ROEM ANIORGITOMSITad TI a MA PUBLICATIONSaN NC NUN PRICING AND RISK MANAGEMENT EMOTO TRUITO MG NICINTITCA TIN Ce Ce OO Cn eC ck ves a exh ow ep nae kh oe edn aden wy deg rd eee. ois sl cence wh epoch comes fe it emt vn yO, Vi ese Fri “oak pt aos sf sig eek ee nats ce amr ey ci ed Ti ak ab pes a exe edn we nag mae, fers no ie pts nd png aero nay ics, wich ny oy hd ‘ngewe ses th phe mpl moe fan Fad et Unt bow conan npr rele soe 5k map, eos d poe invest n est compl syn moto rs, dn kag cs fly ay dies" Ned ar, Ve Pst, ed of esc, Satan ry ics uy “Aanpelersio notin mong syste, Te bok poison xl say began od oa ig mas ees mil ret a he por Ingonrsin”— ao,P,ViePesat of Src Pts of Dts a ot oc Feta Gp onc PLEnergy Derivatives: Pricing and Risk Management Les Clewlow and Chris Strickland Lacima PUBLICATIONSPublished in 2000 by Lacima Publications London, England Email: contactGilacimagroup.com Website: www. lacmgroup com Copytight © Les Clewlow and Chis Strickland All ights reserved. No part of this publication may be reproduced, copied, stored ina retrieval system, or transmitted in any form oF by any means, electronie, mechanical, photocopying, recording. or otherwise without either prior writen permission from the publisher, the copyright owner, or a loence permitting restricted copying in the United Kingdom isued 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 Biddes LimitedContents ABOUT THE AUTHORS PREFACE, ACKNOWLEDGEMENTS. CHAPTER 1_ INTRODUCTION TO ENERGY DERIVATIVES AND FUNDAMENTALS OF MODELLING AND PRICING LL INTRODUCTION TO ENERGY DERIVATIVES 1.2. FUNDAMENTALS OF MODELLING AND PRICING 13 NUMERICAL TECHNIQUES 13.1 The Trinomial Method 1.3.2. Monte Carlo Simulation 14 SUMMARY CHAPTER 2_ UNDERSTANDING AND ANALYSING SPOT PRICES 21 INTRODUCTION 22 MEAN REVERSION 23. SIMULATING MEAN REVERSION 24 THE HALF-LIFE OF A MEAN REVERTING PROCESS 25 STOCHASTIC VOLATILITY 24 SIMULATING STOCHASTIC VOLATILITY 2.7 JUMPS IN THE SPOT PRICE 28 SIMULATING JUMPS IN THE SPOT PRICE. 29 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 3.1 INTRODUCTION 32. ESTIMATING VOLATILITY 3.2.1 Estimation of Volatility From Historical Data 32.2 Estimation of Volaility for a Mean Reverting Process 3.23 Volatility Estimation: Special Issues 3.2.4 Intraday Price Variability ESTIMATION IN ENERGY MARKETS xiContents 33 34 3.2.5. Estimation of Volatility for a Basket 32.6 Implied Volatility STOCHASTIC VOLATILITY MODELS 3.3.1 Estimation and Testing 33.2. Ordinary Least Squares 3.3.3 Maximum Likelihood 334 Testing 3.3.5 Examples from the Energy Commodity Markets SUMMARY CHAPTER 4 ENERGY FORWARD CURVES 41 42 43 44 INTRODUCTION ‘CONSTRUCTING FORWARD CURVES 42.1 Cost of Carry Relationship. 422. Forward Price Bounds for Energies 4.23 Seasonality in Prices FORWARD CURVES IN THE ELECTRICITY MARKET 43.1 Arbitrage Pricing Approach 43.2. The Econometric Approach 433. The Spot Price Modelling Approach SUMMARY CHAPTER 5_ ENERGY DERIVATIVES: STRUCTURES AND APPLICATIONS, 3 52 53 54 38 56 37 58 59 INTRODUCTION EXCHANGE TRADED INSTRUMENTS SWAPS 53.1 Vanilla Swap 532 Variable Volume Swap 533 Differential Swap 534 Margin or Crack Swap 53.5 Participation Swap 53.6 Double-Up Swap 53.7 Extendable Swap CAPS, FLOORS AND COLLARS: SWAPTIONS. COMPOUND OPTIONS - CAPTIONS AND FLOPTIONS. ‘SPREAD AND EXCHANGE OPTIONS. S11 Calendar Spreads 5.72 Crack Spreads 5.73 Exchange Options PATH DEPENDENT OPTIONS 5.8.1 Asian Options ~ Average Price and Average Strike 582. Barrier Options 5.8.3 Lookback Options ~ Fixed Strike and Floating Strike 5.8.4 Ladder and Cliquet Options SUMMARY sears 32 54 56 6B 64 66 66 a 68 6s 0 n n n B B B ” 1" ” 5 1B 16 16 n n B p 80 80 80 8 8 82 83 85 86Contents CHAPTER 6 SPOT PRICE MODELS AND PRICING STANDARD INSTRUMENTS — 89 61 62 63 64 65 66 INTRODUCTION SINGLE FACTOR MODELS Futures and Forward Pricing Option Pricing The Schwartz Single Factor Model Futures and Forward Pricing Option Pricing TWO FACTOR MODELS 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. a 72 13 14 75 16 INTRODUCTION SIMULATION METHODS FOR ENERGY SPOT PRICE MODELS 72.1 Single Factor Simulation 72.2. Example ~ Valuing Spread Options 723. Example ~ Valuing European Energy Swaptions 724 Incorporating Seasonality into Monte Carlo Simulations 125. Computation of Hedge Sensitivities 72.6 Simulation of Multi-Factor Models 72.7 Generation of the Random Numbers 728 Valuing Path Dependent Options Using Simulation BUILDING TRINOMIAL TREES FOR THE ENERGY SPOT PROCESS IMPLIED TRINOMIAL TREES PRICING GENERAL PATH DEPENDENT ENERGY OPTIONS IN SPOT PRICE TREES. 