the difference between the actual price and the strike price by the cap provider. Similarly, a generic floor is a portfolio of European put options with strike prices 77Energy Derivatives: Pricing and Risk Management equal to the floor level and maturity dates equal to the settlement dates of the floor. Assuming the same fixing dates as for the cap and with the fixed floor price (strike price) denoted by K, the value of a standard floor at time r is given by Floor_Price(t) = Soler 1); K, ths te) (5.4) 1 where p(t, F(t, s); K, T,s) is the price of a standard European put option with strike price Kand maturity T on a s-maturity forward contract. A collar is simply a portfolio of a long position in a cap and a short position ina floor. Collars are typically used by energy buyers who wish to hedge against price increases and wish to use the premium on short floors to pay for the cap protection. For example, a buyer who decides they cannot afford to pay the full price of the cap might agree to purchase a cap and finance it through the sale of a floor. The strike prices of the cap and floor can be set to yield a collar at zero cost. 5.5 Swaptions A swaption is a European option on an energy swap. A call option on the swap, also called a payer swaption, with strike price K and time to maturity T, provides the holder of the option the right to enter into a swap, paying the fixed price K and receiving the floating energy spot price. Conversely, a put option on a swap, or receiver swaption, gives the purchaser of the option the right to sell a swap or receive the fixed price K and pay the floating price. If s,; k = 1,..., denote the dates at which the swap cashflows occur, we can write the payoff to the payer swaption as wn (050 FUT) 7 ) (5.5) el Note that equation (5.5) represents the payoff to a cash settled option. If, upon exercise of the option, the option holder enters into a new ‘physical’ swap position, the payments will be exchanged on the swap rest dates and hence the payoff at time 7 must reflect these future cashflows: wa (0, 8S er SJE (T, 54) *) (5.6) where P(T’, s;) is the value of a pure discount bond maturing at time s;,, observed at the maturity of the option, and corresponds to the initial discount function unless we assume stochastic interest rates which are correlated with the energy price. The receiver swaption can be written similarly but as a put option for both cash settled and physical settled options. The value of a cash settled European payer swaption can therefore be written in terms of the discounted expected payoff at maturity: 78Energy Derivatives: Structures and Applications Supn(t;K,T.{s¢},m) = P(t, T)E, [nes (0S crea 2 ) (5.7) = If the payer swaption involves physical settlement then equation (5.7) becomes: Swpn(t;K,T. {s,},m) = P(t, T)E, [mes (0.h 5 P(T, 5: (F(T, x) — *)) (5.8) a Since a swap is a portfolio of forward deals, a swaption is an option on a portfolio of forwards, therefore analytical formulae are typically not available for these instruments and numerical techniques must be used. Sections 7.2 and 8.8 provide a discussion on how swaptions can be priced, in a single factor spot price model and a general multi-factor forward curve model, respectively Swaptions are typically purchased by institutions that anticipate the need for fixed price protection of a swap but are unsure of the direction of future market prices. If the institution purchased a forward starting swap then it would be locked into current swap prices and if the market prices fell would not be able to participate in the savings. By purchasing a swaption the institution can lock into the current market swap prices by exercising the option when markets move favourably. However, if market prices turn out to be lower then the institution can abandon the option and obtain the required swap more cheaply directly from the market. 5.6 Compound Options - Captions and Floptions An option which allows its holder to purchase or sell another option for a fixed price is called a compound option. Compound options occur in energy markets in the form of captions and floptions as well as in their standard form as call or puts on simple calls or puts. A caption is an option on a cap and a floption is an option on a floor. These instruments are therefore options on portfolios of options and as such no simple analytical formulae exist*. Compound options are typically purchased by companies which are anticipating the need for option like protection e.g. a cap or floor in the medium term future and are concerned that market prices may rise between now and the time when they will need to purchase the protection. A typical example is the case of a company which is bidding for a project at a certain price. They are exposed to the risk that market prices will be higher than those on which the bid was based. The company would therefore like to purchase an option which locks in the prices on which the bid was based but only if they are awarded the contract. By purchasing an option on the option, struck at the relevant forward price, they can pay a small premium now to lock into a known price for the future protection. * European versions of standard calls or puts on underlying calls or puts can be priced using the results of Geske (1979) in a Black-Scholes world. 79Energy Derivatives: Pricing and Risk Management 5.7 Spread and Exchange Options A spread option is written on the difference between two prices. Options of this type are popular in the energy industry as many players are exposed to the difference in prices of related energies or are able to sell the energy in different markets. Exchange options are also designed to provide a payout which is based on the relative movement of two energies. We begin by considering two types of spread options. The first is an option written on the spread of two different maturity futures contracts on the same underlying spot instrument. The second is an option on the spread between the prices of futures contracts of two related commodities. 5.7.1 Calendar Spreads If the futures contracts are written on the same underlying energy, but with different maturity dates, then the option is often referred to as a calendar spread option. Let F(t,5,) and F(t, s)) represent the current price of futures contracts on an energy (say oil) that mature at times s; and s) respectively. We define the maturity payoff to a European call spread option with strike K and maturity T(¢< T < 5, < sp) to be: max(0, F(T, s;) — F(T,5) — K) (5.9) The payoff to a European put spread option is defined similarly. Such options are used to protect against changes in the shape of the energy forward curve. The value of the option at time 1 of the call option can be written generally as the following discounted expectation: Calendar_Spread(s; K, T, 51,82) = P(t, T)E,[max(0, F(T, s;) — F(T,s2) — K)] (5.10) These options are valued in Chapters 7 and 8 under single factor and multi-factor model assumptions respectively. 5.7.2 Crack Spreads If the futures contracts underlying the option are written on two separate energies (say natural gas and electricity), then the option is often referred to as a crack spread option. The maturity of the futures contracts can be on the same date or different dates. Options of this type are often used by companies who are exposed to the difference in price between two different energies. For example, one energy might be an input into a process that produces another energy, as is the case for gas fired electricity generators. Other examples include heat spreads (the difference between Number 2 heating oil and crude) and gasoline spreads (the difference between unleaded gasoline and crude). Let F,(7,s;) represent the price of an s,-maturity futures contract on energy a (say natural gas) and F;,(t,s2) represent the price of an sj-maturity futures contract on energy 4 (say electricity). The payoff to a European call option with maturity T and strike K on 80Energy Derivatives: Structures and Applications the spread between these two forward contracts can be represented by the following: max(0, F,(T,s}) — Fy(T,52) — K) (5.11) and so the value of the option at time ¢ can be written generally as: Crack_Spread(¢; K, T, Fa, Fy, 51,82) = P(t, T)E;{max(0, Fy(T,s)) — F(T, 52) — K)] (5.12) We show in Chapter 8 how options of this type can be valued using a multi-energy forward curve model. 5.7.3 Exchange Options The spread options of this section have a payoff which is determined by the difference between two forward contracts. Exchange options, on the other hand, provide a payout which is based on the relative performance of two energy prices. We identify two types of exchange option. The first, often called an ‘out-performance’ option pays off on the better performing energy. For example, for a T-maturing European call out-perfor- mance option on the forward contracts in the previous section, the maturity value can be determined as: alo, (Fotos) FAT») mex(0 es a ') hea : ‘)) ee The out-performance option pays out if one energy price moves more than the other ina given direction, for example if gas prices rise more than oil prices. A second exchange option has a similar payoff to a crack spread option (cf. equation (5.11)), but again on the percentage movements in the energy prices: F,(T,s) F,(T,s) mex(0. ea mu *) cae Both of these options are useful in situations where it is the relative movement of two energy prices which is important. 5.8 Path Dependent Options Although many of the options that we have described so far have complicated maturity payoffs, they have been independent of the path taken by the spot or futures prices over the life of the option. The options that we describe in this section are often known as path dependent options because the payoff to this type of option depends on the path that the underlying energy follows before maturity (or some part of the options life) and not solely on the terminal price*. For the majority of traded path dependent options, if not “ Path dependent options are often treated as synonymous with exotic options but are really a subclass. 81Energy Derivatives: Pricing and Risk Management all, the fixings for the dependency occur at discrete times. In the following examples we assume a consistent set of predetermined dates 4; k = 1,...,m, on which prices are observed. S; will denote the spot price at time 1;. If the path dependency is observed over the whole life of the option then #; = ¢ and /,, = T, but in general it can be a shorter period. Usually the fixing dates will be equally spaced such as hourly, daily, weekly or monthly but they can be a predefined set of ad hoc dates. It is also not uncommon for the spot price fixings to be replaced by observations of particular forward or futures prices. The most well known examples of path dependent options are Asian (or average price) options, barrier options, and lookback options 5.8.1 Asian Options — Average Price and Average Strike Asian options are options whose final payoff is based in some way on the average level of an energy price (spot, forward or future) during some or all of the life of the option. Of the range of exotic options, Asian options in the equity and commodity markets are priced and risk managed as being almost vanilla in that they are now one of the best understood options and are very popular as a means to hedge exposure*. There are two basic styles of Asian option; average price options (sometimes called average rate options due to their origin in the foreign exchange markets) and average strike options®. Several forms of averaging are possible with typically the arithmetic average being specified, a convention that we follow in this section, The payoffs to the four main standard Asian options are summarised in table 5.1, where K is the strike price for the fixed strike options. TABLE 5.1 Payoffs of the Standard Asian Options Name Average Price Call Option: Average Price Put Option: Average Strike Call Option: Average Strike Put Option. > See Levy (1997) for the uses, risk management, and numerical techniques for pricing Asian options in a Black-Scholes wosld © Average price options are also known as ‘fixed strike Asians’ whilst average strike options are also known as floating strike Asians’. 82Energy Derivatives: Structures and Applications For existing Asian options, the averaging period may have started before the current valuation date —in this case the ‘current’ average at the time of contract valuation is needed for pricing. Asian options can be both of European or American style exercise. We give an example of how to price both in sections 7.2 and 7.5 respectively As we have already seen one of the commonest occurrences of Asian options is as the component options in caps and floors. In general the main use of Asian options is hedging an exposure to the average price over a period of time. For example, large buyers of electricity often require to hedge their average fuel cost as the prices they charge customers are based on the average purchase prices. Asian options also fit the risk profile of energy producers who need to meet budget targets based on average prices. 5.8.2 Barrier Options In recent years one of the most significant growth areas in new option types has been in the area of barrier options. Barrier options are standard options that either cease to exist (knock-out) or only come into existence (knock-in) if the underlying price (usually a spot energy price or the price of a forward contract) crosses a predetermined level — the barrier (which we denote by H). Barrier options were originally introduced because they are cheaper than otherwise identical standard options due to the fact that the option can cease to exist or never come into existence. There are four basic types of barrier options — the barrier can be above the current underlying price (up barriers) or below (down barriers) and the option can cease to exist (knock-out) or come into existence (knock-in) when the barrier level is crossed. Each of these four types can be either a call or a put which leads to eight different barrier options. A standard variation of the options just described are options which pay a predetermined cash rebate if an ‘out’ option disappears or an ‘in’ option never appears. As a more detailed illustration we describe a down and out call option. Figure 5.1 shows two possible paths for the spot energy price. Both begin at the same spot price and end at the same spot price which is above the assumed strike price of the option. However, the lower price path crosses the barrier level (H) from above and so in this case would ‘knock-out’ and be worthless. Analytical formulae exist for the basic barrier options in the Black-Scholes-Merton world described in Chapter 1 under the assumption that the crossing of the barrier is continuously checked”. For discretely fixed barrier options we need to use numerical techniques to evaluate prices. The common names and payoffs of the eight basic barrier options are summarised in table 5.2. 7 See Rich (1994) for the mathematical foundations of barrier option pricing. However, his exposition is centred ona generalised BSM world 83Energy Derivatives: Pricing and Risk Management 8 £ < g Fi 8 a Time FIGURE 5.1 Down and Out Call Option Example TABLE 5.2 Payoffs of Standard Barrier Options Name Payoff Down and Out Call max(0, Sp — K Up and Out Call max(0, Sp — K) I paxis, Down and In Call max(0, 57 — K)I minis}. Up and In Call max(0, Sp — K)nax(s,,.5 9124 Down and Out Put = Sr ing S\.Su)> Up and Out Put Vanas(S nS Down and In Put max(0, K ~ Sr) lin Up and In Put max(0,K — Sp}Iinaxis;, Where, the indicator function is defined as follows 1 min(S),...,S,,) > (5.15) Unit Song) > = otherwise The other indicator functions are defined similarly. Barrier options can be either European or American style exercise and we show how to price both types in Chapters 7 and 8. 84Energy Derivatives: Structures and Applications Barrier options are used to obtain cheaper hedges than are provided by standard options. For example, consider an energy producer with a natural long position in the underlying energy. They would potentially be looking to purchase an out-of-the-money put option to hedge the risk of the market price of the energy falling. However, if the market price of energy rises, a standard put option hedge would become worthless as the strike price of the option would be far too low. If instead the energy producer entered into an up-and-out put, it would be less expensive than the standard put whilst still providing the same protection. Furthermore, when energy prices rose sufficiently, the up-and-out put would extinguish automatically and could then be replaced by another with a more appropriate strike price. 5.8.3 Lookback Options — Fixed Strike and Floating Strike With lookback options the payout is a function of the highest or lowest price at which the underlying asset trades over some period during the life of the option. For this reason they are occasionally referred to as hindsight options. There are two basic styles of lookback option; fixed strike and floating strike options®. Table 5.3 summarises the payoffs of the standard lookback options. As with the other path dependent options of this section, the spot fixings are often replaced by observations of particular forward or futures prices. Notice that floating strike lookback options are not really “options” at all as the worst that can happen is that the index value finishes at the minimum point (fora call) or the maximum point (for a put). Figure 5.2 illustrates a possible price path and the payoff to a floating strike lookback call option. The payoff is equal to the difference between the final spot price and the minimum spot price over the life of the option. Analytical formulae exist for the standard lookback options in a Black-Scholes- Merton world when the maximum or minimum is checked continuously, but numerical techniques must be used for discretely fixed examples. Lookback options can be of European or American style exercise. European style lookbacks for energy models are TABLE 5.3 Payoffs of Standard Lookback Options Option Name Option Payoff Fixed Strike Lookback Call max(0,max(S;..., Sy) ~ K) Fixed Strike Lookback Put max(0, K ~min(S),..., Sn)) Floating Strike Lookback Call max(0, Sp — min(S; Sm) Floating Strike Lookback Put max(0,max(S),..., Sn) ~ Sr) ® See Heynan and Kat (1997) for the uses, risk management, and numerical techniques for pricing lookback options in a BSM world. 85Energy Derivatives: Pricing and Risk Management Payoff Spot Price Minimum Time FIGURE 5.2 Example Payoff to a Floating Strike Lookback Call Option priced in section 7.2 via Monte Carlo simulation and American style options can be priced using the techniques of section 7.5. Lookback options tend to be less popular than the other exotics we have described so far because they allow the holder to effectively buy at the minimum or sell at the maximum price and so are relatively expensive. However, floating strike lookbacks do occur in a form in which the period over which the maximum or minimum is observed is a fairly short period e.g. one month. A type of lookback option has also appeared recently in electricity markets which exhibit sudden and relatively large price spikes. These are lookback caps which cap only a relatively small number of the highest prices in a given period. For example a lookback cap might have the same structure as the electricity cap described in section 5.4 but instead of capping every half hour electricity pool price over a certain cap level the product caps only the five highest prices in a given period. 5.8.4 Ladder and Cliquet Options Ladder and cliquet options are discrete level and time versions of lookback options respectively, Depending on the level of discreteness they can be much cheaper than lookback options and are therefore relatively more popular. Ladder options have a predefined set of levels Zz; k = 1,...,n such that if the underlying price crosses a particular level it locks in a minimum payoff equal to the 86Energy Derivatives: Structures and Applications Payoff Spot Price Time FIGURE 5.3 Example Payoff to a Floating Strike Ladder Call Option difference between the level crossed and the strike price. Similarly to lookback options there exist fixed strike and floating strike ladder options for both calls and puts. Figure 5.3 illustrates a possible price path and the payoff to a floating strike ladder call option. Unlike the analogous lookback option, the floating strike ladder option ‘locks in’ the lowest level L crossed and not the absolute minimum. The maturity payoffs to fixed and floating strike ladder calls and puts are shown in table 5.4. Ladder options can be replicated with a portfolio containing the standard European option with the same strike price and for each ladder level a pair of up-and-in options — for example, for the ladder call: a short put with a strike equal to the ladder level (or strike) below and a Jong put with a strike equal to the ladder level. Cliquet options have a predefined set of dates at which the underlying price is observed TABLE 5.4 Payoffs of Ladder Options Option Name Option Payoff Fixed Strike Ladder Call max(0, Sp — Ky (Ly — K) Langs us)>0)90 (En > K) Langs S)>05) Fixed Strike Ladder Put max(0,K ~ Sp, (K — Li) dings, 5y)ety7 9 (K — Ln) Maing \..54) r) of 1% the forward curve increases exponentially at the interest rate less the convenience yield ic. 4% per year, a convenience yield of 5% equal to the interest rate (6 =r) gives a flat forward curve and a high convenience yield (5 > r) of 10% gives an exponentially declining forward curve. The simple shape of these forward curves is clearly not realistic but the model can easily be generalised to allow the market forward prices to be used, we discuss this later in this chapter and also in Chapter 7. 160) Convenience Yield Feo 140 Weare Futures Prices (dollars) 3 8 80- 60. ar) i 2 3 4 5 6 7 8 9 10 Time to Maturity (years) FIGURE 6.1 Futures Prices from Cost-of-Carry for Typical Convenience Yields 90Spot Price Models and Pricing Standard Instruments A straightforward application of Ité’s lemma can be used to show that the volatilities of the futures prices are constant and equal to the volatility of the spot price, i.e.: orlt.s) =o (6.4) where o,(t,s) is the volatility of the s-maturity futures price. We have already seen that forward/futures price volatilities vary with maturity, typically declining relative to the spot price volatility for increasing maturities and also often exhibiting seasonal patterns. We show in the following sections that the model’s description of forward price volatilities can be improved by the introduction of mean reversion and time varying spot volatility. Option Pricing Black (1976) also developed a pricing formula for valuing European futures options. Let K represent the strike price of the option, the value of a European call option at time t, maturing at time 7 on a futures contract that matures at time s is given by: c(t, T,s) = P(t, T)[F(t,s)N(h) — KN(h — oVT = 8} _ ne 3)/K) +: +3e(T = —t) (6.5 ovT In the strict interpretation of the Black model, F(t,s) is the analytical forward price given by equation (6.3) with constant volatility 7 and P(t,7) represents the T-maturity discount factor. In practice this can be replaced by the market price of the forward/ futures contract — this corresponds to making the convenience yield a general function of time in order to make the spot price process consistent with the current market forward prices. This technique is very general and we show in this chapter and Chapter 7 how spot price models can usually be made consistent with the market forward curve in this way. The spot price volatility can also be made time varying to allow for the modelling of seasonal volatility patterns. We describe how this can be done for most spot price models beginning with the Schwartz model which is discussed in the next section. Figure 6.2 illustrates the effect of volatility on the price of a European call futures option. The option is assumed to have a maturity of 3 months with an at-the money strike and original futures price of 100. We plot the European call prices for volatilities of 10%, 25% and 50%. The Schwartz Single Factor Model As we have already noted, the constant volatility assumption of the Black model is not consistent with the empirical observation that longer dated forwards are less volatile then short dated forwards. A more realistic single factor model was introduced by Schwartz (1997) in which he assumes that the spot price follows a mean reverting type 91Energy Derivatives: Pricing and Risk Management European Calll Futures Option Price Seto rac eo eee erga 3 ‘60 70 80 90 100 110 120 130 140 Futures Prices FIGURE 6.2 Option Prices in the Black (1976) Model for Typical Volatilities process (see Chapter 2 for a discussion of simulating mean reverting spot price behaviour and Chapters 2 and 3 for a discussion of parameter estimation): dS = a(— —In $)Sdt + oSdz (6.6) where a is the mean reversion rate—the speed of adjustment of the spot price back towards its long term level j1, ¢ is the spot price volatility and A the market price of energy risk Defining x = In $ and applying Ité’s lemma to equation (6.6), the log price can be characterised by the Ornstein-Uhlenbeck proc dx = a(fi- x)dt + odz (6.7) where ji =p —A-S. This process leads to the following differential equation for all contingent claims whose payoff depends on the level of the energy spot price and time: 1 oS’Css +a(u~ A—In$)SCs+C,-rC =0 (6.8) 92Spot Price Models and Pricing Standard Instruments Futures and Forward Pricing With the appropriate boundary conditions, futures and forward prices with maturity s are equal and given by: | oe o—a(s—t) f er o(S~t} Pee eee Cece a — ge 2a(s-t) F(t,s) = exple InS+(l~e (1 x =) gat e | (6.9) Figure 6.3 shows the sensitivity of the futures price in equation (6.9) to varying the speed of mean reversion. The parameters used for the figure are S = 110, 0 = 0.3, and A = 0 with j fixed at In(100) — thus the current spot price is $10 above the long term level js. The mean reversion rate determines how soon, in terms of maturity, forward prices stabilise at their long maturity level. Notice that the long maturity level of the futures curve is not equal to exp(j:) but is adjusted by an amount which depends on the relative size of a and a: F(t,00) = exp [« =X - (6.10) 112 Mean Reversion Rate eee 07 110 108 1 10 106 3S g Future Prices 3 8 0 1 2 3 4 5 6 a 8 9 10 Time to Maturity (years) FIGURE 6.3 Futures Prices in the Schwartz Single Factor Model for Typical Mean Reversion RatesEnergy Derivatives: Pricing and Risk Management 0.35 Mean Reversion Rate 0.30 0.25 0.20 0.15 ity of Futures Prices 0.10 0.05 0.00! 0 1 2 3 4 5 6 7 8 9 10 Time to Maturity (years) FIGURE 6.4 Volatility of Futures Prices in the Schwartz Single Factor Model for Typical Mean Reversion Rates (7 = 0.3) Ité’s lemma can be applied to equation (6.9) to determine the term structure of proportional futures volatilities in this single factor model, yielding: or(t,s) =e C9 (6.11) Figure 6.4 graphs the effect that the speed of mean reversion has on the term structure of volatility of futures prices. Increasing the speed of mean reversion steepens the attenua- tion of the volatility curve. Equation (6.11) and figure 6.4 illustrate that as the maturity of the forward contract approaches large values the volatility of the contract tends to zero. This attenuation of volatility increases with increasing speed of mean reversion. The volatility structure under the Schwartz single factor model is more realistic than the Black model but still has quite a simple shape. In particular the volatilities tend to zero for longer maturities and this happens for maturities less than one year for mean reversion rates larger than about 7. Although the market volatilities of forward prices decline with maturity they never get close to zero and so the Schwartz model has a problem for pricing options on long maturity forward contracts. We will see in Chapter 8 how a volatility function of similar form is able to capture the attenuation of typical market forward volatility structures. 94Spot Price Models and Pricing Standard Instruments Option Pricing Clewlow and Strickland (1999) show that for the one factor Schwartz model, the price of a European call option maturing at time T on a futures contract that matures at time s is given by: c(t, T,s) = P(t, T)[F(t,5)N(h) — KN(h— Vw)] In(F(t,s)/K) + 4 = Tn In equation (6,12) w? is given by the integral of the forward return variance over the life of the option: h (6.12) ees faeasee wef oe 2a gy f 2 O ,-2als-T) can) =5-(¢ -e 2a (6.13) Ifs = T then the option on a future becomes an option on the spot energy and so the price of a T-maturity option on the spot price c(S, 1, 7) is given by: c(S,4,.T) = P(t, T)[F(t, T)N(h) — KN(h — VW)} (6.14) h In(F(t, T)/K) +3 . vw wa Z (1m) Ina similar way to the Black model the analytical forward price given by equation (6.9) can be replaced by the market price in the option pricing equation allowing the model to be consistent with the market observed forward prices. This is equivalent to allowing the mean reversion level to be a general function of time (see section 7.3 for more details). Furthermore, since the option price only depends on the integral of the forward return variance over the life of the option, the parameters a and ¢ can be made time dependent allowing the inclusion of seasonal patterns. The only complication this introduces is that the integral in equation (6.13) may now need to be evaluated numerically, however, the option pricing equation remains the same. In Chapter 8 we show how analytical volatility functions, such as equation (6.11), can be calibrated to historically estimated market volatilities or implied volatilities. The incorporation of the observable volati- lities, as well as forward prices, is very important in practice because this allows relatively simple models to price options as consistently as possible with the actual behaviour of the underlying market prices. ® Related results have appeared in Jamshidian (1991), Amin and Jarrow (1991, 1992) and Amin, Ng and Pirrong (1995). 95Energy Derivatives: Pricing and Risk Management 45 a —$—_$__—_—— 4o| Mean Reversion Rate a 07 35 4 +10 European Call Futures Option Price Oe agne ere ouae a 0! | 60 70 80 90 100 110 120 130 140 Futures Prices FIGURE 6.5 Option Prices in the Schwartz Single Factor Model for Typical Mean Reversion Rates Figure 6.5 shows how the speed of mean reversion parameter affects the prices of European futures call options. The option has a maturity of 3 months on a forward contract with an original maturity of | year. The strike price of the option is assumed to be 100, the interest rate is 5% and the volatility is 30%. The values for a used in the figure are 0.1, 1 and 10. Finally, for this section, we analyze the effect of the mean reverting process on option prices compared with the Black assumption of constant forward volatilities. Table 6.1 prices options with maturities of 9 months, 1 year, and 1.25 years ona forward with 1.5 years to maturity. The original forward price is assumed to equal 100, interest rates are 5% and the volatility of the spot asset is assumed to be 50% The Black’s prices are derived from equation (6.5) and the Schwartz prices from equation (6.12) with mean reversion parameter a which we vary from 0.01 to 0.25. Table 6.1 shows that as we reduce the mean reversion parameter the Schwartz prices approach the Black prices — a parameter of a = 0 would imply a flat term structure of volatility and the Schwartz prices equal to Black prices. 96Spot Price Models and Pricing Standard Instruments TABLE 6,1 Comparison of Option Prices in the Black (1976) and Schwartz (1997) Single Factor Models Black Schwartz, Vary mean reversion (a) T= 0.75 K 0.01 O1 0.25 95 11.849 11.747 10.885 9.644 100 9.589 9.482 8.577 7.268 10s 7.681 7.574 6.665 5.358 =1 ke : 0.01 Od 0.25 95 12.933 12.831 11.963 10.704 100 10.789 10.682 9.777 8.458 105 8.941 8.834 7919 6.593 T= 1.25 K 0.01 ol 025 95 13.798 13.701 12.872 11.668 100 11.754 11.652 10.791 9.536 los 9.967 9.864 8.992 1.725 6.3 Two Factor Models Some of the deficiencies of the single factor models, namely the very simple shape of the volatility structure and its attenuation to zero with increasing maturity can be remedied by introducing a second stochastic factor into the model. Pilipovic (1997) describes a two-factor mean-reverting model where spot prices revert to a long term equilibrium level which is itself a random variable’. Gibson and Schwartz (1990), Schwartz (1997) and Hilliard and Reis (1998) (HR) all analyze versions of the same two-factor model that allows for a stochastic convenience yield and permits a high level of analytical tractability. The first factor is the spot price process which is assumed to follow the GBM of equation (6.1). The second factor is the instantaneous convenience yield of the spot energy and is assumed to follow the mean reverting process: dd = a5(8 — d)dt + oydz5 (6.15) In equation (6.15) a, and 6 represent the speed of adjustment and long term mean of the convenience yield respectively with os the volatility of the convenience yield. The increments to the Brownian motions driving the spot price and convenience yield are > Pilipovie derives a closed-form solution for forward prices to her model when the spot and fong term prices are uncorrelated, but however doesn’t discuss option pricing in her two-factor framework. 