The Constant Elasticity of Variance Option Pricing Model
|
|
|
- Amie Green
- 10 years ago
- Views:
Transcription
1 The Constant Elasticity of Variance Option Pricing Model John Randal A thesis submitted to the Victoria University of Wellington in partial fulfilment of the requirements for the degree of Master of Science in Statistics and Operations Research April, 1998
2 2
3 Acknowledgements The author would like to thank his supervisors, Peter Thomson and Martin Lally, for their guidance and encouragement. Also, thanks to Credit Suisse First Boston NZ Limited for providing data, and to Edith Hodgen for valuable assistance with the preparation of this document. i
4 ii
5 Abstract The Constant Elasticity of Variance (CEV) Model was first presented in 1976 by John Cox and Stephen Ross as an extension to the famous Black-Scholes European Call Option Pricing Model of 1972/3. Unlike the Black-Scholes model, which is accessible to anyone with a pocket calculator and tables of the standard normal distribution, the CEV solution consists of a pair of infinite summations of gamma density and survivor functions. Its derivation rested on the risk-neutral pricing theory and the results of Feller. Moreover, descriptions of this model in journal and text-book literature frequently contained errors. One difficulty in implementation of the Black-Scholes model is that one of its arguments is an unobserved parameter σ, the share price volatility. Much research has concentrated on estimating this parameter with a general conclusion that it is better to imply this volatility from observed option prices, than to estimate it from stock price data. In the case of the CEV model there are two unobserved parameters, δ 2, with a relationship to the Black-Scholes parameter, and β, which defines the relationship between share price level and the variance of the instantaneous rate of return on the share. Attempts made to estimate this second parameter in the early 1980 s were not altogether satisfactory, perhaps condemning the CEV model to obscurity. A breakthrough was made in 1989 with a paper by Mark Schroder, who devised a method of evaluating the CEV option prices using the non-central Chi-Squared probability distribution, hence facilitating significantly simpler computation of the prices for those with suitable statistical software 1. 1 I use the statistical package SPLUS extensively in this thesis. iii
6 iv This thesis attempts to summarise the development of the CEV model, with comparisons made to the industry standard, the Black-Scholes model. The elegance of Schroder s method is also made clear. Joint estimation of the two parameters of the CEV model, δ and β, is attempted using both simulated data, and a sample of stocks traded on the Australian Stock Exchange. In this section, it appears that significant improvements can be made to earlier estimation methods. Finally, I would like to note that this thesis is primarily a statistical analysis of a financial topic. As a consequence of this, the focus of my analysis differs to that of articles in the financial journal literature, and that of finance texts. Furthermore, the literature on the CEV model is sparse, and in general theorems therein are stated without proof. Some theorems found in this thesis reflect the statistical nature of the analysis and are hence absent from the financial literature which I have surveyed and referenced. Throughout the thesis, I have attempted to make it clear when an idea or proof follows previously established material. Unattributed material is generally that which I have worked on with my supervisors guidance, but which is not found in the references I have used.
7 Contents 1 Introduction to Options The Call Option An example The Value of Call Options GBM and the Black-Scholes Model Introduction Share Price Evolution Simulation of Geometric Brownian Motion The Future Share Price S T Properties of C T - the Exercise Payoff Mean and Variance of C T given S t Simulation of C T The Black-Scholes Formula Properties of Call Price Prior to Maturity Best Predictor of a Future Black-Scholes Price Graphical Examination of C t Use of the Black-Scholes Model The CEV Model Introduction The CEV Share Price Solution Solution of the Forward Kolmogorov Equation v
8 vi CONTENTS Simulation of CEV Share Prices CEV Share Price Series Graphical Examination of S T Properties of C T - the Exercise Payoff The CEV Option Pricing Formula The CEV Solution Reconciling Various Forms of the CEV Solution Computing the Option Price The Absolute CEV Model Computing the General Model CEV Option Prices Behaviour of CEV Prices Use of the CEV Model Data Analysis Introduction Summary of Alternative Methods Estimating β from a share price series β Estimation Strategy The Variance of ˆβ Appraisal of the Estimation Technique Data Analysis Conclusions 141 A Definitions 143 B Proofs for Selected Results 147 B.1 Result B.2 Result B.3 Result B.4 Result
9 CONTENTS vii C Complete List of Shares 155 D SPLUS Code 157 D.1 GBM Simulation D.2 Inversion of the Black-Scholes Formula D.3 CEV Share Price Simulation D.4 Estimation of β Bibliography 164
10 viii CONTENTS
11 List of Figures 1.1 The lower bounds for call option value, and a possible form for the call price A realisation of GBM with initial value S t = $5, and parameters µ = 0.1, σ = 0.3 and 250 subintervals The daily returns for the series shown in Figure 2.1 with estimated and actual mean, and the estimated mean ± 2 standard deviation limits realisations of S T, a future geometric Brownian motion price, with initial value S t = $5, and parameters τ = 1, µ = 0.1 and σ = 0.3, and the theoretical lognormal distribution for S T The unbroken series is a realisation of GBM with initial value S t = $5 and parameters µ = 0.1, σ = 0.3, and 250 subintervals; the second series is the Black-Scholes prices for this share series added to the present value of the exercise price, with K = $5, and r = 0.06 and maturity at T = 1 year; finally the smooth broken line represents the present value of the exercise price ix
12 x LIST OF FIGURES 2.5 The unbroken series is a realisation of GBM with initial value S t = $5 and parameters µ = 0.1, σ = 0.3, and 250 subintervals; the second series is the Black-Scholes prices for this share series added to the present value of the exercise price, with K = $5, and r = 0.06 and maturity at T = 1 year; finally the smooth broken line represents the present value of the exercise price The distribution of call prices obtained using the Black-Scholes formula on a sample of share prices S t+s, where s = 1 year, S t = $5, µ = 0.1, σ = 0.1, K = $5, r = 0.06 and time to maturity τ s of 5 months. Also shown is the theoretical density function The distribution of call prices obtained using the Black-Scholes formula on a sample of share prices S t+s, where s = 1 year, S t = $5, µ = 0.1, σ = 0.1, K = $5, r = 0.06 and time to maturity τ s ranges from 2 years to half a month. Also the theoretical density function for these prices The distribution of call prices obtained using the Black-Scholes formula on a sample of share prices S t+s, where s = 1 year, S t = $5, µ = 0.1, σ = 0.1, K = $5, r = 0.06 and time to maturity τ s ranges from 2 to 75 years. Also the (solid) theoretical density function for these prices, and the (dotted) lognormal density function of S t+s The relationship between β and P (S T = 0), with S t = $5, τ = 1 year, µ = 0.10, σ = δs β/2 1 t = 0.3 in the solid curve and σ = 0.28 in the broken curve
13 LIST OF FIGURES xi 3.2 Firstly, a typical realisation of GBM, with S t = $5, τ = 1 year, µ = 0.10, σ = 0.3 and n = 250 subintervals; secondly, the daily returns for this series with estimated mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against the estimated standard deviation of the daily returns Firstly, a typical realisation of share price, with S t = $5, τ = 1 year, µ = 0.10, σ = 0.3 and n = 250 subintervals and CEV parameter β = 1; secondly, the daily returns for this series with estimated mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against the estimated standard deviation of the daily returns realisations of S T, a future CEV price with β = 1, with S t = $5, τ = 1 year, µ = 0.1 and σ = 0.3, the (solid) theoretical density for these prices, and the lognormal density function with the same parameters The empirical cumulative distribution function of the 5000 realisations of S T shown in Figure 3.4, a future CEV price with β = 1, with parameters S t = $5, τ = 1, µ = 0.1 and σ = 0.3, and the theoretical distribution function for these prices The standard deviation of a future CEV share price S T, with S t = $5, τ = 1 year, µ = 0.1 and σ = 0.3, for 2 β A realisation of a CEV share price with initial value S t = $5, and parameters β = 1, τ = 1, µ = 0.1, σ = 0.3, and 250 subintervals; also the CEV option prices for this share series added to the present value of the exercise price, with K = $5, and r = 0.06 and time to maturity indicated on the horizontal scale; finally the present value of the exercise price itself
14 xii LIST OF FIGURES 3.8 Black-Scholes implied volatilities for Absolute CEV option prices with S t = $5, τ = 1 year, σ = 0.3, r = 0.06 and $4 K $ Out-of-the-money CEV option prices, with β between -2 and 6, σ between 0.4 and 2, and additional parameters S t = $50, K = $55, τ = 0.5 years, and r = At-the-money CEV option prices, with β between -2 and 6, σ between 0.4 and 2, and additional parameters S t = $50, K = $50, τ = 0.5 years, and r = In-the-money CEV option prices, with β between -2 and 6, σ between 0.4 and 2, and additional parameters S t = $50, K = $45, τ = 0.5 years, and r = The log-likelihood surface, l(β, δ), for a simulated series with S t = $5, τ = 3 years, µ = 0.1, σ = 0.3, n = 250 subintervals per year, and β = The cross-section of the log-likelihood surface in Figure 4.1, l(β, ˆδ) for a simulated series with St = $5, τ = 3 years, µ = 0.1, σ = 0.3, n = 250 subintervals per year, and β = 0. In addition, the line β = ˆβ which identifies the maximum e n, given by Equation (4.14), for the simulated series examined previously with S t = $5, τ = 3 years, µ = 0.1, σ = 0.3, n = 250 subintervals per year, and β = Estimates of β, resulting from CEV share price simulation, used for the figures in Table 4.2. Superimposed on these are the density function of an N( β, 1) random variable over the range of the estimates, where β is the sample average of the β estimates Estimates of s( ˆβ), the standard deviation of ˆβ, from Table 4.2 against the true β
15 LIST OF FIGURES xiii 4.6 Sample values e n, defined in Equation (4.14), for the share price of AMC, with superimposed N(0, 1) density function The BHP share series; in addition, the daily returns for this series with moving mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against standard deviation of the daily returns The MIM share series; in addition, the daily returns for this series with moving mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against standard deviation of the daily returns The BOR share series; in addition, the daily returns for this series with moving mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against standard deviation of the daily returns
16 xiv LIST OF FIGURES
17 List of Tables 3.1 A section of MacBeth and Merville s Table 1, of CEV option prices, calculated for the parameters shown, with additional parameters: S t = $50 and r = Beckers δ values (rounded to 4 d.p.) for the Square Root CEV process, found using Equation (3.45), with additional parameters S t = $40, and µ = log(1.05) A section of Beckers Table II, showing Square Root CEV prices, with additional parameters S t = $40, and r = log(1.05) Beckers δ values (rounded to 4 d.p.) for the Absolute CEV process, with additional parameters S t = $40, and µ = log(1.05) MacBeth and Merville s β estimates for six stocks Summary of the β estimates for simulated series, obtained using the maximum likelihood procedure described above, with S t = $5, µ = 0.1, σ = 0.3 and where 3 years data is used p-values for hypothesis test H 0 : β = i, i = 2, 1,..., 6 using the simulated data summarised in Table β estimates for the 44 ASX share series, with estimates of s( ˆβ), the standard deviation of ˆβ, and the resulting confidence intervals obtained by simulation p-values for the test of normality of e n for the 44 ASX share series, where e n is given by Equation (4.14) C.1 ASX Company Codes xv
18 xvi LIST OF TABLES
19 Chapter 1 Introduction to Options 1
20 2 CHAPTER 1. INTRODUCTION TO OPTIONS 1.1 The Call Option A call option gives the holder the opportunity, but not the obligation, to purchase a unit of its underlying asset some time in the future, at a price decided now. If the holder decides to buy the unit of underlying asset, they will exercise the option. The price they pay is called the exercise price. If the option is European the holder may exercise the option only on the exercise or maturity date of the option. American options may be exercised at any time up to and including the maturity date. An option will be either American or European, and it will have the fundamental characteristics: underlying asset, exercise price and maturity date An example In New Zealand, it is possible to buy call options on the shares of Brierley Investments Limited (BIL). These options are traded along with options on other underlying assets on the New Zealand Futures and Options Exchange (NZFOE), information about which can be found on the internet at For example, Investor X could purchase a BIL option traded on the NZFOE, maturing in August 1998, with an exercise price of $1.20. All share options on the NZFOE are American, and so this option allows Investor X to buy 1000 BIL shares for $1200 any time between now and August If Brierley shares trade above $1.20 between now and August, Investor X may choose to exercise the option and receive the thousand shares. Alternatively, X might exercise the option and immediately sell the shares at the current share price (which would be greater then $1.20) for a profit. If the BIL share price does not exceed $1.20 before August, Investor X can (and should) let the option lapse, at no further cost. 1.2 The Value of Call Options Although many traded options are American, and upon exercise yield many shares, for the remainder of this project I will consider only European call
21 1.2. THE VALUE OF CALL OPTIONS 3 options, whose underlying asset is a single share. A call option is a derivative asset, whose value is based on the value of another asset, namely the underlying stock. Since exercise of the call option in the future delivers this share, the price of the option will obviously reflect the present worth of this share and the chances of it being higher than the exercise price on the maturity date. Let me present the following definitions: Definition 1.1. Let S t be the price at time t of a share paying no dividends over the time interval [t, T ]. Definition 1.2. Let C t be the price at time t of a European option. Definition 1.3. Let K be the exercise price of a European option, payable on exercise, for a single share with price S t at time t. Definition 1.4. Let T be the maturity date of the European option, and τ = T t the time until maturity from time t. Definition 1.5. Let r be the risk-free rate, which is payable continuously on an asset whose future value is certain. Definition 1.6. A European option is in-the-money if the current share price S t exceeds K, at-the-money if the current share price is equal to K, and out-of-the-money if the current share price is less than K. On the maturity date T, the European option either expires worthless, or delivers a single share, with value S T, in exchange for a cash payment K. The payoff of the option at maturity can be represented mathematically by the equation: Payoff = C T = { S T K S T > K 0 S T K (1.1) or equivalently C T = max(0, S T K) = (S T K) +.