7.3.1 Pricing Asian Energy Options in a Trinomial Tree 752. Pricing Swing Options SUMMARY CHAPTER 8 FORWARD CURVE MODELS: al 82 INTRODUCTION A SIMPLE MODEL FOR THE FORWARD CURVE ‘THE DYNAMICS OF THE FORWARD CURVE A GENERAL MULTI-FACTOR MODEL OF THE FORWARD CURVE RELATIONSHIP BETWEEN FORWARD CURVE AND SPOT PRICE MODELS. ESTIMATION OF THE VOLATILITY FUNCTIONS: 89 9 90 31 91 93 9s 97 98 100 102 103 lot os 106 107 los 109 109 109 Ho 1g 1g us n7 n7 9 120 123 127 Rs 129 131 13 134 134 13s 137 12 12 13Contents 87 88 89 8.10 sal 8.6.1 Historical Estimation of the Forward Curve Volatility Functions 8.62 Example Analysis of NYMEX Crude Oil Futures 8.6.3. Incorporating Seasonality into the Volatility Funetions PRICING STANDARD EUROPEAN OPTIONS 8.7.1 Options on the Spot, 8.7.2 Options on Forward Contracts PRICING GENERAL EUROPEAN CONTINGENT CLAIMS, Example: Pricing a European Energy Swaption PRICING EXOTIC OPTIONS 89.1 Calendar Spread Options 89.2 Crack Spread Options 89.3. Asian Options 8.9.4 Amercian Options in Forward Curve Models IMPLIED ESTIMATION OF THE VOLATILITY FUNCTIONS 8.10.1 Handling Volatility Skews and Smiles SUMMARY CHAPTER 9 RISK MANAGEMENT OF ENERGY DERIVATIVES 91 92 93 94 95 96 INTRODUCTION DELTA HEDGING OF ENERGY DERIVATIVE POSITIONS GAMMA HEDGING HEDGING VOLATILITY AND OTHER SOURCES OF MARKET RISK FACTOR HEDGING SUMMARY CHAPTER 10 RISK MANAGEMENT OF ENERGY DERIVATIVES 10.1 12 103. los los los 107 108 INTRODUCTION WHAT IS VALUE AT RISK? MODELLING THE MARKET VARIABLES FORECASTING STANDARD DEVIATIONS (VOLATILITY) FORECASTING COVARIANCES AND CORRELATIONS VALUE AT RISK METHODOLOGIES 10.6.1. Variance-Covarianee (Delta VaR) 1062. Delta-Gamma Val 10.63 Monte Carlo Simulation 10.64 Historical Simulation TESTING VALUE-AT-RISK ESTIMATES 10.7.1 Testing the VaR Forecast for Market Variables CONCLUSIONS: CHAPTER 11 CREDIT RISK IN ENERGY MARKETS Ma N12 13 INTRODUCTION WHAT IS CREDIT RISK? ‘THE KEY COMPONENTS OF CREDIT RISK MODELS 11.3.1 Credit Ratings and Credit Rating Transition Probabilities 11.32. Credit Spreads 11.33 Recovery Rates in Default 144 145 146 149 Ist 152 153 155 136 136 136 137 159 160 162 162 163 1683 let 169 mm 13 19 180 180 180 182 183 Is7 188 191 19s 199 202 205 208 208 210 210 210 2u1 au 214 244 is ls WW 1134. A Summary of Major Credit Risk Models (CREDITMETRICS 11.4.1 Modelling Credit Rating Changes 11.42 Simulation of Counterparty Market Returns 11.43. Credit Spreads, Recovery Rates and the Calculation of Derivative Values 1144 Modelling Recovery Rates 11.45 Calculation of the Credit Risk for an Energy Derivative Portfolio CREDITRISK + 11.5.1 Modelling the Occurrence of Defaults CREDIT RISK MEASURES AND MANAGEMENT 11.6.1 Marginal Risk 11.62. Concentration Risk 11.63. Managing Credit Risk 11.64 Assessing the Models CONCLUSIONS REFERENCES INDEX Contents ais 215 216 219 221 221 222 228 230 233 233 234 238 236 23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 sofiware companies worldwide for more than ten years. He is an Associate Research Fellow at the School of Finance and Economies, 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 founded and is principal of Lacima Group. He is an Associate Research Fellow at the Schoo! 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 Administracion, Caracas, Venezuela. Previously, Chris worked for RBC Gilts Ltd, and Kitcat and Aitken, & Co. in London, Heis 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 detivative pricing, interest rate models and numerical algo- rithms. Vineent (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 strategi ‘optimisation of financial and physical transactions, as well as development of vale-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 Masson isa viee 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 @ PhD in Finance from Georgia State University, Atlanta, Georgia. He currently provides analytical support to the risk management group in Enron Energy Services.Preface One of our main objectives in writing Energy Derivatives: Pricing and Risk Management hhas 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, itis 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 1 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. Chapier 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 enerey 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, ete) 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.