MSc in Finance Author: Rikke Knold Christensen, 280702 Department of Business Studies Academic advisor: Elisa Nicolato Stochastic Volatility Models and the use of the Fourier Transformation method as an option pricing tool Aarhus School of Business 01-09-2011 Content CHAPTER 1: INTRODUCTION .................................................................................................................................... 2 1.1 1.2 1.3 1.4 PREFACE ............................................................................................................................................................... 2 PROBLEM STATEMENT .......................................................................................................................................... 3 METHODOLOGY .................................................................................................................................................... 4 DELIMITATION ...................................................................................................................................................... 6 CHAPTER 2: THE BASIC THEORY OF OPTION PRICING .................................................................................. 7 2.1 STOCHASTIC PROCESSES ....................................................................................................................................... 7 2.2 THE BLACK-SCHOLES MODEL ............................................................................................................................ 10 2.3 EXTENSIONS TO THE BLACK-SCHOLES MODEL ................................................................................................... 16 2.3.1 EMPIRICAL EVIDENCE OF A STOCHASTIC VOLATILITY..................................................................................... 16 CHAPTER 3: STOCHASTIC VOLATILITY MODELS ........................................................................................... 21 3.1 GENERAL STOCHASTIC VOLATILITY MODELS..................................................................................................... 21 3.1.1 DERIVING A GENERAL PDE FOR STOCHASTIC VOLATILITY MODELS ............................................................... 24 3.2 THE HESTON STOCHASTIC VOLATILITY MODEL ................................................................................................. 27 3.2.1 THE PROCESS OF THE HESTON MODEL............................................................................................................ 27 3.2.2 PDE FOR THE HESTON MODEL ....................................................................................................................... 30 3.2.3 CHARACTERISTIC FUNCTIONS AND THE FOURIER TRANSFORMATION ............................................................ 31 3.2.4 IMPLEMENTATION OF THE FOURIER TRANSFORM ON THE HESTON MODEL ..................................................... 33 CHAPTER 4: DATA ...................................................................................................................................................... 36 CHAPTER 5: FITTING THE IMPLIED VOLATILITY SURFACE ....................................................................... 38 5.1 5.2 5.3 5.3.1 5.3.2 5.3.3 5.4 5.5 5.6 MODEL CALIBRATION ......................................................................................................................................... 38 SETTING THE INITIAL PARAMETERS ..................................................................................................................... 40 SIMULATION OF THE HESTON PROCESS ............................................................................................................... 43 MONTE CARLO SIMULATION .......................................................................................................................... 44 MILSTEIN SCHEME ......................................................................................................................................... 45 MIXING SOLUTION .......................................................................................................................................... 46 COMPARISON OF SIMULATED AND FOURIER PRICES ............................................................................................ 48 THE IMPLIED VOLATILITY SURFACE OF THE S&P500 INDEX OPTIONS.................................................................. 50 THE HESTON MODEL’ SENSITIVITY TO ITS PARAMETERS ..................................................................................... 53 CHAPTER 6: CONCLUSION ...................................................................................................................................... 59 LIST OF LITERATURE ............................................................................................................................................... 62 1 Chapter 1: Introduction 1.1 Preface In the year of 1973 one of the most recognized models for stock options prices were introduced by Fischer Black, Myron Scholes and Robert Merton (Black and Scholes, 1973). The model, which today is referred to as the Black-Scholes Model, was the first model to present a closed form solution to option pricing. Because of its simplicity and easy implementation the model gained an increasing influence on the way that traders priced and hedged the options traded on the market throughout the years. However, after the stock market crash in 1987, the characteristics of the stock market changed. As a result of this, one of the main implications of the Black-Scholes Model, namely that log-returns are normally distributed, was no longer visible in the market. Empirical studies have later shown that the stock returns from 1987 and forward display a skewed leptokurtic distribution with heavy tails. Furthermore, the stock returns also show clustering in periods, which is evidence of a stochastic volatility. This is a clear violation of the constant volatility assumption underlying the Black-Scholes Model, and stock returns therefore cannot be characterized as constant over time. In order to better understand the way that options are priced, it can be useful to examine the socalled implied volatility surface. The volatility surface is a collection of implied volatilities, which is essentially the volatility that generates market prices when inserted into the Black-Scholes formula. Among practitioners it is known as “…the wrong number in the wrong formula to get the right price…”(Rebonato, 1999) As empirical evidence has shown, as mentioned, the market depicts a stochastic volatility, which means that the implied volatility depends on both the strike price and expiration of the option – and is not constant. The volatility surface and the examination of stock returns can therefore be used to analyse 1) why options are priced the way that they are and 2) how they should be priced if they were priced fairly (Gatheral, 2006, page xxiii). One of the most recognized stochastic volatility models was introduced by Heston (Heston, 1993) in 1993, which much like the Black-Scholes Model provides a (semi)-closed form solution for pricing options. The model is one of the most widely used models that exist on the market, and the model particularly introduces the mentioned (semi)-closed form solution for pricing European call options based on a volatility measure related to a mean reverting square root process 1. The volatility is thereby assumed to be non-constant and to follow a random process. There exist several methods for estimating the stock price of an option through the Heston model, and these methods include 1 The mean square root process was first introduced by Cox, Ingersoll and Ross in 1985 (Gatheral, 2006, page 15) 2 simulation and Fourier Transformation. The different approaches have each their advantages and disadvantages. The simulation approach has become very popular for option pricing as empirical studies have shown that the simulation accuracy increases by the number of simulations performed; however, this also means that the implementation time for the process increases. As a result of this, there is an increasing interest for the use of Fourier Transformation for pricing options. This method provides an estimate of the stock option price by the solution of an integral – which thereby makes it computationally easy compared to the simulation approach. 1.2 Problem Statement The main purpose of this paper is to analyse a chosen model to fit the implied volatility surface of the market and to test the valuation method of Fourier Transformation. In order to analyse the model, the basic theory behind option pricing is first presented and the shortcomings of the BlackScholes Model is analysed. As some of the real world implications of this model are quite severe, the Black-Scholes Model can be considered insufficient for describing the nature of the financial markets (to a certain extent). This raises the need for a model that incorporates stochastic volatility in order to better fit the data observed in the market. The chosen model to be fitted on the implied volatility surface is the Heston stochastic volatility model. Only one stochastic volatility model has been chosen, in order to avoid the paper to become too comprehensive. There will be two valuation methods implemented to the Heston Model: 1) a simulation and 2) Fourier Transformation. This will enable an evaluation and discussion of the use of the Fourier Transformation as an option pricing tool, as it can be directly compared to the simulation approach. This lead to the overall problem statement: How can a model with stochastic volatility be derived to fit the implied volatility surface of the market? To answer the problem statement, the following research questions have been defined: 1) What is the basic theory behind option pricing? 2) What is stochastic volatility and why is it needed? 3) What are the characteristics of the chosen model to be fitted (the Heston Model)? 4) How can the Heston Model be fitted to the implied volatility surface? 3 a. How well is the implied volatility estimated from the Simulation vs. Fourier Transformation b. How well does the Fourier Transformation approach perform as an option pricing tool? 5) How well does the Heston Model capture the behaviour of the market? On the basis of the Heston Model the thesis will be build on the historic data from the S&P500 Index, which consists of 500 large-cap company stock options. The index is one of the most followed indexes and it is widely said to be the best single gauge of the American equities market (http://www.standardandpoors.com/indices/sp-500/en/us/?indexId=spusa-500-usduf--p-us-l--). This makes the index very liquid, which is essential for the analysis and the required data. In the end, the thesis will discuss the differences between the empirical volatility surface and the model fitted volatility surface. The thesis will thereby have analysed the Heston model’s capability to capture the behaviour of the S&P 500 options and discuss the use of Fourier Transformation as an option pricing tool. 1.3 Methodology The main structure of the thesis is divided into two parts – a theoretical part and a part where the theory needed for the analysis of the Heston Model and Fourier Transformation is applied, as shown in figure 1.3.1. Part 1 consists of Chapter 2 and 3 where the theoretical framework of the thesis is introduced. The basic theory behind pricing option is introduced, and the ideas behind stochastic processes, including the Brownian and Geometric Brownian motions is shortly presented. Afterwards, the famous Black & Scholes Model will be outlined and the rationale behind the risk neutral world will be introduced. However, empirical studies have shown that the Black-Scholes Model has several shortcomings, and is therefore insufficient in pricing options to a certain degree. Because of this, the theoretical extensions to the Black-Scholes Model and the existence of stochastic volatility will be discussed, introducing the implied volatility and the existence of the volatility smile. In Chapter 3 the basics for a better option pricing model will be established by the introduction of stochastic volatility models. This alternative to the Black-Scholes Model will be discussed and the 4 chosen stochastic volatility model, the Heston Model, will be presented. After this, the theory behind using Characteristic Functions and Fourier Transformation will be presented in order to solve the valuation problem and fit the Heston Model to the implied volatility surface. Figure 1.3.1 – Methodology of the thesis Theoretical aspect of the thesis Chapter 2 - Basic Option Pricing Chapter 3 - Stochastic Volatility Models Implementation Chapter 4 - Data basis Chapter 5 - The implied volatility surface Conclusion Remarks on the performance of the model and the Fourier Transformation Source: Own contribution The second part of the thesis implements the theory discussed in the previous paragraphs and applies a stochastic volatility model to the market data. First, the data used in the thesis will be presented in Chapter 4, and the chosen underlying index, the S&P500, will be shortly introduced. Thereafter, the data basis for the thesis will be established and the use of a proxy for the risk free interest rate will be discussed. In Chapter 5 the actual fitting of the Heston Model to the implied volatility surface will be conducted. The calibration of the Heston Model parameters will be performed by the use of Fourier Transformation, where the parameters will be fitted to the market data. The evaluation of the Fourier Transformation will then be performed by simulating the Heston process by using a Monte Carlo simulation, and comparing the two methods. In the Monte Carlo simulation the Milstein Scheme will be introduced and applied to the variance process in order to discretize the process. However, the Heston Model is very sensitive with respect to its parameters, and therefore a discussion of this will be performed. As an extension of this the sensitiveness of the parameters will then be tested in order to illustrate the effect of changing the parameters on the 5 implied volatilities. In the end a comparison of the empirical and the Heston fitted volatility surface will serve as an estimate the performance of the Heston model. Finally Chapter 6 will give a conclusion on the thesis and the findings that has been made. 1.4 Delimitation It is assumed that the reader is familiar with some of the basic theory underlying option pricing, and this will therefore only be presented shortly as a recap and a basic for the development of the stochastic volatility models. Furthermore, the analysis will only be conducted for European call options, which means that American options and put options are delimited as they are beyond the scope of the thesis. In the introduction of stochastic volatility models, some alternative models other than the Heston Model will shortly be presented. These models will be presented in different notations compared to the original sources, in order to make the comparison of the models easier. Further analysis and derivation will not be conducted on these models. Since the purpose of this thesis is to fit the Heston Model to the implied volatility surface of the S&P500 index, the primary focus will be on the Heston Model and Fourier Transformation. As the thesis thereby have a practical aspect, the Heston Model and the theory underlying this will be presented in a way that a practical approach and use of the model is possible for the reader of the thesis. 