97Energy Derivatives: Pricing and Risk Management assumed to have correlation coefficient pss. The joint process for the spot and convenience yield lead to the following differential equation for contingent claim prices: 1 22,1 2 : 5 CssS*o" + 3 Cyg05 + CspSpsgro5 + CyS(r — 8) + C5(arg(6 — 6) — Agog) + C, = 0 (6.16) where A; is the market price of risk per unit of convenience yield. Since the convenience yield is not a tradable variable, equation (6.16) depends on investor risk preferences which are embedded in the market price of convenience yield risk, A, Futures and Forward Pricing With interest rates assumed to be constant, futures and forward energy prices are equal Schwartz (1997) derives an expression for the futures price under the assumption that the interest rate term structure is flat and constant at a rate r. + A(t ») (6.17) — (541% _ T6bs8 1 gl—e tae) A(t, s) = (« ota a (s— a4 q% a s o2\ 1 —e-asls-9) + (0 + O05ps5 — a) ——_ re HR, assuming the same processes as Schwartz, derive the futures pricing formula consistent with the observed term structure of interest rates, given in terms of pure discount bond prices by P(t, 5): F(t,s) = Se Mune (6.18) ) Alt) = 6x (Helis) ~ (8 = (095 — apAsy — 05/2 + pseorog05) _ a5 Ho(t.8)° a 7 az 405 1 — e-aals-) Hd) 2 ee (ts) = If we set the initial term structure to be P(t,s) = e~"¢~") then it is easy to verify that the futures pricing formulas given by equations (6.9) and (6.18) are equivalent, Also if we set 0» = G5 = 0 then equations (6.9) and (6.18) reduce to the cost of carry relationship, 98Spot Price Models and Pricing Standard Instruments Mean Reversion Rate of Convenience Yield Futures Price 3 s 80 70: 65 1 2 3 4 5 6 7 8 9 10 Time to Maturity (years) FIGURE 6.6 Futures Prices for the Schwartz Two Factor Model Varying the Speed of Mean Reversion for the Convenience Yield equation (6.3). These results were originally obtained by Jamshidian and Fein (1990) and Bjerksund (1991). Figures 6.6 and 6.7 show the sensitivity of the futures pricing formula (6.18) to the speed of mean reversion of the convenience yield and the correlation between the spot price and convenience yield respectively. For both figures we assume S = 100 and the following parameters which are roughly typical of the values obtained in Schwartz (1997) for oil; r= 5%, 6=5% 5=7%, Ay =2.0, 7 = 30%, 05 = 30%, pss = 0.8, As = 0. For Figure 6.6 we have ay = 0.2, 0.5, 2.0 and for Figure 6.7 we have pss = 0.2, 0.5, 0.8. Figures 6.6 and 6.7 show that the two factor model has a reasonable level of flexibility in the possible shapes of the forward curve (see also Schwartz (1997)). However, the shapes are limited and in particular if the actual curve has more than a single hump or trough then these cannot in general be fitted by the analytical forward curve. In practice, as we have already noted, when pricing options it is not necessary to use the analytical forward curve, the market prices can be used directly. Application of Itd’s lemma to equations (6.17) and (6.18) shows that the volatility of futures contracts are given, in both the Schwartz and Hilliard-Reis framework’s by: ESS Se or(t,s) = Vo + oRHs (1,5) — 2oospseHe(t,s) (6.19) 99Energy Derivatives: Pricing and Risk Management 140- Mean Reversion Rate of Convenience Yield 130: 0.2 120. 110. Futures Price 2 3 3 8 2 8 70 60. 0 1 2 3 4 5 6 7 8 9 10 Time to Maturity (years) FIGURE 6.7 Futures Prices From Schwartz’s Two Factor Model Varying the Correlation Coefficient Tt can be seen from equation (6.19) that the volatility of the futures price depends on the volatilities of both the spot asset and the convenience yield, the correlation between the two, the speed of adjustment of the convenience yield, and the maturity of the futures contract. As time to maturity approaches large values, the volatility of futures returns converges to a fixed value, i.e.: 24% _ 2075Pss op(t,0o) = ie de 7 (6.20) O% Figure 6.8 shows the sensitivity of the futures volatility formula (6.19) to the correlation coefficient under assumptions for pss of 0.75 and 0.9. For both correlations the volatilities attain their long term level, given by equation (6.20) after about 3 years. This figure also shows that these two factor models have reasonably realistic volatility structures with attenuation for short maturities due to the mean reversion but which approach a non-zero level for longer maturities. Option Pricing In order to price European futures options for the two factor model we observe from equations (6.17) and (6.18) that the futures price is lognormally distributed and therefore 100Spot Price Models and Pricing Standard Instruments Mean Reversion Rate of Convenience Yield 0.75 2 0.9 Future Return Volatility 05 1 15 2 25 3 35 4 45 5 Time to Maturity (years) FIGURE 6.8 Futures Return Volatilities for the Schwartz Two Factor Model with Typical Correlations Between the Spot Price and Convenience Yield option prices are given by a Black type formula but with the appropriate integral of the futures volatility over the life of the option (cf. equations (6.5) and (6.12)). Clewlow and Strickland (1999b) shows that a European call option, maturing at time T, with strike price K on a futures contract which matures at time s is given by: c(t, Ts) = P(t, T)[F(4,8)N(h) — KN(A— Viv) (6.21) Fe In(F(t,s)/K) +41? _)- (eneols-1) — oe) x (ermen ae emt) a5 By comparison with (6.5), equation (6.21) can be seen as the Black formula with w? replacing o?(T — 1). The volatility term in the Black formula depends only on the 101Energy Derivatives: Pricing and Risk Management volatility of the spot price and the maturity of the option. However, equation (6.21) shows that the volatilities of both the spot price and convenience yield, the correlation coefficient, the speed of mean reversion of the convenience yield, as well as the maturities of both the option and the futures contract are important determinants of the option price. This formula can be further generalised by replacing the analytical forward prices by the market prices, this can be interpreted as allowing the mean reversion level of the convenience yield to be time dependent. The spot price volatility can also be made time dependent so that seasonal volatility can be modelled — this would require solving the ee of the futures return variance either analytically or numerically as in equation (6.13) The Schwartz 1 Factor Approximation Schwartz (1998) presents a single factor model for energy prices that retains most of the characteristics of his two factor stochastic convenience yield model but is based on the pricing and volatility results of the model represented by equation (6.6). As a conse- quence, the model only requires the numerical computation corresponding to a typical single-factor model. His analysis proceeds as follows: the rate of change in the futures price for the two factor model is independent of the initial values of the state variables’, i.e.: a < O8 pss coe =r-64+-4-S"54% 6.22 FO|es0 20h as (6.22) For the one factor model is determined as: st of carry model defined by equation (6.3) the rate of change (6.23) Therefore if we now define the convenience yield in equation (6.3) as: «9% | Pssaos § 28, 4 Pseoe. 6.24 2ai tap (6.24) the one-factor model will have the same rate of change in futures prices as the two-factor model. In order to match the futures prices from the two models’ Schwartz defines a shadow spot price dependent on the spot asset and convenience yield, Z(S, 6), to give the futures price in (6.3) when the constant convenience yield (6.24) is applied Z(S,6) = lim e'-O"- F(t, 5,8) = Sexp [= : 2] (6.28) 5 ne aah * The mean reversion rate and volatility of the convenience yield could also be made time dependent * The futures price in the two-factor model is given by (6.17). 102Spot Price Models and Pricing Standard Instruments 105 100! —> Long Term Model ‘actor Model = Tw 95, 90 85; 80 Futures Price 75; 70- 65 85 1 2 3 4 5 6 7 8 9 10 Time to Maturity (years) FIGURE 6.9 The Schwartz Two-Factor and Long Term Model Futures Prices The futures price, for the shadow spot price, is then given by: F(Z, 1,8) = Zel"-O9 (6.26) Figure 6.9 shows the long term model fitted to a two-factor model with the following parameters; S=100, r=5%, 6=5%, 6=7%, as =2.0, c= 30%, 75 = 30%, ps5 = 0.8, \= Figure 6.9 illustrates that, for this parameter set, the long-term model approximates the two-factor model well for maturities after around 1.5 years. As well as long term prices being very close, an application of It’s lemma to equation (6.26) shows that these futures price returns have exactly the same volatility as the two-factor model, ie. equation (6.19). This ad hoc adjustment should therefore produce reasonable prices for simple options on long dated contracts (such as a long dated bond with coupons linked to an energy price) but less well for options with maturities less than 2 years or options which depend on changes in the shape of the forward curve. We would not recommend using ad-hoc approximations such as this for managing portfolios of complex energy derivatives as they tend not to be robust to changing market conditions. Option Pricing in the Long Term Model In the long term model, the futures price is lognormally distributed and therefore option prices are given by a Black type formula with the appropriate integral of the futures 103Energy Derivatives: Pricing and Risk Management volatility over the life of the option, Since the futures volatility in the long term model is identical to the two factor models then the prices of standard European options are given by equation (6.21). 6.4 Three Factor Models Schwartz (1997) extends his two factor model of the previous section to include stochastic interest rates. In this three factor model the short term rate is assumed to follow the Vasicck (1977) mean reverting process. The full system of defining stochastic differential equations is therefore given by: dS = (r— 8)Sdt + oSdz d6 = a4(5 — 6)dt + osdz4 (6.27) dr = a,(F — r)dt + 0,dz, where the correlations between the processes are determined by: cov (dz,dz5) =psedt cov (dz,dz,) = ps,dt cov (dz5,dz,) = pydt HR also present a three-factor model in the spirit of Schwartz, but instead of the Vasicek assumption for interest rate evolution they assume that interest rates evolve in an Heath, Jarrow, and Morton (1992) (HJM) world (thus maintaining consistency with the observed yield curve). Under this model instantaneous forward rates are assumed to follow the process: df (t,8) = a(t,s) + Sr oli,s)dz (6.28) i= where f(t, 5) represents the observed instantaneous forward interest rate curve, a(t,s) is determined by no arbitrage conditions (see HJM) and o;(z,) are the associated volatility functions. Equation (6.28) is written in its most general form with n driving sources of randomness. The resulting partial differential equation for contingent claim prices for both Schwartz and HR, is given by: ay 1 7 5 CssS°0" +5 Cis08 +5 Cuo4 + CsoSpsyoos + CspSpsy00, + CoS por sort C5S(r ~ 6) + Cslas( — 6) ~ dvo5) + Cola, ~ A,0,) + C, = 0 (6.29) Tn the HR framework the risk neutralised drift for the interest rate process is substituted out by a term that does not depend on the market price of risk, A,, and so this equation is preference free with regard to interest rate risk. In the Schwartz framework My = 0,(7 1). Futures and Forward Pricing Because interest rates are now no longer assumed constant, futures prices will not equal forward prices. It can be shown that the forward price is independent of the short rate 104Spot Price Models and Pricing Standard Instruments and so is given for the 3 factor model by equation (6.17) for the Schwartz assumptions and equation (6.18) for the HR assumptions. Futures prices, however, are influenced by the short rate and in the Schwartz framework we have futures prices given by: ern) Hae este) ce : + Dolt, ) (6.30) F(t,s) = sexp(-3 oy ra a, a, 5 c where alts) = oeb+ oreess)(C _ o(A(L = eoule-) — = 3 4a; (a,? + oorps-)((1 =e) ~ a, (8 = 1)) a = (1 = e269) — 2a,(s — 2) 4a} (10 24-D) 4 1e-eett-0) (1 er lapt onde 0,0 +a,) _ of (4(L = eal MOP BY gz(t eesti) a3 Le) —a03 (8-1) asa (6—1) mee rae (Lee Nasa (sas ag (5-9) i 2787 (asta;) In order to derive explicit solutions for futures prices, HR restrict the general equation representing the interest rate evolution, (6.28), to a single factor version where volatilities of instantaneous forward rates are equivalent to the Vasicek (1977) model, i.e. n = 1 with oi(t,s) = 0,¢7-¢S—), Under this restriction futures prices are given by®: H(t 1 F(t,s) = SA(t,s)Dy(t,8)Do(t,s)D3(t, eM Tay (6.31) =e Af? spaor)_ at % 4o%s Dy(t,s) = exp ee Ge aan eo ex itt) a5 +0, a, Os Dy(t,s) = exp [222% (s ‘)- | Dales) =exp| (4, = (6-1) — ; ee — e-asle-) 6 = Ho(t,8) ae ae Honjo © This restriction of the HJM model is equivaient to the Hull and White (1990) model, often referred to as the “Extended Vasicek” model 105Energy Derivatives: Pricing and Risk Management Unlike the Schwartz and HR two factor models, the pricing formula (6.31) does not reduce to (6.30) if we assume that the observed term structure is given by P(t,s) =e"), This is because under the interest rate assumption in (6.27) the term structure of rates is given by the Vasicek (1977) formulation. Therefore, the prices given by (6.30) and (6.31) are equivalent if we assume that the observed term structure in (6.28) is the Vasicek term structure generated by the Vasicek short rate process in equation (6.27), Ité’s lemma applied to both the defining sets of equations shows that for both the Schwartz and HR models the volatility of the futures price returns is given by: [02 + o2Hy(t,s)? + oF H(t, 8° — 2oaspssHp(t, \) #2eeppsyFhy(t,8) ~ Loop Hslts)I (ts) or(ts) = (6.32) The volatility term structure therefore depends on the volatility and speed of mean reversion of the 3 random variables and their correlations, as in the two factor case. The volatility of futures returns converge to a fixed value as the maturity of the futures contract tends to infinity: as +842 ~ 22s 2s, Dost (6.33) ag O50, ‘F(t, 00) = \ Option Pricing Miltersen and Schwartz (1998) (MS) extend the 3 factor Schwartz model of this section to the same as that of HR by assuming that interest rates evolve according to the HJM model. MS derive prices for options on futures under general Gaussian assumptions for the interest rate process. Restricting the volatility function further to be the same as the single factor version in HR, MS derive the following closed form solution for pricing European call options on the futures price: elsT.2) = PT) ifr the ev(t Eus/Aysas)y) — ny (MELA 40) (6.34) with w=o(T -1) +20 orp (, ererls-A (eal? — 1) a, a, _ 26088 ( __ @a4ls-8) (eau(T-9) — ») Qs 5 106Spot Price Models and Pricing Standard Instruments +3 (4 2ay(s-1) (e2as(T~1) —1) i ~ag(s—1) (eae T—1) a ») cor 2as as 2 / ron s=i}(Q2o/T-) — J) gals (gaslT=0) gad (: i een rte ermcarecnia te 2) or 2a, a, — 2 257 rP or Apt, i e-a0l-9 (e207) — 1) e-r(s-) (eae(T=) — 1) 4 Sreetonte 9 (elartaT-) — 1) as ieee ay + ay where o — e-a(T-1) sot 1 at wg (ome) a era (ealT-9 Peal g-04(-N(eau(T—8 — g-a(T-1) t- 7 + Q, a, Q, 2a, ops; =m (eax(T—9 — 1) eAlT-)— g-ast—1) (eas(T—1) — e~a(T-1)) ee OROE (apes | eee + Although equation (6.34) looks very complex, comparing it with the previous option pricing equations; (6.5), (6.12) and (6.21), it can be seen to have the same ‘Black’ form but with a more complex expression for the underlying futures return variance over the life of the option. In practice stochastic interest rates have a relatively small impact on the prices of energy derivatives and interest rates can usually be assumed to be deterministic. These three factor models can therefore be viewed as a useful way of assessing the impact of stochastic interest rates without the need to fully implement the model in a risk management system. 6.5 Choosing a Spot Price Model The choice of which model to use depends on the nature of the energy markets and derivatives to which it is to be applied. We have shown that all of the spot price models can be made consistent with both the market forward prices and spot volatilities including any seasonal price or volatility patterns. The key distinguishing feature between the models is the shape of the forward/futures price volatility curve. As we have seen, the prices of options depend on the integral of the underlying forward price variance from its initial value to its value at the option maturity. Since the volatility curve is relatively flat for long maturities then for short maturity options on long maturity forward contracts the Black model could be used. 107Energy Derivatives: Pricing and Risk Management However for situations in which the integral of the forward price variance is over a non- constant portion of the curve a model which reflects the shape of the curve over the critical region should be used. Thus, for short maturity options on short maturity forwards the Schwartz one factor model can be used. For situations in which the whole volatility curve is important, for example for a large and diverse portfolio of energy contracts, the two factor stochastic convenience yield model is more appropriate. It is usually not necessary to use the three factor model as stochastic interest rates will typically have a relatively minor impact on energy derivatives prices’. The inclusion of jumps into these models leads to the loss of the simple analytical solutions for standard options and therefore the numerical techniques which we discussed in Chapter 2 and extend in Chapter 7 are needed. In their paper, HR claim to present a quasi-analytical solution for standard options under the three factor model with jumps. However, this solution appears flawed as it is not consistent with the attenuation of the jumps in the case of simple mean reversion. 6.6 Summary In this chapter we have presented a number of the most popular single and multi-factor spot energy models. For all of these models we have discussed the assumptions and presented the closed form solutions for forward and futures prices and volatilities for constant parameter values, as well as for standard European options, We have indicated how the models are related to each other and how key market data such as observable forward prices and seasonal volatility patterns can be introduced without losing the analytical results. However, for other market factors and for many energy derivatives traded in practice, numerical procedures similar to the techniques introduced in Chapter 1 are required to evaluate prices These techniques are extended to the spot price models of this chapter, in the following chapter. 7 A possible exception would be very long dated contracts. 1087 Spot Price Models: Pricing Path Dependent and American Style Options 7.1 Introduction In this chapter we discuss the pricing of path dependent and American style options for the spot price models presented in Chapter 6. For both these classes of instrument it is rare to find closed-form pricing solutions, and so numerical procedures must be used to obtain the prices and risk management parameters. In earlier chapters we noted that skews and smiles in implied volatilities are also important to take into account in the pricing of energy derivatives. The two main ways in which smiles can be introduced into models — stochastic volatility and jumps - typically result in a loss of analytical tractability for futures/forward prices and option prices and again derivative pricing is achieved via the implementation of numerical techniques. The general numerical procedures we discuss which allow the pricing of complex path dependent and American style derivatives (including contracts with a swing component) are Monte Carlo simulation and trinomial trees, including so called implied trees’. In describing these techniques we build on the framework introduced earlier in Chapter 1. We firstly describe the application of the numerical techniques to the models assuming constant parameters. We then describe how the parameters, in particular the spot price volatility, can be made time dependent in order to incorporate seasonal patterns. 7.2 Simulation Methods for Energy Spot Price Models Recall from Chapter I that the value of an option can be computed as the expectation of the discounted payoff? and that, under Monte Carlo simulation, the expectation is evaluated as the average of all outcomes obtained by simulating many paths, i.e.: Law = (7.1) a ' See Clewlow and Strickland (1998) for further details on the implementation of Monte Carlo simulation and trinomial trees. ? Where the expectation is taken under the appropriate measure. 109Energy Derivatives: Pricing and Risk Management where G; is the estimate of the value of the option with C,, the discounted payoff under the j simulation. In this section we illustrate the technique applied to both single factor and multi-factor models. 7.2.1 Single Factor Simulation We illustrate single factor simulation by using the Schwartz model for the energy price defined by equation (6.6) to value a European call option maturing at time T on an s- maturity forward contract (7 < s) with strike price K. Valuation of a derivative of this kind requires a two-stage process. Firstly, the underlying energy price is simulated to time 7, the maturity of the option. Secondly, the s-maturity futures price is computed at time T and compared with the exercise price of the option to obtain the payoff. In this model interest rates are constant and so the discounting becomes P(t,7) where today is time zero. As we saw in Chapter 6, the process for the logarithm of the spot price in the Schwartz model, x = In S, is given by’: dx = a(ji— x)dt + ode (7.2) Equation (7.2) can be discretised by changing the infinitesimals dx, dt, and dz into small changes Ax, Ar, and Az: Ax = ali x)Ar+oAz (73) The random increment Az has mean zero and a variance of Ar, and can therefore be simulated by random samples of VAve where ¢ is a sample from a standard normal distribution. In order to simulate the full path for the energy spot price we divide the time period over which we wish to simulate, say ¢ € [0,7], into N intervals such that Ar = $. We then generate values of S, at the end of these intervals 1; = #Az, i= 1,...,N, using equation (7.3) as follows: S, = exp(x;) Xp = Xi + (fi — 4) Att oV Bee; where S$; = S,, is the spot price observed at time ;. An important aspect of the spot price models discussed in Chapter 6 is that any required futures or forward price can be evaluated at the option maturity date, 7, as an analytical function of the spot price and parameters of the model. Thus for the single factor Schwartz model, at time 7, for each simulated path, we can compute the value of the s-maturity futures price, F(T,s), via equation (6.9) and then compute the payoff of the cali futures option as Cp = max(0, F(T,s) — K). To obtain the estimate of the call price we simply take the discounted average of these simulated payoffs: (7.4) = P(t,T) M Yo max(0, F(T, s) — K) (7.5) i where F;(T,s) represents the forward price under the j" simulation. * pis defined in equation (6,7). 110Spot Price Models: Pricing Path Dependent and American Style Options S a uw 2 6 what K Ts N 26.90 0.472 2.925 0.000 0.368 2.782 23.20 0.50 1.00 10 dt In(S) P(0,T) call_value SE M 0.05 3.292 0.951 1.616 0.091 1000 t oO 0.05 O41 0.15 0.2 0.25 0.3 0.35 04 0.45 05 & 0.708 0.574 -0.203 -0.006 -0.027 -0.013 0.367 0.629 0.434 -1.140 xt_ [3.292 3.336 3.373 3.342 3.328 3313 3.900 3.318 3.357 3.379 3.271 St_[F(rs)]_ ct 26.34 | 24.41 | 1.22 FIGURE 7.1 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Futures Option Figure 7.1 illustrates a single simulation. The spot energy price and the process parameters are chosen to match the observed crude oil forward curve on 31 March 2000. The spot price is $26.90 with a = 0.472, 4 = 2.925, and ¢ = 0.368. We assume = 0 and 80 ft = hat = 2.782. We price a 6 month (T = 0.5) at-the-money (strike price equal to current futures price implying F = K = 23.20) option on a futures contract with an original maturity of | year (s = 1.0), and assume 10 time steps (N = 10) so Az (dt) = 0.05. Interest rates are assumed to be constant at 10%. The simulation of a single path of the spot price is shown in the box, the first row shows the time, the second row shows the samples from the standard normal and the third line the simulated natural logarithm of the energy spot price at each time step. At the maturity of the option the spot price, represented by ST in figure 7.1, is recovered as 26.34 = exp(3.271), the forward price (F(T,s)) evaluated via equation (6.9), is 24.41 and the option maturity condition (CT), evaluated with a strike price of 23.20 is 1.22. The option price is determined by applying equation (7.5) with M simulations. We call this procedure the simple Monte Carlo method. Figure 7.2 shows the convergence of this simple method — we plot the option price evaluated via simulation for increasing number of simulations from 10 up to 1000. The analytical price of this option, calculated via equation (6.12), is 1.615 and is also shown in figure 7.2. The Monte Carlo estimate after 1000 simulations (call_value) is 1.616 with a standard error (SE) of 0.091. It is well known that in order to get an acceptably accurate estimate of the option price a very large number of simulations have to be performed. Most implementations of Monte Carlo simulation therefore employ variance reduction methods such as antithetic variance reduction, a technique consisting of creating a hypothetical asset which is perfectly negatively correlated with the original asset. Implementation of this technique is very simple and is described in detail in Clewlow and Strickland (1998). For example, consider pricing a European call futures option with strike K. At the same time as we simulate the energy spot price equation (7.4) we also simulate the antithetic variable 5: WiEnergy Derivatives: Pricing and Risk Management 22 ae ee eee ee eee eee — a 24 oa |---- Analytical : — Monte Carlo 19 —— 1.8 17 1.6 15: 14 13 12 1 1.0 0.9 0.8, 0 100-200 300 400 500s D.—s—“ia700'—«BOD-s«—“‘«‘éiSs«é1000 Simulations Option price FIGURE 7.2 Convergence of Simple Monte Carlo Simulation for a European Futures Option exp(X;) Ar+ oVBi(=«:) (7-6) In other words the antithetic path is obtained by replacing ¢; by —e; in the equation for the simulation of the asset where both processes are initialised to the same starting point, Xo = Xp =1n Sp. The simulated payoffs to the option under the two paths are then given for a European call futures option by: x 1 taj — Cry = max{0, F(T, 8, Sp) ~ K} (7.7) Ey; = max{0, F(T,s, 57) — K} (7.8) where F(T,,s, Sp) denotes the s-maturity futures price observed at time T when the spot energy price is trading at S;. We then take the average of the two payoffs as the payoff for that simulation. The antithetic method ensures that the mean of the normally distributed samples is exactly zero which also helps to improve the simulation*, Figure 7.3 illustrates a single simulation for valuing the same option as in figure 7.1 — the samples from the standard normal are now utilised to generate two sample paths (the 4 Note that, not only do we obtain a more accurate estimate from 7 pairs of (Cpj,Cr,) than from 2n evaluations of Cy, ,, but it is also computationally cheaper to generate the pair (Cy, j, Cy,,) than two instances of Cry 112Spot Price Models: Pricing Path Dependent and American Style Options s « B a ° phat K T s N 2690 0472 2925 9.000 0888 «2.782 2320 080 1.00 10 dt Inf) P(O.T) call_value SE ™ 005 3.292 0.981 1.628 0.082 1000 t a 005 04 ois 02025 OS —SS COC © 0.708 0574-0203 0.006 -0.027_ 0.013 0.967 0.629 0.434 1.140 xt [32928308 3.978 aa4e3.a9Bas1S 3900 _asi@ 3.357 3.878 —aa77 Mibar_| 3.292 3222 3.16s 3.172 31633156149 3.110 3.050 3.008 3.097 st_[ Fite) |_ or 2634 | 24.1 [1.20 2.12 [2127 | 000 FIGURE 7.3 Monte Carlo Simulation (with Antithetics) of Schwartz (1997) 1 Factor Model for a European Futures Option rows labeled ‘xt’ and ‘xtbar’). The payoff to the simulation is the average of the payoffs from the 2 paths, i.e. (1.22 + 0.00)/2. The Monte Carlo value is now 1.628 after 1000 simulations with a reduced standard error of 0.052. Figure 7.4 shows the convergence of the antithetic price with 10 to 1000 simulations, as well as the convergence of the simple method repeated from figure 7.2. The Monte Carlo estimate with antithetics has visibly better convergence properties. 22 24 Analytical 2.0 Antithetic Monte Carlo 19 imple Monte Car 18 17 16 15 14 13 12 1 1.0 0.9 08 Option Price 0 100 200 300 400 500 «G00. 700 800. 900.1000 Number of Simulations FIGURE 7.4 Convergence of Monte Carlo Simulation With Antithetics for European Futures Option 13Energy Derivatives: Pricing and Risk Management s o n a o uhat kK Teeeete sa 26.90 0.472 2.925 0,000 0.368 2.782 204 050 075 150 10 dt In(S)_ P(0.T) call_value SE uM 0.05 3.292 0.951 0.676 0.041 1000 t 0 005 04 0145 O02 025 03 035 04 045 O05 € 0.708 0.574 -0203 -0.006 -0.027 -0.013 0.367 0.629 0.434 -1.140 xt_ [3.292 3.338 3.973 3.342 3.928 3313 3.500 3.318 3357 3.379 3.271 ST_|F(tst)[F(ts2)|_ CT 26.34 | 25.32 | 22.89 | 0.39 FIGURE 7.5 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Calendar Spread Option 7.2.2. Example —Valuing Spread Options The simple Monte Carlo technique from the previous section can be easily extended to model more complicated payoff conditions, For example, for the calendar spread option discussed in section 5.7.1, the user evaluates the price of the two futures contracts at the maturity of the option. Figure 7.5 illustrates a single simulation path for such as option. We assume the same process parameters as in figure 7.1 for the Schwartz model and value a 6 month option (J = 0.5) on the spread between futures contacts with original maturities of 9 months (s; = 0.75) and 1.5 years (s) = 1.50). At the maturity of the option the spot price of 26.34 implies forward prices of 25.32 and 22.89 for the forward contracts with 3 months and | year remaining maturities respectively via equation (6.9). The option payoff of 0.39 is obtained after the evaluation of equation (5.9) assuming a strike price of 2.04. The Monte Carlo estimate after 1000 simulations is 0.676 with a standard error of 0.041. 7.2.3 Example —Valuing European Energy Swaptions European swaptions, discussed in section 5.5, can be valued in a similar way to the spread option in the previous section. The payoff of a swaption (see equation (5.5) for a cash settled option) requires the evaluation of m forward prices at the maturity of the option in order to evaluate the swap value. Figure 7.6 shows a single simulation to price a6 month European oil payer swaption which exercises into a 2 year semi-annual swap. The strike price of the option is assumed to be $22.00. In figure 7.6, si = T + i AT; i= 1,... 4, denote the original forward maturity dates that underlie the swap reset dates. At 1 = 0.5 the spot energy price implies forward prices at remaining maturities 0.5, 1.0, 1.5, and 2.0 of 24.41, 22.89, 21.70, and 20.76 respectively. This simulated path finishes in-the- money with an evaluated payoff of CT = 0.44. The Monte Carlo estimate after 1000 simulations is 0.922 with a standard error of 0.056. 114Spot Price Models: Pricing Path Dependent and American Style Options s o ow a shat kK re AT; N 26.90 0.472 2.925 0.000 0.368 2,782 22.00 0.50 0.50 10 at In(S)_P(O,T) call_value SE M 0.05 3.292 0.951 0.922 0,056 1000 t 0 005 Of O15 O02 025 03 035 04 0.45 O05 € 0.708 0.574 -0.203 -0,006 -0,027 -0.013 0,367 0.629 0.434 -1.140 xt_ [3292 3.938 3.373 3.342 3.928 3313 9.300 9318 3.357 3370 9271 ST_|_F(1s1) [F(T,s2)[F(T,s3)[F(T.s4)[_ CT 26.34 | 24.41 | 22.89 | 21.70 | 20.76 | 0.44 FIGURE 7.6 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Energy Swaption 7.2.4 Incorporating Seasonality into Monte Carlo Simulations One of the most important variables which exhibits seasonal patterns is volatility. In Chapter 2 we described how both a simple seasonal pattern or random volatility can be incorporated into spot price models. Random volatility is outside the scope of the models discussed in this chapter but simple seasonally varying spot volatility can easily be incorporated into the models whilst still retaining analytical tractability. In this section we show how this can be done within Monte Carlo simulation of the single factor Schwartz model, but the technique is generally applicable. Generalising the Schwartz single factor model represented by equation (7.2) to time varying spot volatility we have: dx = offi — x)dt + o(t}dz (7.9) The estimation of seasonal volatility patterns is discussed in Chapters 3 and 8, Figure 7.7 illustrates a typical seasonal volatility pattern obtained from Henry Hub natural gas spot prices from 1 September 1999 to I March 2000. Futures and forward prices are now given by the following equation which also relates forward prices at time ¢ to the initial forward curve at time 0: aga) FUT) =F0,1)( ) : ' re (Tu) : u exw|-5 [ Se Tau 4 bern [ ena (7.10) In general the integrals in equation (7.10) may have to be evaluated numerically but the payoff to any option which depends on forward or futures prices can be evaluated at the option maturity date as a function of the simulated spot price and parameters of the model. WsEnergy Derivatives: Pricing and Risk Management 0.8 0.7 0.6 0.5: Spot Volatility 04 0.3 0.2. 