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, Tan 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, Ilia 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 ‘Muy LC cs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 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 fiexible solutions for precise risk management A derivative security can be defined as.a security whose payoff depends on the value of * nupymww.enron.com ® Roaders interested inthe market growth and the development of competitive electricity markets are referred to Kaminski (1997) and Masson (1999),Prergy 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, ete, 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 Sy — 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 Sr and K. The short position, by definition, has the opposite payoff to the long position (i. —Sp + K) as every time the ong 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 oftem refered to as contingent claims as the payout to the security (often referred to asthe ‘maturity payoff) and hence wale, is contingent on other events, 4 For credit purposes, some forward contracts are also marked to market on 5 See Chapter 4 for more details, egular basis.Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Profit ‘Spot Eneray Price (S,) ‘At Maturity of Contact (7) = Loss FIGURE 1.1 Payoff to Long Forward Position that is currently trading at S. If the maturity date of the contract is T, c 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 itis written are related via the ‘cost of carry’ formula’ Fa SelONT=} (aay 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 » Sce Chapter 6 for 2 dscowion of convenience yield in eneray 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 K), 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 payotT to a call option can be described mathematically as follows: max(0,8~ 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 ~ $) (3) 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 Spo Energy Price (5) Valve 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 7, 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 ‘A. Option pay A. Option payott - > x x Long Call Sort Call max(0, 5, ma0, 5, K) Option payott A. Option payor > K K S ‘Short Put smax(0.5,—-®) FIGURE 1.4 Terminal Payotts for 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 eause embedded options to take on ‘complicating chasacterstcs and their impact on pricing. See also Kaminski (1997) 6Iniroduction 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 climinated 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 ofthis 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 5, are assumed to have constant instantaneous drift, 4, and volatility, ¢. The mathematical description of this property is given by the following stochastic differential equation’ aS = pSdt-+ oSdz (4) " Blecrcity cam be stored by hy roclctic schemes by using it to pump water into the reservoir, the electricity can then be recovered by releasing the water through the turbines. Elececty cam also be indively stored by ‘generators in the form of the fuel used to generate it Most model of asset price behavioue for pricing derivatives ae formulated in a continuous time framework by assuming a stochastie dferential equation desenbing the stochastic process Fellowed By the asst 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 dt, and dz is the underlying uncertainty driving the model and represents an increment in a Weiner process during dé. The risk-neutral assumption implies that the drift can be replaced by the riskless rate of interest (ie. =r). Any process describing the stochastic behaviour of the asset price will ead 