22 4 CHAPTER 1. INTRODUCTION TO OPTIONS Prior to maturity, at time t, the option will have value not less than zero, since the option cannot yield a negative payoff, and also not less than the current share price S t less the present value of the exercise price discounted at the risk-free rate Ke rτ. This is established in the following theorem, whose proof is standard and can be found in Hull (1997). Theorem 1.1 (Lower Bounds for Call Option Value). C t max(0, S t Ke rτ ) Proof. Since the option yields a non-negative payoff in the future, the price paid now for the opportunity to receive those payoffs must not be less than zero, Hence C t 0. Consider now two portfolios: Portfolio A, consisting of a single option; Portfolio B, consisting of a share and a liability of K, payable at T. At T, the values of portfolios A and B are: V A T = C T = max(0, S T K) V B T = S T K respectively, and so we see V A T V T B A, and hence it follows that Vt V B t. If the latter relationship were not true, arbitrage profits could be earned by selling portfolio B and investing in portfolio A. Now V B t is given by V B t = S t Ke rτ and hence C t S t Ke rτ. (1.2)
23 1.2. THE VALUE OF CALL OPTIONS 5 Option prices given by any option pricing model should obey the bounds given in Equation (1.2), and will have the same general appearance as the function shown in Figure 1.1. Properties of option prices appear in Merton (1973), and in particular, he proves that call prices must be convex in the share price 1. Option Price C t S t Ke rτ 0 0 Ke rτ Share Price Figure 1.1: The lower bounds for call option value, and a possible form for the call price. From the relationship given in Theorem 1.1 it is immediately apparent that the option price must depend on at least four factors: S t, the current share price; τ, the time to maturity of the option; K, the exercise price of the option; and r, the continuously compounding risk-free rate. 1 Merton (1973), Theorem 10, page 150.
24 6 CHAPTER 1. INTRODUCTION TO OPTIONS In addition, the option price will depend on the stochastic properties of S T. The fact that the call price is not equal to the lower bound given in Theorem 1.1 is due to the random nature of the future share price S T, and in particular: σ, the share price volatility. This parameter, which will be defined more formally later, represents the uncertainty of future share prices. associated with a future share price increases. As this parameter increases, the risk The option price may be thought of as the sum of the lower bound and a premium, where the premium is monotonically increasing in σ. Treating S t as fixed, and considering each of τ, K and r in turn, ceteris paribus, we can anticipate the effect a change in each factor might have on the current option price, using the lower bound for the option price given in Theorem 1.1. As the time to maturity τ increases, the present value of the exercise payment at the risk-free rate r diminishes: lim τ Ke rτ = 0, and so the lower bound C t = S t Ke rτ in Figure 1.1 is translated to the left as the y intercept term decreases, and so the option price C t corresponding to any particular S t must increase to compensate. As the exercise payment increases, there is an opposite effect on the lower bound C t = S t Ke rτ. In this case the y intercept term will increase, thus translating the lower bound to the right, and allowing C t to decrease. Increasing the exercise price therefore decreases the value of a European option. Ke rτ As the risk-free rate r increases, the present value of the exercise payment will decrease as it did when τ was increased, and hence the option value will increase. At exercise, the option delivers the payoff (S T K) +. S T is of course a random variable, since at the current time t we do not know for sure what
25 1.2. THE VALUE OF CALL OPTIONS 7 the share price at T will be. In order to make any statements about the probabilistic properties of S T, assumptions must be made on how the share price evolves through time. The evolution of a general process X t may be described by the following stochastic differential equation (SDE): dx t = µ(x t, t)dt + σ(x t, t)db t (1.3) where dx t may be interpreted as the change in X t over the period [t, t + dt], µ(x t, t) and σ(x t, t) are functions of X t and t, and {B t } is Brownian motion, with initial condition B 0 = 0. This process is too general for our purposes, and I will restrict attention to two specific cases, geometric Brownian motion (GBM), and Constant Elasticity of Variance (CEV) evolution. Definition 1.7. A share price that follows geometric Brownian motion (GBM) is a solution to the following SDE: ds t = µs t dt + σs t db t (t > 0) where µ and σ are constant, and B 0 = 0. Definition 1.8. A share price that follows the Constant Elasticity of Variance (CEV) model is a solution to the following SDE: ds t = µs t dt + δs β 2 t db t (t > 0) (1.4) where µ, δ and β are constant, and B 0 = 0. It is clear from Definitions 1.7 and 1.8 that GBM is a special case of the CEV model, and corresponds to the case when β = 2. The solution to Equation (1.4) has very different behaviour for the three cases β = 2, β < 2, and β > 2. In the first case, the solution to the SDE above is geometric Brownian motion, which is a well known and widely studied process. This process, and the price of a call option over a share price following GBM,
26 8 CHAPTER 1. INTRODUCTION TO OPTIONS are described in Chapter 2. When β < 2, the share price is the original Constant Elasticity of Variance process, and is considered, again with its companion option prices, in Chapter 3. The third case, corresponding to β > 2 is mentioned briefly at the end of Chapter 3, and is applied in Chapter 4 along with both other cases.
27 Chapter 2 GBM and the Black-Scholes Model 9
28 10 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL 2.1 Introduction Analysis of share prices over time would suggest that they are discrete-time, discrete-variable processes. This means that they change at discrete time points, and may take on only discrete values. In practice, however, share prices are generally modelled using continuous-time, continuous-variable processes. The geometric Brownian motion (GBM) process examined in this chapter, and the Constant Elasticity of Variance (CEV) process considered in the following chapter are continuous-time, continuous-variable processes. In addition, both these models are Markovian, meaning that the only relevant information regarding the future of the process is the present value, and that the past is irrelevant. This can be expressed mathematically as follows: P (S t+s < s S u, 0 u t) = P (S t+s < s S t ). This property is consistent with weak form efficiency in the share market, since it implies that all information in prices S 0, S 1,..., S t 1 is encapsulated in the present price S t. The Black-Scholes Model was first presented in an empirical paper by Black & Scholes (1972), which was quickly followed by its derivation in Black & Scholes (1973). This model is based on a number of restrictive assumptions, one of which is that the price of the underlying asset has a lognormal distribution at the end of any finite (forward) interval, conditional on its value at some initial starting point. It will be shown that if we assume that the share price follows GBM, then this condition will be met. In this chapter I will deal initially with properties of geometric Brownian motion, and then examine the Black-Scholes model. 2.2 Share Price Evolution Derivation of the Black-Scholes option pricing equation requires the assumption that the future share price has a lognormal distribution. This condition
29 2.2. SHARE PRICE EVOLUTION 11 is met if share price follows GBM, a process which is defined in Definition 1.7, and is a solution to the following SDE: ds t = µs t dt + σs t db t (2.1) with initial condition B 0 = 0, and where S t is the share price at time t, µ is the continuously compounding expected growth rate of S t, σ is the standard deviation of the instantaneous return on S t, and {B t } is Brownian motion, with E(dB t ) = 0 and Var(dB t ) = dt. Note that both µ and σ are constants, independent of time and the current share price, and that Equation (2.1) is a special case of Equation (1.4), with β = 2 and δ = σ. I show below that the solution of the SDE above has a lognormal distribution. To prove this it is necessary to use a result from stochastic calculus called Itô s Differentiation Lemma. Result 2.1 (Itô s Differentiation Lemma). Suppose that f(x, t) and its partial derivatives f x, f xx and f t are continuous. If X t is given by dx t = µ(x t, t)dt + σ(x t, t)db t then Y t = f(x t, t) has stochastic differential: dy t = (µ(x t, t)f x + f t σ2 (X t, t)f xx )dt + σ(x t, t)f x db t = f x dx t + f t dt σ2 (X t, t)f xx dt. Hull (1997) gives a sketch proof of this result using a Taylor series expansion. Using Itô s Lemma, we can now prove the following well known theorems. Theorem 2.1. Conditional on its value at time t, a share price S T that follows GBM will have a lognormal distribution at T > t, with parameters E(ln S T ) = ln S t + (µ 1 2 σ2 )τ and Var(ln S T ) = σ 2 τ, where τ = T t. Moreover, the form for S T given S t will be: S T = S t e (µ 1 2 σ2 )τ+σ τz where Z is a standard normal variable with mean zero and unit variance.