6 Chapter 2: The Basic Theory of Option Pricing The purpose of the following section is to introduce the basic theory that is necessary for developing a model to fit the volatility surface of the S&P 500 index. The section will more specifically introduce the basic theory behind option pricing, and discuss how it applies to the real world. The section will begin with explaining the basic theory underlying the pricing of options, which will be followed by the introduction of one of the most used option pricing models throughout the years, the Black-Scholes model. In spite of this, the Black-Scholes model suffers from severe real world implications, which makes the model inadequate for option pricing to a certain extent. The discussion of the real world implications will lead to the introduction of the implied volatility and volatility smile, which are empirical proofs that some of the basic assumptions underlying the Black-Scholes model are violated in the real world. This establishes the need for an alternative option pricing model that better describes the evolvement of the stock options on the market. 2.1 Stochastic processes The first step in determining the price of an option is to understand how the underlying of a stock option evolves over time. Any variable that changes in an uncertain way over time is said to follow a stochastic process. There exist two types of stochastic processes, one in discrete time and one in continuous time. The difference between the two is that a discrete time stochastic process can only be measured on specific times while a continuous time process can be measured continuously. One of the most common processes that are used to describe the evolvement of a stock price over time is the so-called Geometric Brownian Motion. The Geometric Brownian Motion is a special case of a Wiener Process or a Brownian Motion2, which has a mean of zero and a variance of 1 each year. In differential form the Geometric Brownian Motion describes the changes in a variable Z as shown below (Hull, 2008, page 266): ππ = ππππ‘ + ππππ Equation 2.1.1 where S is characterized as the stock price at time t, μ is the drift term or the expected return of the stock option and σ is the volatility of the stock price. The term d corresponds to an infinitesimal 2 The Geometric Brownian Motion is also called a Generalized Wiener Process 7 small change of the variable, which means that the process is approximately continuous when dt approaches 0. The Geometric Brownian Motion has the advantage that the stock price will always remain positive as long as the initial value of the stock price is positive. This is an advantage over the standard Brownian Motion where the change in the stock price process, dS will move towards zero as the stock price approaches zero. Another advantage for the Geometric Brownian Motion is that the expected percentage return of a given stock option is independent of the stock price itself. This is due to the fact that the process itself depends on the level of the stock price. It should be noticed that the Geometric Brownian Motion is defined as a continuous-variable, continuous time stochastic process, while stock prices can only be measured in discrete time. The measurement of stock prices and the corresponding changes of these are restricted, because stock options are only tradable when the exchange is open. Therefore, the prices and price changes can only be observed at these points in time (Hull, 2008, page 259). In discrete time the return of the stock price (or in other words the relative change of the stock price) can be defined as the relative change in the stock price over a short period of time, βt, like shown below: βπ π = πππ‘ + πππ Equation 2.1.2 Where the process like before is determined by a drift3, μdt, and a stochastic rate4, σdZ. The term Z is a Generalized Wiener Process, where μ and σ are constants, if the following properties are fulfilled (Hull, 2008, page 261 and 263): PROPERTY 1: The change of βZ during a small period of time, βt, is: βπ = π√βπ‘, π€βπππ π ~ π(0,1) From the first property it follows that the expected value or mean of the variable Z should be equal to zero as derived in the following: πΈ(ππ‘ ) = πΈ(π√π‘) ⇔ πΈ(ππ‘ ) = 0, because ε ~ N(0,1) 3 4 Mean change per unit time Variance per unit time 8 The variance is also easily derived as it can be shown that the variance is equal to t: 2 πππ(ππ‘ ) = πππ(π √π‘) ⇔ πΈ [(π√π‘) ] πππ(ππ‘ ) = π‘ → π = √π‘ As a consequence hereof, the standard deviation of the variable Z is equal to the square root of time. PROPERTY 2: The values of βZ1 and βZ2 for any two different short intervals of time, βt1 and βt2, are independent of each other. It follows from this property, that βZ must follow a so-called Markov process. The Markov property is a special case of a stochastic process, where the only relevant variable for predicting of the future is the present variable (Hull, 2008, page 259). The process therefore implies that all relevant information from the past prices is already incorporated into the present value, and therefore the Markov process is also known as a process without memory. This property is also consistent with empirical evidence from the real world as the competition in a market contributes to ensuring the weak form of the market efficiency5. In addition the process is also assumed to be a martingale process, which means that conditional expectation of the variable Z at any time t, is the value of the stock today. PROPERTY 3: The initial value of Z is equal to zero. π(0) = 0 PROPERTY 4: (Quote, Hull, 2006, page 263): πβπ ππ₯ππππ‘ππ πππππ‘β ππ π‘βπ πππ‘β ππππππ€ππ ππ¦ π ππ πππ¦ π‘πππ πππ‘πππ£ππ ππ πππππππ‘π PROPERTY 5: (Quote, Hull, 2006, page 263): πβπ ππ₯ππππ‘ππ ππ’ππππ ππ π‘ππππ π πππ’πππ πππ¦ ππππ‘ππππ’ππ π£πππ’π ππ πππ¦ π‘πππ πππ‘πππ£ππ ππ πππππππ‘π 5 Weak form for market efficiency is where the information set includes the history of prices and returns, and only this (Campbell et al, 1997, page 22) 9 2.2 The Black-Scholes Model The Black-Scholes model was first introduced by Fischer Black and Myron Scholes in the famous article “The Pricing of Options and Corporate Liabilities” in 1973. This, together with Robert Mertons article “Theory of Rational Option Pricing” (1973), began the biggest breakthrough in option pricing that the world has ever seen. The Black-Scholes Model (also known as the BlackScholes-Merton Model) is considered to be one of the most influential models on the market for option pricing (Hull, 2008, page 277). The model was the first of its kind to suggest a closed-form solution to pricing stock options. Because of the ease of implementation and its close fit to the observed market prices (compared to other models that were on the market at that time), the model gained recognition and quickly became very popular among traders and researchers. As a consequence of this, the model had a huge influence on the way that traders price their options, and still does this day. The partial differential equation of the Black-Scholes Model must be satisfied by the price of any derivative dependent on a stock option (not paying out dividends) (Hull, 2008, page 285). The basic assumptions underlying the Black-Scholes Model are shortly listed below (Hull, 2008, page 286-287): ο· The stock prices follows a Geometric Brownian Motion with a constant mean and variance ο· Short selling of securities is permitted – with the full use of proceeds, and all securities are perfect divisible ο· No transaction costs or taxes ο· No dividends ο· No riskless arbitrage opportunities in the market ο· Continuous security trading ο· Constant risk free rate, r, which is the same for all maturities Many of the abovementioned assumptions underlying the Black-Scholes Model are in fact violated to some degree in the real world. However, the consequences and severity of the violations vary across the assumptions. One of the obvious violations is the one concerning no transaction costs. In the real world there are often taxes on the derivative (this depends on the country in question) as well as transaction costs. As a result of this it must be assumed that this will already be incorporated 10 into the final price of the stock option. The assumption concerning no dividends is without any doubt also often violated in the real world. However, this can easily be relaxed, as the BlackScholes equation can be modified to take dividends into account, see equation 2.2.6. The assumption concerning no arbitrage opportunities is also questionable, as some people earn their living on finding such opportunities in the market. However, whenever an arbitrage opportunity exists, the market will quickly rebalance itself again and therefore the arbitrage opportunity will disappear relatively quickly. It is also assumed that the securities are continuously traded, which means that the delta hedging6 of the portfolio is done continuously. However, this is not exactly true in the real world, where the transaction cost often have an influence on the hedging frequency – the lower the transaction cost, the higher the frequency of hedging, which contributes to making the security trading discrete (Wilmott, 2005, page 146). Lastly the risk free interest rate, r, is assumed to be constant, but in reality the interest rate is stochastic and cannot be known in advance. The basic idea underlying the Black-Scholes Model is simply to perfectly hedge a portfolio, so that risk is eliminated. This is known as a so-called delta-hedge. The final return of the portfolio will then equal the return of the derivative at expiration, no matter how the price of the underlying asset/stock evolves. This is done by constructing a riskless portfolio from a position in the derivative and a position in the stock. The return from the portfolio must then equal the risk free rate in order to fulfil the assumption of no arbitrage opportunities in the market. This is made possible because stock price movements have a similar affect on stock price and the price of the derivative. Thereby the price of the underlying stock will be perfectly correlated with the price of the derivative in any short period of time. This makes it possible to construct a portfolio where a possible gain or loss in the stock position is offset by a gain or loss from the derivative position – thereby eliminating the risk and hedging the portfolio. However, it is important to emphasize that the constructed portfolio is only riskless in a very short period of time, and for the portfolio to remain riskless, it is necessary to continually rebalance the portfolio. This will be explained in more detail in the following, where the construction of the riskless portfolio that enables the BlackScholes differential equation will be derived. The value of a stock option is determined by the stock price (S), the time (t), the standard deviation (σ), the drift (µ), the strike/exercise price (K), the time to maturity (T) and the risk free interest rate 6 Delta-hedging will be explained later 11 (r). In the following, the value of the stock option will be referred to as f(S,t) instead of f(S, t, σ, µ, K, T, r) for simplicity. As mentioned earlier, the value of the portfolio must consist of a long position in the stock option and a short position (β-quantity) in the underlying derivative. The value of the portfolio is then defined as: Π(π‘) = π(π, π‘) − Δπ In the next time interval, t + dt, the value of the portfolio depends on the change in the value of the stock option as well as the change in the underlying derivative. Before rebalancing the portfolio, this can be written as the following: πΠ(π‘) = ππ(π, π‘) − Δππ When determining the price of an option, it is useful to consider the Itô process. The general Itô process is a Generalized Wiener Process and can be defined as (Hull, 2008, page 265): ππ = π(π, π‘) + π(π, π‘)ππ Equation 2.2.1 which states that the change in the underlying asset is determined by a drift, a, and a variance, b. As a result of this, any derivative is a function of the time and the stochastic variable underlying the derivative. In relation to option pricing an important result was found in 1951, where it was shown that a function of a variable and the time follows the following process: ππ(π, π‘) = ππ ππ π2 π 1 ππ‘ + ππ ππ + 2 π 2 (π, π‘) ππ2 ππ‘ ππ‘ Equation 2.2.2 where π 2 = π , since the diffusion term is not depending on S. This above equation is known as Itô’s Lemma. When inserted into the portfolio equation, the value of the portfolio changes then equals: ππ ππ 1 2 π 2π πΠ(π‘) = ππ‘ + ππ + π (π, π‘) 2 ππ‘ − Δππ ππ‘ ππ 2 ππ ππ 1 π2 π ππ β πΠ(π‘) = ( ππ‘ + 2 π 2 (π, π‘) ππ2 ) ππ‘ + (ππ − Δ) ππ 12 When rearranged, the first part of the term can be characterized as the deterministic part or the drift (the term in front of dt) whereas the latter part is a stochastic part, that generates the risk of the portfolio. It is this part that must be eliminated in order to make the portfolio riskless. The risk generating part of the previous equation is therefore defined as: ππ ( − Δ) ππ ππ According to the Black-Scholes the uncertainty (the risk) can be eliminated if the delta quantity in the small period of time from t to t+dt is set to the following (otherwise known as a delta-hedging of the portfolio): ππ ( − β) ππ = 0 ππ β β= ππ ππ which delta-hedges the portfolio. However, as the delta-hedge is not constant (depends on both S and t) it is, as mentioned, necessary to rebalance the portfolio continuously in order to keep the portfolio riskless. However, if continuous trading is allowed (like assumed in the Black-Scholes Model) this is made possible, and the dynamics of the constructed riskless portfolio can be defined as: ππ 1 2 π 2π πΠ(π‘) = ( + π (π, π‘) 2 ) ππ‘ ππ‘ 2 ππ If there exist no arbitrage opportunities in the market, the riskless portfolio must yield the risk free interest rate, r. This essentially means that the investor doesn’t require any excess return to take on the risk of the portfolio. If this is not the case, the return from holding the portfolio is higher than the return of a risk free asset, and thereby an arbitrage opportunity will exist and it will be possible to make money by buying the portfolio for lend money. Therefore it must hold that: πΠ = πΠππ‘ Inserting the equations into the above formula, dividing by the time step, dt, and rearranging will then equate the Black-Scholes differential equation as written below: 13 ππ 1 2 π 2π ππ + π (π, π‘) 2 + π π − ππ = 0 ππ‘ 2 ππ ππ Therefore it can be concluded that for the market to have no arbitrage opportunities, the BlackScholes equation must hold for any derivative depending on a non-dividend paying stock option. There are several