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, FE sO ane rc (13) or I 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 (ie. Cy = max(0, Sp ~ K) for a European call and Cr = max(0, K — Sz) 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(dy) = Ke" N(ds) (1.6) where g, —S/K) + (r+ 4o)(T = 0) ' ott = d)~oVT-1 and where the parameters 5, K, r,t, T, and ¢ 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 8Introduction 10 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(¢,7), the price at time f of a pure discount bond with maturity date 7, the BSM formula can be written: C(t) = SN(d)) — KPC, T)N(ds) (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 ofthe dividend yield, Ifthe asset pays a constant proportional dividend ata 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-*?-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 isentered 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 “onstrike 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 "2 The proportional dividend is interpreted asthe risk Irce rate onthe orsign curreney in options ofthis type. " 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 Aiscrete 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 optionsis 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, ' Tinley (1993) Li and Zhang (1996), Broadie and Glasserman (19972, 19976) and Clewiow and Strickland (1998), amongst others have described methods of extending Monte Carlo simulation tothe valuation of ‘options with early exercise opportunites In Chapter 8 we show how Monte Caro simulation can be applied to American options in eonjunetion with a tee based approach ' 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 (18) Inthe 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 = ude + odz (19) where 1 =r ~ 5 — Jo. Consider a trinomial model of the asset price in which, over a small time interval 1, the asset price can increase by Ax (the space step), stay the same or decrease by Ax, with probabilities p,.P», and py respectively. This is depicted in terms, of x in figure Ls. The drift and volatility parameters of the asset price are now captured in this simplified discrete process by Ax, Py» Pmy and py. It can been shown that the space step cannot be chosen independently of the time step, and that a good choice is Ax = oV3A1. 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 Ar and requiring that the probabilities sum to one, i E(Ax] = py(Ax) + Pm (0) + Pa(-Ax) = vat (1.10) 4dr a —a FIGURE 1.5 Trinomial Model of an Asset PriceEnergy Derivatives: Pricing and Risk Management E{AX)] = ps A2x) + pm (0) + py Ox?) = 7 Ott PAP (1.11) Pot Pm+Pa=t (12) Solving equations (1.10) to (1.12) yields the following explicit expressions for the transitional probabilities: PAr+ AP var an y(SOepae =) (1B) pa = FO AE (ua) _fearevar var (AeA a) (1s) ‘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-N42 FIGURE 1.6 A Trinomial Tree Model of an Asset Price 12Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing Let i 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, denotes the level of the asset price at node (i) then we have = 1=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 fora 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 /) 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 NV correspond to the maturity date in terms of the number of time steps in the tree, i. T= NAt. The values of the option at maturity are determined by the values of the spot price in the tree at time step Vand the strike price of the option Cy = max(0,Syy — KY P= = Itcan 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: oN (1.16) Cy = Dy Cierer + Pm Cents + PACs 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, ie. Se-K} 8) (7) Gig = max{e™"sCossjer + PmChang + PAC, This method also gives us the optimal exercise strategy for the American option since for every possible future state of the world, ie. 