30 12 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL Proof. Consider the transformation Y t = ln S t. Referring to the SDE (2.1), and to Result 2.1, we see that: and so applying Itô s Lemma d ln S t = ln S t t X t = S t µ(x t, t) = µs t σ(x t, t) = σs t f(x t, t) = ln S t dt + ln S t S t ds t + 1(σS 2 t) 2 2 ln S t 2 S dt t = 0 dt + 1 S t (µs t dt + σs t db t ) 1 2 σ2 S 2 t = ( µ 1 2 σ2) dt + σdb t (2.2) Thus, integrating both sides from t to T yields: ln S T ln S t = ( µ 1 2 σ2) (T t) + σ T = ( µ 1 2 σ2) τ + σ(b T B t ) From the properties of Brownian motion, in particular its independent increments: B T B t B T t B 0 = B τ t 1 S 2 t db t since B 0 = 0, and where denotes is distributed as. Hence: ln S T ln S t + ( µ 1 2 σ2) τ + σb τ. (2.3) Conditioned on S t, it is clear that the only random variable in the RHS of the above equation is the Brownian motion term B τ. Using the fact that B t N(0, t) for any t > 0, we obtain the mean and variance of ln S T given S t : E(ln S T S t ) = E(ln S t + ( µ 1 2 σ2) τ + σb τ S t ) = ln S t + ( µ 1 2 σ2) τ + σe(b τ ) = ln S t + ( µ 1 2 σ2) τ dt
31 2.2. SHARE PRICE EVOLUTION 13 Var(ln S T S t ) = Var(ln S t + ( µ 1 2 σ2) τ + σb τ S t ) = σ 2 Var(B τ ) = σ 2 τ In addition, if X is a normal random variable, then the linear combination a + bx, where a and b are both constant, is also normal with mean ae(x) and variance b 2 Var(X). Combining this with the mean and variance above it is clear that ln S T S t N ( ln S t + ( µ 1 2 σ2) τ, σ 2 τ ) (2.4) i.e. S T is lognormal with parameters ln S t + ( µ 1 2 σ2) τ and σ 2 τ. Moreover, since Z = Bτ τ N(0, 1), it follows directly from Equation (2.3) that S T = S t e (µ 1 2 σ2 )τ+σ τz (2.5) as required. Theorem 2.2 (Moments of a Share Price following GBM). The mean and variance of the share price S T conditional on an earlier share price S t, 0 t < T are S t e µτ and S 2 t e 2µτ (e σ2τ 1) respectively. Proof. Note first that the moment generating function of an N(0, 1) variable is E(e sz ) = e 1 2 s2. This can be used to find the moments of the random variable of interest S T S t. ( ) E(S T S t ) = E S t e (µ 1 2 σ2 )τ+σ τz S t ( ) = S t e (µ 1 2 σ2 )τ E e σ τz = S t e (µ 1 2 σ2 )τ e 1 2 σ2 τ = S t e µτ (2.6) ( ) E(ST 2 S t ) = E St 2 e 2(µ 1 2 σ2 )τ+2σ τz S t ( ) = St 2 e 2(µ 1 2 σ2 )τ E e 2σ τz = S 2 t e 2(µ 1 2 σ2 )τ e 2σ2 τ = S 2 t e (2µ+σ2 )τ
32 14 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL Forming the variance using the relationship Var(X) = E(X 2 ) E(X) 2, we obtain as required. Var(S T S t ) = S 2 t e 2µτ (e σ2 1) (2.7) Note that if share price evolution is a deterministic process, i.e. σ = 0, then S T equals the mean of the stochastic process: S t e µτ. Thus using either the expected value, or the deterministic case, the parameter µ can be thought of as the continuously compounding expected rate of return on the share per unit time, and may be modelled for a particular stock using the Capital Asset Pricing Model (CAPM), which features both a market risk premium, and a measure of the riskiness of the particular firm compared to the market. The CAPM describes the expected return on a particular asset: µ j = r + MRP Cov(R j, R m ) σ 2 m where MRP is the market risk premium, R j is the return on asset j, µ j is the expected value of that return, R m is the return on the market portfolio, and σ 2 m is the variance of the market return. Aggregate investor attitudes towards risk affect the size of the market risk premium, which has a direct effect on the size of µ j. The size of this effect is determined by the measure of systematic risk: Cov(R j, R m ) σm 2 which compares the risk associated with the particular asset to the risk associated with the market as a whole. The parameter µ is the only component of the GBM model that reflects investor risk attitudes. The return on an asset j can be modelled using the standard univariate linear regression model: R j = α j + b j R m + ɛ j
33 2.2. SHARE PRICE EVOLUTION 15 where R j and R m are as above, α j and b j are the regression model parameters, and ɛ j is a stochastic error term, with variance σ 2 ɛ j. Applying the variance operator to all terms in the equation above yields the relationship: σ 2 j = b 2 jσ 2 m + σ 2 ɛ j. The variance of the market return, σ 2 m measures the systematic risk in the system, whereas σ 2 ɛ j measures the non-systematic risk. Hence the parameter σ reflects a combination of these, but not investor risk attitudes Simulation of Geometric Brownian Motion Simulation of GBM series, or prices at a particular time in the future is a useful way of analysing the properties of GBM, and later, analysing the properties of Black-Scholes option prices. Because the solution to the SDE (2.1) is known, there is no need to use a numerical method to approximate the solution, but rather the solution can be simulated directly. The solution to the SDE was given in Equation (2.5) and is: S T = S t e (µ 1 2 σ2 )τ+σ τz where Z N(0, 1). Hence in order to simulate a single price at T, a single realisation of Z can be obtained, and the share price computed directly. In order to simulate an entire GBM series, we can divide the interval of interest [t, T ] into n subintervals defined by the times: t = t 0 < t 1 < < t i < < t n 1 < t n = T where the intervals are not necessarily of equal length. Equation (2.5) can also be written to give the share price at time t i conditional on the share price at t i 1 : S ti = S ti 1 e (µ 1 2 σ2 )(t i t i 1 )+σ t i t i 1 Z
34 16 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL hence simulating n realisations of Z, the series can be constructed as follows: i S ti = S t S tj S j=1 tj 1 i = S t e (µ 1 2 σ2 )(t j t j 1 )+σ t j t j 1 Z j j=1 ( i = S t exp (µ 1 2 σ2 )(t j t j 1 ) + σ ) t j t j 1 Z j j=1 ( = S t e (µ 1 2 σ2 )(t i t) exp σ ) i tj t j 1 Z j. (2.8) j=1 A program which simulates GBM using Equation (2.8) is given in Appendix D.1. Share Price Time - years Figure 2.1: A realisation of GBM with initial value S t = $5, and parameters µ = 0.1, σ = 0.3 and 250 subintervals. Figure 2.1 shows a single realisation of geometric Brownian motion with S t = $5, τ = 1 year, µ = 0.1 and σ = 0.3. This particular realisation has
35 2.2. SHARE PRICE EVOLUTION 17 S T = $5.12, which is smaller than its expected value E(S T S t ) = $5e 0.1 = $5.53 but well within a single standard deviation s(s T S t ) = S t e µ e σ2 1 = $1.696 of it. The daily returns for the series {S t } are defined as: R t = ln S t+ t ln S t where t = 1 years is approximately one trading day. Whilst the series 250 {S t } is clearly not a stationary process, Equation (2.2) indicates that the daily returns should be stationary, with a mean (µ 1 2 σ2 ) t and variance σ 2 t. A plot of the daily returns can be seen in Figure 2.2, with an estimate of the mean level and standard deviation shown 1. These have been estimated using a Lowess filter with a smoothing window of 30 observations. The Lowess filter is a robust, centred moving average filter, and was derived by Cleveland (1979). This filter was designed to smooth scatterplots, but has application to the equi-spaced observations of a time series. It estimates the average value at t i using weights from the bisquare function: { (1 x 2 ) 2 x < 1 B(x) = 0 x 1 where x depends on the time t i and the smoothing window, and by making adjustments for outlying values. In this case I have used a smoothing window of 30 days, so that values outside this window are given zero weight, and hence do not affect the estimate. The estimates produced at the ends of the series are unreliable due to back- and forecasting of share price values. These estimates, obtained using estimated share price data, are not shown in the graph. Also shown in Figure 2.2 is the true mean for the daily returns of (µ 1 2 σ2 ) t which is negligible for the parameters chosen 2. We see that the estimated mean does indeed oscillate about the actual mean level, and that 1 The series actually shown are the estimated mean, and the estimated mean ± two estimated standard deviations, from which the estimated standard deviation itself can be recovered. 2 Nevertheless, over a year it compounds to a significant amount.
36 18 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL Log-returns Time - years Figure 2.2: The daily returns for the series shown in Figure 2.1 with estimated and actual mean, and the estimated mean ± 2 standard deviation limits. the estimated standard deviation appears approximately constant. In fact, an autocorrelation plot for the daily return series shows no significant autocorrelation for non-zero lags. This is not surprising, since the series was generated in the first place using realisations of the standard normal variable that should indeed be independent of one another, and hence uncorrelated The Future Share Price S T As discussed in Chapter 1, the value of a European option at t will certainly depend on the share price at the future exercise time T. Since we are able to simulate geometric Brownian motion, we can also simulate the distribution of future share prices by generating many realisations of the same process. The value of this is largely illustrative, since many of the properties of GBM are well known. Suppose a particular share price truly follows GBM, with known µ and σ, then N realisations of the share price at T can be obtained
37 2.2. SHARE PRICE EVOLUTION 19 using the second program seen in Appendix D.1. Simulation of the quantity S T is a useful way of deciding what properties the share price will have at T, and hence what payoff (if any) the option is likely to deliver. In the case of GBM, the theoretical distribution is known, and in this case can be compared to the results of the simulation to appraise the simulation procedure, and help guide the eye. Figure 2.3 shows the result of a simulation of 5000 share prices τ = 1 year in the future, with additional parameters S t = $5, µ = 0.1 and σ = 0.3. The final value of the time series in Figure 2.1 has the same properties as each of the observations shown in the histogram. Superimposed on the observed distribution is the theoretical lognormal density curve with parameters E(ln S T S t ) = ln 5 + ( ) and Var(ln S T S t ) = It is clear from the graph that the fit is very good, particularly when the bars are small. This is expected since the height of frequency histogram bars is a Poisson random variable with mean and variance equal to the expected height of the bars. Hence, the smaller bars heights will have a small standard deviation and should be closer to the curve than the higher bars. The heights of the equi-width bars in the relative frequency histogram shown in Figure 2.3 are proportional to the Poisson heights in a frequency histogram, and so the variability comments hold. A useful means of gauging the success of the simulation procedure is to estimate the mean and variance of a sample of share prices, and compare these to the theoretical values given by Equations (2.6) and (2.7) respectively. These estimates are given by the sample mean and sample variance of and respectively, compared to theoretical values of and Note that while the means differ slightly, the variance estimate is accurate to within 4 decimal places, again testimony to the accuracy with which GBM can be simulated. Note that these estimates do not correspond to the parameters of the lognormal distribution, which are the mean and variance of the log share prices. The maximum likelihood method could be used to fit the best lognormal distribution to the sample values but this has not been done here.