possible ways to solve the Black-Scholes partial differential equation. It can either be solved directly with the closed form solution, or through numeric methods or approximations. One of the possible methods is the risk neutral valuation, which is also known as the martingale approach. In terms of a risk neutral valuation it is important to emphasize that the Black-Scholes equation does not depend on the value of the expected return, µ. The underlying reason for this is that the value of the expected return depends on the risk preferences of the investor. This means that the more risk averse the investor is, the higher the return, µ, that the investor demands also is. However, as the Black-Scholes equation was eliminated for risks by choosing a delta quantity to be hedged, the investor is assumed to be risk neutral. Risk neutral investors do not require a premium for the added risk of an investment, and therefore the market price of risk is equal to zero (Hull, 2008, page 626). The expected return of all investments under a risk neutral assumption is therefore the risk free rate of interest, r. The only risk related to the option pricing will therefore already be incorporated into the volatility. The solution to the Black-Scholes equation is the same, whether or not a risk neutral world is assumed. This means that the solution to the equation will yield the same price regardless if it is done under risk neutral assumptions or under the assumptions of the Black-Scholes Model. This is validated by the Feynman-Kac theorem, which when applied to the Black-Scholes equation gives: π(π, π‘) = πΈπ‘ [π −π(π−π‘) π(π ∗ , π)] = π −π(π−π‘) πΈπ‘ [π(π ∗ , π)] Equation 2.2.3 where the process S* is computed in relation to the expected value of the option as: ππ ∗ = ππ ∗ ππ‘ + ππ ∗ ππ, π€βπππ π ∗ (π‘) = π Equation 2.2.4 Here the term Z refers to a Geometric Brownian Motion under risk neutrality. The only difference from this process and the process under real world assumptions is that the drift term is given by the 14 risk free rate, r, instead of µ, as discussed earlier. To account for dividends, µ should be replaced by r-q. Under risk neutrality the explicit formula for the price of a European call option is defined as the following: π(π, π) = π0 β π(π1 ) − πΎπ (−πβπ) β π(π2 ) Equation 2.2.5 π(π, π) = π (−πβπ) (π0 π (πβπ) β π(π1 ) − πΎ β π(π2 )) π€βπππ, π1 = π π (πβπ) π2 ln ( 0 πΎ ) + ( 2 ) β π π√π π2 = π1 − π√π In the above formulas for the Black-Scholes Model the term S corresponds to the spot price of the underlying asset at time t (in the above formula this corresponds to time t = 0), K is the strike/exercise price, r is the risk free interest rate and σ is the volatility of the underlying asset. The term N(x) is the cumulative probability distribution function for a standardized normal distribution (Hull, 2008, page 291). If the first equation in equation 2.2.5 is rewritten to the second equation, the term N(x) can more easily be explained, as N(d2) then becomes is the probability that the option in question will be exercised in the risk neutral world (Hull, 2008, page 292). The original model as presented by Black and Scholes in 1973 did not take into account the dividends that many options pay in the real world. However, with only a couple of slight adjustments, the model can be modified to do so as well as being able to price American options (this latter will not be further discussed here, as it is not relevant for this thesis). Taking dividends into account the explicit formulas for the Black-Scholes Model are as follows: π(π, π) = π (−πβπ) (π0 β π (π−π)π β π(π1 ) − πΎ β π(π2 )) π€βπππ, π1 = Equation 2.2.6 π π (π−π)π π2 ln ( 0 πΎ )+(2)βπ π√π π2 = π1 − π√π 15 Later in the thesis the above found equations for European call options will be used to value call options on the S&P500 in the mixing solution. 2.3 Extensions to the Black-Scholes Model Empirical studies have in many years underlined the real world implications for the Black-Scholes Model as implied in the preface. In the following paragraph stochastic volatility will be presented. Empirical studies have shown that the volatility of stock option returns are not constant over time, which was one of the main implications of the Black-Scholes Model. Instead empirical data shows that the volatility depicts a curve, a so-called volatility smile, which insinuates the presence of a stochastic volatility in the market, which will be illustrated in the following. This will among others be done by introducing the term implied volatility and implied volatility surface. 2.3.1 Empirical evidence of a stochastic volatility In practice the term implied volatility is defined as the volatility that yields the market price of a given option when inserted into the Black-Scholes pricing equation (Wilmott, 2005, page 183-184). However, as the volatility is the only unknown parameter in the Black-Scholes formula the volatility also becomes a measurement of the market’s expectance of the future volatility. This means that the market price is set on the basis of the volatility that the market expects of the underlying asset/stock in the future. Traders often quote the implied volatility of an option instead of its price (Hull, 2008, page 297), as the implied volatility is usually less volatile than the actual option price itself. This enables the traders to compare the options across strike prices, the underlying, observation times and maturities, in order to estimate an appropriate implied volatility for another option. One of the most discussed assumptions underlying the Black-Scholes Model is the one concerning the constant volatility – this means that in the Black-Scholes model the implied volatility is actually assumed to be constant. In the Black-Scholes framework, the volatility is an increasing function of the stock option price – this implies that a higher volatility, ceteris paribus, will result in a higher value of the option, at least in theory. However, when determining whether a constant volatility fits the observed data on the market, one may look at the empirical evidence from the market data. 16 First, it is natural to explore whether the log returns of a stock options follows a Geometric Brownian Motion and thereby are normally distributed, as was one of the main assumption of the Black-Scholes Model. The normal distribution (also known as the Gaussian distribution) is one of the most widely used distributions in the literature, and its probability density function is given by: π(π₯) = 1 √2ππ π − (π₯−π)2 2π2 Equation 2.3.1.1 The normal probability density distribution is characterized as being symmetric around its mean, µ, and having a skewness of zero and a kurtosis of three (where the kurtosis is a measurement of the peakness of the distribution). Throughout the years there has been conducted several empirical studies in order to test whether or not the market data can be explained by the normal probability density function, and thereby follows a normal distribution. However, empirical studies of financial time-series data have shown that the market data doesn’t exhibit log-normally distributed returns. As an extension of this the implications of the Black-Scholes Model will in the following be illustrated and tested on the observed data for the S&P500 index. From Figure 2.3.1.1 it is clear to see that the daily returns of the S&P 500 index returns doesn’t follow a normal distribution as the observed distribution display fat tails and a high central peak when compared to the normal distribution. Figure 2.3.1.1 Source: Own contribution 17 These are characteristics of mixtures of distributions with different variances, which is in conflict with the Black-Scholes assumption. Furthermore when looking at the log returns of the S&P 500 index options on a QQ-plot of the S&P500 daily log returns compared to the normal distribution, it is clear to see that the distributions differ (figure 2.3.1.2). If the S&P500 log returns Figure 2.3.1.2 did actually depict a lognormal distribution, the blue line in the figure would be straight. However, this is not the case, and it can be seen that the tails of the observed distribution differs from the normal distribution. This suggests that a stochastic variance should be considered in order to explain the development of a stock option price over a period of time. Source: Own contribution Before the so-called Black Monday in 1987, the options market did actually depict a somewhat constant volatility, which is also the assumption underlying the Black-Scholes Model. However, on the day of Black Monday the market experienced an enormous volatility and extreme losses, and therefore the market began to recognize that the existing models were not able to forecast and take into account extreme events and a true stochastic volatility7. 7 This was also known as Crashophia (Hull, 2008, page 395) 18 Figure 2.3.1.3 Figure 2.3.1.3 S&P500 daily shows log the returns plotted over a period of almost 30 years (from January 2nd 1980 to 31st December 2010), and here the extreme rise in the volatility in 1987 is clear. From the graph it is also evident that small movements are followed by small movements and large movements by large. This is evidence of volatility clustering, which implies that the volatility is also auto-correlated (Gatheral, 2006, page 1-3). The autocorrelation Source: Own contribution occurs as a result of the mean-reversion of volatility. In the Black-Scholes framework the implied volatility is assumed constant, and if this were to hold in the real world, the implied volatility of options on a given spot price would be constant across different maturities and strike prices. Empirically this has been shown to be violated in the real world (Lewis, 2005, page 1-2). When plotting the implied volatilities from the S&P500 index options across moneyness (the relationship between the strike price and the spot price) this is clearly evident. If the implied volatility were in fact constant, the figures would depict straight lines. Instead the figures show a smile or a smirk – this is the so-called volatility smile. From the empirical studies of the volatility smile, it has been found that the implied volatility is often higher for an ITM (in-the-money) stock option and OTM (out-of-the-money) option than for an ATM (atthe-money) stock option (http://www.ivolatility.com/help/14.html, visited on 14-08-2011). This is partly due to the non-normality of the returns which contributes in creating the volatility smiles. The fat tails in the true distribution of the returns increase the probability of extreme events. This could for instance be a drop in the returns, and therefore the distributor of an ITM or OTM option must be 19 compensated for the increased potential risk for movements in the underlying stock. This compensation is visible, as the higher implied volatility is, ceteris paribus, the higher the price of the stock option. The most typical volatility smile is depicted where the implied volatility is a decreasing function of the strike price – meaning that a higher strike price should yield a lower volatility. In the figures plotted below, the implied volatility is a function of the moneyness, as mentioned, and the implied volatility is in all cases a decreasing function of the moneyness and thereby of the strike price. Figure 2.3.1.4 – Plot of implied volatilities over log moneyness Source: Own contribution The volatility smile or smirk in equity options can also be explained by the leverage of the underlying company (Hull, 2008, page 395). If the equity in the company declines, the leverage will 20 increase. This makes the equity more risky, as the debt part of the company increases relative to the equity part, which increases the risk of the company defaulting (the risk taking by the stock holders increase). This rationale can be transferred to the implied volatility, as the expected risk in the market will decrease as the price of the stock option increases. This calls for a negatively correlation between the return of the stock option and the volatility of the underlying stock – which is also called the leverage effect. It has now been proven that constant volatility models cannot be used to describe the actual market for stock options, and therefore actors in the market have a demand for new models that incorporate a stochastic volatility. Chapter 3: Stochastic Volatility Models In the following paragraph the theory of stochastic volatility models will be addressed. The section will begin with a short presentation of the various volatility models that exist on the market today. Here the different models’ characteristics will be shortly presented and their real world applications will be discussed. Furthermore the general theory regarding stochastic differential equations will shortly be presented, as this theory will then directly be transferred to the chosen model. After this, the focus of the remaining section will be on the selected stochastic volatility model developed by Heston (1993). The theory behind the model will first be introduced and the general theory behind characteristic functions will be explained. This will lead to the presentation of the Heston model’s characteristic functions and the Fourier transformation that can be used to price options. 3.1 General Stochastic Volatility Models As the insufficiency of the Black-Scholes Model became clear in the market a development to address the need for a better model began which lead to the introduction of stochastic volatility models. In the process of trying to find a model that more closely adjusted itself to the complexity of the real world, several models have been developed throughout the years. Option pricing models in general have some features in common, as they all have to make assumptions concerning the underlying process, the interest rate process and the market price of 21 factor risk (Bakshi et al 1997, page 2003-2004). The choices of assumptions are endless, and vary across the different models that exist on the market, where some are more realistic than others. However, when determining what model is most suitable for the task at hand, it is important to consider the model’s complexity and implemental costs in respect to its accuracy – and whether a more complex model will add significant information to the analysis. The more complex the model is, the more time and resources are needed to make the model “run”. Simply changing a model from constant volatility to a model that incorporates stochastic volatility can approximately double the time needed to run a simulation of the option price. It is therefore an important trade-off that must be considered when the choice of option pricing model must be made. Option pricing models range from constant volatility models, stochastic volatility models and interest rate models etc. where jumps can also be incorporated. As a direct comparison to the BlackScholes Model there exist two alternative models, the jump diffusion option pricing models and the stochastic volatility models8. To provide an overview of the most influential models on the market, the following list below has been composed9. The list is organised after “year of publication” and the models can then be directly compared across the asset process and the volatility process. Table 1: Option pricing models Author(s) Black & Scholes (1973) Hull & White (1987) Stein & Stein (1991) Heston (1993) Merton (1976) Description The Black-Scholes Model Stochastic volatility model Stochastic volatility model Stochastic volatility model Jump diffusion model Process (dS) dS = µSdt + σSdZ dS = µSdt + σSdZ dS = µSdt + σSdZ Process (dσ) + correlation (ρ) None – constant variance dσ = σ(αdt +γdZ) dσ2 = aσ2dt + bσ2dZ, ρ=0 dσ = β(α-σ)dt + γdZ dσ = -κ(σ-θ)dt + bσdZ ρ = 0 dS = µSdt + σSdZ dσ2 = κ(θ-σ)dt + bσdZ, ρ≠0 dS = (α-λκ)dt + σSdZ + Sdp None – constant variance, ρ = 0 Source: Own contribution Remark: The model parameters have been changed so that the models are more easily compared to each other and to the Black-Scholes Model. For further information see the original sources. After the market crash in 1987 the market realized that there was a need for models that took into account the stochastic volatility that were present in the market – as a result of this Hull & White (1987) introduced their model for stochastic volatility. As one of the earliest option pricing models 8 Interest rate models will not be further discussed, as they are not relevant for the thesis It should be noted that the list is far from exhaustive, as interest rate models as well as stochastic volatility models with jump diffusions among others are not included. 