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), Coo. Although we have so far discussed only the valuation of a simple derivative using the tree structure, itis 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 "© Sce Clewlow and Strickland (1998) for an in-depth discussion ofthe implementation oftinomial tees for Serivative pricing, B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 payofT 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 represenited in equation (1.8) in the following way: Migar = Xi + (VAEt olei.ar~ 2)) (1.19) Alternatively, in terms of the original asset price we have the discrete form Shear = Seexp(vAt + ofzesar 2) (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, 2;.., ~ z;,has a mean of zero and a variance of Ar. It can therefore be simulated using random samples from a standard normal multiplied by VAr, ic, VAte where ¢ ~ N(0,1). In order to simulate the spot price we divide the time period over which we wish to simulate, (0,71, into N intervals such that At = TIN, 1; = it, N, Using, for example, equation (1,20) we have Sy, exp(vAt + oVAte,) (121) 5 It is important to note that, since the drift and volatility terms do not depend on the variables S and 1, the discretisation is correct for any time step we choose. This enables us to 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= 1007-6 = 1/(365 x 24) © This sation is hased on Chapter 4 of Clewlow and Strickland (1998) °" Ror example a crack spread that pas the diference between gas nd elect ries. 05,0 = 0.30, At 4Introduction to Energy Derivatives and Fundamentals of Modelling and Pricing ‘Asset Price FIGURE 1.7. tus jon 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, Sz, — K) (1.22) Each payoff is discounted using the simulated short-term interest rate Gyy=exn(= frat), (1.23) In the case of constant or deterministic interest rates equation (1.23) simplifies to: oj = POTICr, 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: 14 ) 7 Oo (1.24)Energy Derivatives: Pricing and Risk Management Therefore C, is an estimate of the true value of the option, Cp, but with an error due to the fact that itis 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 SECs) (1.25) VM where SD(Cq) is the standard deviation of Cy: SE(Co) SD(C,)) ‘The biggest criticism of Monte Carlo methods concems 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. 16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 enerey 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 dillcuit 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 7Energy Derivatives: Pricing and Risk Management prices can then be shifted so as to be positive. Although the analytical results we discuss ‘ill 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 = al —InS)Sdt + oSdz (21) In this mode! the spot price mean reverts to the long-term level S = eat a speed given by the mean reversion rate, « which is taken to be strictly positive. Ifthe spot price is above the long-term level 5 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 5. 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 1 April 1999 to 31 March 2000. 35;— — — ———— ] 33} 344 20} 27} 25) 23 2a ‘Spot Prica (USD/MMETU) 19 47 185: - ___| Date FIGURE 2.1 Henry Hub Natural Gas Spot Price for the Poriod 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 cutve 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 = [alu x) Pldt+ ode (22) Equation (2.2) can be discretised as Ax; = [a(u — x)) — $07] At + oV Are, (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 isa 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, jz, ¢ 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, = 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 discus estimation ofthe mean reversion rate in the following setion In Chapters 6,7 and 8 we show the parameter 4 can be obiained as a time dependent function from the forward price and volatility curve Estimation of volatility s discussed in Chapter 3 19Energy Derivatives: Pricing and Risk Management 120 wf" dy A way ‘on c 8.05 O05 040 or 030 02s. Time (years) FIGURE 2.2 Simulation of a Geometric Brownian Motion Spot Price Path 5 =100,r— 6 = 0.05, 0 = 0.30, At = 1/(365 « 24) 120. 