38 20 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL Relative Frequency Share Price Figure 2.3: 5000 realisations of S T, a future geometric Brownian motion price, with initial value S t = $5, and parameters τ = 1, µ = 0.1 and σ = 0.3, and the theoretical lognormal distribution for S T. 2.3 Properties of C T - the Exercise Payoff The properties of the call value at maturity are intimately linked to those of the share price by the equation: C T = max(s T K, 0) = (S T K) + (2.9) where K is the exercise price of the call option. It is this payoff that investors are interested in valuing. Such a future cash flow could be valued using the equation P t = e λτ E(C T S t ) (2.10) where P t would be the price paid now for the payoff C T, which would be received at time T. This equation features two rates particular to the risk
39 2.3. PROPERTIES OF C T - THE EXERCISE PAYOFF 21 preferences of investors in aggregate: µ, the continuously compounded expected rate of return which features in the expectation, and λ, the discount rate for future cash flows. It will be shown later that both µ and λ can be treated as if they were the risk-free rate r, for valuing this particular future cash flow. Consider the simple transformation Y T = S T K. It is clear Y T has mean E(S T ) K, variance Var(S T ), and Y T +K has a lognormal distribution. Note that Y T itself does not have a lognormal distribution, since the lognormal distribution is defined on the range [0, ) whereas Y T is defined on [ K, ). However, the shape of the distribution of Y T will be identical to that of S T, since we have the relationship P (Y T < y) = P (S T K < y) = P (S T < y + K). Next consider C T = Y + T = max(0, Y T ). It is clear that P (C T = 0) = P (S T K) and so C T will not have a continuous distribution function. Theorem 2.3 (Distribution of C T given S t ). Given S t, C T = (S T K) + has a mixed distribution with density: P (C T c S t ) = f CT S t (c, τ) = F ST S t (K, τ) δ(c) + { 0 c < 0 F ST S t (K + c, τ) c 0 { 0 c < 0 f ST S t (K + c, τ) c > 0 where F ST S t and f ST S t are the cumulative distribution and density functions for the share price S T, and δ(c) is the Dirac delta function. Proof. Since S T given S t has a lognormal distribution with parameters ln S t + (µ 1 2 σ2 )τ and σ 2 τ, given S t : P (S T K c S t ) = P (S T K + c S t ) = P (ln S T ln(k + c) S t ) ( ln(k + c) ln St (µ 1 ) 2 = Φ σ2 )τ σ τ
40 22 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL where Φ(x) = x φ(z)dz is the standard normal cumulative distribution function, and φ(z) is the standard normal probability density function. Thus: F CT S t (c, τ) = P (C T c S t ) { 0 c < 0 = P (S T K c S t ) c 0 { 0 c < 0 = ( ) ln(k+c) ln St (µ Φ 1 2 σ2 )τ c 0 σ τ (2.11) 0 c < 0 f CT S t (c, τ) = P (C ( T = 0 S t ) δ(c) ) c = 0 Φ ln(k+c) ln St (µ 1 2 σ2 )τ c σ c > 0 τ 0 c < 0 = P (S T K S ( t ) δ(c) ) c = 0 1 (K+c)σ φ ln(k+c) ln St (µ 1 2 σ2 )τ τ σ c > 0 τ 0 c < 0 = F ST S t (K, τ) δ(c) c = 0 (2.12) f ST S t (K + c, τ) c > 0 where F ST S t and f ST S t are the cumulative distribution and density functions for the terminal share price S T given by: F ST S t (s, τ) = P (S T < s S t ) ( ln s ln St (µ 1 ) 2 = Φ σ2 )τ σ τ f ST S t (s, τ) = F S T S t (s, τ) s = 1 ( ln s ln sσ τ φ St (µ 1 ) 2 σ2 )τ σ τ and δ(c) is the Dirac delta function defined by: { x 0 x < 0 δ(c)dc = 1 x 0. (2.13) (2.14)
41 2.3. PROPERTIES OF C T - THE EXERCISE PAYOFF Mean and Variance of C T given S t The mean and variance of C T given S t can be found using the density function for C T given in Equation (2.12). An equivalent but simpler method is to note that C T = h(z), where Z is a standard normal random variable. Then the expected value of any function, g, of C T can be found using the relationship E{(g h)(z)} = (g h)(z)φ(z)dz where g h is the composition of g with h. This yields the following. Theorem 2.4 (Moments of C T ). Given S t, C T = (S T K) + has mean E(C T S t ) = S t e µτ Φ(g t ) KΦ(g t σ τ) and mean square E(C 2 T S t ) = S 2 t e (2µ+σ2 )τ Φ(g t + σ τ) 2S t e µτ KΦ(g t ) + K 2 Φ(g t σ τ) where g t = ln S t ln K + (µ σ2 )τ σ τ Proof. Note firstly that from Equation (2.5), it is clear that S T is a function of a standard normal variable Z. Hence, we find that C T too is a function of Z. Note that S T K implies C T = (S T K) + { 0 S T K = S t e (µ 1 2 σ2 )τ+σ τz K S T > K Z ln S t + ln K (µ 1 2 σ2 )τ σ τ = g t + σ τ
42 24 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL and therefore the first moment can be determined thus: E(C T S t ) = g t+σ τ ( S t e (µ 1 2 σ2 )τ+σ ) τz K φ(z)dz = S t e µτ e 1 2 (z σ τ) 2 dz K φ(z)dz g t+σ τ 2π g t+σ τ = S t e µτ φ(y)dy KΦ(g t σ τ) g t = S t e µτ Φ(g t ) KΦ(g t σ τ). The second moment gives us a means of calculating the variance of C T given S t, and is found as follows: E(C 2 T S t ) = g t+σ τ = S 2 t e (2µ+σ2 )τ = S 2 t e (2µ+σ2 )τ ( S t e (µ 1 2 σ2 )τ+σ τz K) 2 φ(z)dz g t+σ τ g t σ τ e 1 2 (z 2σ τ) 2 2π dz 2S t e µτ KΦ(g t ) + K 2 Φ(g t σ τ) φ(y)dy 2S t e µτ KΦ(g t ) + K 2 Φ(g t σ τ) = S 2 t e (2µ+σ2 )τ Φ(g t + σ τ) 2S t e µτ KΦ(g t ) + K 2 Φ(g t σ τ). The variance can be formed as usual using Var(C T S t ) = E(CT 2 S t) E(C T S t ) 2. Thus, from Equation (2.10), an investor may value the option using the expected value, and the market rates µ and λ, to give a price at t: P t = S t e (µ λ)τ Φ(g t ) Ke λτ Φ(g t σ τ). (2.15) where g t is defined in Theorem 2.4 above. Note that as aggregate investor risk attitudes change, both µ and λ will change, and hence it appears that P t will change. However, there is theory to show that the true value of C t should be independent of risk premia on assets, thus any change in µ is offset by an appropriate change in λ when the fair option price is calculated.
43 2.4. THE BLACK-SCHOLES FORMULA Simulation of C T A sample of terminal call values C T can be obtained from a sample of share prices using the simple relationship (2.9). The distribution of such a sample can easily be recovered from the density function of S T shown in Figure 2.3 by repositioning the origin at K, and setting all observations less than K to zero. As before, in the case of C T there is also good agreement between the theoretical and observed mean and variance. For an exercise price of K = $5, the sample yields respective values of and for the estimated mean and variance, compared to theoretical values of and It is also interesting to note the relative frequency of the event C T = 0, i.e. the proportion of occasions on which the option would not be exercised. The theoretical probability that the value of C T will be zero is given by the equation P (C T = 0 S t ) = P (S T < K S t ) = Φ( g t + σ τ) which for the simulation equals Hence, for this particular simulation, the expected number of options that are not exercised is This can be compared to the sample estimate of The Black-Scholes Formula Black and Scholes first presented the Black-Scholes model in an empirical paper (Black & Scholes 1972) with the theoretical underpinnings following in Black & Scholes (1973). Their model considers pricing a European call option, over a stock traded in a market with the following properties: the instantaneous interest rate is known, and constant; the share price follows geometric Brownian motion, from which it follows directly:
44 26 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL the variance rate σ 2 of the return on the stock is constant, and known; no dividends are paid on the share over the life of the option; there are no transaction costs, differential taxes, or short-selling restrictions, and it is possible to trade any fraction of the stock or option. Under these conditions Black & Scholes were able to obtain a price for the option that depends only on the current share price, the time to maturity, and on constants K, r, and σ that are assumed known. The partial differential equation (PDE) for the price of a call over a slightly more general share price process than GBM is established in the following well known theorem. Theorem 2.5 (The Call Price PDE). Suppose S t is the solution to the SDE: ds t = µ(s t, t)s t dt + σ(s t, t)s t db t (2.16) then the price at time t of a call option over a share with price S t must satisfy the PDE: 1 2 σ2 (S t, t)st 2 2 C S + rs C 2 t S + C t rc t = 0. subject to the boundary condition C T = (S T K) + where r is the continuously compounding risk-free rate. Proof. Consider forming a self-financing portfolio at t, of a t units of the stock with value S t, and b t units of the call option with value C t, where a t and b t may be functions of both share price at t and time. This portfolio has value at t: V t = a t S t + b t C t
45 2.4. THE BLACK-SCHOLES FORMULA 27 and over the period [t, t + dt] the change in portfolio value will be: dv t = V t+dt V t = a t+dt S t+dt + b t+dt C t+dt a t S t b t C t = (a t+dt a t )S t+dt + a t ds t + (b t+dt b t )C t+dt + b t dc t = a t ds t + b t dc t. The final step is justified by the assumption that the portfolio is self financing. This means that any change in the quantity of stock held is financed by a change in the quantity of the option held, and vice versa. This yields da t S t+dt + db t C t+dt = 0 as required. The form for ds t is given by Equation (2.16), and dc t can be obtained from it using Itô s Lemma. The option price can be considered a function of two variables: the current share price and time. Hence we can write C t = C(S t, t) where S t is a solution to the familiar SDE ds t = µ(s t, t)s t dt + σ(s t, t)s t db t. Referring to Result 2.1, we see that X t = S t µ(x t, t) = µ(s t, t)s t σ(x t, t) = σ(s t, t)s t f(x t, t) = C(S t, t). Hence, applying Itô s Lemma we can determine the change in C t over the period [t, t + dt]: dc t = C C dt + t S ds t + 1(σ(S 2 t, t)s t ) 2 2 C S dt. 2
46 28 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL The change in portfolio value over the interval [t, t + dt] thus becomes: dv t = a t ds t + b t dc t ( C = a t ds t + b t t ( C = a t + b t S and the portfolio s rate of return: dv t V t = C dt + S ds t σ2 (S t, t)st 2 2 C ) ds t + b t ( C t σ2 (S t, t)s 2 t ) S dt 2 ) 2 C dt S 2 = a tds t + b t dc t a t S t + b t C t ( ) ( ) C C at + b t dst + b S t + 1 t 2 σ2 (S t, t)st 2 2 C dt S 2 = where f t = a t /b t. ( ft + C S ) dst + a t S t + b t C ( t ) C + 1 t 2 σ2 (S t, t)st 2 2 C dt S 2 f t S t + C t Since the only stochastic elements are present in the ds t term, these can be eliminated by choice of f t to form a portfolio whose change in value over the period [t, t + dt] is deterministic. Setting the coefficient of ds t to zero gives dv t V t = f t = C S and so Var( dvt V t ) = 0 and the rate of return on the portfolio becomes: ( ) C + 1 t 2 σ2 (S t, t)st 2 2 C dt S 2 C S S t + C t. (2.17) By this choice of f t, the number of units of stock held per unit of option held, the portfolio has no risk over [t, t + dt], and so by arbitrage theory its rate of return should be the risk free rate r, giving dv t = rv t dt or dv t rv t dt = 0. If this were not the case, investors could borrow (lend) at the risk-free rate r, and go long (short) in the portfolio of stock and option and earn arbitrage
47 2.4. THE BLACK-SCHOLES FORMULA 29 profits. Substituting terms for dv t and V t in the above expression, yields the desired PDE: ( C t σ2 (S t, t)s 2 t 1 2 σ2 (S t, t)s 2 t ) ( 2 C dt r C ) S 2 S S t + C t dt = 0 2 C S + rs C 2 t S + C The solution C t must also satisfy the boundary condition C T = (S T K) + since at T, the option yields the payoff C T. t rc t = 0. (2.18) Theorem 2.6 (The Black-Scholes PDE). The price of a European call option, C t, over a share whose price follows GBM, with time τ until maturity, must satisfy the partial differential equation (PDE): 1 2 C 2 σ2 subject to the boundary condition C + rs S2 S + C t rc t = 0 (2.19) C T = (S T K) + where r is the continuously compounding risk-free rate. Proof. The Black-Scholes PDE follows from Theorem 2.5, with as does the boundary condition. σ(s t, t) = σ It is significant that neither of the PDEs (2.19) nor (2.18) feature µ, the only component of the respective models that describes investor risk preferences. Cox & Ross (1976) use this fact to solve this and other PDEs using a technique called risk-neutral valuation. Cox & Ross suggest that since the PDE does not feature risk preferences, then solution of the PDE
48 30 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL can be achieved by assuming any convenient scenario for investor preferences. In particular, assuming that investors are risk-neutral provides an alternative method of deriving solutions like the Black-Scholes, and Constant Elasticity of Variance formulae, since both µ and λ in the Equation (2.10) can be set to r, and the price P t evaluated. The general validity of this approach was established by Harrison & Kreps (1979). The following theorem was first given by Black & Scholes (1972). Theorem 2.7 (The Black-Scholes Formula). The solution to Equation (2.19), subject to the boundary condition C T = (S T K) + is given by the Black-Scholes Formula: C t = S t Φ(h t ) Ke rτ Φ(h t σ τ) where h t = ln S t ln K + (r σ2 )τ σ τ Proof. Evaluation of the partial derivatives C, C S t and 2 C S 2, where C t = S t Φ(h t ) Ke rτ Φ(h t σ τ) (2.20) and substitution into the PDE indeed shows that the Black-Scholes formula is a solution. It is useful to note the relationship: S t φ(h t ) Ke rτ φ(h t σ τ) = 0. From this it follows that the partial derivatives are: and C S = Φ(h t) 2 C S = φ(h t) 2 S t σ τ C t = S tσφ(h t ) 2 rke rτ Φ(h t σ τ). τ Substituting these derivatives into the PDE (2.19) shows that C t as given by the Black-Scholes formula is indeed a solution of the PDE. Moreover, C t also satisfies the boundary condition, since setting t = T gives C T = (S T K) + as required.