0 0.05 04 015 0.2 0.25 0.30 Time (years) 0.35 04 0.45 0.5 FIGURE 7.7 Typical Seasonal Energy Spot Price Volatility Pattern for Henry Hub Natural Gas s o M a uhat K T s N 26.90 0.472 2.925 0,000 2.782 23,20 0.50 1.00 10 dt In(S)_P(0,T) call_value SE M 0.05 3.292 0.951 2.955 0.169 1000 t 0 005 01 015 02 025 03 035 04 045 05 oft) | 0.506 0.633 0.623 0.669 0.689 0.708 0.606 0.479 0.478 0.472 0.322 e 0,708 0.574 -0.203 -0.006 -0.027 -0.013 0.367 0.629 0.434 -1.140 xt_ | 3202 9.361 3.428 3985 9.370 3.952 3336 3373 3.426 3.457 3.921 ST _[F(ts)| CT 27.69 | 25.40 | 2.20 FIGURE 7.8 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Futures Option with Seasonal Volatility Figure 7.8 shows the same simulation as in figure 7.1 but with the seasonally varying volatility pattern shown in figure 7.7. The simulated (log) spot path is now updated according to the time dependent spot volatility, rather than the constant value used previously. 116Spot Price Models: Pricing Path Dependent and American Style Options 7.2.5 Computation of Hedge Sensitivities The standard hedge sensitivities, delta, gamma, vega, theta and rho can be computed via Monte Carlo simulation by approximating them by finite difference ratios; aC _ C(S + AS) - C(S— AS) iy yc (7.11) a Z IS + AS) = 2c(8) +C(S- AS) (3) vega = OC Clot Boh = Clo ~ Bo) (7.13) theta = 9E xe SEES CO) (7.14) rho = i ee Cet) ai ed) (7.15) where C(S + AS) is the Monte Carlo estimate using an initial energy price of S+ AS, and AS is a small fraction of S e.g. AS = 0.001 S. The other C(.)’s are defined similarly. It is important that every price C(.) should be computed using the same set of random numbers>. In Chapter 9 we show how knowledge of these risk measures can be used to hedge the risk of energy derivative positions. 7.2.6 Simulation of Multi-Factor Models One of the main uses of Monte Carlo simulation by industry professionals is for pricing options under multiple stochastic factors. In this section we show how the Schwartz two factor model of section (6.3) can be implemented via Monte Carlo simulation. Recall that the joint stochastic process for this model is given by: dS = (r ~ 6)Sdt + oSdz (7.16) d6 = a6(6 — 8)dt + o5dz5 (7.17) where the Brownian motions dz and dz; have instantaneous correlation psg. The Monte Carlo procedure is exactly the same as that for the standard European call futures option described in the previous sections except that we simulate the joint processes given above which determine the payoff of the futures option. Equations (7.16) and (7.17) are discretised as follows: AS = (r—8)SAt4 oSAz (7.18) 5 ‘The interested reader is referred to Chapter 4 of Clewlow and Strickland (1998) where the calculation of hedge sensitivities is described in detail. 117Energy Derivatives: Pricing and Risk Management A6 = a4(5 — d)At+ opAzs (7.19) In order to price the option by simulation we need to generate random samples of the variates Az and Az; from a standard bivariate normal distribution with correlation pss This is easily achieved by generating independent standard normal variates ¢, and ¢) and combining them as follows®: Az=e,VAt = Az = (pes t+ 1 — Pee: As we have seen, the best way to simulate a variable following Geometric Brownian Motion is via the process for the natural logarithm of the variable which follows arithmetic Brownian motion. Therefore we have: Xigt = xi + (1 — 6 — 40°) At + ce, VAL (7.20) ea) VB (7.21) Sinn = 6) + 5(6 — AT + 5 (re + Figure 7.9 illustrates a sample simulation for crude oil prices. The process parameters are chosen to match the observed crude oil forward curve on 31 March 2000 resulting in a spot price of $26.90, convenience yield of 26.6% and o = 36.8%. The convenience yield mean reverts back to a level (Sbar) of 17.1% with a speed (a6) of 0.363, it has a volatility (a8) of Sacer 8 6 =a Bhar S_—pSB kK T Os 26.90 0.10 0.266 0.368 0.363 0.171 0.301 0.221 23.15 05 1.00 dt In(S) N M 0.05 3.292 10 1000 call_value SE 2.251 0.139 t OQ 005 01 015 02 0.25 03 0.35 0.4 0.45 05 el 0691 -0.100 0.042 0211 -0535 -0.156 0940 0301 -0.522 2.464 22 0.879 2.244 -0.233 -0.727 0.081 -1.554 -0.860 -0.788-0.714 1.439 xt_[3.292 3.337 3.314 3.205 3.202 3.229 3.198 3.262 3.276 3.225 3.423 6 [0.266 0.332 0.475 0.455 0.405 0.398 0.290 0.245 0.197 0.142 (0.084 st F(T,s) | CT 30.65 | 30.74 | 7.59 FIGURE 7.9 Monte Carlo Simulation of Schwartz (1997) 2 Factor Model for a European Futures Option © ‘The general procedure for generating n correlated normal variates is described in Clewlow and Strickland (1998) where they use the eigensystem representation of the covariance matrix of the correlated variables. 118Spot Price Models: Pricing Path Dependent and American Style Options 30.1% and a correlation (pS6) of 0.221 with the spot energy price. Interest rates are assumed to be flat at 10%. The simulation is again set up to price a 6 month at-the-money European call option ona futures contract with an original maturity of | year with 10 time steps. The spot price path simulation is contained in the boxed region. The first line contains the time, the second two lines contain the independent samples from a standard normal distribution, the third and fourth lines contain the simulated paths for the natural log of the spot price and the convenience yield respectively. The simulated value (call_value) after 1000 simulations is 2.251 with a standard error (SE) of 0.139, The analytical value obtained via equation (6.21) is 2.346. Notice how the price is approximately 50% greater than for the one factor model because of the additional volatility of the convenience yield which leads to the forward curve having a greater overall volatility. In this example we have not fitted the models analytical forward price volatility structure (equation (6.19)) to the actual market volatilities — this is discussed in Chapter 8. 7.2.7 Generation of the Random Numbers A critical part of Monte Carlo simulation is the generation of the standard normal random variables’. Clewlow and Strickland (1998) provide a discussion of the various ways to do this including the sum of twelve standard uniform random numbers, the Box-Miiller transformation, and the polar rejection method. They conclude that the polar rejection method is both faster and more accurate. The method can be stated algorithmically as follows: repeat x, = 2 x standard_ uniform. random_ number — 1 Xx, = 2 x standard_ uniform. random_ number — 1 w= x43 until w > 1 (7.22) { In(w) Vaca 2) = ex, c 2p = €Xa Note that the generation of x; and x, involves independent sampling of uniform random numbers. The polar rejection method generates two standard normal random variables at a time. Therefore in order to obtain maximum efficiency the second random number must be saved for use the next time a standard normal random number is required rather than generating two random numbers every time and throwing one away 7 The generation of the random numbers is typically about 30% of the total execution time of a Monte Carlo. simulation 119Energy Derivatives: Pricing and Risk Management Another way to make the simulations more efficient (variance reduction) is to implement quasi random numbers (also known as low discrepancy sequences). There are many ways of producing quasi random numbers and the interested reader should refer to Chapter 4 of Clewlow and Strickland (1998), Joy, Boyle, and Tan (1996), Niederreiter (1992), and Niederreiter and Shiue (1995) for examples. 7.2.8 Valuing Path Dependent Options Using Simulation In Chapter 5 we discussed many so-called path dependent options - options whose payoff depends on the actual path that the energy price takes over the life of the option. In this section we look at pricing a number of popular types of path dependent option, namely barriers, lookbacks and Asians using Monte Carlo simulation. Barrier Options Recall from Chapter 5 that there are eight basic types of barrier options: down-and-out, up-and-out, down-and-in, and up-and-in versions of both puts and calls. As the valuation principals are consistent across all eight types of options we limit our discussion here to pricing down-and-out calls on the forward energy price. We value a European down and out call option, maturing at time T, which expires worthless if the underlying forward price, F(,s), crosses the barrier on certain predefined reset dates. The payoff can be represented by*: max(0, F(T, 5) — K)lmin(F(4 )..usF (ims))># (7.23) Figure 7.10 shows a sample simulation, We value a down-and-out call option with the S a pw 2% 6 ghat K Te dN 26.90 0.472 2925 0.000 0.368 2.782 23.20 050 1.00 22.04 10 dt In(S) P(0,T) F(0,s) call_value SE M 0.05 3.292 0,951 23.20 1.264 0.083 1000 t o 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 05 & 0.417 -0.877 2114 -0.969 -0.920 1.609 -0.644 -2190 1.049 0.799 xt_ [3.202 3314 3.230 3.393 3299 3.211 3.333 3.267 3.075 3.155 3.212 F(tis) [23.20 23.68 22.55 2530 23.91 22.64 2485 23.86 20.78 22.17 23.30 CT 0.00 FIGURE 7.10 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Barrier Option * ‘The forward price which is checked against the barrier may be different to the forward price determining the payoff to the option. 120Spot Price Models: Pricing Path Dependent and American Style Options same process characteristics as the option of figure 7.1 but with a knock-out barrier contingent on the 1 year forward price. The | year forward price is observed® to be $23.20 and the barrier is set to be 5% below the current forward underlying price (H = $22.04). The option is designed to knock out if the underlying forward crosses the barrier. Although the simulated path finishes in-the-money for a standard call option, the forward price at time 0.4 trades at a level of $20.78 which is below the barrier level and so the option expires worthless. Simulating 1000 payoffs yields a Monte Carlo estimate of the price of 1.264 with a standard error of 0.083 for this particular option. Note that the value of the option is significantly less than the standard option priced in figure 7.1 because of the possibility of the option knocking out. Lookback Options In Chapter 5 it was shown that the terminal payoff to a lookback option is a function of the highest or lowest underlying energy price during the lifetime of the option and that there are four main types which are summarised in table 5.3. As an example we take a floating strike lookback call option with maturity T = ,, where the payoff is determined by the minimum level of the underlying forward price, F(s,s), over the life of the option'®: max(0, F(T, s) — min(F(t),5), 5 F(lm,5))) (7.24) Figure 7.11 shows a sample simulation for pricing a 6 month floating strike lookback call option where the option is fixed on the levels of the forward contract with | year to maturity at the initial time. The process parameters and fixings are as described in figure 7.1 s o H a o hat T s N 26.90 0.472 2.925 0,000 0.368 2.782 0.50 1.00 10 dt In(S)_P(0,T)_F(0,s) call_value SE M 0.05 3.292 0.951 23.20 2.495 0.089 1000 t 0 005 O01 015 02 025 03 035 04 045 05 e “0.200 -0.519 0.503 1.427 1.359 0.164 -1.619 -0.993 0.266 0.246 xt [3.202 3.264 3210 3.241 3.348 3.222 3.225 3.082 2.993 3.010 3.025 F(ts) [23.20 22.92 22.26 22.85 24.72 22.82 23.00 20.81 19.53 19.82 20.10 ct 0.57 FIGURE 7.11 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Lookback Option ° ‘The forward price has been calculated via equation (6.9). 1° ‘The minimum could also be determined over a partial period of the option life. 121Energy Derivatives: Pricing and Risk Management The minimum price of the forward contract occurs at time ¢ = 0.4 under the simulation in figure 7.11 at 19.53. The payoff under the simulation is therefore equal to max(0, 20.10—19.53) = 0.57, With 1000 simulations the Monte Carlo estimate of the price is 2.495 with a standard error of 0.089. The value of this lookback is much greater than the equivalent standard call option valued in figure 7.1 because this option effectively allows the holder to buy at the minmum price. Asian Options Finally in this section we look at options based on the average of energy prices - Asian options. As an example we price a European fixed strike Asian call option with maturity T, where the averaging is applied to the price of a forward contract observed 10 times during the life of the option''. Recall from Chapter 5 that the payoff for the option is determined as the difference between the average of the observed forward energy prices at the fixing times ¢, and the strike price K: 1. max(05, 2 Fes) —K) (7.25) Figure 7.12 shows a sample simulation for the pricing of such an option. We value a 6 month option on the price of a forward contract with an initial maturity of 1 year. The average of the forward prices for the simulation in figure 7.12 is 24.60. The payoff to the Asian option under this simulation is therefore max(0, 24.60 — 23.20) = 1.40. Running 1000 simulations gives a Monte Carlo estimate of 0,891 with a standard error of 0,048. The Asian option is worth approximately half of the equivalent standard call option in s o hn a chat kK a s N 26.90 0.472 2925 0.000 0.368 2.782 23.20 0.50 1.00 40 dt In(S)_ P(0,T)F(0,s) call_value SE M 0.05 3.292 0.951 23.20 0.891 0.048 4000 t 0 005 O01 O15 02 025 O03 035 04 O45 O05 & 0.708 0.574 -0.203 -0.006 -0.027 -0.013 0.367 0.629 0.434 _-1.140 xt_ [3.292 3.338 3.373 3.342 3.328 3.313 3.300 3.318 3357 3.379 3.271 Fits) [23.199 24.045 24.762 24.448 24.398 24.319 24.256 24.755 25.688 26.951 24.414 cT 1.40 FIGURE 7.12 Monte Carlo Simulation of Schwartz (1997) 1 Factor Model for a European Asian Call Option "Partial averaging periods are also popular 122Spot Price Models: Pricing Path Dependent and American Style Options figure 7.1 because the uncertainty in the average is much less than that of the underlying forward price at the option maturity 7.3 Building Trinomial Trees For the Energy Spot Process In this section we describe a general and efficient procedure involving the use of trinomial trees for modelling the spot energy price so that it is consistent with initial futures market data’? and allows the pricing of American and path-dependent options. This method can also be easily extended to incorporate seasonal volatility and/or volatility smiles and skews, extensive details are given in Chapter 5 of Clewlow and Strickland (1998)'°, The ‘fitting’ of the model to the market prices of futures/forwards is achieved by introducing a time dependent mean reversion level into the drift of the spot price. Clewlow and Strickland (1999a) (CS) develop a procedure for building trinomial trees to achieve this consistency between the model and the market. These trees can then be used for pricing American options as well as options with path dependent character- istics. American option valuation requires evaluation of the following expression: 8 C() = Max E, [exn(- | Hud) co) (7.26) BEG [1,7] t where C*(6) is the payoff of the option when it is exercised at date 6 and W[r, T] is the class of all early exercise strategies (stopping times) in [1,7]. The early exercise strategy, and hence the option price, can be easily determined from the tree for the spot energy price. Defining x as the natural logarithm of the energy spot price, x(¢) = In(S(t)), CS show that to achieve consistency with the observed forward curve for the Schwartz single factor model the process for x is given by: dex(t) = [6(0) — ax()]dt + odz(8) (7.27) (0) = ee +aln F(0,1) +20 — et) 3”) The addition of a time dependent parameter in the drift term allows the model to return the market futures prices. The CS tree building procedure consists of two stages. First, a preliminary tree is built for x ming that 4(f) = 0 and the initial value of x is zero. The resulting ‘simplified’ process for this new variable, x, is given by: dx(t) = —aX(t\dt-+ odz(1) (7.28) The time values considered by the tree are equally spaced and have the form iA‘ where i © This section is based on the paper “Valuing Energy Options in a One Factor Model Fitted to Forward Prices” by Clewlow and Strickland (1999a). See also Clewlow and Strickland (19994). "3 Trees can also be extended to incorporate jumps, both in the spot price and volatility, but this is beyond the computational scope of this book. The interested reader is referred to Amin (1994) and Naik (1994) 123Energy Derivatives: Pricing and Risk Management is a non-negative integer. A reasonable choice for the relationship between the space step and the time step is given by: Ax =ov3Ar (7.29) Assume that the value of x at the j-th node at time f; is X;,; where j determines the level of the variable in the tree. The trinomial branching process and the associated probabilities are chosen to be consistent with the drift and volatility of the process (7.28). The three nodes which can be reached by the branches emanating from node (i, ) are (i+ 1, k~1), (é+1, &), and (i+ 1, k-+ 1) where k is chosen so that the value of ¥ reached by the middle branch is as close as possible to the expected value of x at time ¢,,, . The expected value of ¥ at time 1;,; conditional on ¥ = ¥;; is X,; — ak, ;At. Let PujijsPmij> and py; define the probabilities associated with the lower, middle, and upper branches emanating from node (i, /) respectively. These probabilities are given by: 1 [= +x A ex, At — (1+ 2(k J) 4 «| 2 Pasi =3 {Rae FI AS 2 , ax, At +e +249) 40-9] (7.0) oA +023, AP Ax? 1 Paij = 3 Pmi.j = 1 ~ Puij ~ Pai The procedure described so far applies to the process X¥ with 6(t) = 0 and X = 0. The second stage in the tree building procedure consists of displacing the nodes in the first tree in order to add the proper drift and to be consistent with the observed term structure. We can introduce the correct time varying drift, by displacing the nodes at time iAt by an amount a; which are chosen to ensure that the tree correctly returns the observed forward price curve. The value of x at node (i, /) in the new tree equals the value of ¥ at the corresponding node in the original tree plus a;; the probabilities remain unchanged. The key to this stage is the use of forward induction to ensure that all forward prices are priced correctly. In order to determine these time dependent functions we can use pure security (or state or Arrow-Debreu) prices. Defining Q, as the value, at time 0, of a security that pays $1 if node (i,/) is reached, or $0 otherwise, we see that Arrow-Debreu securities are the building blocks of all securities; in particular the price today, C(0) of any European claim with payoff function C(S) at time step i in the tree is given by: (0) = D7, QijCCSi,) (7.31) where the summation takes place across all of the nodes / at time i. The state prices are accumulated according to the relationship: Oty = SO. p77 PUA, (i+ 1)Ad) (7.32) 7 124Spot Price Models: Pricing Path Dependent and American Style Options 28 = = 38 | -=-01131/03/00 —+-Oil 01/03/99 ~Gas 31/03/00 26 y & zs 8 22+ 32 é a B20 38 ° = a Sis 286 ° 6 a p o = » x 8 p iy pee 2 gy eae eae eee 12 13°14 15 16 17 18 19 20 21 22 23 24 Months to Maturity FIGURE 7.13 Typical Energy Market Forward Curves where pj; is the probability of moving from node (i, j") to node (i+1, j) and P(iAt, (i+ 1)At) denotes the price at time /Ar of the pure discount bond maturing at time (i+ 1)Avz. In order to use the state prices to match the forward curve we use the following special case of equation (7.31): P(0,iA0)F(0,iA0) =Laus a (7.33) CS show that the adjustment term needed to ensure that the tree correctly returns the observed futures curve is given explicitly as: iia (2euaneta ia) (734) To illustrate the procedure we have fitted the spot price tree to a number of different market forward curves. Figure 7.13 shows three market curves that are representative of; a downward sloping forward price curve (NYMEX Light, Sweet Crude Oil Futures Contracts, 31 March 2000), an upward sloping curve (NYMEX Light, Sweet Crude Oil Futures Contracts, 31 March 1999), and a forward curve which exhibits seasonality (NYMEX Henry Hub Natural Gas Futures Contracts, 31 March 2000). Two years worth of monthly maturity contracts are used to construct the curves. Figure 7.14 shows the resulting trees with time steps every two months. 125Energy Derivatives: Pricing and Risk Management SAMMI [SEO ENN Shae [SANS | SoH SIM: SONI | Bebe PH: | RPAH: | Eh See. | SRN | oil CSM SM | BB DMM | MRE SSA VA NGA St call. HS | SR Nan meen & 7 Rt £88 8 BS & BSRBRSRRS° NS Tene s (8) eoug ods, (g) e0ud 1ods, (8) eo yods, Maturity (years) 04 0.6 02 FIGURE 7.14 Spot Price Trees Fitted to Market Forward Curves (Downward sloping, Upward Sloping, and Seasonal)Spot Price Models: Pricing Path Dependent and American Style Options TABLE 7.1 Value of European and American Options Calculated from the Tree Options on Spot Options on Future Steps/ Euro Euro Amer Amer Euro Euro Amer Amer Year Call Put Call Put Call Put Call Put 50 2482 2.482 4.320 2.596 1408 2.484 1.429 (2,542 100 2.470 2.470 4.322 2.586 1.413 2.490 1.435 2.547 150 2465 2.465 4.320 2.580 140624831428 (2.540 200 2461-2461 4.319.577 1409 24861431 2.543 250 2.463 2.463 4.321 2,579 L410 2.487 1,432 2.544 300 2464 2.464 4.322 2.580 140924851431 2.543 350 2.465 2.465 4.321 2,581 1.406 2.483 1.429 2.541 400 2.463 2.463 4.321 2.581 1.408 2.484 1.429 2.541 Analytical 2.463 2.463 N/A N/A 1.408 2.484 N/A N/A The volatility parameters used in the tree construction were chosen by best fitting, in a least squares sense, the negative exponential forward price volatility function to sample standard deviations of one years worth of historical daily futures returns. The resulting parameters for the speed of mean reversion and spot price volatility were 0.472 and 0.368 respectively for crude oil, and 1.436 and 0.468 for the gas data. Table 7.1 shows the results of pricing a one year at-the-money (forward) option on crude oil. The tree was constructed to fit the downward sloping forward curve of crude oil on the 31 March 2000 from figure 7.13. Prices for European and American exercise style options on the spot and a 1.5 year forward contract are determined from the tree for different numbers of time steps. The volatility parameters used in the tree construction were chosen by a best fit to sample standard deviations for one year of historical data prior to 31 March 2000. Interest rates are assumed to be 10% We also compare the prices of European options calculated from the tree with the analytical values calculated via equations (6.12) and (6.14). Table 7.1 illustrates that prices calculated from the tree converge to the analytical price. It can also be seen from table 7.1 that there is an early exercise premium associated with both options on the spot price and on the forward price due to the fact that the downward sloping forward curve implies a significant convenience yield on the spot asset. The nature of the construction of the tree implies that hedge parameters can be quickly and easily calculated. If we calculate hedge parameters with respect to some ‘shift’ in the forward curve, then this shift only affects the displacement coefficients — it doesn’t effect the position of the branches relative to the central branch or the probabilities associated with the branches. 7.4 Implied Trinomial Trees In a number of earlier chapters we have discussed the volatility smiles commonly found. in traded options prices, which contain important information about the markets 127Energy Derivatives: Pricing and Risk Management expectation of the future. A key issue for many options traders is how to incorporate this volatility information into numerical schemes in order to be able to value exotic options (the subject of the next section) consistent with the information on the smile embedded in standard option prices. One technique is to build implied trees. Implied trees can be considered as generalisations of the trinomia] trees we have so far described. This generalization can be achieved by making previously constant parameters (such as the probabilities) time dependent and to imply these time dependent parameters such that the tree returns the market prices of the standard options — in this way the structure of the implied trinomial tree will be very similar to that of the constant coefficient trinomial tree. A detailed description of the construction of trinomial trees is beyond the scope of this chapter, but the interested reader is referred to Chapter 5 of Clewlow and Strickland (1998) where implied trees are discussed at great length. 7.5 Pricing General Path Dependent Energy Options in Spot Price Trees'* Having constructed trinomial trees for the spot energy process in the previous section, we outline in this section how to price general path dependent options using the techniques developed in Hull and White (1993) (HW) for a Black-Schole-Merton world and extended by Clewlow and Strickland (1997) (CS) which contains further details and examples of this technique. Assume we wish to price a general path dependent option whose payoff depends on some function G(F(z,s);0 <1 < T,t <5) of the path of the forward price curve. The procedure developed in HW and CS follows a number of steps. Firstly, the user determines the range (ie. the minimum and maximum) of the possible values of G(.) which can occur for every node in the tree. This is achieved by stepping forward through the tree from the origin to the maturity date computing, at each node, the minimum and maximum value of G(.) given the value at the nodes at the previous time step which have branches to the current node and the forward curve at the current node. Secondly, we choose an appropriate set of values of G(.) between the minimum and maximum possible for cach node. In choosing this set of values we note that the nodes which lie on the upper and lower edges of the tree have only one path which reaches them and therefore there can be only one value of G(.). The largest range of values will typically occur in the central section of the tree. The number of values we consider should in general increase only linearly with the number of time steps and also decrease linearly from the central nodes of the tree down to one at the edges of the tree in order to control the computational requirements, Let n;; be the number of values we store at node (i,j) and G;,jx, k = 1... ., my be the values of G(.) where G;,,. is the minimum and G,,jn,, i8 the maximum. Clewlow and Strickland (1998) suggest choosing nj; to be nj; =1 + li ~ abs(j)) (7.33) 14 ‘This section is based on section 5 of Clewlow and Strickland (1999a). 128Spot Price Models: Pricing Path Dependent and American Style Options so that 1; ; will always be one at the edges of the tree and | + ii in the center of the tree In this way we can increase f to increase the accuracy of the approximation by considering more values of G(.). In choosing the actual set of n; ; values for cach node we should consider the distributional properties of the function G(.). This will vary depending on the nature of G(.) and therefore must be considered on a case by case basis The third step in the procedure is to set the value of the option at maturity at every node and for every value of G(.) Cy ik = Clty, Gy, ja); Vik (7.36) Finally, we step back through the tree computing discounted expectations and applying the early exercise condition at every node and for every value of G(.) SORA Cue =e (Pus, Cist rs tue + PmigiCes i,m + Pai jCie tj) (7.37) where f(i,i + 1) denotes the one period forward rate from time step i to time step i+1 and where Cis) jztus Cistjms Ciyi,j-t¢ ate the values of the option at time step i+ 1, given the current G;,;,z, for upward, middle and downward branches of the spot price. These are obtained by computing the value of G(.), given the current value, after upward, middle and downward branches Gi) j41.u, Gist,js Gint,j—-ta> The values Gist jive Gistjms Gini,j-id> and therefore also the option values Ci jeiue Cont,jms Ciai,j-iae Will not in general be stored at the upward, middle and downward nodes and therefore must be obtained by interpolation. For example using linear interpolation we have Ci pelle, 7 Coen ity Corina = Cont + ( \Gratyeta = Gisryors) (738) Gist jl ke i+ Le where k, and k, are such that Gist js1e, S Gist jute S Gist jpis, and k, = k; +1. Thatis the two values of G(.) which lie closest to either side of G,,1,),1, are found and a linear interpolation between these is done to obtain an estimate for Cj; ;1,,. The value of the path dependent contingent claim is read from the tree as the value of Cyo9- 7.5.1 Pricing Asian Energy Options in a Trinomial Tree As a specific example of the generalised methodology outlined in the previous section we price European and American versions of an average price call option, where the average is taken over the spot energy price on the fixing dates 7,,/=1,..., L. Let there be a total of N time steps from the start of the life of the option until its maturity. In order to find the range of values of the average at each node we step forward through the tree from i = 0 to i= N. If we have found the range for all nodes up to time step i-1 then for any node (i,j) the minimum average is determined by the minimum average of the lowest node at time step i— 1 with a branch to the current node and the spot price at the current node. The minimum average is given by 129Energy Derivatives: Pricing and Risk Management Gy if yoy ie a fixing dat Ga ee if 1, =1,, Le. a fixing date (7.39) Git; otherwise where m; is the number of fixing dates which have occurred up to time step i and node (i— 1, j)) is the lowest node with a branch to node (i, j). Similarly the maximum average is determined by the maximum average of the highest node at time step i—1 with a branch to the current node and the energy spot price at the current node Ga Sermeii*S if 1, = 1, ie. a fixing date (7.40) i Gint.jnn otherwise where node (i —1, j,) is the highest node with a branch to node (i,,). Now since the arithmetic average of the spot price is essentially a sum of lognormally distributed prices it will also be approximately lognormally distributed. We therefore choose a log-linear set for the m;,; values of the average at each node (i, f) which gives Giga = Gj18™ (7.41) ie Gye = nGij.1) =I In order to determine the option values of equation (7.37) we first compute what the average would be, given the current average, after upward, middle and downward branches Gjx1, jets Giztjms Fistj-id SuamtSes Gf 5.) = ty,,, ie. a fixing date Giivta : : (7.42) Gi jx otherwise if t.1 = tm,, ie. a fixing date Grss.gm = , (7.43) Gijx otherwise SussSr if 7.) = ty,, ie. a fixing date Lid . (7.44) Gijx otherwise As an example we price European and American versions of a fixed strike average price call option on crude oil with 1 year to maturity and where the terminal payoff is determined by the daily average of the crude oil price during the last month of the life of the option. The valuation date is the 31 March 2000, the tree is constructed to be consistent with the downward forward curve in figure 7.7, using the same parameters as used for table 7.1. Table 7.2 contains the results. Table 7.2 shows the convergence of both the European and American option values as we increase both the numbers of time steps per year and also the number of averages at cach node. 130Spot Price Models: Pricing Path Dependent and American Style Options TABLE 7.2 Convergence of European and American Fixed Strike Average Price Call Options Number of Averages at Each Node of Tree (7) 4 12 20 Steps/Year Euro Amer Euro Amer Euro Amer 30 2.526 2.702 2.488 2.671 2.484 2.667 100 2.575 2.789 2.491 2.711 2.484 2.704 150, 2.629 2.873 2.499 2.737 2.486 2.726 200 2.684 2.953 2.507 2.755 2.488 2.738 250 2.738 3.033 2.520 2.773 2.490 2.746 300 2.792 3.110 2.534 2.793 2.495 2.753 350 2.874 3.225 2.559 2.830 2.505 2711 400, 2.926 3.297 2.575 2.852 2.513 2.780 7.5.2 Pricing Swing Options The term swing option refers to flexibility in the quantity of energy which the holder of the option can receive. The concept being that the volume of energy can ‘swing’ between specified limits. Swing options have arisen because of the uncertainty in the volume of energy, for example gas or electricity, that the end user will consume. Energy suppliers are exposed to this volume risk and swing options were created to allow this risk to be at least partially hedged. Swing options can be thought of as a series of flexible forward contracts or swaps. In general the specification of a swing option will involve the total time period of the contract, the trade dates within that period on which the energy can be bought (or sold) and the purchase (or