115: 110. “Ul gt it ‘Spot Price 800 008 oro O78 020 025 Time yours) FIGURE 2.3. Simulation of a Mean Reverting Path for the Spot Energy Price $= 100,c1 = 100.0, $ = 100, = 0.90, 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 = a(8—xpdr (2.4) Equation (2.4) can be solved by integrating from xo at time 0 t0 x, at time £ to obtain (%-3) = (9- Je" 25) To obtain the hall-life, which we denote by ,,2, at which the distance of x from its long term level is half its initial distance we have $0 ~ 8} = (x — He" (26) Solving for 1,2 we obtain ta =InQ}/a 7) ‘Table 2.1 gives the half-life for a range of values of the mean reversion rate. Itis 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 Hall-Lives T months 0 25 days 100 2:53 days 000 ‘hours 21Energy Derivatives: Pricing and Risk Management 8 8 Spot Price (USD/Barel) a 8 oes ounien eine vase Sunes Date FIGURE (UsD/Barrel) Brent Crude Oil Spot Price for the Period from 1 April 1999 to 31 March 2000 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 e¢ al and Dupire can be represented by the following process for the spot price: AS = pSdi + o(t, S\de 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 oft, S) = 6,8” (2.9) where (3; and (3 > 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) 2 This modelling approach i also discussed in Chapter 7. For a detailed discussion of the implementation of implied trinomia toes seo Cleslow and Steikland (1998), Chapter 5 2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), itis 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 = pSdt + oSdz (2.10) aV = a(P — V)dt-+ eV Vdlw Qn) Equation (2.10) is the GBM model but with a volatility, ¢, 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 squate 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 VV which is proportional to the square root of the variance i.e. the volatility of the spot price. The source of randomness in the variance, dv, is different from the d= 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 iscretised in the following way" Any = (uf) Art oV Bie, ea ~ ryan +eVFBi( vers + y= een) 2.13) where ¢) and ¢» 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 AY, = af! > The spot price process i dlscrtised in trms ofits natural logarithm x. 23Energy Derivatives: Pricing and Risk Management 120 89 — 8.00 005 010 048 020 02s Time (years) FIGURE 2.5 a) _A Simulated Stochastic Volatility Path forthe Spot Energy Price $= 100, ~005,0 030, — 10,V —008,¢ = 05, = 0.0, t= 1/ (305 » 28) price and the variance of the spot price. We first obtain or choose values for js, 0, 4, & P, and Af, Each time step of the simulation then involves generating two independent normally distributed random numbers ey and ¢2, 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 isa 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 tothe Australian electricity markets 24Understanding and Analysing Spot Prices a 005 O70 O78 020 28 Time (years) FIGURE 2.5 (b) _A Simulated Stochastic Volatility Path for the Spot Energy Price $= 100,j1 = 0.05,0 = 0.30,a~ 10, V = 0.08,¢ = 0.5, p = 0.0, At = 1/(365 = 24) a 300: g Eso 100) | “Aa lala al FIGURE 2.6 Australian New South Wales Electricity Pool Prices for the Period from 13 December 1998 to 8 March 1999,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 = wSdt + oSd2 + wSdg (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 +) ~ M(In(1 + 8) — 47,77) (2.15) where fis the mean jump size and + is the standard deviation of the proportional jump size which we call the jump volatility, The jump process (dg) isa diserete 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 | 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, ifthe 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- iffusion model described by equation (2.14) can be discretised as follows: Ax, = (r= 68 — fo Art oV Ate, + (+7 x)(ui > GAD) (2.16) > See Merton (1990) fr s general dicussion of jmp processes applied 10 financial markets © The risk-neutral prooess i obtained by setting x= (7 — 8) Technically #is defined by probaly 1) = 9 de %6Understanding and Analysing Spot Prices where ¢; and cp are independent standard normal random variables and w is a uniform (0.1) random sample. The term (14; < ¢ Ar) is taken to be oneif 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 Ar must be small relative to the jump frequency such that ¢ Ar < 1. Once the parameters d, , 0, and At have been chosen, at each time step, we first generate a uniform random number, u, in order to determine if a jump should occur. [fa jump is to occur