49 2.4. THE BLACK-SCHOLES FORMULA 31 Result 2.2. Let the share price S t be the solution to the SDE: ds t = µs t dt + σ t S t db t where σ t is a deterministic function of time. Then the price of a call option on a stock with share price S t must satisfy the PDE: 1 2 σ2 t St 2 2 C S + rs C 2 t S + C t r C t = 0 with boundary condition C T = (S t K) +, and the call price is given by the Black-Scholes equation, with the substitution: σ 2 = σ 2 1 τ T t σ 2 udu. Proof. The proof of this result is given in Appendix B.1. Equation (2.15) gives the present value at time t, of the expected value of C T, and features two rates µ and λ, which depend on investor risk attitudes. Cox & Ross (1976) evaluate P t with both µ and λ replaced by r giving a price which is identical to the Black-Scholes price. This method of valuing options is called the risk-neutral valuation method, and prices are given by: C t = e rτ E (C T S t ) where E is the expectation operator taken in a risk-neutral world (so that µ = λ = r). This is appropriate because the PDE of interest does not contain any parameters which reflect risk attitudes. Note that Result 2.2 has interesting implications for the interpretation of the volatility parameter σ used in the Black-Scholes formula. Whilst σ 2 is defined as the instantaneous variance of the rate of return on the share price, it can be treated as if it is the average variance over the remaining life of the option, since the same pricing formula results. This framework gives a more intuitive meaning to the parameter σ.
50 32 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL Properties of Call Price Prior to Maturity Prior to maturity, the call price C t and the share price S t are linked by the Black-Scholes formula. This link can be used to derive the probability distribution of the option price at time t + s, where t < t + s < T. At maturity of the option, the relationship between the share and option prices was relatively simple: C T = (S T K) + and yielded a distribution for C T that was very similar to that of S T. If t < T, then the relationship between share and option price becomes the Black-Scholes formula: C t = S t Φ(h t ) Ke rτ Φ(h t σ τ) which yields a more complicated probability function for an option price at a time in the future, but prior to maturity. Theorem 2.8 (Distribution of Black-Scholes Prices). C t+s conditional on S t with s (0, τ), is given by the Black-Scholes formula, and has probability density function: 0 c 0 f Ct+s S t (c, τ) = ( 1 ln BS 1 ) (c;t+s)σ sφ(h )φ S (c;t+s) ln St (µ 1 2 σ2 )s σ c > 0 s t+s BS 1 S where BS 1 S (c; t + s) is the inverse of the Black-Scholes function with respect to S at time t + s, evaluated at c, and h t+s = ln BS 1 S (c; t + s) ln K + (r σ2 )(τ s) σ τ s Proof. First note that C t+s > 0, since the probability of a positive payout is always greater than zero. Therefore for c < 0, P (C t+s < c) = 0. Denoting
51 2.4. THE BLACK-SCHOLES FORMULA 33 temporarily the Black-Scholes price C t+s = BS(S t+s ) where BS is the Black- Scholes formula with single argument S t, consider now the case c > 0: F Ct+s S t (c, τ) = P (C t+s < c S t ) = P (BS(S t+s ) < c S t ) = P (S t+s < BS 1 S (c; t + s) S t) = P (ln S t+s < ln BS 1 S (c; t + s) S t) ( ln BS 1 S = Φ (c; t + s) ln S t (µ 1 ) 2 σ2 )s σ s since ln S t+s N(ln S t +(µ 1 2 σ2 )s, σ 2 s). This distribution function features the inverse of the Black-Scholes function with respect to S t+s. Whilst the form for this inverse cannot be written directly, the Black-Scholes model is a monotonic increasing function of S t, and so computation of the single inverse value can be achieved using a numerical method such as the Newton-Raphson algorithm. This method is particularly useful here since the first derivative of the inverse with respect to c does have an explicit form which is easily evaluated. Differentiating this cumulative distribution function on the range c > 0 gives the form for the density function on the same range. Noting that if y = BS 1 S differentiation: (c; t + s), then y c can be found by using the chain rule for since C S = Φ(h t). So, noting c = BS(y) 1 = BS(y) y y c y c = 1 Φ ( ln y ln K+(r+ 1 2 σ2 )(τ s) σ τ s ) BS 1 S (c; t + s) c = Φ ( ln BS 1 S 1 (c;t+s) ln K+(r+ 1 2 σ2 )(τ s) σ τ s ) = 1 Φ(h t+s)
52 34 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL the derivative of F Ct+s S t (c, τ) with respect to c, with c > 0 is: f Ct+s S t (c, τ) = c F C t+s S t (c) ( 1 ln BS 1 = BS 1 S (c; t + s)σ s φ S (c; t + s) ln S t (µ 1 ) 2 σ2 )s σ 1 s ( 1 ln BS 1 = BS 1 S (c; t + s)σ sφ(h t+s) φ S (c; t + s) ln S t (µ 1 ) 2 σ2 )s σ s giving finally the density function for C t+s : 0 c 0 f Ct+s S t (c, τ) = ( 1 ln BS 1 ) (c;t+s)σ sφ(h )φ S (c;t+s) ln St (µ 1 2 σ2 )s σ c > 0 s t+s BS 1 S (2.21) Φ(h t+s) Note that in order to evaluate this density function, the Black-Scholes function must be inverted with respect to S t. This cannot be done analytically and the equation c = BS(S) must be solved for S numerically. This can be done using a numerical method such as the Newton-Raphson method. Details of this method, and its implementation in this case can be found in Appendix D Best Predictor of a Future Black-Scholes Price Although the moments of the theoretical distribution look impossible to derive using the density function above, it is possible to derive the mean price at a time t + s where 0 < s < τ. It is known that the best predictor of this future price, in a mean square sense, is given by the expected value, conditional on the past and present values of the series. In order to calculate the expected value of C T, the fact that S T given S t is a function of a standard normal random variable Z was utilised. The same procedure can be used here.
53 2.4. THE BLACK-SCHOLES FORMULA 35 Theorem 2.9. The expected value of C t+s, given S t, where s (0, τ), is given by the equation: ( ) E(C t+s S t ) = S t e µs Φ h t + (µ r)s σ τ ( Ke r(τ s) Φ h t σ τ + (µ r)s σ τ ) where h t is as in the Black-Scholes formula, and is given in Theorem 2.7. Proof. The form of S t+s, where 0 < s < τ, is given by Equation (2.5) S t+s = S t e (µ 1 2 σ2 )s+σ sz. Moreover C t+s is given by the Black-Scholes formula (2.20) and so: E(C t+s S t ) = E(S t+s Φ(h t+s ) S t ) Ke rτ E(Φ(h t+s σ τ s) S t ) (2.22) with h t+s a function of the random variable S t+s. Noting then that the Black-Scholes price C t+s is a function of S t+s which is in turn a function of Z, we conclude that C t+s is itself a function of Z and its expected value can be computed using the standard normal density function φ(z) rather than the very complicated f Ct+s S t (c, τ). I will evaluate the two expectations in the above equation separately. The first expectation of interest is E(S t+s Φ(h t+s ) S t ) = z= S t+s Φ(h t+s )φ(z)dz and can be written as two separate components, which I will simplify independently: S t+s φ(z) = S t e (µ 1 2 σ2 )s+σ sz 1 2π e 1 2 z2 = S t e µs 1 2π e 1 2 ( z+σ s) 2 = S t e µs φ(y) where y = z+σ s. Simplifying first ln S t+s, using the same transformation: ln S t+s = ln S t + (µ 1 2 σ2 )s + σ sz = ln S t + (µ σ2 )s σ sy
54 36 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL thus h t+s becomes: h t+s = ln S t+s ln K + (r σ2 )(τ s) σ τ s = ln S t + (µ σ2 )s σ sy ln K + (r σ2 )(τ s) σ τ s ( τ ln St ln K + (r + 1 ) 2 = σ2 )τ τ s σ (µ r)s s + τ σ y τ τ ( ) τ (µ r)s s = h t + τ s σ y τ τ Substituting these expressions into the first expectation on the right hand side of Equation (2.22), and making the change of variables z = y + σ s yields: where E(S t+s Φ(h t+s ) S t ) = y= S t e µs φ(y)φ h t + (µ r)s σ τ ( = S t e µs Φ 2, h t + (µ r)s ; Φ 2 (x, y; ρ) = x σ τ s τ ) τ s τ ( ) y ρz φ(z)φ dz 1 ρ 2 y s τ dy is the cumulative distribution function for a bivariate normal random variable with zero means, unit variances, and correlation coefficient ρ. This distribution function is symmetric about x and y, and so the order of integration can be changed to give an alternative distribution function: ( ) y x ρz Φ 2 (x, y; ρ) = φ(z)φ dz. 1 ρ 2 and in particular Φ 2 (, y; ρ) = y φ(z)φ( )dz = Φ(y) since Φ( ) = 1. This leads to the simplification of the first expectation: ( ) E(S t+s Φ(h t+s ) S t ) = S t e µs (µ r)s Φ h t + σ. τ
55 2.4. THE BLACK-SCHOLES FORMULA 37 Now consider the second expectation in Equation (2.22). I will again simplify the component of the expression separately, but without making the y transformation as done in the previous term: h t+s σ τ s = ln S t+s ln K + (r 1 2 σ2 )(τ s) σ τ s = ln S t + (µ 1 2 σ2 )s + σ sz ln K + (r 1 2 σ2 )(τ s) σ τ s ( τ ln St ln K + (r 1 ) 2 = σ2 )τ τ s σ (µ r)s s + τ σ + z τ τ ( τ = h t σ ) (µ r)s s τ + τ s σ + z τ τ Substituting this into the expectation and interchanging the order of integration as above gives: E(Φ(h t+s σ τ s S t ) = z= φ(z)φ h t σ τ + (µ r)s σ τ τ s τ ( = Φ 2, h t σ τ + (µ r)s σ ; τ ( = Φ h t σ ) τ + (µ r)s σ. τ s τ ) + z s τ dz Hence substituting the two expectations into Equation (2.22), we obtain the expected value of C t+s : ( ) E(C t+s S t ) = S t e µs Φ h t + (µ r)s σ τ as required. ( Ke r(τ s) Φ h t σ τ + (µ r)s σ τ ) (2.23) This formula is intriguing, since it is so similar to the Black-Scholes formula itself, but with the first S t+s replaced by E(S t+s S t ), and the correction term in the Φ terms Graphical Examination of C t Just as the series {S t } can be simulated, so too can the series {C t }. This is done using a share price time series and the Black-Scholes formula. In
56 38 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL addition to parameters selected for simulation of share price evolution, S t, τ, µ and σ, parameters K and r must also be chosen. In this case I have treated the time period over which the share is simulated as the remaining life of the option. Figures 2.4 and 2.5 show three time series, the first {S t }, and the second {C t + Ke rτ } are given by the solid and dotted lines respectively. The third, much smoother curve, is the series {Ke rτ }. The reason for showing the second series instead of simply {C t } is to place both the share and option price series in the same region of the graph, and is facilitated by the relationship given in Theorem 1.1 (C t (S t Ke rτ ) + ). The series {C t } is still evident in the graphs by taking the difference of the series {C t + Ke rτ } and {Ke rτ }. It is apparent that C t is always at least as great as S t Ke rτ as required. It is also clear that the two series are highly correlated, as expected; all movements in {S t } are mirrored by movements in the call price. However this correlation is not linear, since the Black-Scholes formula is non-linear in S t. Figure 2.4 features a share price series that has S T < K, and so the option matures out-of-the-money, and will not be exercised. This is reflected in the option price series, which converges to zero as the share price falls sharply in the latter part of the series. There is a period in the middle of the share series where S t < Ke rτ, and the option price becomes small, but with a reasonable amount of time remaining, it does not disappear altogether. At this stage there is still a good chance that the option will mature in-the-money, and will be exercised. Whilst we see similar share prices at 0.8 and 0.2 years to maturity, as the share price falls from this level, the option price decreases much more quickly in the latter case, since there is little opportunity left for the share price to rise above K. Figure 2.5 shows a share price series that generally increases with time. It is a realisation from the same process as the share price series in Figure 2.4 but there are clear differences between the two. In this case the option
57 2.4. THE BLACK-SCHOLES FORMULA 39 Price Time to Maturity - Years Figure 2.4: The unbroken series is a realisation of GBM with initial value S t = $5 and parameters µ = 0.1, σ = 0.3, and 250 subintervals; the second series is the Black-Scholes prices for this share series added to the present value of the exercise price, with K = $5, and r = 0.06 and maturity at T = 1 year; finally the smooth broken line represents the present value of the exercise price. price plus the present value of the exercise price converges on the share price series, until with approximately 0.2 years to maturity the two series are indistinguishable. In this case the share is so deep-in-the-money (S t K), the probability of the share price at maturity being below the exercise price is so small that a portfolio of the option and a cash asset Ke rτ compounding at the risk-free rate is equivalent to a portfolio containing only the share. Whilst the option price C t is a deterministic function of S t, for s > 0, C t+s will depend on a random variable S t+s and so C t+s is itself a random variable, with density function given by Equation (2.21). It is interesting to examine this density function for a fixed value of s [0, τ]. We have already seen that when τ s = 0 and the option is at maturity, the distribution of call prices is