9 22 that incorporated a stochastic volatility, this model is also one of the simplest. The model is based on similar assumptions as the Black-Scholes Model, and is also based on a Geometric Brownian Motion with a correlation of zero between the variance and asset processes. However, the Hull & White model has the disadvantage that the assumed volatility is not mean-reverting, which empirical evidence has shown to be the case in the market. Stein & Steins stochastic volatility model from 1991 incorporated exactly this, by assuming a so-called Ornstein-Uhlenbeck process (Schoebel et al, 1998, page 1). Even though this model is more precise in the some way than the earlier models, the main disadvantage of the model is that the variance can be negative (negative volatility). This is rejected as true by empirical studies and pure logic. Another model, known as the Heston model was introduced in 1993. This model assumed a correlation between the underlying asset process and the variance process, and the volatility was furthermore assumed to follow a square root process, similar to a CIR-process, see more in section 3.2. Generally it can be said that one of the key disadvantages of all the stochastic volatility models is that they are unable to predict extreme events in the evolvement of the stock option price (Gatheral, 2006, page 50-52 and Bakshi et al, 1997). To correct for this, one can incorporate a jump in the model, which was first introduced by Merton in 1976 (with his jump diffusion model). Throughout the years there have been discussions on the performance of the different models relative to each other. Bakshi et al performed in 1997 an empirical study of some of the most widely used option pricing models in order to test the models ability to price options. The chosen models included a stochastic volatility model (SV), a jump diffusion model and a stochastic volatility random-jump model (SVJ). The models were then tested under three yardsticks10, in order to judge the accuracy and the empirical performance of the given model in comparison to the Black-Scholes Model. The study concluded that taking stochastic volatility into account was the “first-order” importance when improving the Black-Scholes Model (Bakshi et al, 1997, page 2042). Additional, it was shown that adding a random jump to the stochastic volatility model did actually improve the performance of the model, and its ability to price short-term options. However, for hedging purposes it showed that the SV model performed better. For the purpose of this thesis the Heston model has been chosen to fit the volatility surface of the S&P 500 index options. It has been concluded that the accuracy of the Heston model compared to 10 1) Internal consistency of implied parameters/volatility with relevant time-series data, 2) out-of-sample pricing and 3) hedging 23 the Black-Scholes model is significant and as the purpose of the thesis is also to illustrate the Fourier transformation method for pricing options, the choice of including a jump in the Heston model has been considered to be irrelevant. 3.1.1 Deriving a general PDE for stochastic volatility models In the following a pricing equation will be constructed under the assumption of stochastic volatility. The theory behind this will later be transferred to the Heston stochastic volatility model. The approach used in the construction will be based on Gatheral (2006) who uses a methodology closely related to the one used in the derivation of the Black-Scholes Model. The general stochastic differential equation of a stock option is defined as (Gatheral, 2006, page 4): πππ‘ = ππ‘ ππ‘ ππ‘ + √π£π‘ ππ‘ ππ1 Equation 3.1.1.1 where the variance, vt, must satisfy the following: ππ£π‘ = π(ππ‘ , π£π‘ , π‘)ππ‘ + π(ππ‘ , π£π‘ , π‘)ππ2 , Equation 3.1.1.2 where π(ππ‘ , π£π‘ , π‘) = ππ½ √π£π‘ ππ‘ (ππ‘ , π£π‘ , π‘) for simplicity. The term μ in equation 3.1.1.1 is the drift of the stock price return and the term η is the volatility of the volatility (or the volatility of the variance). Many stochastic volatility models also share another common feature, and that is the mean reversion of the volatility. Mean reversion more specifically refers to the effect that the variable has to revert back to a long-run average – the mean reversion of the volatility therefore concerns the drift of the volatility itself or the drift of the underlying stock to which the volatility is related. The speed of this process is referred to as the mean reversion speed. Empirical studies show that mean reversion is present in real world prices, as shown in section 2.3.1, and therefore this feature is vital for the stochastic volatility models to display, in order to capture the true evolvement of the stock prices over time. In order to incorporate a mean reverting volatility into the above general stochastic volatility formula, the drift must be defined more specifically as: π(ππ‘ , π£π‘ , π‘) = π (π − π 2 ). This means that the drift must be depending on the so-called mean reversion speed, κ, which is determined by the average level of the stock variable here called θ. This makes the drift, α, pull the level of the stock variable towards the average, θ, on the long run. Conversely, this also 24 means that the volatility of the underlying stock will be pulled towards the average of the function over time. In equation 3.1.1.1 and 3.1.1.2 there are two Wiener processes, Z1 and Z2, that must be correlated with each other as in the real world – in other words the stock price return and the changes in the variance must be correlated which each other, as written in the following: (ππ1 ππ2 ) = πππ‘ Equation 3.1.1.3 It is here important to emphasize that equation 3.1.1.2 is a general differential equation for pricing stock options. The equation itself is very similar to the one assumed in the derivation of the BlackScholes Model, and in the limit where η ο 0 equation 3.1.1.2 actually transforms into the BlackScholes formula. This is a clear advantage to stochastic volatility models, as practitioners are usually familiar with the theory and intuition underlying the Black-Scholes Model. Hence, the intuition of the stochastic volatility model is closely related to the one of the Black-Scholes formula. It should also be noticed, that there is no assumption made concerning the terms a and b in the general stochastic volatility model – this means that the functional form of the variance in the model is unspecified, and therefore the stochastic volatility models can differ from each other here (Gatheral, 2006, page 4). The only randomness that is present in the Black-Scholes Model is the tradable assets, as the variance is assumed to be constant, as mentioned earlier. This contributes to the market being complete, because the underlying stock can be traded continuously, which hedges the option. However, this is not the case with stochastic volatility models, where a common feature (of several of the stochastic volatility models) is that they are determined by two risk factors, the market risk and the volatility risk (Boswijk, 2001, page 1). This means that the randomness/risk of the model can be traced both to the tradable assets as well as the volatility of the asset’s return (Moodley, 2005, page 7-8). This introduces a new risk measure, however because the volatility cannot be traded in the real world, and thereby hedged, the market is incomplete and this complicates the pricing process. Much like in the derivation of the Black-Scholes Model the random variables must be hedged in order to create a riskless portfolio. This is done by constructing a portfolio Π, which consists of the option that are being priced, a -β quantity of the stock and a -β1 quantity of another asset (where the value depends on the volatility). This gives the following portfolio: 25 Π = π − Δπ − Δ1 π1 The infinitesimal change of the portfolio can then be defined as: πΠ = { − Δ1 { ππ 1 2 π 2 π π 2π 1 2 π 2π + π£π + π£πππ + π } ππ‘ ππ‘ 2 ππ 2 ππ£ππ 2 ππ£ 2 ππ1 1 2 π 2 π1 π 2 π1 1 2 π 2 π1 + π£π + π£πππ + π } ππ‘ ππ‘ 2 ππ 2 ππ£ππ 2 ππ£ 2 +{ ππ ππ1 ππ ππ1 − Δ1 − Δ} ππ + { − Δ1 } ππ£ ππ ππ ππ£ ππ£ In order to make the portfolio riskless, the option must be hedged. This is done by first letting the dv-term be eliminated by setting: ππ ππ ππ1 − Δ1 = 0 → Δ1 = ππ£ ππ1 ππ£ ππ£ ππ£ and then eliminating the dS-term by setting: ππ ππ ππ1 ππ ππ1 ππ ππ£ ππ1 − Δ1 −Δ=0→Δ= − Δ1 →Δ = − ππ ππ ππ ππ ππ ππ1 ππ ππ£ What is left is now the riskless portfolio which is defined as: ππ 1 2 π 2 π π 2π 1 2 π 2π πΠ = { + π£π + π£πππ + π } ππ‘ ππ‘ 2 ππ 2 ππππ£ 2 ππ£ 2 − Δ1 { ππ1 1 2 π 2 π1 π 2 π1 1 2 π 2 π1 + π£π + π£πππ + π } ππ‘ ππ‘ 2 ππ 2 ππ£ππ 2 ππ£ 2 Following the arguments from section 2.2, the return of the riskless portfolio value must now equal the risk free interest rate. This is shown as below: πΠ = πΠππ‘ = π(π − Δπ − Δ1 π1 )ππ‘ 26 Inserting the equations in the above formula and rearranging, gives the following (see Gatheral, 2006, page 4-7 for a more detailed derivation): ππ 1 2 π 2 π ππ π 2π 1 2 π 2π ππ (π + π£π + ππ − ππ + π£πππ + π + − ππ) =0 ππ‘ 2 ππ 2 ππ ππππ£ 2 ππ£ 2 ππ£ Black-Scholes Correlation Volatility Volatility premium , which is the general solution for stochastic volatility models. The solution has been shortened by rewriting the arbitrary function f(St, vt, t) as the term (π − ππ), where the terms a and b refers to the drift and volatility from the variance process that was assumed in equation 3.1.1.2. For clarity the Black-Scholes solution has been highlighted in the equation, as well as the correlation, volatility and volatility premium part. However, as a risk neutral world is assumed, the volatility premium is non-existing and therefore equal to zero. 3.2 The Heston Stochastic Volatility Model The Heston Model was introduced in 1993 by Steven L. Heston and is today one of the most widely used stochastic volatility models on the market. The model was the one of the first models to depict an alternative (semi)closed-form solution to the Black-Scholes Model for the pricing of a European call option. In the following paragraph the process of the Heston model will be introduced and the partial differential equation for the Heston model will be outlined. Later the basic for the calculation of the option prices will be set by the introduction of the characteristic functions and the Fourier Transformation, which will then be implemented on the Heston model. 3.2.1 The Process of the Heston model The Heston model was the first known model of its kind to depict a (semi)closed solution for option pricing after the Black-Scholes Model. The model also differs from other stochastic volatility models, as the development of the underlying asset is assumed to be correlated to the volatility process. This is in compliance with the empirical studies shown in section 2.3.1. This along with the 27 easiness that is related to the implementation of the model has contributed greatly to the attractiveness of the model throughout the years. The Heston model is compiled of two partial differential equations that will be discussed in the following. The Heston model is suitable for pricing stock options, as it also builds on the generalized Wiener Process as introduced earlier in this thesis in section 2.1. In the Heston model the stock price is assumed to develop in accordance to the following diffusion: πππ‘ = ππ‘ ππ‘ ππ‘ + √π£π‘ ππ‘ ππ1 Equation 3.2.1.1 The stock price is dependent on a drift and a variance, which is also similar to the Black-Scholes Model with the exception that the volatility now depends on the time (and it no longer a constant). It is here noticeable that the instantaneous variance, vt, is based on a square root process similar to the so-called CIR-process, first developed by Cox, Ingersoll and Ross in 1985. They introduced a model that described the evolution of interest rates, and the instantaneous interest rate was said to follow a CIR-process, which was a square root process (for more information see the direct source). Following from this, the variance in the Heston model must therefore satisfy the following stochastic differential equation (Heston, 1993, page 328): ππ£π‘ = π (π − π£π‘ )ππ‘ + π√π£π‘ ππ2 Equation 3.2.1.2 The terms κ, θ and vt above describe the mean-reverting volatility of the process, as with the general stochastic volatility model mentioned in the previous section. The mean speed of reversion, π , determines the relative speed of the volatility or the weight that the long-run variance and current variance are given. The average level of the stock, π, is the long-run variance that the drift pulls the volatility towards. The vt term is the current variance, while π is the volatility of the volatility (the last will be elaborated on later). The variance will furthermore always remain positive as long as: 2ππ − π2 > 0 Equation 3.2.1.3 In accordance to the general stochastic volatility model presented in section 3.1.1, the terms Z1 and Z2 are Wiener processes that must be correlated with each other. This is shown in the following: 28 (ππ1 ππ2 ) = πππ‘ Equation 3.2.1.4 In the above equation the term ρ is the correlation coefficient between the return of the underlying stock and the changes in the variance. This relationship explains the before mentioned phenomenon of the leverage effect. This correlation has proven to be a great advantage to the Heston model as this is also present in empirical studies that have been performed over the years. The correlation, which is often negative, will ensure that the volatility for example will rise if the value of the underlying asset falls dramatically. In addition the variance is also mean-reverting, which