we generate the normally distributed random number 2 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 sulficiently 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 = aly In S)Sdt + oSde + Sg (2.17) iy AY WAM, os i . N\A vay!" Spot Price 0.00, 0.05 0.10 015 020 0.25 “Time (years) FIGURE 27 _A Simulated Jump-Ditlusion Path for the Spot Energy Price $= 100,r = 0.08, = 0.30, =0,¢ = 100,7 = 0.02, At = 1/(365 = 24) 2Energy Derivatives: Pricing and Riske Management 8.00 005, 010) 0.15 020 0.25 Time (years) FIGURE 2.8 A Simulated Mean Reverting Jump-Ditfusion Path for an Electricity Spot Energy Price $ = 20,1 = 2000, = 20,1 = 30.0, «= 0, ) = 250, = 1.2, At = 1/(368 x 24) 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 = of ~ x)dt +0 (2.18) This can be discretised as follows: Ax, = ay bax +96, (2.19) where ap = aXAr 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 ix essentially the same as equation (2.2) but with the -1/2 0° incorporated into x 28Understanding and Analysing Spot Prices and x; in the presence of noise (represented by as,). Therefore, if we regress observations. of Ax, against x, we can obtain estimates of ag = axArand a, = —aAr as 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 x. 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 Ar 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): 49 = 0.0398(0.0152) a = 0.0428(0.0168) which implies x=25 a= 111(4a) This value of a and standard error implies a half-life for price shocks of between 16 and 38 days. 04, : os 005 “05 6 or 08 09 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)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 ofa 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, itis 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 Halthour Poo! Price Returns for 13 December 1998 to FIGURE 2.10 Australian New South W: 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 A. 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 ease 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: 6 = number of jump returns / time period of data 2.20) = average of jump returns e2n +) = standard deviation of jump returns 222) 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-Ditfusion Parameters of NSW Electricity eration sD Jumps 6 ‘ y Y 0.1183 37 219 0.000189 0578 2 out rez ane 0.000107, 0.4957 3 0.1088 9 S75 0.000197, 0.4683 4 0.1076 9 539.9 0.000696, 0.4595 5 0.1065 i) ox 0.001022 0.4513 6 0.1088 136, 662.1 0.001105 0.4462 7 0.1080 163, ols 0.001348 aaatt 8 0.1043 170 ms 0.001271 0.4368 ° 0.1041 in 00 0.001830 0.4354 0 o.i081 in 7300 =0.001480 0.8354 SD is the standard deviation of the returns after emoving 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 & 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:8), that a given set of log price changes, Ax, would hhave 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;9), we obtain a sum of the logs of the probabilities for the individual observations. If we maximise L(A.x;@) with respect to the parameter set @ we obtain the set of parameters which maximises the likelihood that we would have observed the set of price changes in our data set. Phe density funetion of changes in » is given by eo 0a pane) =>§ OA n (Ax; RAt-+jIn(1 + %),0°At +i92) (2.19) a ‘where n(x,b) is the normal density with mean a and variance b and the vector © = (&,¢, 4, >). Assuming we have m spot price returns, Axy,...,Axy, the logarithm of the likelihood function is given by InL(Ax,@) = )Inp(Ax;,0) (2.23) 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: F = 0.0000, « = 0.1938, ¢ = 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 2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 asa random variable with time varying parameters. If we assume that volatility follows a predictable (ie. non-random) seasonal pattern then this is equivalent to replacing the constant volatility parameter, ¢, by a function of time, o(0), which repeats each seasonal éycle. We ean estimate the function o(0) from historical data by using a roling 30 day sample standard deviation for example”. Figure 2.11 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 Qn, 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 clectricity 