58 40 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL Price Time to Maturity - Years Figure 2.5: The unbroken series is a realisation of GBM with initial value S t = $5 and parameters µ = 0.1, σ = 0.3, and 250 subintervals; the second series is the Black-Scholes prices for this share series added to the present value of the exercise price, with K = $5, and r = 0.06 and maturity at T = 1 year; finally the smooth broken line represents the present value of the exercise price. a mixed distribution given by Equation (2.12). Alternatively, for a fixed s, as T the value and distribution of a call option will converge to the value and distribution of the share. This is financially sensible, since as the share pays no dividends, the only difference between the option and the share is the exercise payment, and is consistent with putting τ s = in both the Black-Scholes formula at t + s and in the density function f Ct+s S t (c, τ s) given by Theorem 2.8. As the exercise date is moved further and further into the future, the present value of the exercise payment shrinks to zero. Hence, an option with infinite life has the same characteristics as the share. At time t + s, the option price distribution and its limiting properties can be examined graphically. The first series of graphs in Figure 2.7 shows
59 2.4. THE BLACK-SCHOLES FORMULA 41 the behaviour of f Ct+s S t (c, τ) as the remaining life of the option decreases to zero (s τ). The sequence of six graphs has time to maturity decreasing from two years to approximately two weeks. The histograms in the figure represent the distribution of simulated values of C t+s, with s = 1 year, and the superimposed curve, the theoretical distribution of such prices. As the time to maturity of the option decreases, the sequence of graphs shows the convergence of the density function to the mixed distribution of C T (where s = τ). We see the number of simulated option prices near zero increase, and the density function approach the continuous part of the mixture density (the truncated, translated lognormal density) and a spike at c = 0 with infinite height, and area P (S t+s < K). Again, there is good agreement between the histograms and the theoretical density functions, particularly when the value of f Ct+s S t is small. The closeness of the observed and the expected distributions stems from our ability to simulate {S t } with only sample error. Hence, a good fit in the S t+s distribution should result in a good fit to any derivative distribution. In this way, simulation provides a means of confirming that the density functions I have derived are correct, with a close fit indicating success. The density function, when τ s is small, has an interesting characteristic which is not always apparent, and does not feature for the choice of parameters used in Figure 2.7. By decreasing the volatility parameter σ, however, the effect is observed. Figure 2.6 shows the results of a new simulation of 5000 share prices, and the density function of its call prices with 5 months to maturity. The density function in this case has an interesting saddle effect, created as the probability of an option maturing out-of-the-money increases, but the probability of it expiring at-the-money is still moderate. Figure 2.8 shows a sequence of graphs with τ s increasing from 2 years to 75 years, with two curves superimposed, and can be used to demonstrate the behaviour of the call price distribution as time to maturity increases. The solid curve is the theoretical density function of the call prices, and the second
60 42 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL f(c) Figure 2.6: The distribution of call prices obtained using the Black-Scholes formula on a sample of share prices S t+s, where s = 1 year, S t = $5, µ = 0.1, σ = 0.1, K = $5, r = 0.06 and time to maturity τ s of 5 months. Also shown is the theoretical density function. c broken curve is the lognormal density function of S t+s, and the limiting form of f Ct+s S t (c, τ). The convergence of the option prices to the call prices is apparent as the sample distribution and theoretical curve move along the c axis towards the density function of the share price, until at τ s = 75 the two curves are virtually indistinguishable. 2.5 Use of the Black-Scholes Model Geometric Brownian motion is a process whose transition probability density is known, whose moments are easily determined, and which can be accurately simulated with minimal computing resources. In addition, the formula for a European call option over an asset whose value follows GBM is known, and is easily obtained, verified and evaluated. The Black-Scholes option price
61 2.5. USE OF THE BLACK-SCHOLES MODEL 43 depends on only five parameters: S t, τ, K, r and σ, of which the first three are directly observable, and the fourth is relatively easy to estimate. The model prices also have properties which are easily derived. The above analysis gives the density function for call options both at and prior to maturity. The expected value of option prices at a time in the future, up to and including maturity can also be determined. Black-Scholes prices are also easily simulated. Why then is an alternative model needed, when the GBM and Black-Scholes framework provide so many benefits? The answer to this question is that unfortunately Black-Scholes model prices do not correspond to those observed in the market. Even before the theoretical derivation was published, Black & Scholes (1972) noted that the model tends to overprice options on high variance stocks and under price options on low variance stocks. They also raise questions regarding the estimation of the final parameter σ. A series of empirical studies followed, attempting to obtain estimates of σ, which, when used in the Black-Scholes formula will more closely approximate future market option prices. These studies include those by Latané & Rendleman (1976), Chiras & Manaster (1978), and Beckers (1981). These authors, and others, discover that market option prices yield better estimates of volatility experienced over the remaining life of the option than estimates obtained using time series data of the share price. The estimates obtained from market option prices are called implied volatilities, and provide further evidence that the GBM assumption of the Black-Scholes model is not appropriate. Implied volatilities are calculated using a numerical method in much the same way as BS 1 S (c; t + s) was calculated for use in Equation (2.21). The implied volatility is the solution to the equation C mt = BS(σ; S t, τ, K, r) where C mt is the market option price at time t, and BS(σ; S t, τ, K, r) is the Black-Scholes function with single argument σ. The implied volatility ˆσ is
62 44 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL the value of σ that satisfies the equation above. It can be interpreted as an estimate of the forward average volatility discussed in Result 2.2. If the Black-Scholes GBM assumption is correct, options over the same underlying asset, with the same maturity date, should yield identical implied volatilities. In fact if share price volatility σ is indeed constant, any option over a particular share should yield the same implied volatility. In practice this is not observed, and a smile relationship is seen between strike price K and the Black-Scholes implied volatility, and a similar relationship between time to maturity τ and implied volatility. Based on these phenomena, it appears that an option pricing model which incorporates changing volatility is required. Such a model is considered in the following chapter.
63 2.5. USE OF THE BLACK-SCHOLES MODEL Months to Maturity 18 Months to Maturity f(c) f(c) c 12 Months to Maturity c 6 Months to Maturity f(c) f(c) c 3 Months to Maturity c 0.5 Months to Maturity f(c) f(c) c c Figure 2.7: The distribution of call prices obtained using the Black-Scholes formula on a sample of share prices S t+s, where s = 1 year, S t = $5, µ = 0.1, σ = 0.1, K = $5, r = 0.06 and time to maturity τ s ranges from 2 years to half a month. Also the theoretical density function for these prices.
64 46 CHAPTER 2. GBM AND THE BLACK-SCHOLES MODEL 2 Years to Maturity 5 Years to Maturity f(c) f(c) c 15 Years to Maturity c 30 Years to Maturity f(c) f(c) c 50 Years to Maturity c 75 Years to Maturity f(c) f(c) c c Figure 2.8: The distribution of call prices obtained using the Black-Scholes formula on a sample of share prices S t+s, where s = 1 year, S t = $5, µ = 0.1, σ = 0.1, K = $5, r = 0.06 and time to maturity τ s ranges from 2 to 75 years. Also the (solid) theoretical density function for these prices, and the (dotted) lognormal density function of S t+s.
65 Chapter 3 The CEV Model 47
66 48 CHAPTER 3. THE CEV MODEL 3.1 Introduction Empirical analysis of observed option prices, and observed phenomena like the volatility smile, led researchers to formulate option pricing models that included non-constant volatility. In response to observations of an inverse relationship between share price and share price volatility, documented by Fischer Black (1975), the Constant Elasticity of Variance (CEV) model was derived. Cox (1996) suggests that the CEV model is the simplest way to describe such an inverse relationship. In fact, the CEV model was derived after a direct request from Fischer Black to John Cox, for a share price evolution model that includes an inverse dependence of volatility and the share price, as described in Cox (1996). The standard CEV model assumes that share price S t evolves according to the stochastic differential equation: ds t = µs t dt + δs β 2 t db t (t > 0, β < 2) (3.1) where µ and δ are constants, and with initial condition B 0 = 0. Unlike GBM, a solution to Equation (3.1) can become negative unless otherwise constrained. This is clearly inappropriate for an asset price, so an absorbing barrier must be imposed at zero, such that if S t = 0 then S u = 0 for all u > t. Note that a reflecting barrier at zero is not sufficient, since once a firm s share price falls to zero, the firm will be bankrupt, and the share price cannot become positive again. The absorbing barrier, however, overcomes this problem and is consistent with bankruptcy of the firm. As mentioned above, the stochastic process for S t assumed by Black and Scholes has β = 2 in Equation (3.1). Emanuel & MacBeth (1982) consider the above process when β > 2. Hence, the option price over an underlying asset whose price is a solution to the general SDE introduced in Definition 1.8 is known for any value of β. The case when β > 2 is mentioned briefly in Result 3.5 and estimated in Chapter 4, but this chapter focuses primarily on the case β < 2.
67 3.1. INTRODUCTION 49 The variance of the stock s instantaneous rate of return, given data to time t is: ( ) dst Var S t S t = δ 2 S β 2 t dt (3.2) which, in the Black-Scholes case, is constant over time. In general, when β < 2, the variance of the instantaneous rate of return decreases as stock price increases, thus loosely satisfying empirical observations. The rate of return on a firm s equity, in the absence of debt, R a, can be written in the presence of debt, as a weighted average of the rates of return on its debt and equity, R d and R e respectively: R a = S V R e + B V R d(1 α) where S is the present value of equity, B the present value of debt, V = S +B the present value of the firm, and α reflects the tax scenario 1. Rearranging the relationship for R e, and taking the variance of both sides, with random variables R e and R a, and all other quantities assumed constant, yields: ( ) dst Var(R e ) = Var S t S t ( = Var(R a ) 1 + B ) 2. (3.3) S This is the desired relationship between share price and the variance of the stock s rate of return, since as S t increases relative to B t, the variance of its rate of return (the constant volatility parameter in the Black-Scholes case) will decrease, and vice versa, provided Var(R a ) remains fixed. as S t Moreover, 0, under the CEV model, the share price volatility will tend to infinity, as Equation (3.3) dictates. However, as S t, in such a way that Bt S t 0, the CEV model indicates that share price volatility should become zero, whilst financial theory says that it should converge to a positive constant Var(R a ), where this quantity remains fixed. This fact is reflected in 1 In a Modigliani and Miller tax world α = T c, the corporate tax rate, but in a Miller world, α = 0. For further details, see Copeland & Weston (1988) and the references therein.