selling) price which may be fixed or floating. The contract will also specify the limits on the quantity of energy involved — typically a minimum and maximum volume per trade date and an overall minimum and maximum for the contract. The difficulty in pricing swing options is that the optimal decision on the quantity of energy to buy (or sell) on each date depends not only on the energy price but also on the quantity of energy already bought (or sold). Thus swing options can be characterised as path dependent American style options. The approach required to price these type of contracts is similar to that described in section 7.4 in which the quantity of energy becomes the path dependent ble. To illustrate the pricing approach for swing options we use the example of the ‘Take- or-Pay’ option (ToP). The ToP contract specifics a series of dates, 4; i= 1,...,m and prices, K;, at which the underlying energy can be bought together with a maximum. quantity, Vinax, on each date and an overall minimum quantity, Visin, which must be 131Energy Derivatives: Pricing and Risk Management purchased in order to avoid a penalty, g. The dates are typically spread over a one year period and the penalty is usually a specified fraction of the final date purchase price applied to the volume deficit. On each date the decision involves a trade-off between the profit or loss which would be achieved by purchasing energy at the specified contract price relative to the actual market price and the need to purchase sufficient energy overall to minimise the penalty payment at the energy of the contract. Assume we have built a tree consistent with the spot energy price process where i indicates the time step and j indicates the spot price level as described in section 7.3. The first step is to determine the range of values of the path dependent volume of energy purchased. In fact it is more convenient to choose the path dependent variable to be the volume of energy which remains to be purchased in order to avoid the penalty. The range of values which this can take is from zero (if enough energy has already been purchased to avoid the penalty) up to the minimum required to avoid the penalty (if no energy has been purchased). Next we choose a discrete set of values for the volume of energy between the minimum and maximum, Vj; ;4;k = 1,..., ai, the greater the number of values, 7,,, the more accurate will be the computed price of the contract. The value of the contract at maturity, i = N, (the final purchase date) can then be computed. The final decision is relatively simple because the penalty amount is known with certainty. If the spot price is greater than the purchase price, K,,, then the payoff is strictly increasing in the volume purchased and the maximum quantity of energy, Vinax, should be purchased. If the spot price is less than K,, but greater than (1 — z) K,, then purchasing a quantity up to that required to avoid the penalty or the maximum possible, whichever is the smaller is optimal. This is because the loss on the purchase of the energy is more than compensated by the reduction in the penalty payment. If the spot price is less than (1 — g) K,, then no purchase of energy is the optimal decision. The optimal volume to take, V*y,;4, and the value of the contract, C*y,;,., can be summarised as follows: Vmax for Swe > Ky Vrije = § min(Vy jes Vmax) for Ky > Sy,jx > (1-2)Ky (7.45) 0 for (1—2)Ky > Sy jn 20 CN = (Sw ak — Kur) Viv pie — 2K max(0, Vey, jae ~ Vive) (7.46) Finally, we step back through the tree computing the discounted expectations of the contract value at each node and computing the optimal purchase decision at the purchase dates. The optimal purchase decision at each node and for each value of the remaining volume can be computed by searching over the range of possible purchase volumes (zero to Vinax) for the volume which maximises the sum of the discounted expectation'® and the value of the current purchase. At the root node (i = 0, j = 0) the '> The discounted expectation depends on the current volume purchase decision, since this affects the volume remaining to be purchased in the future to avoid the penalty payment, 132Spot Price Models: Pricing Path Dependent and American Style Options volume remaining to be purchased to avoid the penalty is known and therefore the value of contract only needs to be computed for the associated value of k. For a new contract the volume remaining will be V,yin, whereas for an existing contract this may not be the case. Take-or-Pay contracts often have an added complexity, referred to as a make-up provision. These type of ToP contracts are usually multi-year or have multiple penalty dates. The penalty paid on each penalty date can be credited against future purchases above the minimum volume to be purchased in the following year. This extra flexibility introduces another path dependent state variable which can be captured by the volume remaining which can have a penalty credit applied to it. The solution procedure for this more complex problem is basically the same as for the simple case, however, at each node and for each pair of values of the path dependent variables the optimal decision must be found by searching a two-dimensional grid of the variables. Solving these types of problems is therefore correspondingly more computationally demanding. 7.6 Summary In this chapter we have discussed two of the most popular techniques used by practitioners for pricing energy derivatives using spot price models, when there are no analytically tractable solutions. Monte Carlo simulation is shown to be a flexible technique for pricing many types of path dependent option, which can easily be extended to handle seasonality and to implement multi-factor models. We have illustrated the technique with examples showing the pricing of futures options, spread options, and energy swaptions, as well as a number of path dependent options including barrier, lookback and Asian options. For pricing options that have early exercise opportunities we have shown how to construct a trinomial tree for a mean-reverting spot model fitted to the observed forward curve. Although the implementation of this technique is more difficult than simulation, it allows the pricing of American style options in addition to European pricing. Path dependent options, including swing contracts, can also be handled relatively straightforwardly in this framework 1338 Forward Curve Models 8.1 Introduction Historically the majority of work on modelling energy and commodity prices has been focussed on stochastic processes for the spot price and other key variables, such as the convenience yield and interest rates (examples include; Schwartz (1997), Gibson and Schwartz (1990), Hilliard and Reis (1998), Miltersen and Schwartz (1998) ). However, this approach has some fundamental disadvantages — firstly the key state variables, such as the convenience yield, are unobservable and secondly the forward price curve is an endogenous function of the model parameters and therefore will not necessarily be consistent with the market observable forward prices. Figure 8.1, which reproduces figure 4.1, illustrates three market forward energy curves. Each curve is a plot of the first 24 monthly contracts with prices taken from NYMEX. The two crude oil curves (left- hand scale) show that the forward curve moved from backwardation into contango over the period | March 1999 to 31 March 2000. The third curve is the Henry Hub natural gas contract (right hand scale) observed on the 31 March 2000 and clearly shows the strong seasonality of natural gas prices associated with the Northern hemisphere’s seasons. For derivative pricing, therefore, many industry participants require the forward curve to be an input into the derivative pricing model, rather than an output from it, as is the case with the constant parameter versions of the spot price models in Chapter 6. An approach based on modelling the entire forward price curve with multiple sources of ‘uncertainty was introduced by Jamshidian (1991b) and used in Cortazar and Schwartz (1994) to analyse the pricing of Copper Interest-Indexed Notes'. This approach uses all the information contained in the term structure of futures prices in addition to the historical volatilities of futures returns for different maturities. A similar approach has been put forward by Joy (1998a, 1999b) in which the author adds jumps modelled as a Poisson process. The model that he proposes is applied to the power markets and the jumps are assumed to have limited duration to reflect their transient nature in the markets. The framework of this chapter is based on these approaches and draws on material from the paper of Clewlow and Strickland (1999b)’. In this chapter we show how the multi-factor model can be developed in stages and we then apply it to oil, gas and electricity markets. We show how the multi-factor model can be calibrated to observable market data, and derive formulae for the pricing of standard energy " See also Amin, et al (1995). ? See aiso Clewlow and Strickland (1999c). 134Forward Curve Models 287 - ——e +38 | 0131/03/00 | | be | —-oilot/ogeg | 736 _ asl “Gas 31/03/00 | . g 3 $ 22 32 g a a | z 5 20 3 8 ° 2 Bia! 28% 5 | § 16} 126 144 anand lea 12 [27 10+ 2 7°2°3'4'5 67 8 9 10 11 1213 14 15 16 17 18 19 20 21 22 23 24 Months to Maturity FIGURE 8.1 Market Energy Forward Curves derivatives. These pricing formulae are then extended to enable the pricing of path dependent derivatives. We also discuss how the model can be adapted to handle seasonal volatility and volatility smiles. 8.2 A Simple Model for the Forward Curve We will use the expression forward curve models to mean models that explicitly model all the forward energy prices simultaneously rather than just the spot price as with the models in Chapters 6 and 7. To illustrate, we begin with a simple single factor model of the forward curve which we represent by the following stochastic differential equation: di ( cn = oe (TY do (1) (8.1) The inputs to the model are the observed forward curve F(t,T) which denotes the forward price at time ¢ for maturity date 7, and ae~“(?—) which is the single “factor” or volatility function associated with the source of risk dz(t). Notice that equation (8.1) has no drift term — because futures and forward contracts have zero initial investment, their expected return in a risk-neutral world must be zero, implying that the process describing 135Energy Derivatives: Pricing and Risk Management 0.45 - | 04 o4 0.05: 0 08 i 15 2 25 3 35 4 45 5 Maturity (years) FIGURE 8.2 A Simple Negative Exponential Volatility Function for Forward Prices their evolution has zero drift®, The volatility function of equation (8.1) has a very simple negative exponential form which we illustrate in figure 8.2. For this volatility function short dated forward returns are more volatile than long dated forwards — information occurring in the market today has little effect on, say, the 5 year forward price but can have a significant effect on the one month forward price. The parameter values used for the figure are a = 1.0 and ¢ = 0.40. Here o represents the ‘overall’ volatility of the forward curve whilst a tells us how fast the forward volatility curve attenuates with increasing maturity. With an a of 100% we see that the one month forward has a volatility of about 37%, decreasing to roughly 2% for the 3 year forward Figure 8.3 shows a graphical representation of the evolution of the forward curve described by equation (8.1) with a volatility function of this type. The lower curve, denoted by F(s,7), is the original observed market forward curve, whilst the upper curve, F(t + dt,7), represents the curve after a small time step dt where there has been a positive shock to the system (dz(r) > 0). In this case the whole forward curve shifts up, with each point a multiple of ce~“'7~") of the shock. The volatility function is not restricted to have the parameterised form of equation * If we wish to work with the real world process we can add a drift to the models in this chapter without affecting the modelling results. However, for derivative pricing we will work in a risk neutral world. 136Forward Curve Models (8.1). Instead we can allow the function to be very general: dF(t,T) | Fay Me (8.2) where o(t,7) would be read as “the time ¢ volatility of the 7-maturity forward price return”. We can determine from market data what the form of a(7, T) should be. As we shall see later in the chapter we can also allow the forward curve to be affected by multiple sources of uncertainty. 8.3 The Dynamics of the Forward Curve In order to determine empirically the form of the volatility function(s) we can look at the historical evolution of the market forward data. Figures 8.4 and 8.5 show the evolution of the NYMEX oil and natural gas futures curves respectively between January and June 1999 for the first 24 monthly contracts. For each figure the x-axis displays the maturity structure of the forward curve, the y-axis the date on which the curve is observed and the z-axis the market level of the futures price. Both the oil and gas forward curves can be seen to move in a similar way to that illustated by figure 8.3 with the shorter maturity forwards being more volatile than the longer maturity forwards. However, it is also clear that the real movement of the forward FLT | + F(t+dt, 7) Forward Price 0 05 1 15 2 25 3 3.5 4 45 5 Maturity (years) FIGURE 8.3 Example Evolution of the Forward Price CurveEnergy Derivatives: Pricing and Risk Management (sn) ep FIGURE 8.4 Evolution of NYMEX Oil Data (January 1999 to June 1999) curve is probably more complex than is represented by equation (8.1). In particular, the seasonality in the gas forward curve can be clearly seen in figure 8.5. By looking at the annualised standard deviations of the forwards returns for each maturity, we can obtain an estimate of the overall volatility structure of the forward curve. Figure 8.6 shows the annualised standard deviation of returns for both oil and gas for the monthly contracts going out 2 years using daily data over the period from April 1998 to March 2000. These curves can be used as the general volatility function o(7, 7) in equation (8.2). Figure 8.6 shows empirically that short maturity forward prices are more volatile than longer maturity forward prices. Although volatility declines with maturity we see that it levels off at a non-zero level. In figure 8.7 we fit the simple negative exponential function of equation (8.1) to the first six months of the empirically observed natural gas volatility curve by minimising the sum of the squared differences between the two curves, We see that, in order to capture the strong mean reversion implied in the declining volatility for the short dated contracts (the implied mean reversion rate is 1.9), the simple model forces the volatility for longer maturity forwards to go to zero too quickly. This leads the model to underestimate the volatility of the longer maturity gas forward prices and so would give a downward bias to option values that depend on this part of the curve. An important observation from figures 8.4 and 8.5 is that forward prices of different maturities are not perfectly correlated — the curves generally move up and down together 138Forward Curve Models FIGURE 8.5 Evolution of NYMEX Natural Gas Data (January 1999 to June 1999) but they also change shape in apparently quite complex ways. One method that can be used to determine the set of common factors that drive the dynamics of the forward curve is principal components analysis (PCA) or eigenvector decomposition of the covariance matrix. This procedure can be utilised to simultaneously identify the number of important factors and estimate the volatility functions. The technique involves calculating the covariances between every pair of forward price returns in an historical time series, similar to those shown in figures 8.4 and 8.5, to form a covariance matrix. The eigenvectors of the covariance matrix yield estimates of the factors driving the evolution of the forward curve. We will give a detailed example of how to perform this analysis in section 8.6. As we shall see later these factors can be used as the volatility functions in a multi- factor model of the forward curve. Figure 8.8 shows the three most important volatility functions for the NYMEX crude oil futures curve using data from April 1998 to March 2000. Figure 8.8 shows the typical pattern obtained from PCA - generally there are risk factors which act to “shift”, “tilt” and “bend” the forward curve. The most important factor “fn 01”, is positive for all maturities, implying that a positive shock to the system causes all prices to “shift” up (although by slightly different amounts). The second most important factor, “fn 02”, is a “tilt” factor which causes the short and long maturity 139Energy Derivatives: Pricing and Risk Management 50%. 45% -= NYMEX Crude Oil 40% + NYMEX Natural Gas 35%: 2 = 30% 5B S ~ 2 25% 8 20% 6 15%: 10% 5% ari 4 8 12 16 20 24 Maturity (months) FIGURE 8.6 Overall Volatilities for NYMEX Crude Oil and Gas Futures Contracts 50% 45% -* NYMEX Natural Gas ~~ NYMEX Natural GasFitted Negative Exponential 40% 35% 30% 25% 20% Overall Volatility 15% 10% 5% 0% 0 4 8 12 20 24 Maturity (months) FIGURE 8.7 NYMEX Natural Gas Market and Fitted Model Volatilities 140Forward Curve Models 40%: indi = fn 02 30% == {03 >< Overall 20% Volatility 3 z 3 & 10%. MC TS 8 4 5 6 7 8 8 10 11 12 13 14 15 16 17 18 19 20 Bt OD 23 Maturity(months) FIGURE 8.8 Three Main Volatility Functions for NYMEX Crude Oil Data from April 1998 to March 2000 ends of the curve to move in opposite directions. The third factor, “fn 03”, is a “bending” factor which causes the short and long ends to move in the opposite direction to the middle section. The “overall” volatility, which is the combined effect of all the three factors, is also shown in figure 8.8. For many energies, as we have seen for natural gas and electricity, seasonality in the forward price volatilities is an important feature in the evolution of the forward curve, We show how this important concept can be incorporated into the PCA in section 8.6.3. A natural question to ask is how many factors do we need to realistically model the evolution of the forward curve? Principal component analysis can also give us insights into this and in section 8.6 we return to this issue. The implication of the discussion of this section is that to effectively describe the evolution of the energy forward curve we need more than a single factor. We can amend the model described by equation (8.2) by adding additional sources of risk and volatility functions. For example, a 3 factor model is given by: aF(t,T) _ FUT) a(t, T)dz,(t) + o2(t, T)dza(t) + o3(t, T)dz3(1) (8.3) where the volatility functions 9;(.), 92(.), and 93(.) would describe the “shifting”, “twisting”, and “bending” functions in figure 8.8, with d2(.), dza(.), and dz3(.) representing independent sources of uncertainty (Brownian motions). We show in detail how to empirically estimate the volatility functions in section 8.6. 141Energy Derivatives: Pricing and Risk Management 8.4 A General Multi-Factor Model of the Forward Curve For a general multi-factor model the behaviour of the forward curve can be represented by the following equation: aT) > (1, T)dzt) (8.4) In this formulation there are n independent sources of uncertainty which drive the evolution of the forward curve. Each source of uncertainty has associated with it a volatility function which determines by how much, and in which direction, that random shock moves each point of the forward curve. The g;(t, T) are therefore the n volatility functions associated with the independent sources of risk dz((f). In practice we would usually set n = 1, 2, or 3. 8.5 Relationship Between Forward Curve and Spot Price Models Intuitively, a model that describes the evolution of the whole forward curve is implicitly saying something about the front end of the curve, which is simply the spot energy price, and so the forward curve models of this chapter must be related to the spot price models of Chapter 6. In this section we show exactly what that relationship is. We can integrate the defining stochastic differential equation (8.4) to obtain the following solution: F(1,T) = F(0, T)exp [= {-3 if “ou, T)2du + f “oil, nas} (8.5) iI This equation expresses the forward curve at time / in terms of its initially observed state (time 0) and integrals of the volatility functions. The spot price is just the forward contract for immediate delivery and so the process for the spot price can be obtained by setting T = 1, ie. a lr Pt S(0) = F(t,0) = F(0,) exp [= {- sh o(u,t)°du +f oi(u, vaco}| (8.6) 2Jo 0 =I Equation (8.6) can then be differentiated to yield the stochastic differential equation for the spot price: 210.9 $F fafa any [AED aah ts i= dS(0) SO) : (87) + So ai(t, dei(0) 142Forward Curve Models The term in square parentheses in the drift can be interpreted as being equivalent to the sum of the deterministic riskless rate of interest r(¢) and a convenience yield 6(7) which in general will be stochastic. Also, since the last component of the drift term involves the integration over the Brownian motions, the spot price process will, in general, be non- Markovian — i.e the evolution of the spot price will depend upon its past evolution. One special case of the general model is the simple single factor model described by equation (8.1). Recall that for this model we have n = 1 and o,(t,T) = ae~#7~9, Clewlow and Strickland (1999a) evaluate (8.7) with this volatility function and show that the resulting spot price process is given by: dS(t) dlnF(0, ft) @ So 7 far Fn FO, 500) + al ee] ai + oa (8.8) This implies: oa (u(t) — an S(d)]at + odz(t) (8.9) where w(t) = pa +aln F(0,0 + 7 (1-e*#"). This single factor forward curve model is therefore just the single factor Schwartz (1997) model of Chapter 6, however with a time dependent drift term. It is this term in the drift which allows the model to now fit the observed forward prices. Note also that this particular form of the forward curve volatility function results in a “Markovian” spot price process — as the dependence in the drift on the path of the Brownian motion disappears. The relationship between the forward curve model and the spot return model also shows that the mean reverting behaviour of the spot price is directly related to the attenuation of volatility of the forward curve. Recall also from Chapter 6 that by setting c = 0 we obtain the Black (1976) model, this is therefore a special case of the general model in equation (8.4) with o(t, 7) =o and n= 1. 8.6 Estimation of the Volatility Functions Although, in the previous section, we discussed special cases which arise as restrictions of the general model, the main advantage of the forward curve modelling approach is the flexibility that the user has in choosing both the number and form of the volatility functions. Similar to other derivative markets these can be chosen in one of two general ways; historically, from time series analysis; or implied from the market prices of options.Energy Derivatives: Pricing and Risk Management 8.6.1 Historical Estimation of the Forward Curve Volatility Functions In this section we show how to obtain the general volatility functions from historical forward curve data. The method that we describe allows the user to determine both the form of the volatility functions as well as the number of factors to use. The general multi- factor forward curve model of this chapter can be represented, after an application of Ito’s lemma, in logarithmic form, as: ois, T}dte Se oilt T)dz,(1) (8.10) =i = This can be discretized for small time changes Ar as i(t, t+ 7) Az; (8.11) Amn F(t,t+ 7) =-$>oai(t,¢+ 5) > Equation (8.11) implies that changes in the natural logarithms of the forward prices with relative maturities 7;, 7 = 1,...,m are jointly normally distributed. We can compute the sample covariance matrix of these forward prices in the standard way: ,_1¢ 2 8 = yO Ou — ENR - FH) (8.12) ia where there are N samples (k =1,..., N) of x and x, which are defined as: Xie = In(F (th, te + 7))) — N(F (ty — At, ty — At+7))) (8.13) and where %;, X; are the sample means. The time interval At would usually be one day. The discretised volatility functions, o((¢,¢+ 7); i= 1,..., n and j = 1,..., m, are recovered by eigenvector decomposition of the covariance matrix. The decomposition yields the set of independent factors which drive the evolution of the variables under- lying the covariance matrix. [t decomposes the covariance matrix which we denote by &, into n eigenvectors v; and associated eigenvalues \, such that y= rar! (8.14) where ML UR o| pal % oO) len em i and where superscript T here denotes transpose. The columns of I’ are the eigenvectors. The eigenvalues represent the variances of the independent “factors” which drive the 144,Forward Curve Models forward points in proportions determined by the eigenvectors. The volatility functions are then obtained as 9,(1,1 + 7%) = uWX. 8.6.2 Example Analysis of NYMEX Crude Oil Futures We use the methodology of the previous section to determine the volatility functions for NYMEX crude oil futures over the period from the ] April 1998 to 31 March 2000, using the first 24 monthly contracts for the analysis. The first step is to construct a time series of forward price returns according to equation (8.13). Table 8.1 shows the returns for the first and last four observations in the data set. Next we compute the sample covariance matrix by applying equation (8.12). Table 8.2 shows the results for the first and last four contract months Table 8.3 summarises the results of the eigenvector decomposition. The first 24 columns of the data, “fn 01”, ..., “fn 24”, represent the eigenvectors, with the last column, headed 4, containing the eigenvalues. Each row (excluding the last column) corresponds to the indicated forward price describing how much that particular forward is influenced by the factor associated with each column. TABLE 8.1 Returns to NYMEX Crude Oil Futures Contracts Date F(Im) Fm) Fm) F(4m) FQIm) F(22m) © F23m) — F(24m) 02/04/98 0.0128 0.0125 0.0123 0.0115. 0.0000 0.0000 0.0000 9.0000 03/04/98 0.0158 0.0142 0.0128 0.0114. 0.0051 0.0046 0.0046 0.0046 06/04/98 -0.0344 0.0312 -0.0288 0.0266. 0.0160 -0.0155 —0.0149 0.0149 07/04/98 0.0150 -0.0140 -0.0125 -0.011! 0.0017 0.0017 0.0017 0.0017 28/03/00 0.0255 0.0244 -0.0218 0.0190 -0.0158 0.0164 —0.0169 0.0175 29/03/00 -0.0239 -0.0184 -0.0121 —0.0067... 0.0243 0.0249 0.0255 0.0267 30/03/00 0.0094 0.0077 0.0066. 0.0063. 0.0061 0.0061 0.0062 0.0062 31/03/00 0.0075 0.0118 0.0116 0.0113 0.0093 0.0089 0.0085 0.0085 TABLE 8.2 Covariance Matrix For NYMEX Crude Oil Futures Film) F(Qm) FGm)— Fm). FQIm) F(22m) FQ3m) FQ4m) F(im) 0.1356 0.1153 0.1042 0.0959... 0.0494 0.0486 0.0464 0.0466 FQm) 0.1153 0.1068 0.0975 0.0901 eee 0.0486, 0.0478 0.0457 0.0457 FGm) 0.1042 0.0975 0.0904 0.084, 0.0471 0.0464 0.0445 0.0437 F(4m) 0.0959 0.0901 0.0834 0.0783 0.0453 0.0446 0.0428 0.0436 F(21m) 0.0494 0.0486 0.0471 0.0853, (0.0384 0.0381 0.0369 0.0367 F(22m) 0.0486 0.0478 0.0464 0.0446 0.0381 0.0381 0.0367 0.0367 FQ3m) 0.0464 0.0457 0.0445. 0.0428 =... (0.0369 0.0367 0.0399 0.0358 FQdm) 0.0466 0.0457 0.0437 0.0436. 0.0367 0.0367 0.0358 0.0392 145Energy Derivatives: Pricing and Risk Management TABLE 8.3 Eigenvector Decomposition for NYMEX Crude Oil Futures fo0l fn02_— fn03 fn 04 fo 21 fn 22 fn 23 fn 24 x F(lm) 0.2926 0.5138 -0.7253 -0.1404 = 0.0060 —0.0048 -0.0036 0.0033 1.2388 Fm) 0.2764 ~0.3545 0.0605 0.1155. 0.1435 0.1755 0.0194 0.0191 0.0888 F(3m) 0.2592 0.2612 0.1739 0.2544 0.3941 0.4317 —0.1889 0.1656 0.0107 F(4m) 0.2452 -0.1860 0.1627 -0.0388 0.1060 0.0955 0.5185 0.4368 0.0060 F(2Im) 0.1625 0.2257 -0.1309 0.2140 0.0047 0.3128 —0.0050 0.0706 0.0001 FQ2m) 0.1609 0.2346 -0.1511 0.2075 0.0317 -0.2012 0.0110 —0.0214 0.0001 F(23m) 0.1554 0.2476 —0.2672 0.6795 0.0467 0.0475 —0.0099 0.0050 0.0000 F(24m) 0.1568 0.2601 -0.2419 -0.2925 ... -0.0026 -0.1135 0.0033 —0.0341 0.0000 TABLE 8.4 Volatility Functions for NYMEX Crude Oil Futures fn0l fn02_—fn03_—__fn 04 fo21 fm 22 fn 23 fn 24 Film) 0.3257 -0.1531 -0.0751 -0.0109 0.0000 0.0000 0.0000 0.0000 Fm) 0.3076 -0.1056 0.0063 0.0089 _ 0.0011 0.0013 0.0001 —-0.0001 Fm) 0.2885 -0.0778 0.0180 0.0197 =0.0030 =0.0031 =0.0008 0.0006 F(4m) 0.2729 -0.0554 0.0169 —0.0030 - 0.0008 0.0007 0.0022. —0.0015 Fim) 0.1808 0.0673 -0.0136 0.0166 0.0000 0.0022 0.0000 0.0002 FQ2m) 0.1791 0.0699 -0.0157 0.0161 : 0.0002 -0.0014 0.0000 —0.0001 FQ3m) 0.1729 0.0738 -0.0277 0.0526 0.0004 0.0003 0.0000 0.0000 FQ4m) 0.1745 0.0775 -0.0251 -0.0227 _ 0.0000 0.0008 0.0000 -0.0001 By multiplying the eigenvectors by the square root of the eigenvalue we obtain the resulting volatility functions. The columns of table 8.4 indicate the result of this calculation. The volatility functions illustrated in figure 8.8 are just the first 3 columns of this table. The eigenvalues resulting from the eigenvector decomposition tell us the importance of each eigenvector and hence the number of factors that we should typically be using in the general model. Figure 8.9 plots the eigenvalues for the NYMEX crude oil futures data (the column headed “X” in table 8.3) in descending order of size. The figure shows that the first eigenvector (correspondingly volatility function) is the most important, explaining 91.1% of the total variation in the evolution of the futures curve. Together the first 2 factors explain 97.7% of the total variation with the first 3 factors explaining 98.4%, Factors 4 and above are not significant and so a 3 factor model is sufficient to explain the evolution of the NYMEX oil data over this period. 8.6.3 Incorporating Seasonality into the Volatility Functions As we have described earlier, gas and electricity forward curves exhibit strong season- ality. Seasonality in the volatility functions can be represented as the product of a time 146Forward Curve Models Value 0: ee 12°34 5 6 7 8 9 10 11 12 1314 15 16 17 18 19 20 21 22 23 24 Eigenvalue FIGURE 8.9 Eigenvalues for NYMEX Crude Oil Futures. From April 1998 to March 2000 dependent spot volatility function and maturity dependent volatility functions. The general equation (8.4) can therefore be specialised to cope with seaonality as follows: cen? sO Lol — tdz\(t) (8.15) where os(f) denotes the spot price volatility at time ¢ and o;(T —#) the n maturity dependent volatility functions. In this way, the maturity structure of the volatility functions is normalised by the spot volatility. In order to estimate this model we first estimate the spot volatility, for example, by using a rolling 30 day sample standard deviation. We then divide the daily forward price returns by the spot volatility estimate for that day ~ the calculations for the covariance matrix and the eigenvector decom- Position are then obtained as before with this normalised set of returns. We use the methodology of this section to determine the volatility functions for NYMEX natural gas futures contracts and NYMEX California/Oregon Border (COB) electricity futures over the period 1 April 1998 to 31 March 2000. Figure 8.10 shows the natural gas “spot” daily return standard deviation using the shortest maturity future with a rolling window of 30 days returns. Figure 8.11 shows the result of applying the principal component analysis to the adjusted NYMEX natural gas futures data, with the first three factors shown. Comparing figure 8.11 with figure 8.8 we see that, although the levels of the volatility functions differ between the crude oil and natural gas, qualitatively the “shifting”, 147Energy Derivatives: Pricing and Risk Management 0.050 0.045 0.040 0.035 0.030 0.025 Standard Deviation 0.020 0.015 0.010 01/04/98 01/07/98 30/09/98 3012/98 31/03/99 30/06/99 29/09/99 29/12/99 29/03/00 Date FIGURE 8.10 NYMEX Natural Gas Spot Dally Return Standard Deviation Sane Bo Volatility Function Value ° R oF FO 11 12 13 14 15 16 17 18 19 20 21 22 23 ~ Maturity(months) FIGURE 8.11 Seasonally Adjusted NYMEX Natural Gas Volatility Functions 148Forward Curve Models 5.0. 45: 4.0 35: | 3.0 Value 2.5, | 2.0 1.0 | : ee ee eee 123-4 5 6 7 B 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Eigenvalue FIGURE 8.12 Eigenvalues for NYMEX Natural Gas Futures from April 1998 to March 2000 “tilting” and “bending” factors are common to both and in fact to all energy contracts that we have studied. Also notice that the zero crossing points for the tilting and bending of the curves are in different maturity segments. For example, the tilting factor crosses at 9 months maturity for oil but as early as 4 months for natural gas. Figure 8.12 shows the eigenvalues obtained via the PCA for natural gas. Comparing figure 8.12 with figure 8.9 which shows the results of the oil analysis we see that for natural gas more factors are slightly significant. The first eigenvector now explains 91.7% of the evolution of the futures curve, the first two together 96.8%, rising to 98.1% with the addition of the third. To explain 99% we would need to use a five factor model for natural gas compared with only four factors for crude oil. Figures 8.13 and 8.14 show the analysis for NYMEX COB electricity futures. The oscillations in the volatility functions beyond nine months maturity and the slow decline in the eigenvalues is indicative of noise in the futures prices which are fairly illiquid beyond three months maturity. 