is also evident in the market. The mean-reverting process is the term π (π − π£). In the correlation shown above in equation 3.2.1.4 the terms Z1 and W are both are Wiener processes that follow a standard normal distribution with mean zero and variance one (N(0,1)). The process of Z2 can be described as a function of the process Z1 and an independent Brownian Motion W, as written below: π2 = ππ1 + √1 − π2 π Equation 3.2.1.5 This implies that the processes Z1 and Z2 will be fully dependent on each other, whenever the correlation coefficient is either 1 or -1. However, if the correlation coefficient is instead equal to zero, the process of Z2 will instead be dependent on the process of W – making Z1 and Z2 independent of each other. One of the advantages of the Heston model is that it can be applied to a various number of distributions – while the Black-Scholes model assumes a normal distribution, where the only adjustable parameters are the mean and variance. The Heston model is able to “handle” different distributions because the correlation factor, ρ, has an effect on the heaviness of the tails of a distribution and the volatility of volatility factor, η, has an effect on the kurtosis on the distribution. This means that the parameters of the Heston model can be adjusted in order to better fit the market data in question. However, this also means that the correlation coefficient have a direct effect on the volatility smile. A positive correlation (where π > 0) will make the distribution of the returns positively skewed, resulting in a thin left tail and a fat right tail. This is due to the fact that the left tail of the probability density function is associated with low variance and therefore will not be spread, whereas the right tail will (Heston, 1993, page 336-338). This is also coherent with the 29 rationale behind the theory – a positive correlation between the underlying stock and the variance means that an increase in the asset price will, ceteris paribus, result in a higher variance which will affect the asset price in a positive way. In practice most indexes though have a negative correlation, and the smile is skewed. Therefore the opposite occurs. The Heston model furthermore differs from other stochastic volatility models in the simplicity of its implementation. The solution to the Heston model is based on characteristic functions, which enables the implementation to be faster than with other stochastic volatility models, which for example may use Monte Carlo simulation. The advantage of the characteristics functions lies in the technique that only requires a numeric solution of integrals, which makes the Heston solution a (semi)closed form solution (the semi part due to the complexity of the numeric solution). The theory underlying characteristic functions will be explained more deeply in section 3.2.3. A disadvantage of the Heston model is that it poorly prices options with a short maturity. The underlying reason for this is that the Heston stochastic volatility model is based on a Geometric Brownian Motion, and therefore the model cannot predict extreme events on the option market. Furthermore, the stochastic volatility that is to be estimated is not observable in the market. This makes the estimation process difficult, and the model is in addition very sensitive to its parameters. This makes the calibration of the model parameters vital for the model to be able to price the options accurate. This also means that the more realistic the model is required to be, the more complex the calibration of the parameters should be (Mikhailov et al, 2008). 3.2.2 PDE for the Heston Model The general stochastic volatility formula presented in section 3.1.1 is fairly easy to transfer to the Heston model. The value of any asset must according to the Heston model satisfy the following partial differential equation (PDE) where the term π(ππ‘ , π£π‘ , π‘) = π (π − π£π‘ ) (the mean-reverting process) and π(ππ‘ , π£π‘ , π‘) = π√π£π‘ . When inserting the values of a and b from the Heston model into the solution for general stochastic volatility models, see equation 3.1.1.2, the following solution the Heston model is presented: ππ 1 π2 π ππ π2 π 1 π2 π ππ + 2 π£π 2 ππ2 + π ππ π − ππ + π£πππ ππππ£ + 2 π£π2 π2 π£ + (π (π − π£) − πππ) ππ£ = 0 ππ‘ Eq. 3.2.2.1 30 However, as mentioned earlier, it is assumed that the option valuation is in the risk neutral world. The process of the underlying stock under risk neutral assumptions is defined as: πππ‘ = π β ππ‘ ππ‘ + √π£π‘ ππ‘ ππ1 Equation 3.2.2.2 where the drift has been replaced by the risk free interest rate, r (and r-q when taking dividends into account). Here it is important to notice the term λ(S, v, t), which is the price of volatility risk. As the pricing is assumed to be risk neutral, the volatility risk term λ must equal zero. This eliminates the term of ληρ in equation 3.2.2.1. 3.2.3 Characteristic Functions and the Fourier Transformation In the world of option pricing there exists numerous ways of estimating the fair value/price of a stock option. These methods include finite difference method, simulation etc. As an alternative method of pricing options, the mapping of the characteristic function of the density function has been recognized for its easy and less complex computation. Empirical studies have shown that the distribution of the stock option doesn’t follow a Gaussian distribution in the real world. However, the use of characteristic functions and Fourier transformation doesn’t require the distribution of the underlying to be known, and therefore there has been a growing interest for the use of this method. In addition the method also offers a fast computational advantage when compared to many of the other methods available. The main idea underlying this method is to take the integral of the option payoff function over the probability function. The probability function is then retrieved by inverting the Fourier transform. The characteristic function defines the probability distribution of a random variable, and every function has its own characteristic function from which the density function can be computed (Schmelzle, 2010, page 2). The general definition of a characteristic function with respect to u is given by (Schmelzle, 2010, page 8): +∞ ππ (π’) βΆ= πΈ[π ππ’π₯π |π₯π‘ = 0] = ∫−∞ π ππ’π₯π ππ (π₯)ππ₯ Equation 3.2.3.1 31 π where π₯π = πππ ππ . This is also known as a Fourier transform, which is basically used to transform 0 a complex function to another function of the same variable. The characteristic function is defined for arbitrary real numbers u, where i is an imaginary number π = √−1. The f(x) is the probability density function, and the stochastic process appears for −∞ < π’ < ∞. The relationship between the probability density functions and the characteristic function is said to be “one to one”. This means that one can derive the one from the other, and the relationship between the two is illustrated below (Schmelzle, 2010, page 7-11): 1 +∞ ππ (π₯) = β± −1 [ππ (π’)] = 2π ∫−∞ π −ππ’π₯π ππ (π’)ππ’ Equation 3.2.3.2 This relationship is also what makes the use of characteristic functions in option pricing popular, as it is possible to derive the probability density function from the characteristic function, even though the density function is not known in closed-form11 (Cherubi et al, 2010, page 32). This is possible, as the probability density function can be expressed in terms of an integral in which the characteristic function is a part of. This is done by the inversion of the Fourier transform. One of the main advantages of the Fourier transformation is that under certain conditions the inner and scalar products will be the same in the Fourier transforms. This is especially an advantage when the characteristic function is use to describe a distribution. This is expressed as the Planchard Theorem or Parseval’s Theorem as written below: ∞ ∞ ∫−∞ π(π₯)πΜ (π₯)ππ₯ = ∫−∞ πΜ(π₯)π (π₯)ππ₯ Equation 3.2.3.3 The Fourier transform has several properties however these will not be expressed in detail in the thesis. For more information see the original article by Schmelzle, 2010. 11 The only known density functions in closed-form are the normal distribution, the Cauchy distribution and the inverse Gaussian distribution (Cherubi et al, 2010, page 34) 32 3.2.4 Implementation of the Fourier transform on the Heston model The Heston model was the first model to introduce the Fourier transformation methods to option pricing (Kahl and Lord, 2010, page 1). The Fourier transformation can be applied to a given characteristic function in order to derive the probability density function of the distribution, as explained earlier. The characteristic function of the normal distribution is known as the following (Gatheral, 2006, page 57): ππ (π’) = πΈ[π ππ’π₯π ] = π −0.5π’(π’+π)π 2π = π −0.5π 2 ππ’π−0.5π 2 π’2 π Equation 3.2.4.1 which is also the characteristic function for the Black-Scholes model, as it assumes normally distributed log-return of the underlying stock. In order to predict a future price of a stock, which by definition is uncertain, the process depends on the probability distributions. The value of a typical European call option can be expressed in terms of the probability of it being exercised at-the-money. This gives the following definition (Schmelzle, 2010 page 17 and Kahl and Lord, 2010, page 1): π π π π π‘ πΆ(π0 , πΎ, π) = πΈπ‘ π [(ππ − πΎ)+ ]π −ππ = πΈπ‘ π [(ππ β 1{ππ‘ >πΎ} )] − πΎ β πΈπ‘ π [(1{ππ‘ >πΎ} )] = π(π‘,π) β π πΈπ‘ π [(1{ππ‘ >πΎ} )] − πΎ β π[(ππ > πΎ)] Equation 3.2.4.2 = πΉ(π‘, π) β ππ [(ππ > πΎ)] − πΎ β π[(ππ > πΎ)] where the latter equation is similar to the Black-Scholes formula. When pricing a European call option, the value depends on the probability function of the stock price measure being larger than the exercise price. The basic theory of option pricing says that the price of a call option at maturity is the expected positive value of its payoff discounted back to today, as shown below: πΆ(π, πΎ, π, π, π) = πΈ π [(ππ − πΎ)+ ]π −ππ Equation 3.2.4.3 Where the process of the underlying stock behaves as the following: 33 ππ = π0 π (π−π)π β π π₯π Equation 3.2.4.4 Inserting this in equation 3.2.4.3 yields the following, where the constant π0 π (π−π)π has been put outside the parenthesis: πΆ(π, πΎ, π, π, π) = πΈ [π0 π = πΈ [(π π₯π − (π−π)π (π π₯π − + πΎ π0 π (π−π)π + πΎ π0 π (π−π)π ) ] π −ππ ) ] π0 π (π−π)π π −ππ From Gatheral (2006, page 58) the price of a call from a characteristic function is given by: 1 +∞ ππ’ πΆπΊπ΄ππ»πΈπ π΄πΏ (π, πΎ, π) = π − √ππΎ π ∫0 1 π’2 + 4 π π π [π −ππ’π ππ (π’ − 2)] Equation 3.2.4.5 Where r = q = 0. However, this is a simplification to the real world, and the formula needs to be rewritten in terms of r = q ≠ 0, so that both the risk free interest rate and a dividend yield is incorporated into the formula. This is done by producing the following call price, where the spot price is assumed to be zero: +∞ πΎ 1 ππ’ π −ππ’π πΆπΊπ΄ππ»πΈπ π΄πΏ (1, , π) = 1 − β πΎ ∫ π π [π π (π’ − )] √1 π π π’2 + 1 2 π0 π (π−π)π 0 4 Then accounting for dividends in the real spot price is done by the following equation of the call price: πΆ(π0 , πΎ, π, π, π) = π0 π −ππ β πΆπΊπ΄ππ»πΈπ π΄πΏ (1, πΎ π0 π (π−π)π , π) For simplicity the term k is defined as: π = πππ ( πΎ π0 π (π−π)π ) This yields the following function for the price of the call option: 34 +∞ 1 ππ’ π πΆ(π0 , π, π, π, π) = π0 π −ππ β 1 − √πΎ ∫ π π [π −ππ’π ππ (π’ − )] π π’2 + 1 2 0 4 To find the price of the call option, the characteristic function must now be inserted into the above formula, and solved. The characteristic function of the Heston model is defined as: ππ (π’) = π πΆ(π’,π)π+π·(π’,π)π£π‘ Equation 3.2.4.6 With vt = v0 as the initial variance and the following defined as: 2 πΆ(π’, π) = π {π − π − π2 πππ ( 1−ππ −ππ 1−π −ππ )} Equation 3.2.4.7 1−π π·(π’, π) = π − ( ) 1 − ππ −ππ π½ ± √π½ 2 − 4πΌπΎ π½ ± π π±= =: 2 2πΎ π π = √π½ 2 − 4πΌπΎ πΌ=− π’2 2 − ππ’ 2 π½ = π − πππ − ππππ’ πΎ = π− π= + π + πππ’, j= 0,1 π2 2 It is here important to emphasize that the initial conditions of C and D must be equal to zero, so that the initial parameters are not evaluated at infinity – this will cause the minimization function to choose the initial values as estimates of the parameters. In order to test the functionality of the Fourier Transformation applied to the Heston model, it is first applied to the standard Black-Scholes model. This enables the Fourier Transformation method to be verified, as the Black-Scholes solution to the Fourier Transformation can be verified by the direct solution of the Black-Scholes pricing equation. The characteristic functions differ with the process at hand, and the characteristic function for the Black-Scholes model is defined as: 1 ππ (π’) = πΈ[π ππ’π₯π ] = ππ₯π {− 2 π’(π’ + π)π 2 β π} Equation 3.2.4.8 35 Chapter 4: Data The data used in the analysis of the Heston model consists of index option data that has been extracted from Wharton Research Data Service (WRDS). The option market data more specifically consists of the data from the S&P 500 index (Standards and Poors), which is a gathering of 500 large-cap common stocks in a value weighted index. The index is traded on the NASDAQ and the New York Stock Exchange (NYSE), and is by some considered to be an ideal proxy for the total market in the US as the index covers about 75% of the U.S. Equities12. All data used in this thesis of the S&P 500 index has been downloaded from the website http://wrds- web.wharton.upenn.edu/wrds/process/wrds.cfm after opening a free account. The data more specific consists of daily closing prices of the S&P 500 index with matching dividend yield and highest closing bid- and lowest closing ask-prices. The selected time period to analyse the Heston model across is from January to March 2010. However, the actual plotting of the volatility surface will not be done on the entire period, but the attached code is easily modified for this. The chosen data set consists of one monthly observation day, which has been chosen to be the 3rd Thursday of each month. This underlying reason for this is that the S&P500 index options expires at the third Friday of the month (http://www.randomwalktrading.com/main/index.php?option=com_content&view=article&id=216 &Itemid=338 visited at 08/31/2011). As an input in the analysis, it is also required to find a measurement for the risk free interest rate, as it is to be used as the discounting rate as well as the drift in the process of the underlying stock (because of the assumption made of the risk neutral world). In this thesis the risk free rate used in the analysis is assumed to be constant – which means that the assumed interest rate structure is assumed to be flat. This is a simplification of reality, as empirical studies have shown that the risk free term structure is far from flat – in reality the interest rate yield curve is usually upward sloping. Often treasury bills and treasury bonds have been used as the risk free rate. The reason for this is that it is assumed that the chance of a government defaulting on an obligation, which is 12 Source: http://www.standardandpoors.com/servlet/BlobServer?blobheadername3=MDTType&blobcol=urldata&blobtable=MungoBlobs&blobheadervalue2=inline%3B+filename%3DFactsheet_SP_500.pdf& blobheadername2=ContentDisposition&blobheadervalue1=application%2Fpdf&blobkey=id&blobheadername1=contenttype&blobwhere=1243931099288&blobheadervalue3=UTF-8 visited on 08/15/2011 36 denominated in its own currency it relatively low (close to none-existing) (Hull, 2008, page 74). Therefore, this is considered to be near the closest an investor can come to making a risk free investment. However, the treasury rates are considered to be artificially low by practitioners as a result of tax and regulatory issues. Instead, derivatives traders often use LIBOR rates as proxy for the risk free interest rate. The LIBOR rate13 is the interest rate at which large banks in London are willing to make a large wholesale deposit with other banks (Hull, 2008, page 74) – this is essentially the same as borrowing money to the bank in question. As a result of this, it is required that the financial institution satisfies certain creditworthiness, and it is required typically that it at least have an AA-rating. As such, one can say that the LIBOR rate is not totally risk free, as there will always be a chance, however very little, for the bank to default on the loan. In this thesis it has be chosen to find an estimate of the risk free interest rate using the BlackScholes Model. As mentioned earlier, the only unknown parameter in the Black-Scholes formula is the volatility, which cannot be directly observed in the real market. However, when inserting the implied volatility into the Black-Scholes Model, the market price of the stock option can be found. As a result of this, the market price can be defined as (Wilmott, 2005, page 184): πΆ(π0 , πΎ, π) = πΆπ΅π (π0 , πΎ, πππππ (π0 , πΎ, π), π) Equation 4.1 As the implied volatility is a known parameter through the data collection in this thesis, the risk free interest rate can be found by “backing out” from the Black-Scholes formula. The dividend yield can also be found through the same way as the risk free interest rate. However, for simplicity it is assumed that the dividend yield is constant, and therefore it will serve as an input in the calibration of the risk free interest rate. The dividend yield extracted from WRDS is the daily dividend yield, and the dividend yield used in the analysis is calculated as an average dividend yield of all S&P 500 stock options. The risk free interest rate is then found by minimizing the difference of the BlackScholes Model prices with the known market prices, CM, with respect to r, as seen below: min = min πΆπ΅π (π0 , πΎ, π, πππππ , π, π) − πΆπ (πΎ, π) π 13 π Equation 4.2 The LIBOR rate is only quoted for maturities up to 12 months. 37 As an estimate of the market prices in this minimization process, the average bid-ask price is used, which have been used in many empirical studies throughout the years. The average bid-ask quote has been calculated from the following: π΄π£πππππ ππππππ = ππππ +ππ ππ 2 Equation 4.3 The initial value14 of the risk free interest rate is set equal to an annualized rate of an American Treasury bill15. This is done in the hope that the estimation of the risk free rate will be closer to the true risk free rate than otherwise. However, as it turns out from working with it, changing the initial estimate of the risk free interest rate does not change the estimation process and the final estimated value of the interest rate – this means that the minimization process performed is stable. This approach gives a proxy for the risk free interest rate, however, it is important to emphasise that the estimate is based on a model (the Black-Scholes Model) which fails to describe the market data, as some of its assumptions are too constrained. Furthermore, the use of an average dividend yield can also have an effect on the estimated interest rate, however smaller than the previous one. In spite of the acknowledgement of the disadvantages of the method applied to find an estimate of the risk free interest rate, it is assumed that the interest rate backed out from market prices and the Black-Scholes formula it a “good enough” proxy for the real risk free interest rate. This should of course also be seen in relation to the fact that interest rate processes is not the main focus in the thesis. It is therefore assumed that the calibrated risk free interest rate is reliable and can be used in the further analysis of implied volatility surface. Chapter 5: Fitting the implied volatility surface 5.1 Model Calibration In order to fit the Heston model to the implied volatility surface of the S&P500 index options, it is first necessary to calibrate the parameters of the model. There exist several methods for this, 14 Initial risk free interest rate is set to 0.33 (the 1-year Daily Treasury rate from 01/15/2010) http://www.treasury.gov/resource-center/data-chart-center/interestrates/Pages/TextView.aspx?data=yieldYear&year=2010 visited on 08/29/2011. 15 38 however empirical studies performed by Bakshi, Cao and Chen (1997) have shown that simply fitting the implied parameters of the Heston model to the ones observed in the market, is not sufficient enough in order to make the model fit the market. Another approach that empirically has been used to estimate the parameters of the Heston model is to use the so-called inverse problem, which is solved by finding the parameters of the model that produce the right market price. The purpose of the calibration of the parameters is simply to make the Heston model fit as closely as possible to the market data, and thereby reducing the error margin between the estimated model price from the Heston model and the market price observed. The following parameters in the Heston model need to be calibrated or estimated: πππππππ‘πππ (π) = π , π, π£, π, π The calibration itself is performed by minimizing the following non-linear least-squares optimization problem, as written below (Mikhailov et al, 2008, page 76): p M 2 min SSE(p) = min ∑N i=1[Ci (K i , Ti ) − Ci (K i , Ti )] , where i = 1,2 … N p Equation 5.1.1 p In equation 5.1.1, the terms πΆππ and πΆππ are the model price and the price observed in the market respectively. The term p refers to the set of calibrated parameters values, and N is the number of options used for the actual calibration. In order to ensure that the process is positive, the following equation must hold that: 2ππ£Μ > π, where π = π Furthermore, the following conditions should be set for the initial parameters: The speed of mean reversion, π , should be non-negative 0<π The long-run volatility, π, should be non-negative 0<π The initial variance should be non-negative 0 < π£π‘ The volatility on volatility should be non-negative 0< π The correlation coefficient should be in the interval −1 < π < +1 39 This should minimize the time needed to calibrate the model, as the intervals (that needs to be searched) are smaller with theses constraints on the initial parameters than without. 5.2 Setting the initial parameters In order to calibrate the Heston parameters to the empirical data from the S&P 500 index options, it is necessary to determine a set of initial parameters. The initial values can be chosen in several of ways, and differ in the literature. This thesis will only use one of these methods and this is chosen to be an historical estimate of the initial values of the Heston model parameters. As presented earlier in the thesis, the variance process of the Heston model is defined as the following: ππ£π‘ = π (π − π£π‘ )ππ‘ + π√π£π‘ πππ‘ The historical data chosen to form the basis of the estimate of the initial parameter values is daily data in the period from January 2010 to March 2010. The variance process is then constructed from historical prices in a month, so that the time step of the variance process is monthly. This is done in accordance to Hull (2008, page 282-283) The variance process is discretized in the following: π£π‘+1 − π£π‘ = π (π − π£π‘ )Δπ‘ + π √π£π‘ √Δπ‘π(0,1), π€βπππ π(0,1) ππ π‘βπ πππππ π‘πππ ππ‘+1 Similar to the method used in short interest rate models (Bond & Interest Rate, Elisa Nicolato, handout page 19), the variance process can be found by performing a regression on the change in the variance process, as written below: π£π‘+1 − π£π‘ = π πΔπ‘ − πΔπ‘√π£π‘ + ππ‘+1 40 Here the estimate of the first term in the regression (π (π − π£π‘ )Δπ‘) becomes πΜ = π πΔπ‘ and the second term is estimated as πΜ = −π Δπ‘16. By finding the coefficients corresponding to these terms it is possible to isolate π Μ and πΜ. This gives the initial values of the two terms for later use in the calibration. Following the method of the short interest rate estimation the residuals of the above regression are saved, and the squared residuals (the error term in previous equation) is to be regressed against the variance, vt. The squared residuals can be defined as: 2 ππ‘+1 = (π√ππ‘ πππ‘ )2 Running a regression will yield the following result, where the Brownian motion will be captured in the error term ut+1. Here it is important to notice that the term d is not discretizing the variance, but is instead simply the coefficient in front of vt. 2 ππ‘+1 = π + π β π£π‘ + π’π‘+1 2 ππ‘+1 = π2 Δπ‘ β π£π‘ π~π2 Δπ‘ The term c is not relevant for the initial parameters, and is therefore ignored. Performing the regression yields an estimate for η2, and by taking the square root of this estimate, the initial parameter value for η is found. It is here important to emphasize that this has yielded a negative π2 , which is a violation of the basic conditions (taking the square root of a negative number is simply not possible). However, as the parameter is only meant as an initial estimate of the real parameter, it is assumed that the estimated value is in absolute terms – making it positive. The correlation parameter, ρ, is the correlation between the price process and the variance process. This coefficient is found by using basic statistical calculations. The basic statistical formula for the correlation coefficient between two variables, X and Y, is defined as: 16 Linear regression as y = a + bx 41 π= πππ£(π,π) √ππ΄π (π)β√ππ΄π (π) Equation 5.1.2 It is therefore necessary to find the covariance between the two time series, the spot price and the variance. From equation 2.2.4 in section 2.2 the risk neutral process of the underlying stock was defined. The logarithm of the spot price, xt = ln St, can be found by applying Itô’s Lemma on the underlying process, which yields the following: 1 ππ₯π‘ = ((π − π) − π£π‘ ) ππ‘ + √π£π‘ ππ1 2 Here the term Z1 is a Brownian motion under risk neutral assumptions. In order to find the change in the logarithm of the spot prices, this is transferred to the change in spot prices. The change in the log spot price process and the variance process are written below: 1 ππππ‘+1 − ππππ‘ = ((π − π) − π£π‘ ) Δπ‘ + √1 − π2 ππ΅1 + πππ΅2 2 π£π‘+1 − π£π‘ = π (π − π£π‘ )Δπ‘ + π √π£π‘ ππ΅2 For the process of the underlying stock the change in spot prices should be calculated by using one observation each month in order to match the frequency of the variance process (which is also one variance a month). However, in order to make the value of the spot price used in the calculation more valid, an average of the spot prices for a whole month is calculated, to match the variance of that same month. Then, finding the covariance between the two time series yields the following: πππ£(π, π) = πππ£ (√1 − π2 ππ΅1 + πππ΅2 , √π£π‘ ππ΅2 ) = πΈ[ππ √π£π‘ Δπ‘ β π 2 (0,1)] πππ£(π, π) = ππ √π£π‘ Δπ‘ After finding the covariance between the two time series the following equation can be used to find an initial estimate of the π parameter needed in the calibration of the Heston model parameters. Rearranging the previous equation, yields the following equation for π: 42 π= πππ£(π, π) √ππ΄π (π) β √ππ΄π (π) = ππ √π£π‘ Δπ‘ √Δπ‘ β π√π£π‘ √Δπ‘ The only parameter missing for the initial values is now the initial value for the variance. This is set equal to the long-run variance, π, in the Heston model, and will afterwards change in accordance to its process. The initial parameters are set to: Parameter πΏ π½ ππ πΌ π Initial value 0.7958723 0.03581966 0.03581966 0.1168394 -0.6678461 The results of the calibration are presented in the following table. It is here evident that the calibrated parameters differ with the market data for which is has been fitted. Overall it can be said that the calibrated parameters are very similar and nothing sticks out. Parameter 01/15/2010 02/19/2010 03/19/2010 5.3 πΏ 0.82910108 0.82776839 0.82877590 π½ -0.71738058 -0.72926088 -0.80367335 ππ 0.49492123 0.50368195 0.41344328 πΌ 0.07578355 0.07855661 0.07488894 π 0.03954365 0.04094678 0.03216978 Simulation of the Heston process As a performance measurement of the Fourier Transformation it has been chosen to illustrate the Heston model by the use of Monte Carlo simulation. As simulation is widely recognized for the use in option pricing, the accuracy of the Fourier Transformation pricing will be compared to the simulated prices. In the following, the theory behind Monte Carlo simulation will therefore be introduced. 43 5.3.1 Monte Carlo simulation There exist many methods for option pricing, including the finite differences method and Monte Carlo simulation. Each method varies in form and their implementation. The basic idea underlying Monte Carlo simulation is to value a derivative by generating random numbers of some uncertainty and probability density. The procedure was first introduced by P.P. Boyle in 1977, and its relatively easy implementation and simple structure have since its introduction been a contributing factor to its popularity (Glasserman, 2004). The Monte Carlo approach is based on the assumption that the value of the derivative is equal to the expected value of the derivative in the future discounted back to time zero, like the following equation states: π(π, 0) = π −ππ πΈ[π(ππ , π)] Equation 5.3.1.1 The simulation can be divided into 4 steps which are listed in the following: 1. Generate and simulate N number of N(0,1) outcomes under the risk neutral assumption 2. Approximate and calculate N number of terminal values 3. Calculate N number of the final payoff and then calculate the arithmetic average of the payoff at maturity 4. The price of the option today is then found by discounted this value with the risk free interest rate One of the advantages of simulation as an option pricing tool is its accuracy and easy implementation. This makes it a fine tool for controlling the performance of other models and methods. However, the accuracy of the simulation process depends very much on the number of simulations, and the convergence rate which can be described as: π √π where N is the number of simulations. This also means that in order to minimize the simulation error by half, the number of simulations should be four times higher (Empirical Finance notes, 2010, chapter 4, slide 21). One should also keep in mind that the higher the number of simulations, 44 the longer the time before the process has been simulated. The chosen number of simulations in the analysis is 10,000. Unfortunately, the Monte Carlo simulation sometimes contains a bias that can make the simulation less accurate. However, this can be minimized as shown in the following section. 