68 50 CHAPTER 3. THE CEV MODEL the compound option pricing model derived by Geske (1979) but not in the CEV model, since under the CEV model, as share price goes to infinity, the volatility of the share price will go to zero. Therefore whilst the CEV model is theoretically motivated by the relationship between the firm s return on equity and the value of equity outlined above, it does not fully describe this relationship. The CEV model acquired its name from the fact that the elasticity of variance of the rate of return on S t is constant. This means that the ratio of any proportional change in S t and the resulting proportional change in the variance of the rate of return on S t will be constant. The proportional change in S t over a period dt will be simply dst S t and the proportional change in the variance of the rate of return on S t over the same period dt, conditional on S t, will be dvar(ds t /S t ) Var(dS t /S t ). Taking the ratio of these two quantities, and noting the variance of the rate of return on S t given by Equation (3.2), yields the desired relationship: / dvar(ds t /S t ) dst = dvar(ds / t/s t ) Var(dSt /S t ) Var(dS t /S t ) S t ds t S t = (β 2)δ 2 S β 3 t dt = β 2. where again, all variances are conditional on S t. / δ 2 S β 2 dt Hence, a 10% increase in S t will result in an increase of 10%(β 2) in the variance of the rate of return on S t. Note that since β < 2 this will actually be a decrease in the volatility as required. 3.2 The CEV Share Price Solution Unlike the Black-Scholes case, for the general CEV model it is not possible to derive a form for S T similar to the one we saw in Equation (2.5). According t S t
69 3.2. THE CEV SHARE PRICE SOLUTION 51 to Cox & Ross (1976), the CEV option pricing formula can be derived using the same technique that Black and Scholes used to obtain their formula. This involves taking a transformation of the process C t, chosen so that its governing PDE and boundary condition have a known solution. Both the Black-Scholes PDE and the appropriate PDE for the CEV option price can be transformed into the famous heat equation of physics, and solved using standard mathematical techniques. Cox and Ross used the CEV and other models as examples to pioneer a pricing method called risk-neutral pricing alluded to in Section 2.4. In order to price a future payment, P T, whose solution satisfies a PDE independent of investor risk preferences, it may be assumed that the underlying process has instantaneous mean r, and all future payments may be discounted at r. Hence the value of P T now will be P t = e rτ E (P T F t ) where the expectation E is a risk-neutral expectation, taken under the assumption that the underlying process has instantaneous mean r, and F t is available information up to and including t. In the case of an option price, the future payment we are interested in is P T = (S T K) +, where S T is the underlying stochastic process. The implications of Cox and Ross result are that if we can find a density function that describes the evolution of P T between t and T, or in this particular case, of S t to S T, then we may be able to evaluate the (risk-neutral) expected value above. The Kolmogorov Equations can be used in this context. They are a set of two partial differential equations which describe the transition probabilities of a Markov diffusion process. The lognormal density function given in Equation (2.14) is an example of a transition probability density for a process following geometric Brownian motion. Cox & Miller (1965) define the Kolmogorov equations on page 215, which describe properties of a continuous time, continuous variable stochastic process X(t). These equations are defined in the following result.
70 52 CHAPTER 3. THE CEV MODEL Result 3.1 (The Kolmogorov Equations). Suppose the continuous time, continuous variable process X(t) is a solution of the SDE: dx(t) = β(x, t)dt + α(x, t)db(t) with initial condition X(t 0 ) = x 0. Then X(t) has instantaneous mean β(x, t) and instantaneous variance α(x, t). In addition, X(t) has a transition probability density p(x 0, t 0 ; x, t) which satisfies the Kolmogorov equations, and whose interpretation is given by the following equation: P (a < X(t) < b X(t 0 ) = x 0 ) = b The Kolmogorov equations are the forward equation: a p(x 0, t 0 ; x, t)dx. 1 2 p {α(x, t)p} {β(x, t)p} = 2 x2 x t (3.4) in which the backward variables x 0 and t 0 may be considered constant and enter by way of the boundary conditions, and the backward equation: 1 2 α(x 0, t 0 ) 2 p + β(x x 2 0, t 0 ) p = p (3.5) 0 x 0 t 0 in which the forward variables x and t enter only through boundary conditions. In the CEV case we are again interested in the function f ST S t (s, τ), which we may use to examine the properties of the future share price S T, given S t. It appears that this probability density function might be the solution to the forward and backward Kolmogorov equations. S t is a continuous time, continuous variable process, which satisfies the equation ds t = µs t dt + δs β 2 t db t with initial condition B 0 = 0. The process S t has instantaneous mean and variance E(dS t S t ) = µs t and Var(dS t S t ) = δ 2 S β t to find its transition probability density f ST S t (s, τ) = p(s t, t; s, T ) respectively, and we intend
71 3.2. THE CEV SHARE PRICE SOLUTION 53 where S t = s t is given. One point of concern is the fact that the process S t has an absorbing barrier at zero, imposed in order to ensure that the share price cannot become negative. Cox and Miller consider this case on page 219, and state that an additional boundary condition must be imposed when solving each of the Kolmogorov equations. In particular, for an absorbing barrier at a, the additional boundary condition for the forward equation is: where x < a < x 0, yielding for s < 0: p(x 0, t 0 ; x, t) = 0 f ST S t (s, τ) = 0. It follows from Cox and Miller s condition, that if x a: P (X(t) < x X(t 0 ) = x 0 ) = x p(x 0, t 0 ; u, t)du = 0 since for x < a < x 0, the density function p is identically zero. Thus in order to determine the density function for S T = s given S t = s t we must solve the backward and forward Kolmogorov equations given by: and s 2 {δ2 s β t f} s {µs tf} f T = 0 (3.6) 1 2 δ2 s β 2 f t s 2 t respectively, subject to the boundary conditions: + µs t f s t + f t = 0 (3.7) f ST S t (s, τ) = 0 if s < 0 f ST S t (s, 0 S t = s t ) = δ(s s t ), where δ(x) is the Dirac delta function. The first of these conditions follows from the absorbing barrier, and the second from the fact that f is a transitional density function.
72 54 CHAPTER 3. THE CEV MODEL Solution of the Forward Kolmogorov Equation Feller (1951) solves the class of partial differential equations of the form: u t = (axu) xx ((bx + c)u) x (3.8) where u = u(t, x), and a, b, c are constants, and a > 0. This equation is equivalent to the Fokker-Planck, or forward Kolmogorov equation seen in Equation (3.4), with α(x, t) = 2ax and β(x, t) = bx + c. In order to apply Feller s results to find the CEV solution, a transformation must be made to the process {S t } to obtain a partial differential equation of the required form, and the instantaneous mean and variance of this new process obtained. The transformation made by Cox (1996), and by Emanuel & MacBeth (1982), is Y = S 2 β, and is applicable for all β 2. We can use Itô s Lemma to obtain the instantaneous mean and variance of the new process Y t. By Itô s Lemma, since and Y t = f(s t ) = S 2 β t, dy t = f S ds t δ2 S β t ds t = µs t dt + δs β 2 t db t 2 f S 2 dt = (2 β)s 1 β t (µs t dt + δs β 2 t db t ) δ2 S β t (2 β)(1 β)s β t dt = ( µ(2 β)y t δ2 (2 β)(1 β) ) dt + δ(2 β) Y t db t. Thus, the process Y t has instantaneous mean: and variance: β(y t ) = µ(2 β)y t δ2 (2 β)(1 β) α(y t ) = δ 2 (2 β) 2 Y t. The transition density f = f YT Y t (y, τ) will satisfy the forward Kolmogorov equation: 1 α(y 2 t) 2 f 2 Y + β(y t) f = f t Y t t (3.9)
73 3.2. THE CEV SHARE PRICE SOLUTION 55 which is of the form specified by Feller when β < 2, with a = 1 2 δ2 (2 β) 2 b = µ(2 β) c = 1 2 δ2 (2 β)(1 β) and can be used in his result to obtain the transition density function of Y t. This function is given by Feller in his Lemma 9, as: f YT Y t (y, τ) = b a(e bτ 1) exp ( ) ( ) b(y + Yt e bτ c a ) ye bτ 2a I1 a(e bτ c a 1) Y t ( ) 2b(e bτ yy t ) 1 2 a(1 e bτ ) (3.10) where a, b and c are given above, and I ν (z) is the modified Bessel function of the first kind of order ν. In order to obtain the transition probability density for S T conditional on S t, we must again make a change of variables: F ST S t (s, τ) = P (S T < s S t ) = P (S 2 β T < s 2 β S 2 β t ) = P (Y T < s 2 β Y t ). Differentiating with respect to s we obtain the density function: f ST S t (s, τ) = s F S T S t (s, τ) = (2 β)s 1 β f YT Y t ( s 2 β, τ ). Substituting into this final equation the constants a, b and c, and making the replacements y = s 2 β and Y t = S 2 β t Feller, we obtain the density function for S T given S t : where k = f ST S t (s, τ) = (2 β) k 1 2 β ( x z 1 2β ) 1 2 2µ δ 2 (2 β)(e µ(2 β)τ 1) x = and s > 0. This density function of S T 1 in the density function given by 2 β e x z I 1 2 β (2( x z) 1 2 ) (3.11) ks 2 β t e µ(2 β)τ z = ks 2 β (3.12) conditional on S t, is a function of S t and the time elapsed, τ, between t and T, but also depends on CEV
74 56 CHAPTER 3. THE CEV MODEL parameters µ, δ and β. It describes the statistical properties at T, of a general process satisfying Equation (3.1), where β < 2, given its value at t. Abramowitz & Stegun (1968) give the power series expansion of the modified Bessel function seen in the density function above in their Equation (9.6.10) as: I ν (z) = ( 1 2 z)ν n=0 ( 1 4 z2 ) n n!γ(ν + n + 1). (3.13) Using this identity, the Bessel function in the density function above may be rewritten: ( ) I 1 (2( x z) 1 2 ) = 1 (2( x z) β 2 2 β 2 ) = ( x z) β n=0 n=0 ( 1 4 (2( x z) 1 2 ) 2 ) n n!γ(n β ) ( x z) n n!γ(n β ) giving the alternative form for the transition density function: f ST S t (s, τ) = (2 β) k 1 2 β ( x z 1 β ) 1 2 β e x z ( x z) n n!γ(n ) 2 β = (2 β) k 1 2 β e x z Probability of Bankruptcy n=0 n=0 The density function given above is valid only for S T x n+ 1 2 β z n+ 1 β 2 β n!γ(n β ) (3.14) > 0. The absorbing barrier imposed at S u = 0 where t < u T prevents the share price from becoming negative, and results in a positive probability of bankruptcy under the CEV model. The probability of bankruptcy given below is well known, but is often stated with the unnecessary assumption µ = r. Result 3.2. G(x, ν 1) = 1 g(y, n + ν) where G(y, ν) and g(y, ν) are the survivor and probability density functions at y for a gamma random variable with shape parameter ν, and unit scale parameter. n=0