8.7 Pricing Standard European Options In this section we turn our attention to the pricing of derivatives in the forward curve modelling framework. We begin with the pricing of standard European options on the spot energy price and then extend the analysis to pricing options on forward contracts*. + This section is based on section 3 of Clewlow and Strickland (1999b). 149Energy Derivatives: Pricing and Risk Management 1.27 | “ein0t +n 02 te —+n03 ++ O1 0.8) ity Function Value © g Volat ° ‘Maturity(months) FIGURE 8.13 NYMEX COB Electricity Futures Volatility Functions from April 1998 to March 2000 3.0) --—_—- ——__——- ————_—_—— 25. 2.0 ob seseees sae os 9s. 72°34 5° 6 7 8 8 1011 12 13 14 15 16 17 18 19 20 21 22 23 24 Eigenvalue FIGURE 8.14 Eigenvalues for NYMEX COB Electricity Futures from April 1998 to March 2000 150Forward Curve Models 8.7.1. Options on the Spot From standard risk-neutral pricing results the price at time ¢ of any contingent claim on the spot energy price with payoff at time T, C(¢, S(4), is given by the expectation of the discounted payoff under the risk neutral measure: C(t, S() = E,[P(t, T)C(T, S(T))) (8.16) where P(t,7) is the T-maturity discount factor. For a standard European call option with strike price K and maturity date T on the spot energy price, c(t, S(”); K, T), we have c(t, S(;K,T) = E,[P(t, T) max(0, S(T) ~ K)] (8.17) From equation (8.6) the natural logarithm of the spot price at time T, given the forward curve at time /, is normally distributed as follows: insiry~ [nero -$30{ fata ra, ¥{focarea}| (8.18) fer (Jr wat (Jr The distribution for the spot energy price is therefore determined solely by the current forward price for the future date T and the assumptions regarding the volatility functions, 0,(.). Since interest rates are assumed to be deterministic and spot prices are lognormally distributed, we can use the results of Black and Scholes (1973) to obtain the following analytical formula for a standard European call option: c(t, S();K,T) = P(t, T)F(t, T)N(h) — KN(h— v'w)] (8.19) where OTK) he eT ; n= RECDIO He, -r{/ lu TP} The corresponding pricing formula for a standard European put option, p(t, S(t); K, T), can be easily obtained by put-call parity: P(t, S(Q;K,T) = Plt, T)|KN(—h+ Vw) - F(t, T)N(-A)) (8.20) For the single factor restriction of section 8.2, equation (8.18) can be evaluated explicitly with the distribution of the spot price satisfying: op T) oS) eT) 20) 2017) In S(T) winFe.7) = [1 e Js[l e | (8.21) European call and put options are then given by equations (8.19) and (8.20) respectively 151Energy Derivatives: Pricing and Risk Management with w replaced by we g [1 e2ner—9 (8.22) A special case of equation (8.2) is the case of o(¢, 7) = 0. This is the restriction of Amin et al. (1995), and in this case w = 0>(T — 1) 8.7.2 Options on Forward Contracts We now derive the price at time ¢ of a European call option with strike price K that matures at time 7 on a forward contract that matures at time s, which we denote by c(t, F(t,s); K, T,s). Options are again priced using the standard methods. At date ¢ the European call option price is the expected discounted payoff: c(t, F(t,8); K, T,8) = E,[P(t, 7) max(0, F(T,s) — K)] (8.23) From equation (8.5) the natural logarithms of the forward prices at time T are normally distributed: In F(T,s) ~m(ineeg -S9{ f“onstas}. Sof [nln ora} (8.24) i=l f eee ! Using this result it is straightforward to show that the solution to equation (8.23) is given by: et, F(t,s);K,T,s) = P(t, T)[F(t,s)N(h) — KN(h- Vw)] (8.25) where nF s)/K) aw 8 Df [ ovusPa} The formula for standard European put options can be obtained by put-call parity. Clewlow and Strickland (1999a) show that for the single factor restriction of this chapter European call futures option prices are calculated via equation (8.25) with w given by: T w= f o(u,s)Pdu 1 r 2 ff oe 2" ly (8.26) i ace (er 7) _ g-2ale- °) 2aForward Curve Models 25: — 0 Cap at $2.2 20 + Cap at $2.4 e 15 8 < o é 10} 5 0 ee OoiereouHuEedl 6 Gs iOceeeioseereia iors Guter Corsa iaoaT os Maturity (months) FIGURE 8.15 Natural Gas Cap Prices with Monthly Maturities Armed with these analytic results it is possible to price other standard energy derivatives. For example, in Chapter 5 we showed that an energy cap can be viewed as a portfolio of call options on futures contracts. Figure 8.15 shows the prices of caps with monthly maturities from 1 month until 23 months for natural gas under the general model (see equation (5.3)). Each cap price is the sum of the monthly caplets up to the cap maturity. The individual caplets have been priced using the observed NYMEX forward curve for natural gas for 31 March 2000 (illustrated in figure 8.1) and the three volatility functions derived from our historical analysis (illustrated in figure 8.11). 8.8 Pricing General European Contingent Claims In this section we show how to price general European contingent claims for the multi- factor forward curve model described by equation (8.4). For example, consider a European option, with maturity 7, on a portfolio of forward prices with maturities ona set of dates s,,k = 1,...,m(T {f o,(u, Sy)oudu, wpa (8.28) i=1 To efficiently sample under this covariance matrix we compute the orthogonal repre- sentation of the covariance matrix (or PCA) which returns the m eigenvectors w; and associated m eigenvalues A; such that y=Tart (8.29) where Wry 2 Wha A 0 1. 0 Wap Wag Wa 0». 0 r= | 21 22 ‘20 a ee 2 Wont Wind se Wm OR Oct. Let M be the number of simulations and ¢), i = 1, ..., m be independent standard normal random numbers. Clewlow and Strickland (1999b) show that the expectation in equation (8.27) can be represented by; C(t {ee FG) }) = P(t, T) adsl (Ty {en F(t 54) Y(t, Tsp) k = 1.2.5 mp) (8.30) where Yj(¢, 7,5) = exp ~35 ble HE feu Notice that because of the Gaussian nature of our framework we effectively jump straight to the end of the life of the option under each simulation, rather than the more usual practice of simulating at a large number of small discrete time steps until maturity. This formulation also gives a natural way to trade off accuracy with speed by truncating the integration space dependent on the sizes of the eigenvalues. The size of eigenvalues indicate the relative importance of the corresponding eigenvector factors in reproducing 154Forward Curve Models the covariance matrix D. Therefore by truncating the eigenvector representation we reduce the dimensionality of the Monte Carlo integration and thus reduce the computa- tion time required but at a cost of reducing the accuracy. Example: Pricing a European Energy Swaption Recall from section 5.5 that the time value of an energy swaption, with maturity date T, to swap a series of floating spot price payments on dates s,.k = 1,.. . ym, for a fixed strike price amount K can be written as: Swpn(t;K,T, {s¢},m) = P(t, nfm (0. {23° cr so} 7 ‘| (8.31) r= From equation (8.30), we have: Swpn(t; K,T, {s,},m) = P(t, Oy [mes (0. 35 F(T, 5) *) iat a F(t, sp) Y(t, T, 54) -*)| (8.32) where the exponential martingale Y,(1, Ts.) is defined in equation (8.30) and the &), i=1,..., mare new samples for each j. Table 8.5 shows the prices of swaptions on natural gas, using the same gas data (forward curve and volatilities) as for the previous cap example. The maturity of the options range from | month to 1 year. Each of these options exercises into swaps with maturities of 1, 3, or 6 months (“Swap Tenor”) — each swap has monthly resets. All swaptions have an exercise price of K = $2.20. Notice that, because of the seasonality observed in natural gas prices, option prices do not necessarily increase with either option maturity or swap tenor. TABLE 8.5 Example Natural Gas Swaption Prices Swap Tenor Im 3m 6m Im 0.7528 0.7633, 0.7872 Option 3m 0.7784 0.7753, 0.8549 Maturity 6m 0.8709 0.9280 0.7954 lyr 0.5316 0.5182 0.5117 155Energy Derivatives: Pricing and Risk Management 8.9 Pricing Exotic Options In this section we extend the analysis of the previous sections to pricing a range of exotic or path dependent options under the multi-factor forward curved model of this chapter. We begin by pricing the two types of spread options described in section 5.7. We also consider the pricing of Asian options and American style options in this section’. 8.9.1 Calendar Spread Options Recall from section 5.7.1 that calendar spread options are written on the difference between futures contracts with different maturity dates on the same underlying energy with payoff condition given by equation (5.9). Comparing equation (5.9) with equation (8.27) we see that the pricing of a calendar spread option fits in with the general pricing methodology of section 8.8. Therefore we can use equation (8.30) to evaluate the solution to the expectation: Calendar Spread(t; K,T,,51,52) M = PUT) > Imax(0, FUT 51) ~ F(T, 8) - K)| a5 j=l 3 M, = Plt, n> [max (0, F(t, 51) Y(t, T, 81) — F(t,2) ¥j(¢, 7,52) ~ K)] = 8.9.2 Crack Spread Options Recall from section 5.7.2 that if the futures contracts underlying the option are written on two separate energies (say natural gas and electricity), then the option is referred to as a crack spread option. Because option contracts of this type depend simultaneously on forwards related to separate energies then, in order to value options of this type, we require the joint distribution of the two forward price curves. We can extend the multi-factor forward model of equation (8.4) to simultaneously model a number of different energy forward curves®. The following stochastic differential equation describes the joint evolution of energy forward curves (indexed by ¢) for an n-factor model: aon > eit, T)dzi(0) (8.34) * See Chapter 5 for definitions and examples of these and other path dependent derivatives. © This methodology is used in Chapter 10 for VaR calculations for energy portfolios that contain multiple energies. 156Forward Curve Models where e = oil, gas, electricity, etc. Increments in the different energy forward curves are expressed in terms of a consistent set of volatility functions, o,;(t, 7), which describe how all the forward curves evolve together. For example, the first volatility function describes how the primary factor affects all the energy forward curves. Under this model all forward price returns are jointly normally distributed. Option prices are again computed via Monte Carlo simulation. In order to simulate the joint behaviour of the energy prices on which the portfolio depends we must compute the volatility functions ¢,((¢,7) from the covariance matrix of all the relevant forward prices. An example of this joint estimation is shown in Chapter 10 for calculating value-at-risk for multiple energy portfolios. If F,(t,,) represents the price of an s;-maturity futures contract on energy a and F,(t,s2) represents the price of an s)-maturity futures contract on energy 6 then the price of the option, as represented by equation (5.12), can be evaluated as follows; Crack Spread(t; K, T, F,(t,s1), Fs(t, $2), 81,52) M = P(t,T) wy, (max(0, F,(T, s;) — F,(T, 52) — K)] (835) M, = P(t, Nay {max (0, Fy(t,51) ¥(t, 7,51) — F(t, 82) Y(t, T, 2) — K)] ja Note that the calculation of the exponential martingales now depends on the elements of the decomposed joint covariance matrix derived from the evolution of the two energy forward curves. 8.9.3 Asian Options As an example of an option with a terminal payoff dependent on the path taken by the spot instrument or a forward contract, valued in the forward curve framework, we show how to price an Asian option. The methodology we describe is equally suited to barriers, lookbacks, and other exotic options. We consider the following fixed strike European Asian call option. We assume the maturity of the option to be s,,, with the payoff determined by the average of the s,,-maturity futures contract observed at times 51,..., Sm With s; < ... < 5,,. The last observation is therefore also the spot price at time s,,. This option is an example of the average price call option in table 5.1 and so the value can be written as the following discounted expected payoff: Asian(t; K, Sn, {5¢}) = P(t, 5 JE, [nss(o {3 Fasa} - | (8.36) fa Note that, if the average were taken over the futures contracts with a constant time to maturity As (for example 1 month) on each observation date the payoff would be 157Energy Derivatives: Pricing and Risk Management written as: Asian(t; K, Sp, {s¢}) = P(ts5p)E, om (0. {2 Sra +a0| i s) (8.37) m In order to evaluate the expectation in equation (8.36) (or 8.37) we need a more general representation than (8.27) because the maturity payoff is determined by observables at times other than maturity. We now consider any contingent claim to be a series of possibly contingent cashflows, C;,.(s,, ©, {¢)F (s,, 57)}). occurring on dates %,k=1,..., m and depending on forward prices with maturity dates also on the set of dates 5, with face values ¢ and also on the parameter vector ©. The price of a contingent claim of this form is given by CUt,0, (64},{5)) = E, [SE amapedore. cores} (8.28) at The expectation is taken over the mx m-dimensional normal distribution of the correlated natural logarithms of the forward prices In(F(s,,s))). In order to perform Monte-Carlo simulation to evaluate (8.38) we must therefore compute the m! x m/’ covariance matrix © (where m' = m2): Bq = Covfln(F(4,5;)),In(F(te,s4))] = S> { | a oilusionunsiyu} (8.39) isl In order to efficiently sample under this covariance matrix we again compute the orthogonal representation of the covariance matrix which now yields the m/! eigenvectors ww; and associated m/ eigenvalues 4. Using this result, equation (8.38) can be written as: M. Cea} ub =p [Sremetne lortortesnsat (8.40) a where ¥,(t)5.5)) = exp [- sy {wba} > {revie}] ia, Applying (8.40) to our example of the Asian option of equation (8.36) we therefore have: M im Asian(t; K, 5p, {54}) = Plt ooh om {esr so) ¥snse)} - x) Ma ta (8.41) Other path dependent options can similarly be handled in this way. For example, for European barrier options with the barrier fixed against the level of the s,,-maturity futures contract observed at dates 5),...,5,, we can use the above methodology to 158Forward Curve Models calculate the forward observations F(s,,s,,) for k = 1,..., m. The values are then compared to the barrier level to determine if the barrier has been crossed’, Similarly for lookback options, the forwards on the reset dates can be calculated using this method and the maximum or minimum determined from these observations dependent on the option characteristics* 8.9.4 American Options in Forward Curve Models Although American options are not usually considered exotic (unless we are looking at the early exercise of some of the path dependent options already considered in this section) their values are typically difficult to evaluate in the simulation setting of this chapter. The problem arises because simulation methods generate trajectories of state variables forward in time, whereas a backward dynamic programming approach is required to efficiently determine optimal decisions for pricing American options. We outline in this section how, by supplementing the Monte Carlo integration of previous sections with Markovian spot price trees in order to generate an early exercise strategy, we can efficiently obtain approximations for prices of American style derivatives. The price of an American style contingent claim reflects the extra premium derived from the ability to exercise early. Equation (8.38) can be further generalised to reflect the optimal choice of early exercise strategy by allowing the parameter vector © to now include the early exercise strategy The single factor model of section 7.3 can be considered a one factor Markovian restriction of the general model of this chapter. Recall from Chapter 7 (equation (7.27)) that the logarithm of the spot price follows the process: dx(t) = [0(t) — ax(t)|dt + ode() (8.42) where the time dependent function is determined by the initial forward price curve. A trinomial tree was constructed in section 7.3, consistent with both the current forward price and forward volatility curves. Similarly, we can construct a trinomial tree consistent with the initial forward curve and the overall volatility function in the multi-factor model. The trinomial tree constructed in this way allows us to determine an approximate early exercise strategy for American contingent claims. We use the approximate early exercise strategy obtained in this way as an input into our multi-factor non-Markovian Monte Carlo simulation. If we represent this early exercise strategy by the parameter vector © we can obtain a lower bound approximation to the price of the American contingent claim under the full model using (1,8, Loa} {5}) = Es] So Plt 5) Cals, Cer lours9}) (8.43) il 7 See Chapter 5 for a description of barrier options and a definition of typical option payoffs, * See Chapter 5 for a description of lookback options and a definition of typical option payoffs. 159Energy Derivatives: Pricing and Risk Management Furthermore we can obtain an arbitrarily good approximation by maximising the value of the contingent claim with respect to the early exercise strategy. 8.10 Implied Estimation of the Volatility Functions Many practitioners argue that the most important issue in the use of any model for pricing and hedging derivatives is that of calibrating the model to market data. Calibration is the process of choosing model parameters so that the prices returned by the model coincide with the observed market prices. Calibrating an energy model is analogous to choosing, or implying, a volatility parameter for the Black-Scholes-Merton model, when valuing say stock or index options, that equates the model price with the market price. However, in the Black-Scholes-Merton model a single variable carries all the volatility information, but in forward curve models it is the set of volatility functions. Exact implied calibration to all market prices simultancously is often not possible, depending on the volatility functions chosen, and so the procedure usually involves choosing the model parameters so that prices returned by the model are as close as possible to the market prices of options. This closeness of fit is often measured in a ‘least- squares’ sense. Define the set of market prices to be Cyrarker,i 3 1 = 1,..., N and the corresponding model prices be Cynodet,;(@) where © is the set of parameters of the option pricing model. To calibrate the model prices to the market prices we solve the following problem: y : 2 ssa {3 (Se Ct (®)) | (3.44) oF Canker For clarity we will illustrate the implied estimation process with the single factor model of equation (8.1), but the extension to the general multi-factor model is straightforward. The data we use for our calibration exercise is Brent crude oil futures options prices traded on the International Petroleum Exchange (IPE) in London for 15 July 1999. Although these options have American style exercise, we assume for this example they have European style exercise, so that we can use equation (8.25) and the corresponding put option pricing formula. The first section of table 8.6, “Market Prices”, shows the prices for Brent crude oil call and put futures options for maturities between September 1999 and February 2000, and for strike prices ranging from $13.00 to $21.00 on the 15 July 1999. The maturity dates for both the options and the futures contracts were obtained from the IPE calendar, with the future expiring approximately | week after the option maturity. Missing prices imply that the option did not trade on the 15 July 1999. The second section of table 8.6, “Model Prices”, shows the model prices obtained via equation (8.25) and the corresponding put price. Recall that model prices are determined once the volatility curve is specified, which for this model requires specifying @ and o (the vector in equation 8.26). We employed a standard numerical search routine to find a and o by minimising the squared percentage errors across all of the options in table 8.5. 160.Forward Curve Models TABLE 8.6 Calibration of the Schwartz One Factor Model to IPE Brent Crude Oil Futures Options Prices Market Prices Calls Puts Strike Price — Sep-99 Oct-99 Nov-99 Dec-99 Jan-00 Feb-00 Sep-99 Oct-99 Nov-99 Dei 99 Jan-00 Feb-00 13.00 0.00 0.03 13.50 5.58 0.00 14.00 0.00 0.00 0.04 0.06 0.14 14.50 0.00 0.01 0.06 009 0.20 15.00 3.97 0.01 0.02 0.10 0.13 0.28 15.50 3.52 0.02 0.03 16.00 3.00 0.04 0.06 027° 0.51 16.50 2.67 0.07 O11 17.00 2.20 2172.28 2.27 012 0.19 17.50 1.77 195 203 1.76 0.19 0.31 0.72 0.98 0.88 18.00 1.38 144 159 167 1.74 (149 030 0.46 0.72, 0.94 1.191 18.50 102 114 1.29 142 1471.25) 0.440.66 092 1191.42 1.37 19.00 0.72 0.90 1.06 1.20 1.25 104 = 0.64 0.92 «119-147 L.70 1.66 19.50 0.50 0.70 0.86 1.01 1.05 0.86 0.92 122 149 178 200 1,98 20.00 0.37 0.54 «0.70 0.85 (0.88 1.29 1.56 1.83 20.50 0.40 21.00 0.29 Model Prices Calls Puts Sep-99 Oct-99 Nov-99 Dec-99 Jan-00 Feb-00 Sep-99 Oct-99 Nov-99 Dec-99 Jan-00 Feb-00 13.00 6.01 584 5.67 548 5.26 5.06 0.00 000 001 002 0.04 0.06 13.50 5.52 5.36 5:19 5.02 4.81 4.63 0.00 0.00 0.02 0.030.06 0.09 14.00 5.02 487 4.73 456 438 «4.21 = 06.00 -0.01.- 0.03 -0.06 0.10 0.14 14.50 4.53 439° 4.27 4.12 3.96 3.80 © 0.00 0.02, 0.05 0.09 O15 0.20 15.00 4.04 3.92 3.82 3.70 355 342 0.00 0.03 «0.09 -O.140.22 0.28 15.50 3.55 3.46 3.39 3.29) 3.17 3.06 = 0.01. 0.06 «0.14 «0210.30 0.38 16.00 3.06 3.01 2.97 2.90 281 271 0.02 010 021 030 O41 0.50 16.50 2.59 259 259 254 247° 239 0.04 0.17 0.30 042 055 0.65 17.00 2.14 2192.22 2.21 216 2.10 0.09 0.26 «0.42 0.56 «0.70 0.82 17.50 1.72 1.83 189 1.90 188 183 0.16 0.38 0.57 0.73 0.89 1.01 18.00 1.34 150 159 1.62 1.62 1.59 0.28 0.54 0.76 «0.92 10 1.24 18.50 1.011.211.3337 «1381.37, 0440.74 0.97) LIS 1341.48 19.00 0.74 (0.96 1.10 115 118 117 0.66 0.98 1.22 1411.60.75 19.50 0.52 0.75 089 0.96 1.00 1.00 093 126 150 169 189 204 20.00 0.35 «(0.580.720.7984 0.85 1.26.87 1.81-2.00 2.35 20.50 0.23 «0.44 «(0.58 0.65 «0.70 0.72.63, 1.92 2.15 2.34 2.69 21.00 0.14 0.32 0.46 «0.53 0.58 060 204 230 251 2.69 3.03 161Energy Derivatives: Pricing and Risk Management In this way the calibration is a “best fit’ to the full set of market data. The values found in this way yield a = 0.741 and o = 0.355. In practice, we would potentially fit the model to the prices of other energy derivatives at the same time. For example, simultaneously fitting to oil swaptions data would enable us to better fit the covariance structure of the forward contracts. 8.10.1 Handling Volatility Skews and Smiles The importance of the concept of the implied volatility smile was introduced in Chapter 1 and its incorporation into spot price models has been discussed in Chapters 2, 3, 6 and 7. Incorporating volatility smile information into the multi-factor forward curve models of this chapter can also be achieved with varying levels of complexity Probably the simplest approach is via a simple extension to the calibration procedure in the previous section. Once the volatility function(s) have been calibrated we simply compute the required volatility scaling factor for each strike price and option maturity which makes the model price exactly equal to the market price. Interpolation between these scaling factors can then be used to price options with strikes and maturities not observable in the market. However, this approach does not help in the pricing of exotic options consistent with the market prices of the standard options. A simple way to price exotic options is to compute the value of the exotic option in a trinomial tree fitted to the multi-factor model as described in section 8.9.4 and in a implied trinomial tree also fitted to the market prices of standard options. The difference between these two prices can be taken to be the pricing differential due to the volatility smile. This can then be added to the price obtained for the exotic under the full multi- factor model using the techniques in section 8.9. 8.11. Summary In this chapter we have developed a general framework for the pricing and risk management of energy derivatives based on the observed forward curve for energy prices. The framework is designed to be consistent not only with the market observable forward price curve but also the volatilities and correlations of forward prices. We have shown that there is a direct relationship between the forward curve models and the spot price models of Chapter 6. Estimation of the volatility functions of the forward curve model from both historical and cross-sectional market data was discussed including the incorporation of seasonality and volatility smiles. Although based on very general volatility functions, we develop closed-form solutions for standard European options on the spot energy price and forward prices. For options not amenable to analytical solutions we also develop a general Monte Carlo framework for pricing a wide range of European style energy options. The approach can also handle the pricing of exotic or path dependent options, and we illustrated this with examples related to calendar and crack spreads, and Asian options. Finally, we indicated how American options can be priced using a combination of trinomial trees and Monte Carlo simulation 1629 Risk Management of Energy Derivatives 9.1 Introduction Although energy price risk management is in the early stages of development compared to the financial markets, there is an increasing move of energy companies towards regular buying and selling of forwards, swaps, caps, and other option contracts. In this chapter we discuss the day-to-day management of price risk for institutions that trade energy derivatives and who are then faced with the problem of managing the risk of those positions. In Chapters 10 and 11 we look at the risk management issues of calculating Value at Risk and measuring credit risk for energy derivative portfolios. We can think of risk management as the immunisation of risk — this is achieved, for example, by setting up portfolios that contain positions in the underlying energies and energy derivatives in such a way that the portfolio is not affected by small changes in the price of the underlying energy and other key variables. We will see that the sensitivities of components of the portfolio to changes in their valuation parameters provide the key information for risk management. For an energy company that sells an option position the most complete hedge is an exactly offsetting position, e.g., long calls bought against short calls sold. This long position is then held as a static hedge - meaning that the hedge need not be changed through time. However, in most circumstances this hedge may not be practical or profitable. Although it may be impossible to trade the offsetting position in this way, option positions can be created synthetically to give the same payoff. For example, as a first step the hedger makes their derivative portfolio immune to small changes in the underlying energy price in the next small interval of time. This process is known as delta hedging and the position is said to be delta-neutral. With a dynamic hedge the position will be neutral over a short horizon and the hedge must be rebalanced or rolled through time. Often option positions are also gamma and vega hedged, making the portfolio relatively insensitive to large changes in the underlying energy price and to changes in the volatility of the energy respectively. We also look in this chapter how hedging can be accomplished within the full multi-factor model of the previous chapter. 163Energy Derivatives: Pricing and Risk Management 9.2 Delta Hedging of Energy Derivative Positions The process of delta hedging an option involves dynamically trading a position in the underlying energy contract in such a way that, over each small interval of time between trades, the change in the option price is offset by an equal and opposite change in the value of the position in the underlying. In this way the hedged portfolio - consisting of the option and the position in the underlying — does not change in value, and is therefore immune to the risk of the underlying price change. Figure 9.1 shows the relationship between the price of a standard European call option, c(F), on an energy forward contract and the price of that underlying forward. The diagonal straight line at the price F, is the slope of the option price curve and is approximately given by the ratio of the change in the option price Ac for a small change in the underlying forward price AF, centred on the current price. This ratio, Ac/AF, is known as the de/ta of the option and measures the sensitivity of the option price to changes in the underlying (here the forward price). The delta of the option tells us the number of units of the underlying such that the position in the underlying will change by the same amount as the option for a small change in the underlying unit price. Thus in order to offset a change in the option price we must take a position in the underlying which is equal to the negative of the option delta. A specific example will make this clear. Suppose we have written a call option. The delta of a short position in a call option is equal to —Ac/AF and therefore to delta hedge c(F) Ft <—____> AF FIGURE 9.1 Sensitivity of a Call Option Price with Respect to Forward Price 164Risk Management of Energy Derivatives this position we must buy Ac/AF of the underlying forward contract. If P denotes the value of the hedged portfolio, then we have! Ac p=- . chap (9.1) If the forward price were to change by a small amount AF the change in the value of the portfolio will therefore be Ac AP=—Ac+ 7 AF =0 (9.2) The change in the portfolio value is therefore equal to zero and so the portfolio is hedged against small changes in the underlying price and the portfolio is said to be delta neutral. Theoretically, in order for the hedge to be perfect we must consider the changes in F to become very small which leads to AC ac —| =— = del a Bayo OF > lta (9.3) In equation (9.3) C is now taken to be the price of any general energy derivative. Since delta changes continuously as the underlying forward price changes, this implies that we must continuously trade the underlying contract in order to keep our position in the underlying equal to delta and our portfolio hedged. In reality it is not possible or necessary to trade the underlying continuously but only when the underlying has moved by a significant amount. We discuss the issue of the frequency with which the hedge should be rebalanced later. If the delta hedging process is continued to the option expiry date then the hedger is left with an overall position in cash and the underlying which exactly offsets the payoff of the option. If the option finishes out—of-the-money both the cash and underlying position are zero. If the option finishes in the money then the cash and underlying position is equal to the difference between the underlying price and the strike price (ic. the payoff to the option) Consider the simple one-factor forward price model of section 6.2. Recall that the price of a standard European call futures option, c(/, F, K, T, s), with strike price K, and maturity Tis given by c(t, F,K,T,s) = P(t, T)|F(t,s)N(h) — KN(h— Va) (94) where _ nF, us +30 oa e (omen a eto~a) 1 Time subscripts on the forward and option prices have been dropped for ease of exposition. 