5.3.2 Milstein Scheme In order to perform the Monte Carlo simulation the stochastic processes need to be converted from continuous time to discrete time. However, when performing a Monte Carlo simulation a discretization error appears from the estimation of the time-discretization of the stochastic differential equations (Glasserman, P, 2004, page 339-436) - the error arises since the majority of the models implemented can only be applied approximately. However, as the discretization time interval is shortened, the discretization error is reduced. The discretization itself can be done in various ways, and the most popular of them is the so-called Euler Scheme. However, when generating the variance process according to Euler, negative variances may be produced if the random numbers generated from normal standard distribution are large and negative. This is inconsistent with real market data, as the variance cannot go negative. As an alternative method the so-called Milstein Scheme can be applied. The Milstein Scheme takes method into account a higher order of the Taylor expansion than the Euler Scheme, and therefore the method performs better compared to the Euler discretization. More specifically the Milstein Scheme increases the accuracy by adding a second-order Taylor expansion through the means of Itô’s Lemma. When applied to the geometric Brownian motion in the Black-Scholes model, the following equation is the result of the conversion from continuous time to discrete time, βt17: 1 ′ π ππ πππ (π 2 − 1)Δπ‘, π€βπππ π = π(0,1), 2 1 = ππ + (π − π)ππ βπ‘ + πππ √βπ‘π + π 2 ππ (π 2 − 1)Δπ‘, 2 ππ+1 = ππ + π β ππ βπ‘ + π β ππ √βπ‘π + ⇒ ππ+1 In the above equation the a and b terms are the drift and volatility of the process corresponding to the risk free interest rate minus the dividend yield and the variance σ in the Black-Scholes Model. The conversion from continuous time to discrete time is generally illustrated by the stochastic 17 b'(x) is the first derivative of b(x). 45 process of dS being transformed into a stepwise time measurement of βt. The change in the stock return in the time period, βt, is then determined by the above equation. When applying the Milstein Scheme to the variance process in the Heston model, the following equation arises: π£π+1 = π£π + π (π − π£π )Δπ‘ + π√π£π √Δπ‘π + π2 Δπ‘(π 2 − 1) , 4 π€βπππ π~π(0,1) Which can be rewritten as the following (Gatheral, 2006, page 22): 2 π π2 π£π+1 = (√π£π + √Δπ‘π) + π (π − π£π )Δπ‘ − Δπ‘ 2 4 It should here be noted, that only the variance process should be discretized by the Milstein Scheme, as it has been chosen to perform the mixing solution on the processes, as explained in the following. 5.3.3 Mixing solution The solution to estimating the option price can also be found through the so-called mixing solution. The mixing solution was first introduced by Hull & White (1987), who presented a solution with no correlation between the stock price changes and the volatility changes. However, Romano and Touzi later understood the importance of this correlation coefficient and thereby extended the model in 1997 to incorporate a correlation between the stock price changes and volatility changes. Lewis (2000) then further extended the theorem to handle payoff functions and generalized stock price volatility coefficients (Lewis, 2002) (with no correlation), whereas the theorem presented by Romano and Touzi applied to call and put options. The basic idea underlying the mixing solution is to give an alternative theoretical and computational method to solve option models (this only applies to certain advanced models, including some stochastic volatility models). The mixing solution calculates the option values by a weighted sum of constant volatility prices or a mixture of explicit Black-Scholes prices with a calculated effective spot and effective volatility (where the latter approach relates to call and put options). The 46 advantage of the mixing solution is therefore that the solution to the model is just to take the expectation over the volatility process. The mixing approach can be divided into three basic steps as illustrated below (Lewis, 2005, page 97): 1. Separate the stock price evolution into two processes, independent of each other. One of the processes should be independent of the volatility process 2. Integrate the variance and mean 3. Insert the integrated variables from step 2 into the Black-Scholes formula (mixing formula) The solution for the option price process in a risk neutral world can be defined as the following: ππ = π0 π 1 π π (π−π)π− ∫0 π£π‘ ππ‘ +∫0 √π£π‘ ππ΅π‘ 2 Equation 5.3.3.1 which is similar to the Black-Scholes explicit solution in continuous time. As the process is written in continuous time, the mixing solution can be applied to it. The Brownian motion can again be expressed as π΅π‘ = πππ‘ + √1 + π2 ππ‘ , π€βπππ πππ‘ πππ‘ = 0. If the stock price process is divided into two independent processes, separating the Brownian motions, the following equation appears: π 1 π πππ (π−π)π− ∫0 (1−π2 )π£π‘ ππ‘+∫0 √(1−π2 )π£π‘ πππ‘ 2 ππ = ππ π Where the effective spot in the T is defined as: πππ ππ π 1 π 2 π π£π‘ ππ‘+∫0 π√π£π‘ πππ‘ = π0 π −2 ∫0 It is here noticeable that the effective spot is affected by the values of the variance and the Brownian motion Wt, and at the same time is independent of the movements in Zt. The effective variance can be defined as: πππ π£π = 1 π ∫ (1 − π)2 π£π‘ ππ‘ π 0 47 For simplicity the integrals of the mean and the variance are defined as the following: π πΌπ1 = ∫ π£π‘ ππ‘, π πΌπ2 = ∫ √π£π‘ πππ‘ 0 0 In order to simulate the process of the effective spot and variance, the differential equations must be discretized finding the dynamics of the integrals. This can be done by discretizing the variance process by the Milstein scheme, as discussed in section 5.3.2. The distribution of the log spot price, xt, can be described as a mean-variance mixture of the normal distribution as shown below: πππ π₯π‘ ~(π − π)π + πππππ πππ + √ππ£π π(0,1) which is similar to the Black-Scholes Model that assumes a normal distribution. As a result of this, it is possible to express the price of a call option as the expectance of the Black-Scholes price by inserting the effective spot and variance in the equation: πππ πππ πΆ(π0 , π£π , π) = πΈ[πΆπ΅π (ππ , π£π , π)] Equation 5.3.3.2 The simulation process for the Heston model should then be performed as the Black-Scholes simulation with the exception of the simulation of the effective spot and variance. 5.4 Comparison of simulated and Fourier prices In the following paragraph the direct comparison of the simulated call prices and the Fourier prices from the S&P 500 index options will be conducted. The simulation approach is known for its accuracy for estimated option prices, as the simulated prices converges to the true market prices as the number of simulations is increased. Therefore the simulation approach is suitable for evaluating the performance of the Fourier Transformation method for option pricing. By comparing the two methods, it is possible to give a rough estimate of the performance of the Fourier transformation. To test the precision of the Fourier method in pricing options with the Heston Model, it is first necessary to ensure that the valuation is performed correctly. For simplicity this is done with the 48 Black-Scholes Model for which there exist a closed-form solution for. This means that it is possible to check the result from the Fourier estimated call prices with the closed-form solution from the simulated process. However, as the Black-Scholes Model is not the focus of the thesis, the reader is referred to the appendix, where the code for simulation and Fourier Transformation prices exists. The simulation is now performed on the Heston model using the mixing solution, as explained earlier. When comparing the prices from the simulation with the Fourier transform, it is evident that the two methods are similar for short maturities. Figure 4 – Plot of prices across log moneyness Source: Own contribution Remark: Green lines are prices calculating via the Fourier Transformation, and red lines are simulated through the mixing solution 49 Even though the two methods appear similar, it is easy to see that they differ in some places. This is especially the case for longer maturities where the call prices differ. This is evidence that the Fourier method is not able to completely replace the simulation approach, however, this should be viewed in the light of the parameters chosen for the Heston model. As mentioned in the calibration section in 5.1, the calibration of the parameters is very sensitive with respect to the initial parameters chosen. Depending on the initial values the minimization process will find the closest minimum to the objective function given the parameters, however, there is no guarantee that the minimum is a local minimum, and not a global minimum. This is coherent with the importance of finding the true parameters of the Heston model that best fit the market data. Nonetheless, is has been assumed that the found minimum is the global. Later it will be illustrated how changing the parameters in the Heston model can make a big difference on the prices estimated from the model. Furthermore the fact that the options with a shorter maturity are priced more closely to the simulated prices can also be a product of the data used for the calibration. As the dataset contains more option that are in the money, these options will receive a higher weight in the calibration than the remaining options – which will mean that these options will be priced more accurate. 5.5 The implied volatility surface of the S&P500 index options In the following section the implied volatility surface of the S&P 500 index options (the market data) will be compared to the estimated implied volatility surface of the Heston model from the simulation and the Fourier Transformation. As one of the main implications of the Black-Scholes model is the constant volatility assumption, it makes no sense to plot the implied volatilities given in the Black-Scholes model, as it will only give a flat line. However, the two valuation methods can still be compared with respect to the implied volatility found through the methods for the Heston model. Here it is also possible to compare the outcome directly to the market fit, as the implied volatilities is a part of the downloaded dataset. As can be seen in figure 5.5.1, the Fourier method applied to the Heston model gives a poor fit compared to the actual implied volatility for short maturities. This is also what was expected. As the time to maturity increases the implied volatility from the Fourier Transformation fits more closely to the one observed in the market – and actually has a better fit than the simulated implied volatility (it should here be noted that the simulation can be improved by increasing the number of simulations). Therefore it seems that for maturities other than short maturities, the Fourier Transformation method is reliable. 50 Figure 5.5.1 – Plot over the implied volatilities across log moneyness Source: Own contribution Remark: The blue line is the imported implied volatility from the market data, the green line is the Fourier calculated implied volatility and the red line is the simulated implied volatility Gathering the implied volatilities across moneyness and time to maturity gives an image of the implied volatility surface – in a 3D plot. Comparing the Heston model with the plotted implied volatilities observed in the market gives an overview of how the Heston model fits to the market data. The implied volatility from the market has been backed out of the Black-Scholes model, as it is known to be the wrong number in the wrong formula that gives the right price. As can be seen from figure 5.5.2, the Heston model clearly does a better fit that the Black-Scholes model (which is obvious as it assumes constant volatility, as mentioned). As the plots are constructed as a function of the log moneyness, the area in the middle refers to at-the-money options (ATM), the area to the 51 right to out-of-the-money (OTM) options and last the area on the left which represents in-themoney (ITM) options. Figure 5.5.2 Source: Own contribution Remark: For maturities: 7, 36, 64, 75, 92, 155, 166, 246, 258, 337, 350 days to maturity Figure 5.5.3 Source: Own contribution Remark: For maturities: 7, 36, 64, 75, 92, 155, 166, 246, 258, 337, 350 days to maturity 52 It appears as if the implied volatilities from options with a longer time to maturity have a wider spread than the ones closer to the exercise date. This is more clearly seen on the volatility surface on the market implied volatility in figure 5.5.3. Furthermore it can be seen that the fit is worse as the time to maturity is shortened – here the market implied volatilities are nicely organized, where the implied volatilities from the Heston model are not. This is coherent with the empirical evidence of the Heston model, as it does not incorporate jumps. The fit of the Heston model is relatively close to the market implied volatility surface, and therefore it can be concluded that the model can be fitted nicely. It should here be noted that the deviations can be the result of the calibration process finding a local minimum instead of a global minimum, which means that there can exists parameters for the Heston model that fits the implied volatility surface even better. 