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 itis 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 ° tn practice this “raw” estimate of de seasonal volatility funtion shouldbe smoothed, for example by Sting ‘smooth functional form such asa polynomial (ce Preset a (1992) *"For example in the UK and Ausiealian spt lett Sted on a half hourly bass 33Energy Derivatives: Pricing and Risk Management 0% — oriowwe 2189 TaG7ee —2ecRS —TeTONe —G7/I2e9 2810100 1609100 Date FIGURE 2.11 Henry Hub Natural Gas Seasonal Spot Volatility Over the Period trom 1 April 1999 to 31 March 2000, es oe 3000} os] “\ 1500) Mean Reversion Rate 1000! ro 2 Week FIGURE 2.12 Mean Reversion Rate of UK Electricity Pool Prices on a Weekly Basis for January 1999 to September 1989 34Jamo Frequenoy 3 2100 1900 700 1500 0 Understanding and Analysing Spot Prices rr 3 2 Wook 3 30 % FIGURE 2.13 (a) Jump Frequency of UK Electricity Pool Prices on a Weekly Basis for January 1899 to September 1999 ben vy FIGURE 2.13 (b) _Jump Volatility of UK Electricity Pool Prices on a Weekly Basi January 1989 to September 1998 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 istorical estimates. However, for electricity data itis 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 astochastic 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 was 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. "Tn practice the historically estimated seasonal patterns should be smoothed in the same manner as for seasonal voatlity cycles by fitting a smooth function such as a polynomial (see Pres eal. (1992)). 363 Volatility Estimation in Energy Markets Vince Kamins! 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). ‘© The volatility definition and measure should capture the key features of energy 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 asa 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 betow. 0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 isa 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, Spatt (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 riven 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 t0 the cast 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 pre 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 readjustment 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 Volatility In this section we look at issues concerning the estimation of volatility including both historical and implied measures 39.Brergy 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,/S,_ — 1. The logic of the approach described above is that for relatively small x, In(1 + x) * x. Taking the natural log of S,/S,.1 is equivalent to taking the natural log of 1 + 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 (Syin/Sren-1) * (Sren-1/Sran-2) X =X (Syai/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 Jog returns for the sub periods Step 2. Caleulate standard deviation of the logarithmic price returns. Step 3. Annualise the standard deviation by multiplying it by the correct factor. Asa first approximation the annualisation factor depends on the price data frequency. In the case that the data is monthly, the factor is y/12; for weekly data it is y/52. For the daily data available for each calendar day one has to use y/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 distributed) random variables, the variance of the sum of 1 random variables, 21,%2).01 %q #8 equal to the sum of the variances (if price returns are uncorrelated): Says tcthy = Oy +O to OS, G4) 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 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 98s Sat Say Sar Say where s represents the multiplicative seasonality coefficient, common to both prices, Sy. corresponds to the deseasonalised price for period 1. 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) $ of the form dS = a(S — S)dt + odz (3.2) If the starting price level is (0), the expected price at time f will be given by 5+ (S(O) ~ S)exp(~at). As the term exp(—at) goes to 0, as 1 becomes large, the ong-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 ¢ 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 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, = ap +018, +5, G3) 41