165Energy Derivatives: Pricing and Risk Management For this model the delta of the option is therefore given explicitly by OC(t,F, K,T,s) Gy 7 PDN) (9.5) For the same standard European call futures option but evaluated under the general multi-factor model of Chapter 8 the delta is given by equation (9.5) with wade{ footw rau} (9.6) i=l Figure 9.2 plots the delta profile for two at-the-money European call forward options on NYMEX crude oil with maturities of 3 months and 6 months on an underlying forward with nine months to maturity. The forward price underlying the option is assumed to be 23.97 which was the price of the nine month forward on 31 March 2000. We use a three factor model with the volatility functions taken from the historical estimation in section 8.6. When the forward price is low relative to the strike price, the delta of the option is close to zero reflecting the low probability of the option finishing in the money. As the forward price increases, the delta of the option increases. For at-the-money options, where the strike price is close to the forward price the delta is approximately 0.5. For high forward prices the delta is close to one as the probability of finishing in-the-money 1.0 0.9 1 08 07 0.6 0.5: Delta 0.4 0.3 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 Forward Price FIGURE 9.2 Delta Profile for European Call Futures Options 166Risk Management of Energy Derivatives rate [xk 0.05} 24 L Volatility Functions | [Date Bivo3/00 03/04/00 t ot o2 03 0.50 0.4918 | 6.0000 0.3257 0.1531 | -0.0751 0.75 0.7418 0.0833 | 0.3076 | -0.1056 | 0.0063 0.1667 | 0.2885 | -0.0778 | 0.0180 0,8) 23.97 23.32 0.2500 0.2729 -0.0554 | 0.0169 0.3333 | 0.2613 | -0.0395 | 0.0189 all -1.6210 -1.2930 0.4167 0.2525 -0.0273 | 0.0231 elta 0.5190 ~0.4576 | 0.5000 | 0.2448 | -0.0151 | 0.0208 0.5833 | 0.2367 | -0,0049 | 0.0165 sition in F 0.5190 0.5190 0.6667 | 0.2298 | 0.0046 | 0.0147 0.7500 | 0.2257 | O.0115 | 0.0147 all 0.3280 0.8333 | 0.2188 | 0.0237 | 0.0083 F -0.3373 | 0.9167 | 0.2143 | 0.0243 | 0.0142 ror. 0.0093 1.0000 | 0.2087 | 0.0322 | 0.0078 FIGURE 9.3 Delta Hedging a European Call Option on a NYMEX Crude Oil Futures Contract is high. Figure 9.2 also shows that the delta becomes steeper as the option maturity decreases as the probability of the option finishing in-the-money becomes more sensitive to small changes in the forward price close to the strike price. In figure 9.3 we show the calculations for delta hedging a short position in a six month (T = 0.5), at-the-money option on a 9 month NYMEX crude oil futures contract (s = 0.75) on the 31 March 2000 using the three factor version of the general multi-factor model. The strike price of the option is set equal to the price of the nine month futures contract (to two significant figures), K = 24.00. The inputs to the model are the observed forward curve on the 31 March 2000 and the three volatility functions derived from historical analysis (columns headed ol, ¢2, and 03). Using equations (9.4) and (9.5) the value of the short option position is ~1.6210 and the delta is —0.5190. This delta implies that the writer of the option should take a long position of 0.5190 units of the nine month futures contract — this is indicated by the “Position in F” row in figure 9.3. Figure 9.3 also shows the performance of the delta hedge over the weekend with the hedge statistics evaluated for the closing price of the nine month futures contract on 3 April 2000. The value of the futures contract decreases to 23,32, the short option position therefore increases to —1.2930. The option position has increased in value by 0.3280 and the position in the futures contract has decreased in value by 0.3373. Therefore the error in the hedge is the difference between changes in the values of the two positions, i.e. —0.0093. Figure 9.4 graphs the delta hedge’s performance for a wide range of future prices. The hedge can be seen to work well close to the current 167Energy Derivatives: Pricing and Risk Management futures price (Dcall denotes the change in the option price with DF denoting the change in the futures price) but declines in effectiveness as the forward moves away from the current forward price. Figure 9.4 helps to answer the question of how often should the delta hedge be rebalanced. The answer is not terms of a time interval but in terms of how much the underlying price has moved from the level at which the hedge was established. Alternatively, we could set the hedge rebalancing points or triggers in terms of an upper limit on the hedging error, The hedge rebalancing trigger points are not the only factor to consider. In practice, every time the hedge is rebalanced, costs are incurred in trading in the underlying asset. These costs, for example bid-ask spreads and liquidity costs, can be significant in energy markets where even the futures markets can have relatively wide spreads and liquidity problems. Efficient hedging therefore requires an appropriate trade-off between risk reduction and trading costs. This trade-off can be analysed by simulating proposed hedging strategies using historical time series of market prices and Monte Carlo simulation. Now consider hedging a standard European swaption, Swpn(t:K,T,AT,m), with strike price K that matures at time T on a swap with m resets of period AT. Recall from Chapter 8 that there does not exist a closed-form solution for the price of a European swaption in the general multi-factor model. Therefore there will also not be a closed form solution for the delta. The natural hedge instruments are the underlying forward contracts of the swap. The delta positions are therefore given by the sensitivity of the 4— - eal i — 0F | +> > Derror er ae 24 25 26 +++. 2 Future Price FIGURE 9.4 Performance of a Futures Option Delta Hedge Over a Weekend for a Range of Futures Prices 168Risk Management of Energy Derivatives option price with respect to the forward contracts which must be evaluated numerically: OSwpn(t; K,T, AT, m) OF (AS) (9.7) For a swap with a large number of reset dates, delta hedging using all of the underlying forwards would be impractical or even impossible if they do not trade on a futures exchange. We show in section 9.5 how to deal with this problem. 9.3 Gamma Hedging The reason for the declining effectiveness of the delta hedge as the underlying price moves away from the level at which the hedge was established can be understood from figure 9.4. The option is a non-linear (curved) function of the futures price whilst the value of the delta hedge, a fixed position in the futures contract, is a linear function Therefore, as the underlying price moves by larger amounts the differences in the value of the option and delta hedge are greater and do not exactly offset. Another way to view this problem is that, as the underlying price changes, the delta of the option changes and so the position in the underlying required to be delta neutral changes — i.e. the delta hedge is sensitive to changes in the underlying asset. This problem is most severe for options whose strike price is close to the current underlying price. As figure 9.2 illustrates, for options which are far out-of-the-money the option price does not change very much when the underlying price changes (delta = 0) and for options which are far in-the-money the option price changes in a linear way (delta = 1). We can solve this problem by neutralising the sensitivity of our delta hedge to changes in the underlying price. In order to do this we need to make the sensitivity of the delta of the hedged portfolio equal to zero. The sensitivity of delta to changes in the underlying asset is known as gamma. The calculation of gamma can be performed in a similar way to delta, as a finite difference approximation or as the second derivative of the option price with respect to the underlying price: CUR+AF)-C(F) C(A)-C(F-AF) AFC. (25588) - COG ) (9.8) gamma = 55 = AF E JAF +0 A low value of gamma implies that the option delta changes only slowly and so a delta hedge would not need to be rebalanced very frequently for it to remain delta neutral. A large gamma implies that the delta changes rapidly for changes in the underlying price, and so the hedge would have to be adjusted frequently. Figure 9.2 showed that the delta for a European call futures option remained relatively unchanged for deep in-the-money and deep out-of-the-money options, but could move quickly for at-the-money options. Figure 9.5 plots the gamma for the same two options as used in figure 9.2. For standard European futures options, gamma can be obtained analytically in an 169Energy Derivatives: Pricing and Risk Management S z Gamma 0.08 0.06 0.04 0.02 10 12 «#414 «16 «18 «620 «22246 BCD 2 34s K_——«B Forward Price FIGURE 9.5 Gamma Profile for European Call Future Options analogous form to the Black-Scholes version: Pet, F,K,T,s) _ Plt, T)n(h) OF(i,sy ——— F(t,8) Se where n(.) is the standard normal density function and w is given as in equation (9.4) for the single factor model of section 8.2 and as in equation (9.6) for the multi-factor model of section 8.4. In order to neutralise the gamma of a portfolio we must use another option since the gamma of a forward or futures contract is zero. Let Crarger represent the value of the option we are trying to hedge (the ‘target’ option) and Ajarger aNd Dyarger the delta and gamma of the target option respectively. Now, let Cheages Mnecige ANd Vpedge represent the price, delta, and gamma of the option which we use to hedge the target option (the ‘hedge’ option). In order for our option position to be simultancously delta-gamma neutral we first choose a position 9 in the hedge option such that the gamma of the portfolio is zero, i.e. we require: Dvarget + BV hedge = 0 (9.10) Equation (9,10) implies that the position that has to be taken in the hedge option to make the portfolio delta-gamma neutral is 8 = —P varget/T hedge: The combined position in the target and hedge options, by construction, has a gamma of zero but a non-zero 170Risk Management of Energy Derivatives rate K 0.05 24 Fri Mon Volatility Functions Date __ 3170300 | 3/04/00 a_ | o | 3 | Target Hedge Target_Hedge 0.3257 |-0.1531 |-0.0751 ir 050 | 0.25 0.4918 | 0.2418 0.3076 | -0.1056 | 0.0063 | 5 0.75 0.75 0.7418 | 0.7418 0.2885 0.0180 | 0.2729 | .0.0554 | 0.0169 F(0,s) 23.97 | 23.97 | 23.32 | 23.32 0.2613 | -0.0395 | 0.0189 | | 0.2525 | -0.0273 | 0.0231 call -1.6210 | 1.0936 1.2930 | 0.7740 0.2448 | -0.0151 | 0.0208 bate 0.5127 -0.4576 | 6.4192 0.2367 | -0.0049 | 0.0165 | amma 0.1399 0.0955 | 0.1436 0.2298 | 0.0046 | 0.0147 | | 0.2257 | 0.0115 | 0.0147 b 0.6591 0.6591 0.2188 | 0.0237 | 0.0083 [Residual Delta| -0.1811 | | 0.2143 | 0,0243 | 0.0142 Position in F O.1811 | 0.1811 0.2087 | 0.0322 | 0.0078 | \scall(Target) 0.3280 call(Hedge) -0.2106 et -OLLT7 ror -0.0003, FIGURE 9.6 Delta-Gamma Hedge for European Call Option on a NYMEX Crude Oil Futures Contract residual delta since, in general, the delta’s of the target and hedge option positions will not exactly cancel. We can neutralise the delta of the combined portfolio of the two options by taking a position in the underlying asset equal to the negative of the residual delta. This delta hedge position will not affect the portfolio’s gamma because the underlying asset has a gamma of zero. Note that the gamma must be eliminated first and then the residual delta. Figure 9.6 shows the calculations for setting up a delta- gamma hedge for the same target option as in figure 9.3 The short position in the target option has a value of ~1.6210, a delta of -0.5190 and a gamma of -0.0922. For the hedge option we use a three month maturity European call with otherwise the same characteristics as the target option. This option has a price of 1.0936, a delta of 0.5127, and a gamma of 0.1399. In order to neutralise the gamma of the portfolio the user buys —(—0.0922)/(0.1399) = 0.6591 of the hedge option. The residual delta of the portfolio is (0.5190) + 0.6591 x (0.5127) = -0.1811. Therefore a long position of 0.1811 futures contracts must be taken to eliminate the delta of the portfolio. Figure 9.6 also shows the performance of the hedge over the weekend. From the 31 March 2000 to the 3 April 2000 the futures price decreases from 23.97 to 23.32 causing the short position in the target option to increase in value by 0.3280. The change in the long hedge position decreases in value by 0.2106 and the long position in the future 171Energy Derivatives: Pricing and Risk Management 4 — Deall(T) 3 =~ Deall(H) — DF 7 22+ Error Future Price FIGURE 9.7 Performance of a Futures Option Delta-Gamma Hedge Over a Weekend for a Range of Futures Prices decreases by 0.1177, leading to a hedging error of —0.0003 to the 4'* decimal place. Figure 9.7 shows the performance of the delta-gamma hedge for a range of values of the underlying futures contract. By comparison with figure 9.4, the delta-gamma hedging error is significantly smaller then the delta hedging error for a wide range of futures prices. Since the hedging error is small for a wide range of futures prices the hedge needs to be rebalanced much less frequently. However, trading costs in options markets are typically much greater than in the futures markets, therefore it is still important to compare the improvement in the hedge gained by gamma hedging with the additional cost involved in trading in the options market. 9.4 Hedging Volatility and Other Sources of Market Risk In addition to the risk due to changes in the spot or forward prices, the other major source of market risk comes from changes in volatility. We have seen throughout the previous chapters that, apart from price, the volatility or uncertainty in the price of the underlying energy is the key determinant of the value of derivatives. Thus, changes in volatility are an important source of risk in a derivatives portfolio. Volatility can be hedged using a similar technique to the gamma hedging described in the previous section. The sensitivity of an option or portfolio to changes in volatility is called vega and 172Risk Management of Energy Derivatives can be calculated in a similar way to delta: ac _ (oe + Aor) = Clor - an) Oat ae Aap (9.11) Aor 0 In order to neutralise vega we must take a position in an additional hedging option such that the vega of the hedging option exactly cancels the vega of the original target portfolio. The steps involved in constructing a delta-vega hedge are identical to those in the previous section for a delta-gamma hedge but with gamma replaced by vega. However, in many cases a trader may want to neutralise delta, gamma and vega. This requires trading in two different hedging options, say ‘hedgel’ and ‘hedge2’. Since options have both gamma and vega, we must neutralise both gamma and vega simultaneously by solving the following two equations: Prarget + Aredger! nedger + Breage! hedge? = 0 (9.12) Vearget + Preaget Vredget + Predger Viedge2 = 0 (9.13) The solution to equations (9.12) and (9.13) is: Thedge2 Viargct — Tranget “hedge? Dredge = (9.14) Theaget Viedge2 — Phedge> Vnedge r, Viedger ~ Ty ee Pig = pte hedge Tete ‘target (9.15) edge! Viedge? ~ Vhecige? Mnedget With these solutions for the positions in the hedging options the residual delta can be calculated to obtain the position required in the underlying energy asset 9.5 Factor Hedging In this section we describe a general approach to hedging a portfolio of energy derivatives. The method is based on the multi-factor model described in Chapter 8. The set of factors, either estimated from historical data or implied from the market prices of options, represent the risks to which a portfolio of energy derivatives is exposed. For example, figure 9.8 (which reproduces figure 8.8) shows the three main factors (volatility functions) for the NYMEX crude oil forward curve. A portfolio of energy derivatives; for example caps and swaptions, is exposed to the tisk that the crude oil forward curve may change by a combination of random proportions of these “shifting”, “tilting” and “bending” factors. In order to hedge the energy portfolio we need to take positions in a set of hedge instruments (typically futures and futures options) such that the hedged portfolio is immunised against the possible changes in the forward curve described by the volatility functions The first step in the process is to work out how the portfolio changes in value if the forward curve were to be shocked by each of the volatility functions separately. The 173Energy Derivatives: Pricing and Risk Management 40% 30% 20% Volatility 8 0% -10%: 20% 0 123 4 5 6 7 B 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Maturity (months) FIGURE 9.8 Three Main Volatility Functions for NYMEX Crude Oil Forward Curve 20.5 -* Original Curve) —-fn tT up ~~ fn 1 dn | oefn 2 up fn 2dn insu inach 20.0 o aw 3 2 Forward Price = & 18.0 175 17.0 a é 10 12 Maturity (months) FIGURE 9.9 Shifting the NYMEX Crude Oil Forward Curve by its Three Main Factors 174Risk Management of Energy Derivatives appropriate shocks to apply to the forward curve are obtained by discretising the multi- factor model of equation (8.4) taking each volatility function in turn: AF(0,8) = F,(0,s)o/(0,8)Az) ; 3 (9.16) The size of the shock Az, should be chosen to give a typical movement of the curve over the hedging period, Az, for example a one standard deviation move ie. VAt is reasonable. We discuss the choice of the hedging interval Ar and the shock Az; later, but for now we take a hedging interval of one week to give relatively large movements in the curve. The curves of changes described by equations (9.16) are applied in both positive and negative amounts to the current forward curve, F(0,s). Figure 9.9 illustrates the result of this for the NYMEX crude oil futures curve on the 31 March 2000, In figure 9.9 ‘fnl up’ and ‘fn1 dn’ denote the original forward curve after an up shift determined by fl and a down shift respectively. The other shifts are defined similarly. The next step in the process is to compute the changes in the value of the portfolio between the downward and upward shifts of the forward curve for each factor. If P represents the value of the portfolio, we have: AP) = PF.) — PFio()) § t= In equation (9.17) the U subscript indicates the up shifted forward curve and D indicates the down shifted curve. The calculation of the changes in the portfolio value require complete revaluation of the portfolio based on the shifted forward curves. The three changes in the portfolio can be hedged using three different forward contracts F(0,s;) choosing appropriate positions ; in these contracts such that the overall change in the hedged portfolio P;; is zero for each factor. That is we solve the following set of equations for the unknown positions /3;: 3 (9.17) AP) = AP; + 8; AFO,s), + B2 AFO.s2); + Bs AF(O,S3): = 0 APi12 = APs + 6, AF(O,5)2 + 8 AF(O,S3)2 + 3; AFUO,s3)2=0 (9.18) AP iz = AP; + 8; AF(O,5))3 + 82 AF(O,52)5 + 83 AF(O,s3); = 0 Equations (9.18) are a set of three linear equations in the three unknowns, 3; ; i = 1,...,3, which can be solved easily and exactly as a matrix equation. An alternative and more general solution method is simply to minimise the sum of the squared hedging errors (AP) minimise { (APi1)’+ (APu2)°+ (APu3) 5182s} (9.19) A specific example will help to clarify the method. We assume the portfolio to be hedged contains a single European one month, at-the-money crude oil swaption on a six month swap with monthly resets. We hedge the three crude oil factors identified in section 8.6 using the nearest, fourth nearest and seventh nearest futures contracts. Figure 9.10 175Energy Derivatives: Pricing and Risk Management shows the calculations for constructing the factor hedge. The hedge valuation date is 3] March 2000, the swaption has one month to maturity (J = 1/12) and exercises into a six month swap with monthly resets (¢7 = 1/12). The floating payments are swapped for a fixed (strike) price (K) of $18.78. The price of the swaption, assuming interest rates of 5% is derived via equation (8.30) to be 0.5817. The boxes headed ‘Up Shifted Curves’ and ‘Down Shifted Curves’ contain the forward curves with the positive and negative shifts represented by equation (9.16). Below these are the revaluations of the portfolio (swaption), P, and the price of the futures contracts to be used as hedging instruments. The changes in the portfolio (AP), az 0.1385 {Volatility Functions Up Shifted Curves: Down Shifted Curves s Fs) | tm m2 ta folup f2up fnSup | | tntdn fr2dn_tn3 dn ssi] 0.0000 | 19.38 | 0.8257 -0.1831 -0.0751 20.2541 18,9690 19.1783} | 186059 19.7910 19.5817 0.0833 | 19.28 | 0.3076 -0.1056 0.0068, 20.1013 18.9979 19.2967| | 18.4587 19.5621 19.263, 0.1667 | 19.12 | 0.2685 0.0778 0.0180 19.8840 18.9139 19.1677) | 18.3660 19.3261 19.0723, #4 0.2500 | 18.97 | 0.2729 -0.0854 0.0169 19.6869 16.8243 19.0143] | 18.2631 19.1187 18.9267, 0.3393 | 1882 0.2613 -0.0395 0.0189 19.5011 187170 18.8692) | 18.1389 18.9230 18.708) 0.4167 | 18.69 | 0.2525 -0.0273 0.0281 19.9436 16.6193 18.7497] | 16,0364 18.7607 18.6303, 0.6000 | 1869 | 0.2448 0.0181 0.0208 19.2203 18.8512 18.6436| 117.9697 18.6286 18.5364 0.5833 | 1849 | 02367 -0.009 0.0165 19,0961 18.4776 18.5323) | 17.8839 18.5024 18.4477, 0.6867 | 1839 | 0.2208 0.0048 0.0147 18.9753 18.4018 16.4274) | 17.8047 18.3762 16.3526 0.7500 | 1830 | 02287 0.0115 0.0147 12.8719 18.3201 18.9374] |17.7281 18.2700 18.2626 0.6393 | 18.21 | 02188 0.0237 0.0083 | 18.7617 18.2698 18.2908 | 17.6583 18.1502 18.1692] o.o1e7 | 1813 | 0219 0.0043 ora | 18.6680 18.1910 18.1658 |17.6920 18.0890 18.0944 1.0000 | 18,08 | 0.2087 0.0322 0.0078 48.5819 18.1405 18.0706] [17.5381 17,9795 16.0404) Portfolio P| sss: 0.5334 0.6074] |o2967 0.6324 0.5666 | Swaption F(0,s1) [20.2541 16.9690 19.1783] [18.6059 19,7910 19.6817, 1878 F(0,s2) | 19.6869 16.6043 19.0143| [18.2591 19.1157 18.9257 r ocass | rina |io 72095512 vasean| [7.7 vasoes 1st 0.0833 fat 0.05, fol tn2_tma ap _| 0.6914 -0.0990 0.0508 eo 0.5817 AF(s1) | 1.7483 -0.8220 0.4034 AF(s2) | 1.4338 -0.2915 0.0866 ‘AF(s9) | 1.2606 _-0.0777_0.1074 apH_| 0.0000 0.0000 0.0000 beta_| 0.0029 0.2881 0.2048 FIGURE 9.10 Delta Factor Hedging a European Swaption 176Risk Management of Energy Derivatives pate 31/03/00 0.1386 Volatility Functions. Up Shifted Curves ‘Down Shifted Curves s F (0,8) fot tn2 fn3 foiup fm2up__ fn3 up fntdn_fm2dn_tn3 dn (sf 0.0000 | 19.38 | 03257 -0.1691 0.0751 20.2541 16.9690 19.1783] | 18.5059 19.7910 19.5817 Ti] 0.0893 | 1928 | 0.3076 -0.1056 0.0083 20.1013 18.9979 192907| | 18.4687 19.5621 19.2635 0.1667 } 19.12 | 0.2885 -0.0778 0.0180 19.8840 18.9139 19.1677 18.3560 19.3261 19.0723 82 0.2500 | 18.97 | 0.2729 -0.0554 0.0169 19.6869 18.8243 19.0143 18.2531 19.1157 18.9257 0.3333 | 18.82 | 0.2613 -0.0395 0.0189 19.5011 18.7170 18.8692 18.1389 18,9230 18.7708 ote? | 18.69 | 0.2525 0.0273 0.0231 19.9436 18.6193 12.7497) | 18.0964 18.7607 18.6909 #4 05000} 1.59 | 02448 -0.0151 0.0208 19.2203 18.6512 12.6436] | 17.9607 12.6288. 10.5964 0.5833 | 18.49 | 0.2367 -0.0049 0.0165 19.0961 18.4776 18.5323 17.8839 18.5024 18.4477 0.6867 | 18:39 | 0.2298 0.0048 0.0147 12.9759 18.4018 10.4274] | 17.2007 12.3762 10.9526 07500] 183 | 0.2257 0.0118 0147 ve.a719 18.3201 18.3374 |17.7281 182709 18.2626] 0.8303} 18.21 | 02188 0.0237 0.0088 12,7617 18.2698 12.2008] | 17.6585 18.1502. 18.1892 ote7 | 18.13 | 02143 0.0243 aor4e 12.6580 18.1910 12.1655] | 17.5920 14,0600 18.0944 1.0000 | 18.06 | 0.2087 0.0322 0.0078 18.5819 18.1405 _ 18.0796. 17.5381 _17.9795 18.0404 | Porttotio K t | posse: 05334 0.6074] [0.2967 0.6324 0.5566 | Swaption Fio.st) |20.2541 18,9600 19.1783) [18.5059 19.7910 19.5817! k 18.78 "'ros2) | 19.5969 18.0243 19.0143] | 18.253) 19.1157 18.9057 Tr 0.0833, F(0,83) | 19.2203 18.5512 18.6436 17.9597 18.6288 _ 18.6364 iT 0.0833 19.12 | 0.0833 |¢(K1,T1)| 1.1435 0.5831 0.7094 0.3579 __0.7957__ 0.6602 6 18.97 | 0.1667 |etka.t2)| 13368 0.8475 0.9447 | | 0.5926 0.9990 0.8986 ate 0.05, 18.69 | 03999 {oxa.ra)| 1.5703 1.1645 1.2927 | | 0.8965 1.2986 1.1702 po os817 | fntup fdup_tndup| [tnidn fndn tnd dn aP_| 0.4065 -0.0463 0.0267 | [0.2860 0.0508 _-0.0261 ctk1,11)| 0.6845 arist) | a.741 0.4110 -0.2017| [-og7a1 0.4110 0.2017 e(K2,72)| 0.9215 aF(s2) | 07169 _-0.1457 0.0443) |-0.7169 0.1457 0.0443 | o(k3,T3)] 1.2012 AF(s3) | 0.6303 0.0388 00536 | [0.6303 0.0388 0.053 c(t) 0.1014 0.0249) |-0.9266 0.1112 _-0,0243 ‘sc(2) 0.0740 0.0232 | [-0.9289 0.0776 0.0229] ‘sc(@) -0.0367 0.0315 | [.0.3148 0.0373 _-0.0311! APH | 0.0000 0.0000 0.0000 0.0000 __0.0000 0.0000 | beta [-0.0199 0.3250 0.1831] [01584 0.9190 -1.1609 | FIGURE 9.11 Delta-Gamma Factor Hedging a European Swaption the hedge instruments (AF(.)) and the hedged portfolio (APH) are shown towards the bottom centre of figure 9.10. In the box labelled ‘beta’ is the solution for the positions to be held in the hedge futures such that the changes in the hedged portfolio are zero for all three factor shifts. The approach we have just described can be seen as a general form of delta hedging. It 177Energy Derivatives: Pricing and Risk Management os - - - 4 == No hedge 04 + Delta hedge 0.3: "| Change in Hedged Portfolio Value ° 400: WSN SSSSNRRESSS SSPRRRSSERRERGBESSS ‘Simulation Number FIGURE 9.12 Performance of Factors Hedges allows us to delta hedge the factors driving the entire forward curve. The method suffers from the same problem as the simple delta hedge discussed in section 9.2. Since we are hedging a portfolio of derivatives, which change in a non-linear way with respect to the forward curve, with instruments that have linear payoffs, the hedge is only good for small changes in the forward curve. We can assess the hedging error in a similar way to that used in sections 9.2 and 9.3, by varying the size of the changes in the forward curve via Az. Using this approach we can determine how far we can allow the forward curve to move, in terms of Az, before the factor hedge must be rebalanced. This also indicates the size of Az which should be used in calculating the factor hedge. ‘The factor delta hedge can be improved ina similar way as for the simple delta hedge ~ by using standard European futures options to gamma hedge the factors. Continuing with our previous example we now introduce as hedge options three European futures at-the-money call options with maturities 7), T, and T; and strike prices Ki, Ky and K3. Figure 9.11 shows the calculations for the delta-gamma factor hedge. We use the same six shifted forward curves but we now consider the shifts relative to the current forward curve. This gives six changes in the portfolio and the six hedge instruments (three futures and three futures options). We compute the six positions, 3; ; i=1,...,6, in these hedge instruments so as to minimise the sum of the squared changes in the hedged portfolio. To demonstrate the performance of both the delta and delta- gamma factor hedges, figure 9.12 shows the changes in the value of the hedged portfolio for 100 simulated changes in the forward curve under the combined three factors for the cases of no hedge, the delta factor hedge and the delta-gamma factor hedge. 178Risk Management of Energy Derivatives Over these 100 simulations the unhedged portfolio changes range between ~31 cents and +42 cents, narrowing to -0.01 cents and +7 cents for the delta hedged portfolio and -0.08 cents and +0.05 cents for the gamma hedged portfolio 9.6 Summary In this chapter we discussed the risk management of portfolios that contain energy derivatives. Firstly, we introduced the basic concepts of delta, gamma and vega hedging for a single option position. We then described how the multi-factor forward curve model developed in Chapter 8 could be used to generalise the delta and gamma hedging methods to handle a complex portfolio of energy derivatives. This factor hedging approach was illustrated with the example of hedging a crude oil swaption using futures and futures options. Finally, we illustrated how the effectiveness of the factor hedge could be assessed via Monte Carlo simulation. 17910 Value at Risk 10.1. Introduction One of the key risk management concepts in use in financial markets today is Value-at- Risk (VaR). Value-at-Risk is a measure of the maximum potential change in value of a portfolio of financial instruments with a given probability over a pre-set horizon. VaR answers the fundamental question facing all risk managers: How much can I lose with a given probability over a chosen number of days? The concept of VaR is a very appealing one ~ it can be developed for any kind of portfolio and can be aggregated across portfolios containing different kinds of instruments. One of the major advantages of VaR is that it has become a widely accepted industry standard with support from such bodies as the Group of Thirty, The Bank of International Settlements, and the European Union. Our objective in this chapter is to give the reader sufficient understanding of the implementation of the main VaR approaches as they relate to energy markets to be able to build a basic VaR measurement system. We begin our discussion of VaR, in the following section, by introducing the basic concepts of VaR. In the following two sections we look at the basic model of energy market variables returns used in VaR and methods of forecasting the key parameters from historical data - the volatilities and correlations of energy spot and forward prices. We then go on to describe in detail the implementation of the four main VaR approaches — delta, delta-gamma, historical simulation and Monte Carlo simulation. For each of these methods we give a numerical example of calculating the VaR for an energy derivatives portfolio containing crude oil and natural gas futures options. The testing of the accuracy of the VaR forecasts is the subject of the penultimate section and the final section contains concluding remarks. Throughout this chapter we reference the RiskMetrics' Technical Document (J. P. Morgan (1996)) which is an excellent source of further information on Value-at-Risk methods. We also recommend Duffie and Pan (1997) and Shimko (1997a, 1997b). 10.2 What is Value at Risk? VaR is a measure of the level of loss which would not be expected to be exceeded over a chosen time period (the VaR horizon which we denote 7;) with a chosen confidence level (or probability which we denote p) under normal market conditions. Let us consider the | RiskMetrics is a registered trademark of J. P. Morgan. 180Value at Risk key components of this statement carefully. VaR is stated as a level of loss for a chosen confidence level and horizon — e.g. USD 2 million over ten days at 95%. The VaR is quoted in terms of percentage loss. This means that we would expect our loss on the portfolio, over a ten day period, to exceed USD 2 million only 5% of the time or once in every twenty independent ten day periods. Note that the VaR figure is not an absolute worst case or maximum loss and that the VaR is specific to a particular portfolio, the horizon and the confidence level chosen. In particular the VaR will be greater for higher confidence levels — the reasons for this will become clear later in the chapter. The calculation of the VaR implicitly assumes that the portfolio can be liquidated at the VaR level without additional costs, this implies normal market and trading conditions. If the market has just undergone a large correction then the costs, in terms of spreads, of liquidating any reasonably sized portfolio could be significant. Thus we can re-state the definition of VaR as the loss that is expected to be exceeded with probability I-p ina T, year period under normal market conditions. Underlying VaR is the concept of a probability distribution of portfolio values at the VaR horizon. Figure 10.1 illustrates a probability distribution for a portfolio over a ten day horizon which has a current value of $10 million and a VaR of $2 million at a confidence of 95%. The probability, or confidence, of getting a particular range of values for the portfolio is given by the area under the curve in Figure 10.1 — this is roughly the portfolio value 0.4 0.35 0.15 O41 0.05 5 6 7 eneseeonneenag 1 12 13 14 15 Portfolio Value (millions) FIGURE 10.1 ilustration of a Typical Portfolio Distribution 181Energy Derivatives: Pricing and Risk Management range multiplied by the average probability over that range. Since we are interested in losses above a certain level, we are interested in the far left end of the curve which is called the left-hand tail. The area under the curve from its left-most point (zero) up to $8 million (that is the $10 million current value less a $2 million loss) is approximately 5% (this is called a percentile) and therefore we are 95% confident we will not get a loss worse than $2 million in the next ten days with this portfolio. Because VaR looks at the risk of the entire portfolio, it is an integrated way to deal with different markets and different sources of risk. It allows all these different factors to be combined into a single number — one which is a good indicator of the overall risk level. However, it is not a sufficient risk measure for risk management and risk control. It is also important to emphasise that VaR is only an estimate of the potential loss. One of the great strengths of VaR is that it provides a ‘global’ view of risk, and as such has the potential to be used as the basis of integrated risk management. An energy trading firm may simultaneously trade many energies such as oil, gas, and electricity as well as having cash and foreign exchange positions. Because of the substantial correla- tion between energy prices, the overall risk to the firm should be much less if all the energies are treated together in one book, rather than divided into separate books which are treated individually. VaR provides a way to net the price risks across different energy markets and so achieve more efficient controls — position limits, capital requirements, etc. New deals can be assessed for their impact on the firms total risk using the VaR approach (see Garman (1997) for efficient ways to achieve this type of analysis). Finally, a main use of the VaR value is as an estimate of the level of capital required to support the trading or other operations of the company. In this way it can be used to assess the efficiency with which the capital is being used (see Shimko (1997a)). 