5.6 The Heston model’ sensitivity to its parameters The Heston Model is very sensitive with respect to its parameters, and therefore an examination of the effects of the parameters will be conducted. This will illustrate the effect the parameters have on the spot price process. In order to ensure a high level of accuracy for the option prices calculated from the Heston model, it is vital that the calibrated parameters fit the market data as perfectly as possible. This was also evident from the calibration of the Heston model where the calibrated parameters differed with the datasets – which means that they were fitted to a particular dataset and therefore they are not constant. As a result of the sensitivity of the parameters it is relevant to look at each calibrated parameter and illustrate the effect a change will have on the implied volatilities – thereby showing the sensitivity of that parameter18. The default parameters are set to the first calibrated parameters (01/15/2010): Parameter Default value πΏ π½ ππ πΌ π 0.82910108 0.07578355- 0.03954365 0.49492123 -0.71738058 The mean reversion of speed, π : 18 The following section is based on Heston, 1993, page 335-339 and lecture notes from Advanced Financial Modeling 53 The mean reversion speed factor determines the speed at which the volatility is pushed towards its long-run volatility, θ. The mean reversion speed is also a determinant for the volatility clustering in the returns of the spot. When plotting the implied volatilities and moneyness for the Heston model, with the only changed parameter being π , it is clear to see that π determines the spread of the volatility, and that the difference in implied volatilities is greater for option with longer maturities. This is because the coefficient, ceteris paribus, will have a greater impact the longer the time horizon (time to maturity). Figure 5.6.1 – Plot of implied volatilities across log moneyness with kappa changing +/-0.3 Source: Own contribution The long-run volatility, π: As mentioned earlier the long-run volatility is the mean of the variance process, for which the variance process moves as time go by. The long-run volatility is an increasing function of the price of the option, so as the volatility increases, the price and the implied volatility of the option also increases. Changing the long-run volatility has a great impact on the implied volatilities, and for 54 longer time for maturity the effect is greatest. However, as it is evident from figure 5.6.2 the implied volatilities changes level, but the process remains stable, as the mean reversion speed is held constant. Figure 5.6.2 – Plot of implied volatilities across log moneyness with theta changing +/-0.1 Source: Own contribution The initial volatility, vt: The initial volatility determines the current level of the volatility, and is by the mean-reversion speed pushed towards the long-run volatility. Changing the parameter by +/-0.01 has a visible effect on all the plots in figure 5.6.3. However, for the options with the shorter maturity the effect of altering the parameter has the greatest effect. This is because the long-run volatility will have the biggest impact on the current volatility over a period of time, ceteris paribus. 55 Figure 5.6.3 – Plot of implied volatilities across log moneyness with vt changing +/-0.1 Source: Own contribution Basically there are two parameters that directly have an effect on the volatility smile and thereby the volatility surface. These parameters are the volatility of volatility (volatility of the variance), η and the correlation coefficient, ρ. The volatility of volatility, π: As the volatility of volatility parameter has an effect on the kurtosis of the distribution of spot prices, it can be used to better fit the distribution that the market data follows. This is an advantage of the Heston model, compared to for example the Black-Scholes Model, whose kurtosis is equal to zero. Whenever the volatility of volatility is equal to zero, the volatility is deterministic as the variance doesn’t changes – this makes the distribution normal. When altering the volatility of the variance by +/-0.1 the OTM options and ITM options are affected. As the correlation coefficient 56 between the stock price process and the variance process is negative, a positive increase in the volatility of volatility will decrease the implied volatility for ITM options whereas the implied volatilities for the OTM options will increase. The effect on altering the parameter doesn’t seem to affect ATM options to the same extent. Figure 5.6.4 – Plot of implied volatilities across log moneyness with eta changing +/-0.1 Source: Own contribution The correlation coefficient, π: The correlation between the stock process and the variance process affects the skewness of the volatility smile and of the spot returns. As explained earlier, the introduction of the correlation coefficient in the option pricing model has contributed a great deal to the popularity of the Heston model, as it is similar to the real world. The Heston model is very sensitive with respect to the correlation parameter, as it directly has an effect on the skewness of the distribution. 57 For stock options the correlation between the price and variance processes are normally negative, as explained earlier. This means that the volatility is a decreasing function of the spot price, and as the spot price falls, the volatility increases. Figure 5.6.5 – Plot of implied volatilities across log moneyness with rho changing +/-0.1 Source: Own contribution Following the analogy from earlier, this will result in the left tail being “spread out”, as it is associated with high volatility. At the same time the right side of the distribution, representing the low volatility, will remain the same, and will therefore not be spread. As it can be seen from figure 5.6.5, the price of an OTM option will increase as the correlation is set up by +0.1, and the opposite for OTM options. 58 All in all it can be concluded that the Heston model is very sensitive with respect to its parameters. As a result of this it is vital for the accuracy of prices it produces, that the parameters are calibrated to the market data. Chapter 6: Conclusion The main objective of the thesis was to fit the Heston Model to the implied volatility surface of the S&P 500 index options. To perform this analysis, a part of the basic theory underlying option pricing was first presented. This included stochastic processes and the Geometric Brownian Motion which is often used to describe stock option data. After this, the Black-Scholes Model was introduced in order to establish the need for a stochastic volatility model on the market for pricing stock options. Several of the assumptions underlying the Black-Scholes Model are violated in the real world, however the violations differ as some of the assumptions can be relaxed, or the model can be adjusted (as with dividends and American options). In order to test the empirical implications of the Black-Scholes Model, an analysis of the empirical stock returns of the S&P 500 index options was then conducted. The analysis was intended to test whether the log returns of the stock options were normally distributed as well as if the data depicted a constant variance or not. The analysis showed that the empirical stock returns displayed a high central peak and fat tails, which are evidence of a non-normal distribution. The log returns of the stock options further showed a deviation from the normal distribution in the tails. This is evidence of a stochastic volatility in the market data. Furthermore, when plotting the stock returns from 1980 to 2010 it showed a non-constant variance, as the returns had been subject to extreme events throughout the years – beginning with the so-called Black Monday in 1987. As a result of the analysis of the empirical data, the basis for introducing a stochastic volatility was made. Some of the most influential stochastic volatility models and the basic jump-diffusion model were then shortly introduced in order to provide an overview on the existing models that exists on the market. The chosen model, the Heston Model, was then introduced and its characteristics were outlined. One of the main reasons for the model’ popularity is that the model incorporates a correlation between the underlying asset process and the variance process, which is evident in the market. The Heston Model is also characterised as being mean-reverting. Furthermore, the model has a (semi)-closed form solution by the means of a numeric solution to an integral (Fourier 59 Transformation), which makes it less computational complex compared to other models. This has increased the model’ popularity as the time needed to find the option prices is far less than with the common Monte Carlo simulation. However the Heston model is still considered to have a semiclosed form solution as the solution of the integrals in the Fourier Transformation can be relative difficult to incorporate. Therefore, the secondary purpose of the thesis is to evaluate the performance of the Fourier Transformation for option pricing. In order to estimate the performance of the method, the Heston Model was also simulated by a Monte Carlo simulation, which has been known to converge to the true market price as a function of the number of simulations chosen. Although the Heston model contains many advantages compared to other models, it still has the disadvantage that it poorly fits the implied volatility surface of options with shorter maturities. In the analysis the Heston Model was calibrated in order to fit the implied volatility surface of the S&P 500 index options. This calibration process, which minimizes the square difference between market and model price, is very sensitive with respect to the initial estimate of the parameters, as there is no guarantee that the used method will find a global minimum and not a local minimum when solving the calibration process. However, it is assumed that the calibrated model parameters are global minimums. The calibration was conducted for three months, and the estimates were shortly compared to each other, which showed that they were very similar. In order to evaluate the performance of the Heston model and its use of Fourier Transformation for option pricing, the mixing solution was applied to the Heston model. By simulated an effective variance and calculating an effective spot price, the solution of the Heston model is then simply to insert the two values into the standard Black-Scholes Model which will then yield the price of the option. The simulated prices were then compared to the prices obtained from the Fourier Transformation, and it showed that the approaches were almost identical for short maturities. However, when comparing the implied volatilities of the two approaches with the implied volatility backed out from the market prices it turned out that the Fourier Transformation was actually good at describing and pricing options with longer maturities. The implied volatility from the Fourier Transformation was actually closer to the market implied volatility than the simulated was. Therefore it can be said that for option with longer maturities the Fourier Transformation approach is relevant, as it is relatively easy to implement and faster compared to the Monte Carlo simulation. The Heston model was then compared the implied volatility surface of the market. The analysis here showed that the Heston model deviated from the market data for shorter maturities, but for longer maturities the fit was good. However, it should also be mentioned that the calibration of the Heston parameters is not 60 guaranteed to produce a global minimum. As a result of this it can be that there are parameters for the Heston model that more closely can be fitted to the implied volatility surface. As the Heston parameters are very sensitive for the pricing process, an analysis of the parameters changing were conducted, all else being held equal. This showed that the Heston model is indeed sensitive with respect to its parameters, and therefore it is an important aspect to consider when determining the use of valuation approach. One needs to leverage the computational fast advantages and the accuracy of the Heston model and its solution in the form of Fourier Transformation with the time difficulties and continuously calibration of the parameters (that are not stable). 61 List of Literature Academic Articles: Bakshi, Gurdip; Cao, Charles; Chen, Zhiwu (1997), Empirical Performance of Alternative Option Pricing Models, The Journal of Finance, Vol. LII, No. 5 Bergomi, Lorenzo (2005), Smile Dynamics II, Risk, 18 Black, Fischer and Scholes, Myron (1973). "The Pricing of Options and Corporate Liabilities"., Journal of Political Economy, 81 (3) page 637–654 Carr, Peter and Madan, Dilip B (1999), Option valuation using the fast Fourier transform, Journal of Computational Finance, page 61-73 Campbell, John Y, Lo, Andrew W and MacKinlay, A Craig (1997), The Econometrics of Financial Markets, Princeton Cox, J.C., Ingersoll, J.E. & Ross, S.A. (1985), A Theory of the Term Structure of Interest Rates, Econometrica: Journal of the Econometric Society, vol. 53, no. 2, pp. 385-407. Heston, Steven L (1993), A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options, The Review of Financial Studies, Vol. 6, No. 2 , pp. 327-343 Hull, John and White, Alan (1987), The Pricing of Options on Assets with Stochastic Volatilities, The Journal of Finance, Vol. 42, No. 2, pp. 281-300 Kahl, Christian and Lord, Roger (2010), Fourier Inversion Methods in Finance, First version: 31st August 2009, this version: 26th January 2010 Merton, Robert C. (1973), Theory of Rational Option Pricing, Bell Journal of Economics and Management Science, 4 (1): 141–183 Merton, Robert C. (1976), Option pricing when the underlying stock returns are Discontinuous, Journal of Financial Economics, 3, page 125–144. Mikhailow, Sergei and Nögel, Ulrich (2008), Heston’s Stochastic Volatility Model Implementation, Calibration and Some Extensions, Wilmott Magasine Renault, Eric and Touzi, Nizar (1996), Option Hedging and Implied Volatilities in a Stochastic Volatility Model, Mathematical Finance, Vol. 6, NO. 3, page 279-302 Romano, Marc and Touzi, Nizar (1997), Contingent Claims and Market Completeness in a Stochastic Volatility Model, Mathematical Finance, Vol. 7, No. 4, page 399–410 62 Schmelzle, Martin (2010), Option Pricing Formulae using Fourier Transform: Theory and Application Schoebel, Rainer, and Zhu, Jianwei (1998), Stochastic Volatility With an Ornstein-Uhlenbeck Process: An Extension, www.uni-tuebingen\uni\wwm\papers.html Stein, E. and J. Stein (1991), Stock Price Distributions with Stochastic Volatility: An Analytic Approach, Review of Financial Studies, 4, page 727–752. Books: Cherubini, Umberto; Lunga, Giovanni Della; Mulinacci, Sabrina and Rossi, Pietro (2010), Fourier Transform Methods in Finance (The Wiley Finance Series) Gatheral, Jim (2006), The Volatility Surface. A Practitioner’s Guide, Wiley Finance Glasserman, Paul (2004), Monte Carlo Methods in Financial Engineering, Chapter 6, pages 399-343 Hull, John (2008), Options, Futures and other Derivatives, Prentice Hall, 7th edition Lewis, Alan L (2005), Option Valuation under Stochastic Volatility, Finance Press, 2nd printing Rebonato R. (1999), Volatility and Correlation, John Wiley, Chichester Wilmott, Paul (2005), Paul Wilmott Introduces Quantitative Finance, John Wiley, Chichester Academic reports Moodley, Nimalin (2005), The Heston Model: A Practical Approach with Matlab Code Others: Bond & Interest Rate, Elisa Nicolato, handout Empirical Finance notes, 2010, chapter 4, slide 21 Advances in Financial Modeling, 2010, notes Used as inspiration: Kivila, Liis and Ucar, Sibel (2009), Stochastic Volatility Models with Application to Volatility Derivatives, Thesis, Aarhus School of Business 63 Internet sources: (all visited last time on 08/31/2011) http://www.ivolatility.com/help/14.html http://www.randomwalktrading.com/main/index.php?option=com_content&view=article&id=216& Itemid=338 http://www.standardandpoors.com/indices/sp-500/en/us/?indexId=spusa-500-usduf--p-us-l-http://www.standardandpoors.com/servlet/BlobServer?blobheadername3=MDTType&blobcol=urldata&blobtable=MungoBlobs&blobheadervalue2=inline%3B+filename%3DFact sheet_SP_500.pdf&blobheadername2=ContentDisposition&blobheadervalue1=application%2Fpdf&blobkey=id&blobheadername1=contenttype&blobwhere=1243931099288&blobheadervalue3=UTF-8 http://www.treasury.gov/resource-center/data-chart-center/interestrates/Pages/TextView.aspx?data=yieldYear&year=2010 http://wrds-web.wharton.upenn.edu/wrds/process/wrds.cfm 64 Appendix See attached CD-rom 65