10.3 Modelling the Market Variables The first step in evaluating the risk of a portfolio, as with pricing and hedging, is to define how the underlying market variables - energy prices, interest rates, exchange rates, etc. might change in the future.” This requires that we specify a model for these market variables. The model used in the basic VaR methods is that market variable returns are normally distributed, also called the random walk model — we discuss the shortcomings of this model in the sections on historical and Monte Carlo simulation and in our conclusions in section 10.9. The simple mean reverting spot price model of chapter 2, the models of chapter 6 except for the model which incorporates jumps, and the general multi-factor model in chapter 8 are consistent with the assumption of normally distributed underlying market variables®. As usual we define returns as the change in the natural logarithm of the market variable over a certain period of time, ie. if Sar 2 In this chapter we will focus on energy spot and forward prices as the underlying market variables. > Under the mean reverting spot price model, stochastic convenience yield and/or stochastic interest rates and. multi-factor forward curves models, both spot prices and forward or futures prices are normally distributed. However, we must be careful that the volatility is calculated in an appropriate way — we indicate this at the appropriate points in this chapter (see also Chapter 3), 182Value at Risk and S, are the market variable values at times t — Ar and ¢ then the return, which we denote by x,, is given by: x, = In(S,) ~ In(S,.5,) (10.1) The random walk model can be represented as follows: X= Hy b oe5 er ~ ID N(O,1) (10.2) This simple model has two parameters associated with each individual market variable - the mean js, and variance o,? of the normal distribution. The mean is usually assumed to be zero as it does not generally have a large effect on VaR estimates*. The square root of the variance is the standard deviation which is often referred to as the volatility. It has been claimed in numerous articles that this model is inappropriate for commodities and energies because it does not allow for mean reversion. This is only true if the model is interpreted in a naive way where the underlying market variabes are forward prices with fixed maturity dates. We saw in chapter 8 that forward prices with fixed maturity dates have increasing instantaneous volatility as they approach maturity and that this is directly related to the mean reversion in the spot price. However, forward prices of constant time to maturity will generally have approximately constant volatility®. Furthermore, we saw in chapter 8 that the volatility of a constant maturity date forward price over a specified time period, in the sense of the standard deviation of the return distribution, is given by the time average of the instantaneous forward price volatility over that time period. Thus equation (10.2) can even be applied to fixed maturity date forward price returns if we compute ¢, as the average volatility over the range of the maturities which the forward will have in the time period Av. We highlight these points again in the relevant sections of this chapter. The model represented by equation (10.2) also has a parameter associated with each pair of market variables, the correlation, which measures how closely the two variables move together. A correlation of 1 indicates the variables move exactly together, a correlation of zero indicates there is no association between their movements and a correlation of —1 indicates the two variables move in exactly opposite directions. In the following sections we discuss how these parameters can be estimated from historical market data. 10.4 Forecasting Standard Deviations (Volatility) Since the VaR estimate is a forward looking estimate we need to forecast the values the parameters of our market model should take for the chosen VaR horizon. The simplest and most widely used method of forecasting the standard deviation of returns is the Simple Moving Average (SMA) method (also known as the rolling (sample) standard * See RiskMetrics Technical Document, Chapter 5, Section 5.3, page 90. 5 Neglecting the time varying component which we associate with the time varying spot volatility. 183Energy Derivatives: Pricing and Risk Management deviation). Imagine we have a history (or time-series) of returns of a market variable, for example a spot price or a constant maturity forward or futures price®, x;, i= 1,...N, with x, being the oldest return and xy being the most recent return. The SMA estimate of the standard deviation & (where * signifies an estimate of the parameter @) is given by: (10.3) This method of estimation has the disadvantage that all the returns, used to compute the estimate, have equal importance or weight. Thus a large return, which may be several months old, will affect the estimate just as much as a large recent return, even though the older return is associated with an event whose effect would have declined significantly. The Exponentially Weighted Moving Average’ (EWMA) estimator solves this problem by giving older returns exponentially less weight. The EWMA estimator is given by: (10.4) (10.5) Substituting the approximation® in equation (10.5) into equation (10.4) gives the formula for the EWMA used in RiskMetrics: = {e —r) Stes — py (10.6) The EWMA estimator depends on the parameter 4(0 < < 1) which is called the decay factor and which determines the relative weight given to older returns. Figure 10.2 illustrates the weights which are applied to each sample return for a range of decay factors. For decay factors close to 1, for example 0.99, the weighting on each return declines only slowly giving recent returns roughly the same weighting. For lower values of \ the © It is important that we use forward or futures prices which have a constant time to maturity because, as we saw in Chapter 8, apart from seasonal effects, their volatility would be expected to be approximately constant. ? This is the method used by RiskMetrics (see the RiskMetrics Technical Document, Chapter 5). ® Note that this approximation is only good for typical values of A. As \ approaches 1 the approximation fails 184Value at Risk a = jot Lambda 0.09 08st] f > 0.93 1 = 0.95 (haan + 0.97 0.07 ~~ 0.99 Weighting Factor 100 «9080 70 60 50 40 30 20 10 0 ‘Age of Return (periods) FIGURE 10.2 EWMA Weighting Factors for a Range of Typical Decay Factors (\) weightings decline much faster resulting in more recent returns contributing relatively more to the estimate. Note also that the number of returns which effectively contribute to the EWMA estimate depends on A, for a value of 0.91 the weights beyond the 100” return are less than 10-°, whereas for \ = 0.95 the weights only fall below 10° at the 170" return. Figure 10.3 illustrates the difference between the SMA (using a 30 day sample size) and EWMA forecasts (A = 0.94) on Brent Crude Oil spot prices for 1998 Notice how both the SMA and EWMA forecasts jump on the 23 March 1998 due to the spot price going from $13.04 to $14.80 per barrel. The SMA forecasts then stays roughly constant for the following 30 days until the large return drops out of the 30 day moving window at which point the SMA forecast drops suddenly back to a lower level. In contrast the EWMA forecast slowly declines from the peak on the 23 March down to the lower level. The level of the decay factor used in figure 10.3 is that recommended by RiskMetrics for daily time series Another attractive feature of the EWMA estimator is that it can be written recursively in terms of the previous days estimate if we assume that the sample mean of returns is zero. Thus the estimate on date /+1 is given in terms of the estimate on date i by ee pre? (1 d)x? (10.7) The decay factor \ determines how fast the effect on the estimate of large returns, such as that on the 23 March in figure 10.3, declines. How should the decay factor be chosen? 185Energy Derivatives: Pricing and Risk Management 80% rent 30d SMA Brent EWMA | 70% 60% 50% Volatility Estimate 40% 30%. 20%. 04/0218 06/04/98, 06/06/98, 06/08/98 06/10/98 06/12/98 Date FIGURE 10.3 Comparison of a SMA and EWMA Standard Deviation Forecasts on Brent Crude Oil Spot Prices for 1998 Riskmetrics uses 0.94 for daily forecasts and 0.97 for monthly forecasts”. These optimal decay factors are obtained as the weighted average of the optimal decay factors for each market variable in the RiskMetrics data set. The optimal decay factors for each market variable are obtained by minimising the Root Mean Squared Error (RMSE) between the EWMA variance forecast for period i+ | (given returns up to and including period i) and the actual squared return for period i+1. Table 10.1 illustrates the results of computing the optimal decay factor for Brent crude oil and Henry Hub natural gas daily spot price returns for the periods from January 1995 to February 1999 and January 1998 to February 1999. The optimal daily forecast decay factor for crude oil is much higher than the RiskMetrics value. Furthermore, the optimal daily forecast decay factor for natural gas is very different for the two periods. This is primarily due to the period of very high volatility in natural gas prices in the first quarter of 1996. This analysis indicates that it is important to independently obtain the optimal decay factors for energy data and not use the RiskMetrics standard values. The time varying or seasonal behaviour of volatility is significant in many energy markets and it is therefore important to not only compute the optimal decay factor for each time series over a number of seasonal cycles but also to interpret the results. The RMSE indicates the average error in the EWMA variance forecast over the chosen e.g. A = 0.99 gives weighting factors up to approximately 20% different form the correct values. 186,Value at Risk TABLE 10.1 Optimal EWMA Decay Factor for Crude Oil and Natural Gas (Computed by Minimisation of the RMSE of the Variance Forecast) Brent Crude Oil Henry Hub Natural Gas Data Set RMSE a RMSE a Jan 95-Feb99 ——_(0,001162 0.971131 0.056728 0.779144 Jan 98—-Feb 99 0.001861 0.979190 0.010669 0.915702 period. If this error is a significant fraction of the actual average variance for the period this indicates that a single decay factor may not appropriate for the full seasonal cycle and one should look at estimating separate decay factors for the different “seasons”. 10.5 Forecasting Covariances and Correlations The SMA and EWMA estimators can be used to construct covariance and correlation forecasts in a similar way to the forecasts of volatility. Imagine we have two time series of returns x),,and x2,;, i= 1,... N, for two different market variables, for example crude oil and natural gas spot prices. The covariance forecast for the SMA and the EWMA are given by: SMA: v : SE Gaye = 1) 224 = oa) (10.8) EWMA: (10.9) x a 1 r= (1-A) ON My ~ a) ay - Assuming 1) and j2; are zero, the EWMA covariance forecast can also be written in recursive form: Faeer = AGtay + (1 Almaare, (10.10) The standard definition of correlation is the ratio of the covariance of the two time series to the product of the standard deviations of the individual time series. The correlation forecast is therefore given by: (10.11) 187Energy Derivatives: Pricing and Risk Management 50% — Brent - Nat Gas SMA Cov 40% — Brent - Nat Gas EWMA Cov 30% 20% Covariance 10% 0%: 10%. -20% 01/01/95 01/07/95 01/01/96 01/07/96 01/01/97 01/07/97 01/01/98 01/07/98 01/01/99. Date FIGURE 10.4 Comparison of a SMA and EWMA Covariance Estimators on Brent Crude Oil Spot Prices and Henry Hub Natural Gas Spot Prices for 1998 Figure 10.4 illustrates the SMA and EWMA estimators of the covariance between Brent crude oil spot prices and Henry Hub natural gas spot prices for 1998. Notice, in figure 10.4, how the EWMA forecast is more stable than the SMA forecast. Figure 10.5 shows the corresponding correlation forecasts. 10.6 Value at Risk Methodologies In this section we describe the four main methods for computing VaR; Delta, Delta- Gamma, Historical Simulation and Monte Carlo Simulation. For each method with give a detailed numerical example using the same example portfolio containing an oil and natural gas futures option. Our objective here is to describe the key implementation steps and to highlight the pros and cons of these main approaches to VaR. The two main steps for calculating VaR are: — All positions must be marked to market (valuation) — Estimate the future variability of the market value (risk evaluation). The following simple examples illustrate the basic approach to computing VaR and will aid in establishing the key concepts. The examples use standard deviations and correlations of market variable returns and the assumption that these returns are normally distributed, 188,Value at Risk Correlation 01/0195 01/07/95 01/01/96 01/07/96 01/01/97 01/07/97 1/01/98 0/07/98. 01/0199. Date FIGURE 10.5 Comparison of a SMA and EWMA Correlation Estimators on Brent Crude Oil Spot Prices and Henry Hub Natural Gas Spot Prices for 1998 Example 1: A Single Underlying Market Variable - The Crude Oil Price Consider a USD-based energy company with a USD 100 million spot crude oil position. They wish to calculate their VaR over a 1-day horizon such that there is a 5% chance that the realised loss will be greater than the VaR level (i.e. a confidence level of 95%). The 5% probability level may be a choice made by the energy company which reflects its appetite for risk or may be imposed by regulatory requirements. The first step in the calculation is to compute the exposure to market risk (i.e., mark-to-market the position). Asa USD- based company, the exposure is equal to the market value of the position in the base currency i.e. USD 100 million. The next step is to evaluate the risk of the position. In order to measure the risk we need an estimate of how much the crude oil price can potentially change over the selected horizon (in this case 1 day). The standard deviation of the daily returns of the crude oil price, measured historically, provides a measure of the size of the possible daily movements", {Imagine that the company calculates the daily standard deviation of the crude oil price to be 2.5%. Assuming that the returns on crude oil are normally distributed then, for the chosen confidence level of 95%, the VaR is given by ~1.65 ° See section 5.3.2 of the RiskMetrics Technical Document, Fourth Edition. ‘0 By computing the historical standard deviation over the same returns period as the VaR horizon we are 189Energy Derivatives: Pricing and Risk Management times the standard deviation'', that is ~1.65 x 2.5% or ~4.125%. This implies that the crude oil price would not be expected to drop more than 4.125% in one day 95% of the time. In USD, the VaR of the position is equal to the market value of the position times the VaR percentage i.e. $100 million x —0.04125 = —$4.125 million. This means that 95% of the time, or 95 days out of 100, the energy company would not expect to lose more then $4,125 million in that 24 hour period. Example 2: Two Underlying Market Variables — The Crude Oil Price and the USD/ BOL FX Rate Consider a Venezuelan Bolivar (BOL) based energy company with a USD 100 million spot crude oil position, What is the VaR over a 1-day horizon such that there is a 5% chance that the realised loss will be greater than the VaR level (i.e. a confidence level of 95%)? The exposure is the same as in the previous example but the company now also has foreign exchange risk. In order to evaluate the risk of the position we first calculate the VaR of the individual positions. The crude oil risk is the same as in the previous example but now we state it in the new base currency by converting at the current exchange rate of 500 BOL/USD giving 500 x —4,125,000 = BOL —2062,500,000. Imagine that the company calculates the daily standard deviation of the BOL/USD exchange rate to be 0.5%. Assuming that the returns on the exchange rate are normally distributed then the VaR for the foreign exchange position is —1.65 x 0.005 x 500 x 100,000,000 or BOL —412,500,000. Now, the total risk of the crude oil position is not simply the sum of the spot oil price risk and the FX risk because of the correlation between the return on the oil price and the exchange rate. If the correlation is estimated to be —0.25 the total risk of the position is given by: Total VaR = \/VaR%,, + VaRzy + 2pon.rx VaR VaRpy Total VaR = 2062, 500, 000? + 412, 500, 000? + 2 « (—0.25) * 2062, 500, 000 « 412, 500, 000 Total VaR = BOL 1999,668,000 This is the simplest method of computing VaR and is called Variance-Covariance or Delta VaR. The application of this method to portfolios containing derivatives is described in detail in the next section. In general, an energy portfolio will generate cash flows at a large number of different future dates. If we were to represent these cash flows explicitly we would potentially need a very large correlation matrix whose size would vary depending on the particular portfolio. Furthermore, there is generally a limited set of reliable correlation information (Le. liquid forward/futures prices) in many energy markets. Therefore the implicitly taking into account the effect of mean reversion and also approximately jumps (see Chapter 8). 190Value at Risk RiskMetrics datasets provide volatility and correlation information for a predefined set of standardised maturities called the RiskMetrics Vertices ~ 1, 3, 6, 12 months, 2 to 5, 7, 9, 10, 15, 20 and 30 years. In RiskMetrics, cash flows not falling on these maturity dates are handled by mapping them onto the standard maturities. The mapping process involves splitting a cash flow into two equivalent cash-flows on the two standard maturities either side of the actual cash-flows maturity. Equivalence is defined by the two cash-flows having the same market value and risk (volatility) as the original cash flow (see Section 6.2, RiskMetrics Technical Document). An alternative is to use the volatility function or eigen vector representation of the volatility and correlation information. This allows us to obtain the volatilities and correlations for any pair of forward/futures contracts by interpolating from the volatility functions (see chapter 8 for details). The approach of only explicitly representing a limited number of maturities is important in practice for reducing the amount of data which must be analysed and stored in a VaR system. In the following sections we describe in detail the implementation of Delta, Delta- Gamma, Historical Simulation, and Monte Carlo Simulation. 10.6.1 Variance-Covariance (Delta VaR) The Variance-Covariance approach is that used in the two examples in the previous section. It assumes that the returns for all the market variables (e.g. energy prices, exchange rates, interest rates) are normally distributed, With this assumption the future possible values of the market variables can be captured by the Variance-Covariance Matrix (VCM) of the market variables returns. The VCM is computed from historical returns of forward/futures price data as described in section 10.4 (the RiskMetrics datasets gives this information for a wide range of key market variables although the energy market information is limited). Derivatives are handled by representing them in terms of a Delta equivalent position in the underlying asset (the Delta-Gamma VaR approach described in the next section is more accurate but more complex to imple- ment). The major advantage of this approach is that in many markets it is widely used and accepted as the basic method for evaluating VaR for portfolios without a large component of options. It is also the simplest method to understand and implement, and is the least computationally demanding. The disadvantages are that the underlying model of the market is not very good, particularly for energy markets (see section 10.9 for a discussion of the problems with the assumption of normally distributed returns) and it is not suitable for portfolios of options. Also, the volatilities and correlations are based on past history and may not reflect the future behaviour of the market (in fact this is true of all the main VaR approaches — we discuss a way to correct this problem in the concluding section 10.9). However, it is useful to study the details of Delta VaR before looking at the more advanced and complex methods. The first step is to represent the portfolio changes in terms of the changes in the underlying market variables which we are modelling, 191Energy Derivatives: Pricing and Risk Management Representing the Porifolio Returns in Terms of the Underlying Returns Consider a portfolio with value P containing N instruments of value V;, i= L,..., N. The value of the portfolio is simply the sum of the individual instrument values: PHVi + Wot... by (10.12) The change in the value of the portfolio, AP, over the VaR horizon is given by AP =AV, + AV, +... AVy (10.13) where AY; are the changes in the value of the individual instruments over the VaR horizon. The return on the portfolio is therefore given by APIP = (AV, + AVa +... AVN) + V2 +... Vy) (10.14) If the instruments are all underlying market variables (spot or forward/futures prices) then V, = S; and we can write the portfolio return in terms of the returns on the individual instruments as follows: AP/P = wy X1 + w2X2 +... WH XW (10.15) Where the returns are given by the approximation!” x, = AS{/S'and w; = S,/P are called leverage weights (see also Section 6.3.1.2, RiskMetrics Technical Document). Note that the portfolio return is not simply equal to the sum of the returns of the individual instruments (x;). If the instruments are derivatives or non-linear instruments then we represent the change in the value of the instrument in terms of a delta (34) position in the underlying market variable: OV Ave AS 10.16 as (10.16) The return on the portfolio becomes ov, OV; AP/P = oe wisi + ag wake + : (10.17) '! We call this NSD — the number of standard deviation for a particular confidence level. '2 The length of the return interval for which the approximation is good depends on the level of spot volatility, the rate of mean reversion and the volatility of long maturity forward prices, For typical energy parameter values, for example spot volatility of 35%, mean reversion rate of 10 and long maturity forward volatility of 192Value at Risk Computing Delta VaR We now have the portfolio return represented in terms of a lincar combination of underlying market variable returns. Since we are assuming that the returns are normally distributed we can write down the standard deviation of the portfolio returns in terms of the standard deviations of the underlying market variables returns. The VaR estimate is then simply the portfolio standard deviation multiplied by the number of standard deviations (NSD) appropriate to the chosen confidence level. The portfolio standard deviation is given by op = DRET (10.18) av, av, Vy sh 2 Wa2,. BS, "NON (10.19) {1 pia Pw ee tigy R=| Pan (10.20) (PN. PN2 1 | Where the o; are the volatilities of the underlying market variables'®, the p;,; are the correlations between the market variables and the superscript 7 denotes the transpose of the matrix. The VaR estimate for the portfolio is then simply given by VaR = NSD op (10.21) Example: Calculation of Delta VaR for a Portfolio of Two Options on Crude Oil and Natural Gas Futures Consider a portfolio containing a one year at-the-money option on a two year oil future and a three month at-the-money option on a six month natural gas future on the 30 March 2000. We wish to calculate the Value-at-Risk for a one working day horizon with a confidence level of 95%. We assume interest rates are flat at 5%. Using our standard. notation, with subscripts 1 and 2 denoting the oil and gas contracts respectively, we have the following: K, = 21.21, T; = 1.00, s; = 2.00, Fy = 21.21 Ky = 2.99, Tz = 0.25, 52 = 0.50, Fy = 2.99 T), = 0.003968, p = 95%, NSD = -1.65 We fit a three factor version of the multi-factor forward curve model described in chapter 8 to two years of historical NYMEX crude oil and natural gas futures data from | April 20% the approximation is usually good for intervals up to a month and is certainly good for daily returns. ™ We note again that the volatilities must be computed to account for the variation of forward price return. 193,Energy Derivatives: Pricing and Risk Management 1998 to 31 March 2000. Using this model the prices and deltas of the two options are: Cy = 0.6981, Cp = 0.1246 Delta; = 0.4921, Deltay = 0.5146 The leverage weights are therefore calculated to be: Ww = 25.77923, ws = 3.630484 The standard deviations of the two underlying forwards at 7), and their correlation are: 01 = 0.004461, 7) = 0.012015, pio = 0.254295 We then evaluate the © vector and R matrix via equations (10.18) and (10.19): & = (0.056589, 0.022449} 1.0 0.254295 R= 9.254295 1.0 Which gives the standard deviation and VaR of the portfolio op = 0.065972, VaR = —1.65 x 0.065608 = —0.108514 Figure 10.6 shows the computed portfolio return distribution, which is simply normally 8, a a aaa aaaseuitsnausee aioaaisuuesaue I | \ 7 i \ 8) i a wo “9 Probability Density S 3 02 “04 0 o4 02 03 Return FIGURE 10.6 Delta VaR Example Portfolio Return Distribution 194Value at Risk distributed with a standard deviation of 0.065608, for comparison with the methods discussed in the following sections. 10.6.2 Delta-Gamma VaR The Delta~Gamma VaR method is an extension of the Delta VaR method described in the previous section which takes better account of the non-linear dependence of option prices on their underlying market variable. Figure 10.7 illustrates the difference between Jinear and non-linear instruments. The value of a linear instrument as a function of the underlying variable can be represented by a straight line which is equivalent to a fixed amount of the underlying. For a non-linear instrument the relationship is a curved line and can’t be represented as simply a fixed amount of the underlying variable. The change in the value of an option (V) in terms of the changes in the underlying market variables can be represented by a Taylor series expansion: ov 1eV ov ov IV 2 AV = gs as+ Mar AoE ass ee (10.22) In equation (10.22), S- is called theta and is called gamma. When Delta VaR is used the option is converted into its delta ea position and only the first term \2—— Non-linear Linear Instrument value (V) Underlying Market Variable (S) FIGURE 10.7 Illustration of the Difference Between Linear and Non-linear Instruments 195Energy Derivatives: Pricing and Risk Management in equation (10.21) is used giving the approximation for the change in the option value in equation (10.15). In Delta-Gamma VaR we retain all of the terms shown in equation (10.21)'4 which gives a better approximation for the change in the instrument value as a function of the underlying asset: ov ov 1eV AV RS AS + zax% y? 10.23 Osetra 2 oe Note that the theta term, $/ Ay, is also included but makes relatively little contribution to AV for typical VaR horizons. Equation (10.23) is often called the delta-gamma-theta approximation because of the inclusion of the theta term. Representing the Portfolio Returns in Terms of the Underlying Returns As for the Delta VaR method we have the following equation for the return on the portfolio (see equations (10.12) to (10.14)): AP/P = (AV, + AV, +... AV y)/(Vi + V2 + Vy) (10.24) Substituting for the change in the instrument values in equation (10.24) using equation (10.23) the return on the portfolio becomes! avy i AP/P= Faun +454 Fa ei Six +ym iT, ) +. (10.25) Note that equation (10.25) applies to the case of market instruments which are under- lying market variables since in this case 3¢= 1 and 2% = 8% = 0 Computing Delta-Gamma VaR Under the delta-gamma-theta approximation for the change in the instrument values the portfolio return is no longer a linear combination of underlying market variable returns as in equation (10.17). This in turn means that the portfolio return distribution is no longer normally distributed (which is true in reality and is the reason for using the more accurate delta-gamma-theta approximation). We cannot therefore compute the VaR estimate using the simple matrix multiplication we used for Delta VaR. A number of authors have suggested ways in which the delta-gamma-theta approximation can be used to actually compute a VaR estimate. The method proposed in the RiskMetrics Technical Document (4'" Edition, December 1996, section 6.3) is to compute the first four moments of the portfolio distribution and then to fit a general distribution from the Johnson family (Johnson (1949) using a version of an algorithm due to Hill, Hill and volatilities over the VaR horizon (see Chapter 8). '4 Technically speaking we retain all terms up to second order 196Value at Risk Holder (1976). No details of this approach are discussed in the RiskMetrics Technical Document and it would certainly be rather complex to implement with no guarantee of significantly increased accuracy. In an earlier RiskMetrics Monitor (First Quarter 1996) it is proposed that an approximation to the VaR estimate can be calculated in terms of the portfolio moments using the Cornish-Fisher expansion (see Johnson and Kotz (1970)). However, it is shown in the same article that this method has rather poor accuracy compared with the true VaR estimate computed by simulation. An excellent discussion and analysis of these and other methods can be found in Pritsker (1997). The main point of the delta-gamma-theta approximation is that it allows us to compute the change in the value of an option without having to use the full option pricing model which may be relatively computationally intensive. In the full Monte Carlo simulation approach, which we discuss in the following section, the most computationally demanding part is the re-pricing of the options positions under each simulation. Our recommended Delta-Gamma approach is to use Monte Carlo simula- tion of the underlying market variables (which we discuss in section 10.6.3) together with the delta-gamma-theta approximation for the change in the option prices. Example: Calculation of Delta~-Gamma VaR for a Portfolio of Two Options on Crude Oil and Natural Gas Futures TABLE 10.2 Example of Delta-Gamma Monte Carlo Simulation VaR Calculations Sim. FI F2 cl oe Port APP 1 21.173 2.967 0.6791 0.1139 0.7931 0.0361 2 21.325 2.953 0.7553 0.1069 0.8622 0.0480 3 21.056 2.938 0.6242 0.1002, 0.7244 0.1195 4 21.220 2.965 0.7024 = «0.1127 O.81S1 0.0093 5 20.147 2.958 0.6667 0.1092 0.7759 0.0570 6 21.228 2.939 0.7062 0.1007 0.8069 -0.0193 7 21.260 2.971 0.7220 0.1159 0.8379 0.0184 8 2.147 2.931 0.6670 = 0.0971 0.7641 0.0713 9 2.137 2.960 0.6618 0.1104 0.7723 -0.0614 10 21.261 2.957 0.7226 0.1089 0.8315. 0.0107 491 3.065 0.6666 = 0.1680 0.8346 0.0144 492 2.949 0.6799 0.1053 0.7882 0.0457 493, 2.940 0.6901 0.1008 0.7910 0.0386 494 2.988 0.7330 © 0.1245 0.8575 0.0422 495, 2.957 0.7185 0.1089 0.8274 0.0057 496 2.970 0.7036 = 0.1153 0.8189 —0.0047 497 3.035 0.7350 0.1497 0.8847 0.0753 498 3.056 0.6793 0.1622 0.8416 0.0229 499, 2.987 0.7118 0.1240 0.8358 0.0159 500 3.003 0.6555 0.1322 0.7877 —0.0426 197Energy Derivatives: Pricing and Risk Management We use the same portfolio, parameters and model as for the Delta VaR example in section 10.6.1. The prices and sensitivities of the two options are: Ci = 0.6981, C2 = 0.1246 Delta, = 0.4921, Deltas = 0.5146 1.2417 Gamma, = 0.2056, Gamma, Theta, = —0.1979, Theta, = -0.1965 Table 10.2 shows an example of Delta~Gamma Monte Carlo simulation results. In Table 10.2, the columns headed “FL”, “F2”, “C1”, “C2”, “Port” and “AP/P” show the simulated futures prices, option prices computed using the delta-gamma-theta approx- imation, portfolio values, and portfolio returns respectively. The portfolio returns are defined as the difference between the simulated portfolio value and the current portfolio value divided by the current portfolio value. The full simulated data set illustrated in Table 10.2 contains 500 daily portfolio returns - the VaR estimate for the chosen 95% confidence level is given by the (1-0.95) x 500 sorted return ie. the 25'" sorted return which for this set of simulated data is —0.110426. Figure 10.8 shows the resulting portfolio return histogram. Notice that even for the relatively short one day horizon where Delta VaR should be quite accurate the VaR estimate under Delta-Gamma Monte Carlo VaR is 1.8% higher. 8. - 7 Probability Density b a o e 010 0.088 0.07 see se meet else aa ie e928 Besse eB ESSE EERE S See cce sg3S88838555558 Return FIGURE 10.8 Delta-Gamma VaR Example Portfolio Return Distribution 198,Value at Risk We answer the question is this a more accurate estimate in the following section in which we compute VaR by full Monte Carlo simulation. 10.6.3 Monte Carlo Simulation The Monte Carlo VaR method involves assuming a model for the joint behaviour of all relevant market variables and uses this to simulate future possible values of the portfolio. The model assumed is typically the same as for the Delta and Delta- Gamma VaR methods i.e. that returns are normally distributed. However, the advan- tages of this approach are that it allows for the accurate modelling of market behaviour such as the addition of jumps and incorporation of the knowledge of future events such as changes in the operation of a market. It also leads to a more accurate representation of the dependence of the prices of options on the market variables via the use of trading models. The problems with this approach are that it is relatively complex to understand and implement, and very computationally demanding. The Monte Carlo approach requires the combined simulation of the spot and forward price curves underlying our portfolio at the VaR horizon. We can use the multi-factor model described in chapter 8 to model all the energy forward curves underlying the portfolio simultaneously. Recall from section 8.9.2 that the instantaneous returns in the different energy forward prices can be expressed in terms of a consistent set of volatility functions: dF.(t,T) > owt, T)dz;(t) (10.26) ial F(t) where e indexes the different energy forward curves e.g. oil, gas, electricity, ete. The volatility functions o,(t,7) describe how all the energy forward curves evolve together!®. Under this model all forward price returns are jointly normally distributed and so in terms of the model for the underlying market variables it is identical to the Delta and Delta-Gamma VaR methods. However, it differs in that the dependence of non-linear derivatives on the market variables is more accurate. This is because option pricing models are used to compute the changes in the option prices for the each of the simulated states of the underlying forward curves. This is in contrast to the Delta and Delta- Gamma methods in which simple approximations are used to obtain the values of options as a function of the underlying forward curves. Tn order to simulate the joint behaviour of the energy prices on which the portfolio depends we must obtain the volatility functions ¢,{t,7) from the covariance matrix of all the relevant forward prices. RiskMetrics supplies the following energy price volatility and correlation information: 5 ‘The same approximation for returns, x, = AS/S, as in Delta VaR is used 1© An alternative approach would be to use the volatility functions for each energy forward curve. This would 199Energy Derivatives: Pricing and Risk Management im 3m 6m ly Natural Gas v v v v Heating Oil v v v v Unleaded Gas v v v x Light Sweet Crude ¥ v v v Therefore in order to evaluate long dated or other energy based instruments, further data must be obtained. This will typically be historical time series data which must be from reliable data sources and subjected to rigorous checks. Using time series of daily returns the sample covariance matrix of the energy forward curves can be obtained from which the volatility functions can be obtained using Eigen Analysis (see section 8.6). The volatility functions allow us to compute the covariance matrix of the forward curves at the chosen VaR horizon (T;): Tr Yq = Cov[In(F,, (Th, S«))n(Fe,(T.8)))] = >> { | elt Se) Fey sya} (10.27) The VaR horizon may be longer than the original time series return period and therefore the original sample covariance matrix cannot be used. Furthermore, it is not correct to simply scale the original sample covariance matrix using the square root rule due to the time varying volatility of forward prices (i.e. mean reversion in the spot price) and also the fact that the volatility functions may not be Markovian (see chapter 8 for more details). We can then use Eigen Analysis (or for greater efficiency Cholesky Decom- position'’) to obtain the “factors” required for simulating the returns at the horizon 7), As we have seen in chapter 8 Eigen Analysis decomposes the covariance matrix as follows: 2=Tal? where UL Uy Um 4 0 0 0 =| Una | #075 0.0) ae and A= O00 toe 0 0 0 An The columns of I’ are the eigenvectors and m is equal to the total number of forward prices to be simulated. We can then use these volatility functions to simulate the underlying forward prices at the VaR horizon many times and compute the changes in the value of the portfolio. The sample of portfolio changes give us a portfolio return distribution and thus the VaR estimate. be equivalent to assuming the correlation between the different energy forward curves was zero. 200Value at Risk TABLE 10.3 Example of Monte Carlo Simulation VaR Calculations Sim. FL F2 cl Q Port DP/P 1 21.173 2.967 0.6791 0.1139 0.7931 0.0361 2 21.325 2.953 0.7553 0.10690.8622 0.0480 3 21.056 2.938 0.6242 0.1003 0.7245 —0.1195 4 21.220 2.965 0.7024 = «0.1127, 0.8151 0.0093 5 21.147 2.958 0.6667 0.1092 0.7759 0.0570 6 21.228 2.939 0.7062 0.1007 0.8069 — 0.0192 7 21.260 2971 0.7220 0.1159 0.8379 0.0184 8 20.147 2.931 0.6670 = (0.0971 0.7641 = —0.0713 9 21.137 2.960 0.6618 = 0.1104 0.7723 0.0614 10 21.261 2.987 0.7226 © 0.1089 0.8315 0.0107 491 2.147 3.065 0.6666 0.1680 0.8346 0.0144 492 21.174 2.949 0.6799 0.1053 0.7852 —0.0457 493 21.195 2.940 0.6901 0.1009 0.7910 0.0386 494 21.281 2.988 0.7330 0.1248 0.8575 0.0422 495 21.253 2.957 0.7185 0.1089 0.8274 0.0057 496 21.223 2.970 0.7036 = «0.1153 0.8189 0.0047 497 21.285 3.035 0.7350 © 0.1497 0.8847 0.0753 498 20.173 3.056 0.6793 0.1622 O.8415 0.0228 499 21.239 2.987 O78 0.1240 0.8358 0.0159 500 21.123 3.003 0.6555 0.1322 0.7877 0.0426 Example: Calculation of Monte Carlo Simulation VaR for a Portfolio of Two Options on Crude Oil and Natural Gas Futures The details of the portfolio are identical to the example for Delta VaR in section 10.6.1 The model for simulating the forward prices is the same three factor model used in the previous examples in sections 10.6.1 and 10.6.2. Table 10.3 shows a typical set of results. In table 10.3, the columns headed “FI”, “Cl”, “C2”, “Port” and “AP/P” contain the simulated futures prices, option prices, portfolio values, and portfolio returns respectively. The simulated option prices are obtained using the simulated futures prices and the three factor model estimated from historical data (see section 10.6.3). The portfolio returns are defined as in the previous section i.c. difference between the simulated portfolio value and the current portfolio value divided by the current portfolio value. The full simulated data set illustrated in Table 10.3 contains 500 daily portfolio returns — the VaR estimate for the chosen 95% confidence level is given by the (10.95) 500 sorted return ic, the 25'" sorted return which for this set of simulated data is —0.110420. Figure 10.9 shows the resulting portfolio return histogram. The VaR estimate under full Monte Carlo simulation is almost identical to that computed using the Delta-~Gamma approach in the previous section. Full simulation can typically take between 100 and 1000 times longer to compute than Delta~-Gamma Monte Carlo because of the full re-valuation of the option positions. Therefore, Delta-Gamma 201Energy Derivatives: Pricing and Risk Management Probability Density o zy ~~ eo a a -0.069 1 ean 9 oo a oO+ +t OM - ODMR WH un oO nh oD 239488 B88 8 asics geeegase 858228 8S 238 225 8 sso gog¢g GGG THO es iGO OHO HO TG TaiGh = Return FIGURE 10.9 Monte Carlo Simulation VaR Example Portfolio Return Distribution Monte Carlo is the recommended approach, particularly for shorter horizons. We saw in chapters 2, 3 and 7 that many energy prices exhibit jumps and stochastic volatility and that these factors can easily be incorporated into Monte Carlo simulations. Monte Carlo VaR therefore allows the incorporation of these important effects in a straightforward way. 10.6.4 Historical Simulation The idea behind Historical Simulation VaR is to use historical time series’ of market variables to obtain a distribution of portfolio returns. It therefore differs from Monte Carlo simulation which uses a model of the market variables and a set of random numbers to generate a portfolio distribution. For example, using the last two years of daily prices for the underlying market variables of our portfolio, we could compute the daily changes in the value of the current portfolio that would have occurred if the historical “changes” in the daily prices repeated themselves starting from today’s values. This would give us approximately 500 daily returns for the current portfolio. Under the assumption that this past history of changes is a good forecast for the range of changes the portfolio could have in the future, we can compute a VaR estimate at our chosen confidence level. The advantages of this approach are that it is simple to understand, intuitive and 202Value at Risk relatively straightforward to implement. It can be applied to all instruments and all market risk types — all that is required is sufficient history of the relevant market variables. It allows the use of the pricing models used for trading to revalue option positions. The approach is particularly valuable if the user does not believe that the distributions of market returns can be described by normal distributions or tractable alternatives as is likely for energy markets. The disadvantages of historical simulation are that it is computationally demanding due to the need to revalue the whole portfolio'®. Furthermore, the portfolio distribution and thus the VaR estimate is completely determined by the distribution of the underlying market variables over the historical time period selected which can lead to unreasonable behaviour of the VaR estimate. For example, imagine we decide to use two years of daily changes and there is a large change in the underlying market variables two years ago. As we move forward in time and this particular day drops out of the historical data it is likely there will be a large change in the VaR estimate. A subtler problem is that the VaR estimate obtained will depend on the range of the market variables over the historical period used. If this period is too short then this may not reflect the range of possible values of the market in the future. Ideally we need a period of data which includes all possible changes in the market variables which could ever occur. Therefore, the period of data should be at least as long as the typical business cycle of the market. alculation of Historical Simulation VaR for a Portfolio of Two Options on and Natural Gas Futures We use the same portfolio as for the previous examples — the parameters values are given in the example in section 10.6.1. The prices of the two options!? are: C, = 0.6981, C = 0.1246 Table 10.4 shows a typical set of calculations. The first three columns in Table 10.2 contain the historical data — date and NYMEX closing futures prices on that date. The columns headed “FI” and “F2” contain the simulated futures prices and, since the VaR horizon is one working day, the simulated futures prices can be obtained by multiplying the current futures prices by the ratio of the historical futures price on consecutive days, as follows; F(T) = Fo) Ld (10.28) Fil) Where f(1,) is the historical futures price on date 1}. The columns headed “C1”, “C2”, “Port” and “AP/P” contain the simulated option prices, portfolio values, and portfolio 7 See RiskMetrics Technical Document, Appendix E '® We can use the delta-gamma-theta approximation method to achieve more efficient evaluation in the same way as for Monte Carlo VaR. 203Energy Derivatives: Pricing and Risk Management TABLE 10.4 Example of Historical Simulation VaR Calculations Date — Oil(2y)_Gas(6m) FL F2 cl C2 Port DP/P 01/04/98 17.52 2.568 21.210 3.029 0.6974 0.1465 0.8438 0.0256 02/04/98 17.52 2.604 21.307 2.977 0.7460 0.1186 0.8646 0.0508 03/04/98 17.6 2.595 20.897 2.982 0.5533 0.1215 0.6748 —0.1798 06/04/98 17.34 2.591 21.247 3.106 0.7156 0.1935 0.9090 0.1049 07/04/98 17.37 2.694 21.381 3.006 0.7845 (0.1337 0.9182 0.1161 08/04/98 17.51 2.711 21.077 2.953 0.6336 0.1070 0.7406 —0.0999 09/04/98 17.4 = 2.68 21.210 2.987 0.6974 0.1238 0.8212 -0.0019 10/04/98 17.4 2.68 21.039 2.837 0.6164 0.0606 0.6770 ~0.1771 13/04/98 17.26 2.545 21.099 3.005 0.6442 0.1331 0.7773 —0.0553 14/04/98 17.17 256 21.482 3.010 0.8387 0.1362 0.9748 0.1848 20/03/00 20.92 2.84 21.291 2.933 0.7381 0.0980 0.8361 (0.0162 21/03/00 2864 21.423 3.012 0.8068 =—0.1372 0.9440 0.1474 22/03/00 2.893 20,899 3.017 0.5543 0.1400 0.6942 ~0.1562 23/03/00 2.93 21.302 3.025 0.7433 0.1444 0.8877 0.0790 24/03/00 2913 21.474 2.972 0.8343 0.1161 0.9504 0.1551 27/03/00 2.973 20.749 3.046 0.4926 0.1566 0.6492 -0.2110 28/03/00 3 20.841 3.014 0.5300 0.1382 0.6682 —0.1878 29/03/00 2.961 21.783 2.948 1.0127 0.1048 1.1175 0.3582 30/03/00 2.932 returns respectively. The simulated option prices are obtained using the simulated futures prices and the three factor model estimated from the same historical data, The portfolio returns are defined as the difference between the simulated portfolio value and the current portfolio value divided by the current portfolio value. The full data set in Table 10.2 contains 523 dates which gives 522 daily portfolio returns, The VaR estimate at the 95% confidence level is therefore the (1-0.95) x 522 sorted return i.e. the 26" sorted return which is found to be —0.234053. Figure 10.10 shows the resulting portfolio return histogram. Comparing figures 10.6, 10.8, 10.9 and 10.10 it is clear that the shape of the historically simulated portfolio return distribution is somewhat different than the approximations obtained by Delta, Delta-Gamma and Monte Carlo VaR ~ there are two distinct reasons for this. Firstly, the return distribution of a portfolio containing call options is positively skewed due to the leverage effect of the options — small positive returns on the underlying lead to much larger positive returns for the portfolio. This can be seen in the right-hand tail of the histogram in Figure 10.10. Secondly, the historical futures price return distributions are fat-tailed leading to fat-tails in the portfolio return distribution. The left-hand tail in Figure 10.10 shows this quite clearly and it is also reflected in the VaR estimate (which depends on the shape of the left-hand tail of the distribution) being approximately twice has big as the other estimates. 204Value at Risk 2 3 2 | 5 a 2 5 a 8 8 & 24 1 CSS RS TIS ZS es Sesneesvooeage Pee ee¢esgesgegeg¢eG¢e¢ Gs SGS5G5056S5 E556 = Return FIGURE 10.10 Historical Simulation VaR Example Portfolio Return Distribution 10.7 Testing Value-at-Risk Estimates It is important that the VaR estimates are compared with the resulting market and portfolio returns to check how often actual returns exceed the VaR forecast — this process is called backtesting. The Bank for International Settlements (BIS) Basle Committee on Banking Supervision (Basle) has issued standards required for back- testing”. Basle defines three zones depending on the number of exceptions in the previous year for a 99% confidence and a one day horizon. i.e. the number of times in the previous 250 working days the returns exceeded the expected number for the computed VaR level (ie. for a period of 250 working days at a confidence level of 99% over one day, the expected number of returns exceeding the VaR level is 2.5 days). Zone | or Green Zone: Four or less exceptions — the model is acceptable. Zone 2 or Yellow Zone: Five to nine exceptions — certain actions are required such as increasing the safety multiplier. Zone 3 or Red Zone: Ten or more exceptions — the model is unacceptable and must be substantially revised or changed. © We use the same three factor model as used in the example in section 10.6.1 205Energy Derivatives: Pricing and Risk Management 0.15;— — — ~— ~ 04 0.05 Return ° -0.05 04 | 0.15 15/05/96 23/08/96 01/12/96 11/08/97 19/06/97 27/09/97 05/01/98 18/04/98 24/07/98 01/11/98 09/02/99 Date Key: Diamonds are daily returns, solid lines are 90% VaR Bands 0.15: : oA 0.05: Return 2 -0.05! -0.1 0.15 15/05/96 29/08/96 01/12/96 11/03/97 19/06/97 27/09/97 05/01/98 15/04/98 24/07/98 01/11/98 09/02/99 Key: Diamonds are daily returns, solid lines are 98% VaR Bands FIGURE 10.11 Daily Returns, 90% and 98% VaR Bands for the Crude Oil Nearest Maturity NYMEX Futures Contract 206Value at Risk 0.25: —_—_—_——— ——_--- -0.25+ — a Y5DSBE _2508N6 O12 110997 19100197 270907 OSIOINE 15048 2407NB O1/11/98 o90200 Date Key: Diamonds are daily returns, solid lines are 90% VaR Bands 0.25: 02 0.15) O41 0.05 Return -0.15 0.2 -0.25 15/05/36 23/896 01/2/96 11/0997 191067 27/09/97 05/1/98 1510498 24/0798 O1/11/98 09/02/99 Date Key: Diamonds are daily returns, solid lines are 98% VaR Bands FIGURE 10.12 Daily Returns, 90% and 98% VaR Bands for the Natural Gas Nearest Maturity NYMEX Futures Contract 207Energy Derivatives: Pricing and Risk Management In the following section we test the VaR forecasts for various underlying market variables. 10.7.1 Testing the VaR Forecast for the Market Variables RiskMetrics VaR forecasts, for the market variable returns, are given by the bands of a confidence interval that is symmetric around zero. For example, the scale factors —/+1.65 give the 90% confidence interval or VaR level representing the Sth/9Sth percentiles of the standard normal distribution respectively. Under this assumption we expect 5% of the standardised observations” to lie below — 1.65 and 5% to lie above 1.65. To test these VaR forecasts, for each day we compute the volatility forecast, 4, using equation (10.6) with the RiskMetrics standard decay factor of 0.94” and compare the VaR forecast at the 90% level ie. —/+1.65 6 and at the 98% level i.e. —/+2.33 6. Figures 10.11 and 10.12 show the results, for the NYMEX crude oil and natural gas shortest maturity futures contracts respectively, over the period from July 1996 to September 1998. In Figures 10.11 and 10.12 we expect that 5% of the returns will fall below —1.65 and 5% will fall above 1.65 G. Similarly, we expect that 1% of the returns will fall below —2.33 6 and 1% will fall above 2.33 &. For the crude oil Futures contract we find that 5.07% fall below -1.65 6, 6.64% fall above 1.65 6, 2.10% fall below —2.33 6 and 1.57% fall above 2.33 6. For the natural gas Futures contract we find 6.47% fall below --1.65 6, 5.07% fall above 1.65 6, 2.10% fall below —2.33 & and 1.57% fall above 2.33 &. The conclusion is that at the 90% level the VaR estimates are quite good but at the more extreme 98% level they underestimate the VaR considerably. 10.8 Conclusions In this chapter we discussed the main approaches to calculating VaR which are currently used and applied them to energy derivatives portfolios. The Delta and Delta~Gamma VaR methodologies rely for their relative simplicity on the assumption that the under- lying market variables are normally distributed. As we have seen in the earlier chapters (see chapters 2, 3, 6, 8, and 9) this assumption may not be particularly good for energy markets — jumps in the spot price and random volatility lead to non-normal distribu- tions for energy spot and forward price returns. Although the implementation of the Monte Carlo VaR approach, studied in this chapter, was also based on the assumption of normally distributed returns, we showed in chapters 2, 3 and 7 how Monte Carlo simulation can be extended to incorporate jumps and stochastic volatility. The evalua- tion of the VaR estimate in the Monte Carlo approach is not affected by adding these extra factors. 2” See the BIS website at www.bis.org and also the RiskMetrics Technical Document, Appendix F. 2! The standardised returns are the mean centred return normalised by its standard deviation. 22 We start computing the volatility forecasts after 90 days, as before this date we don’t have enough data to 208Value at Risk In a recent study Marshall and Siegel (1997) tested the implementation and use in practice of one particular VaR model by assessing the variation in VaR estimates produced by different commercial implementations of the same model of value-at-risk. The authors asked a number of leading risk management system vendors to assess the value-at-risk of an identical portfolio of instruments of varying complexity according to the RiskMetrics methodology. The portfolios tested ranged from ‘linear’ instruments such as bonds, through interest rate swaps and FRA’s to ‘non-linear’ Foreign exchange options and interest rate caps and floors. The results show that for the “‘linear” instrument classes the VaR estimates were fairly consistent across vendors. However, for the more complex instruments the variations were very large — for example, for interest rate swaps the total VaR had a mean of U$ 306,663 and a standard deviation of US 66,648 (21%), for foreign exchange options ~ a mean total VaR value of U$ 763,317 and standard deviation of US$ 198,829 (25%), and interest rate caps had a mean VaR of U$ 483,130 and standard deviation of U$ 115,194 (28%). It is therefore important that the relevant people in the institution are appropriately trained so they fully understand the implementation of VaR. Furthermore, the VaR system whether internally developed or purchased from a system vendor should be fully tested by independent consultants with proven expertise in VaR and derivatives risk management. It is clear from our analysis and others that the simple VaR methods do not provide the accuracy required for energy markets. However, the more accurate methods such as Historical and Monte Carlo simulation are very time consuming and complex to implement. Our recommendation would be to develop the VaR implementation in stages, beginning with the simplest Delta VaR. From this it is possible to build up via Delta-Gamma VaR to one of the more sophisticated analytical or semi-analytical methods”. An implementation of Historical Simulation VaR, which is relatively straightforward to implement, should always be available as a comparative check for the other methods. In this chapter we have described in some detail the implementation of the key components of a VaR calculation for all of the main methodologies. This information together with the RiskMetrics Technical Document should provide the reader with the necessary tools to begin the process of building an appropriate VaR system for their needs. Implementing VaR is a difficult and time consuming exercise, it is, however, an essential part of the risk management process. compute a forecast. * For example, RiskMetrics Mixture of Normals method (see RiskMetrics Monitor Second Quarter 1996) or RiskMetrics Generalised Error Distribution method (see RiskMetries Technical Document, Appendix B). 20911 Credit Risk in Energy Markets 11.1 Introduction Credit Risk is the uncertainty in the value of a portfolio due to the fact that the counterparties to the contracts in a portfolio may be unable to meet their financial obligations. This uncertainty is usually summarised in terms of a discrete index — the credit rating, which can be said to reflect the credit quality of the institution — i.e. the likelihood of the counterparties meeting their contractual obligations. A major problem with measuring credit risk is the lack of comprehensive credit quality and default data. Whereas VaR methodologies are an exercise in extracting the key properties of market risk from reasonably abundant data, the lack of credit risk data means that measuring credit risk is more an exercise in finding relatively simple models which explain the few key observables. The measurement and management of credit risk is one of the major challenges facing energy companies today. It is not possible to cover every detail of this complex subject in a single chapter, indeed one could fill many books on this subject alone. It is therefore our goal in this chapter to describe the key concepts and tools in sufficient detail that the reader can make intelligent choices regarding the type of approach which best suits their needs and implement a basic credit risk management system. 11.2 What is Credit Risk? In general terms, credit risk is the risk of a loss in the value of the portfolio of instruments held due to changes in the credit quality of the counterparties to the instruments. Credit quality measures the probability that the counterparty will be able to meet its financial obligations. The lower the probability that the counterparty will be able to pay, the lower the value of the instrument to the holder relative to its value if payment was certain. One of the major problems with modelling credit risk is that major decreases in credit quality, particularly default, are relatively rare but cause very large losses. This leads to the distribution of the portfolio values under credit risk having a shape which is difficult to model. Figure 11,1 illustrates typical] market and credit risk portfolio return distributions. In Value-at-Risk models it is reasonable, at least as a first approximation, to model the underlying variables and the portfolio return distribution as normal. In credit risk modelling the key variables and especially the portfolio return cannot be approximated accurately as normally distributed. 210Credit Risk in Energy Markets © & © ® Probability Density ° ° ° fe esemee tees ener eae a a ® BR & B 8 -6 <4 -2 0 2 4 Retum FIGURE 11.1 Typical Market and Credit Risk Distributions 11.3 The Key Components of Credit Risk Models Credit risk can be thought of as having two main components: Credit quality risk: This is the risk due to the uncertainty in the future credit quality of the counterparties to the portfolio, and is reflected in the fact that the credit rating of an institution is not fixed but changes depending on its performance in changing economic conditions. Since changes in credit quality imply changes in the likelihood of the company meeting its obligations, this impacts on the value of the instruments held with the counterparty. Default risk: This is the risk that an institution may actually default on its obligations resulting in a relatively large loss for the counterparty. In the following sections we discuss the basic observables and data which feed into the majority of credit risk models. 11.3.1. Credit Ratings and Credit Rating Transition Probabilities A credit rating is a discrete “value” on a pre-defined scale which summarises, in the opinion of the organisation calculating the rating, the overall ability (and willingness) of an institution to meet its financial obligations (i.e. its creditworthiness), Many organisa- tions and companies publish or sell credit ratings of banks, corporations, and other 211Energy Derivatives: Pricing and Risk Management TABLE 11.1 Moody’s Investors Services and Standard and Poor’s Credit Rating Scales Moody 18 Moody 8 S&P 18 S&P 8 1 Aaa Aaa AAA AAA 2 Aal Aa AAt AA 3 Aad A AA A 4 Aad Baa AA- BBB 5 Al Ba At BB 6 A2 B A B 7 AB Caa A- ccc 8 Baal D BBB+ D 9 Baad BBB 10 Baa3 BBB- I Bal BB+ 12 Ba2 BB 13 Ba3 BB- 14 BI B+ 15 B2 B 16 B3 B- 17 Caa ccc 18 Dd D financial entities and here we use Moody’s Investors Service (Moody) and Standard and Poor’s (S&P) credit ratings as examples. Both Moody and S&P provide credit ratings on 18 level and 8 level scales. Table 11.1 shows these four rating scales with the labels which are assigned to each level. Rating level 1 represents the highest level of creditworthiness on each scale and the credit worthiness decreases down the scale to the lowest level which is default. To give an idea of the relative meaning of the various levels we a present a summary of the description of the major Moody's rating levels: Issuers rated Aaa offer exceptional financial security. While the creditworthiness of these entities is likely to change, such changes as can be visualized are most unlikely to impair their fundamentally strong position. Issuers rated Aa offer excellent financial security. Together with the Aaa group, they constitute what are generally known as high grade entities. They are rated lower than Aaa entities because long-term risks appear somewhat larger. Issuers rated A offer good financial security. However elements may be present which suggest a susceptibility to impairment sometime in the future. Issuers rated Baa offer adequate financial security. However, certain protective elements may be lacking or may be unreliable over any great period of time. Issuers rated Ba offer questionable financial security. Often the ability of these entities to meet obligations may be moderate and not well safeguarded in the future. 212Credit Risk in Energy Markets Issuers rated B offer poor financial security. Assurance of payment of obligations over any long period of time is small Issuers rated Caa offer very poor financial security. They may be in default on their obligations or there may be present elements of danger with respect to punctual payment of obligations. Issuers rated Ca offer extremely poor financial security. Such entities are often in default on their obligations or have other marked shortcomings. Issuers rated C are the lowest rated cl of entity, are usually in default on their obligations, and potential recovery values are low. Credit ratings are assigned after a detailed analysis of the company but many energy companies are not yet rated. Individual energy companies must therefore calculate their own internal ratings. It is generally not feasible for a company to do the depth of analysis performed by the rating organisations, but the general approach is still valid and can be checked by applying the process ‘blind’ to rated companies and checking the agreement with published ratings. Associated with each credit rating scale are transition probabilities representing the likelihood that a financial entity with a particular credit rating will retain the same rating or change to a different rating over the next year. The study of credit rating changes is often called migration analysis and many rating organisations also publish or sell this information. This information is obtained from the historical analysis of the changes in ratings in one year of a large set of companies. Credit rating transition probabilities are one of the key inputs in the CreditMetrics model', one of the main credit risk models discussed in this chapter, and the information can be obtained from their web site”. Table 11.2 shows a typical credit rating transition probability matrix for the S&P 8 level scale. For a given column each row indicates the probability over a one year period that a company with the credit rating at the top of the column will change to the rating given in the first column. For example, a company with a rating of AA (second column) has a probability of 0.0031 (0.31%) of changing to BBB over the next year. Notice that for each rating by far the most likely outcome as indicated by the transition probability is for the rating to remain the same. The next most likely event is a single level upgrade or downgrade and multiple upgrades are very unlikely although not impossible. Careful study of table 11.2 reveals its origin in historical analysis via minor anomalies such as the zero probability of a CCC rated company becoming AA rated but having 0.21% chance of becoming AAA rated. 1 CreditMetrics is a registered trademark of J. P. Morgan 2 See www.creditmetrics.com under DatasetsEnergy Derivatives: Pricing and Risk Management TABLE 11.2 Credit Rating Transition Probability Matrix To/From AAA AA A BBB BB BHCC HaiD) AAA 0.9082 0.0826 0.0074 0.0006 0.0011 0.0000 0.0000 0.0000 AA 0.0065 0.9088 0.0769 0.0058 0.0005 0.0013 0.0002 0.0000 A 0.0008 0.0242 0.9130 0.0523 0.0068 0.0023 0.0001 0.0005 BBB 0.0003 0.0031 0.0587 0.8746 0.0496 0.0108 0.0012 0.0017 BB 0.0002 0.0012 0.0064 0.0771 0.8116 0.0840 0.0098 0.0098 B 0.0000 0.0010 0.0024 0.0045 0.0686 0.8350 0.0392 0.0492 ccc 0.0021 0.0000 0.0041 0.0124 0.0267 0.1170 0.6448 0.1929 D 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 11.3.2 Credit Spreads Credit spreads are the differences between the interest rates by which the market discounts credit risky cashflows and the interest rates with which the market discounts certain (default free) cashflows. The idea is that if you are promised a particular cashflow at a specified time in the future but you are not certain that it will be delivered you will calculate the present value of that cashflow to be less than if you considered it certain to be delivered. The lower value can be viewed as equivalent to discounting the cashflow at a higher rate than the markets riskless or default free interest rate. Credit spreads are obtained by comparing the interest rates implied by the market prices of credit risky corporate bonds with the interest rates implied by government bonds (which are considered to have no credit risk ~ at least for developed countries). Credit spreads can also be obtained from the CreditMetrics web site. 11.3.3 Recovery Rates in Default In the event of a default of a counterparty, a firm will only receive a fraction of the full amount owed by the counterparty, the fraction is termed the recovery rate, The product of the recovery rate and the total amount owed is equal to the amount which the firm recovers as a result of liquidation or restructuring of the defaulting institution. Recovery rates should take account of the priority (seniority) of the instrument relative to the other obligations in the defaulters portfolio and also any kind of collateral or similar sort of security. Historical recovery rate data indicates that there is significant variation in recovery rates across all levels of seniority. Given this uncertainty, it is important that the uncertainty of recovery rates is modelled if possible and that the sensitivity of the model outputs to the recovery rate assumptior nalysed*, We discuss the modelling of recovery rates in section 11.4.4. * This is referred to as stress testing and is discussed in section 11.6.4 214