75 3.2. THE CEV SHARE PRICE SOLUTION 57 Proof. A proof of this result is given in Appendix B.2. Theorem 3.1 (Probability of Bankruptcy for CEV Stocks). The probability of bankruptcy under the CEV model, with β < 2, will be G( x, 1 2 β ) where G(y, ν) is the survivor function at y for a gamma random variable with shape parameter ν and unit scale parameter, and x = are as previously defined. ks β 2 t e µ(2 β)τ and k = 2µ δ 2 (2 β)(e µ(2 β)τ 1) Proof. First evaluate f ST S t (s, τ)ds. Noting that z = ks 2 β, we find: ( ) 1/(2 β) s = z k ds = k 1/(2 β) 1 2 β z 1+1/(2 β) d z. Making this transformation, and combining with the alternate form for the density function in Equation (3.14) yields: f ST S t (s, τ)ds = (2 β) k 1 2 β e x z = e x z n=0 n=0 x n+ 1 2 β z n+ 1 β 2 β n!γ(n ) 2 β ( k 1 2 β z 1 β 2 β 2 β x n+ 1 2 β z n n!γ(n (3.15) )d z 2 β In addition since β < 2, z is a 1-1, increasing function of S T and when S T = 0, z = 0. Hence: P (S T = 0) = 1 = 1 = 1 = 1 f ST S t (s, τ)ds s=0 e x z z=0 n=0 1 n+ x x 2 β n=0 n=0 e Γ(n ) 2 β g( x, n β ) x n+ 1 2 β z n n!γ(n β )d z e z z n d z n! d z )
76 58 CHAPTER 3. THE CEV MODEL since g( z, ν) = e z z ν 1 Γ(ν) is the gamma probability density function on z [0, ). where Finally, using Result 3.2 we can determine the probability of bankruptcy: P (S T = 0) = 1 n=0 g( x, n β ) = G( x, 1 2 β ) (3.16) G(y, ν) = y e x x ν 1 Γ(ν) dx is the survivor function for the gamma random variable with shape parameter ν and unit scale parameter. Hence, the probability of bankruptcy between now and T > t, for a firm whose share price is a solution to the SDE (3.1), depends on the parameters µ, δ and β, the initial share price S t, and on the length of the time period [t, T ] in question. Analysis of the relationship between β and the probability of bankruptcy indicates that for a fixed T, P (S T = 0) is insensitive to changes in share price at t, but sensitive to changes in the size of µ and δ. As we would expect, an increase in µ decreases the probability of bankruptcy for any β while an increase in volatility (via the parameter δ) increases the probability of bankruptcy. Figure 3.1 shows the relationship between β and P (S T = 0), with S t = $5, τ = 1 year, µ = 0.10, σ = δs β/2 1 t = 0.3 in the solid curve and σ = 0.28 in the broken curve. Making small changes in δ result in very large movements in the curve, as seen in the graph. Here a 6.67% decrease in σ (from 0.3 to 0.28), via the parameter δ, results in a 29.64% decrease in P (S T = 0) when β = 2, and larger percentage decreases for larger values of β, with a 97.10% decrease in P (S T = 0) when β = 1. Note that since k = when β = 2, the probability of bankruptcy given by Equation (3.16) generalises to the Black-Scholes case where P (S T = 0) =
77 3.2. THE CEV SHARE PRICE SOLUTION 59 P(bankruptcy) Figure 3.1: The relationship between β and P (S T = 0), with S t = $5, τ = 1 year, µ = 0.10, σ = δs β/2 1 t = 0.3 in the solid curve and σ = 0.28 in the broken curve. beta 0. This result is given by Markov s Inequality: P (X x) E(X) x. 1 Letting X be a gamma random variable, with shape parameter, and unit 2 β scale parameter, then: 1 G( x, ) = P (X x) 2 β E(X) x = 1 2 β / = δ2 (e µ(2 β)τ 1) S 2 β t = 0 2µS 2 β t δ 2 (2 β)(e µ(2 β)τ 1) 1 when β = 2, and so lim β 2 G( x, ) = 0 as required. 2 β
78 60 CHAPTER 3. THE CEV MODEL Since the probability of bankruptcy is positive for β < 2, the mixed distribution of S T given S t can be written as a linear combination of a step function, and a continuous distribution function: ( ) ( ( )) 1 1 s f G x, + 1 G x, ST S t (u,τ)du 2 β 2 β 0 1 s 0 1 G( x, F ST S t (s, τ) = 2 β ) 0 s < 0 Theorem 3.2 (Mean and Variance of S T (3.17) given S t ). The mean and variance of the CEV share price S T conditional on an earlier share price S t are E(S T S t ) = S t e µτ and { } S t e µτ k β Γ(n+1+ 2 β n=0 g( x, n + 1) ) Var(S T S t ) = S Γ(n+1+ te µτ β < β ) S t µ 1 e µτ δ 2 (e µτ 1) β = 1 respectively. Proof. The expected value of S T given S t is given by: E(S T S t ) = s=0 sf ST S t (s, τ)ds. The form of f ST S t (s, τ)ds is given in Equation (3.15) and s = ( z/ k) 1/(2 β), so we can write: ( ) 1 2 β E(S T S t ) = e x z z=0 z k n=0 ( ) 1 2 β e = x k x x n n! = S t e µτ n=0 0 x n+ 1 2 β z n n!γ(n β )d z e z z n+ 1 2 β Γ(n β )d z where the final equality follows from the definition of x, and properties of the gamma probability density function.
79 3.2. THE CEV SHARE PRICE SOLUTION 61 The expected value of S 2 T given S t is found in much the same way: ( ) 2 E(ST 2 2 β S t ) = e x z x z=0 z k n+ 1 2 β z n n!γ(n )d z n=0 2 β ( ) 1 2 β = k x k 1 e x x n e z z n+ 2 2 β 2 β n! n=0 0 Γ(n )d z 2 β = S t e µτ k 1 2 β g( x, n + 1) Γ(n ) 2 β Γ(n ) 2 β n=0 which yields a formula for the variance: { Var(S T S t, β) = S t e µτ k 1 2 β g( x, n + 1) Γ(n ) } 2 β Γ(n ) S te µτ. 2 β n=0 (3.18) The mean square above can be simplified in the Square Root CEV case, where β = 1, to yield a variance formula that does not feature an infinite summation. Setting β = 1, E(S 2 T S t) can be simplified to: E(ST 2 µτ S t, β = 1) = S t e k 1 µτ = S t e k 1 µτ = S t e k 1 Γ(n + 3) g( x, n + 1) Γ(n + 2) g( x, n + 1)(n + 2) n=0 n=0 { x n=1 = S t e µτ k 1 ( x + 2) = St 2 e 2µτ µτ + 2S t e k 1 since when β = 1, x/ k = S t e µτ. Hence: n 1 e x x (n 1)! + 2 n=0 Var(S T S t, β = 1) = E(S 2 T S t, β = 1) E(S T S t ) 2 = S 2 t e 2µτ + 2S t e µτ k 1 S 2 t e 2µτ = 2S t e µτ δ2 (e µτ 1) 2µ = S te µτ δ 2 (e µτ 1) µ e x x n n! }
80 62 CHAPTER 3. THE CEV MODEL since when β = 1. k = 2µ δ 2 (2 β)(e µ(2 β)τ 1) = 2µ δ 2 (e µτ 1) It is interesting that the mean of all CEV share prices S T given S t is the same, regardless of β. The mean also corresponds to the expected value under GBM, however the formula for the variance is not as attractive. The relationship between Var(S T S t ) and β can be seen in Figure 3.6 later in this section Simulation of CEV Share Prices In the case of GBM, where the share price S T could be written as a function of a standard normal variable, using SPLUS we could generate {S t } directly. Whilst in some cases it is possible to generate random variables with nonstandard density functions using a uniform random variable on the interval [0, 1], in this case the non-linearity of the density function in the argument s makes this technology difficult to apply. Thus, in order to simulate S T, or in particular the time series {S t }, it will be necessary to compute an approximate solution of the continuous time SDE (3.1). An appropriate method is the Euler scheme, described in Kloeden & Platen (1992) and Mikosch (1994). This method forms a discrete approximation of the continuous time stochastic differential equation, writing: S ti = S ti 1 + µs ti 1 (t i t i 1 ) + δs β 2 ti 1 (B ti B ti 1 ) (3.19) where the period of interest [t, T ], is partitioned into n subintervals: t = t 0 < t 1 < < t i < < t n 1 < t n = T not necessarily of equal length.
81 3.2. THE CEV SHARE PRICE SOLUTION 63 The solution can be simulated by appealing to the independent increment property of Brownian motion: B ti B ti 1 B ti t i 1 B 0 = B ti t i 1 N(0, t i t i 1 ) since B 0 = 0 and B t N(0, t). Hence fixing S t, and generating n increments from a normal distribution with zero mean and unit variance, the Brownian motion increments can be obtained: B ti B ti 1 = t i t i 1 Z i where the Z i are independent N(0, 1) variables. These increments can then be used to construct the series {S t } using the approximation in Equation (3.19). Thus a single share price is given by: S ti = S t + = S t + i (S tj S tj 1 ) j=1 i j=1 µs tj 1 (t j t j 1 ) + δs β 2 tj 1 ti t i 1 Z i for i = 1, 2,..., n. Therefore, it is clear that using this Euler scheme, an individual share price at T > t can be calculated only by first calculating all previous share prices. Recall also that the Euler method discretises a continuous process, and so the approximation will be best when t i t i 1 is very small for all i. These two facts introduce computational difficulties, for the following reasons: to improve the quality of the approximation, n, the number of subintervals, should be made large, so that the time increments are small; however, if n is large, many increments must be calculated in order to compute elements of the series {S ti }, when i is large. Thus increasing n decreases the speed of the computation.
82 64 CHAPTER 3. THE CEV MODEL Having fixed the number of subintervals, the Euler scheme provides the discrete time series {S ti }. The continuous time process is then approximated by linear interpolation of the points (t i 1, S ti 1 ) and (t i, S ti ), for i = 1, 2,..., n. The function that I have used to simulate CEV share prices can be seen in Appendix D.3. A parameter of this function is the volatility of the share price, σ, which assuming Black-Scholes evolution, is equal to the standard deviation of the rate of return on S t. When the volatility is estimated empirically based on historical share price data, it is usually computed using the formula ˆσ 2 = 1 n i=1 (r i r) 2 t n 1 where r i = ln S ti+1 ln S ti, r = 1 ri, and t = t n i+1 t i for all i, as given in Hull (1997) on page 233. This is simply the annualised variance of the daily returns on the share. Estimating volatility in this way assumes that the share price is following geometric Brownian motion, or at least that the volatility is independent of share price, since any intra-data relationship is lost by taking a global average. In the CEV case volatility is a function of share price as seen in Equation (3.2). This relationship can be used to translate an empirically sensible volatility into a value of δ according to the share price s β. Both Beckers (1980) and MacBeth & Merville (1980) compare CEV and Black-Scholes option prices for various share price processes. In order to make valid comparisons, these processes must be aligned in some way so that the differences in the share price properties (and hence in the option price properties) are due only to the form of the variance, rather than that the processes are intrinsically different in nature. If realisations of the processes did not look similar, we could not expect option prices based on these processes to be comparable. Since there are systematic differences between CEV processes with different β, any alignment process is restricted to a single point in time. Beckers chooses to align the share price processes at S T,
83 3.2. THE CEV SHARE PRICE SOLUTION 65 by choosing δ so that S T has the same variance regardless of β. MacBeth and Merville align the processes at the beginning of the simulation period. More specifically, MacBeth & Merville align different processes by setting the variance of the rate of return of each process to be the same at the beginning of the simulation period. GBM is the solution to the SDE: ds t S t = µdt + σdb t. Using MacBeth and Merville s alignment procedure, a comparable CEV process would satisfy the SDE: ds t S t = µdt + σs 1 β 2 S β 2 1 t db t t 0 so that for t = t 0 both processes would have the same SDE, and in particular the rate of return on each process would have the same initial variance. Hence MacBeth and Merville obtain δ using the equation: δ = σs 1 β 2 t 0. (3.20) Beckers adopts a different approach. Instead of aligning the processes at the beginning of the simulation period, Beckers determines the appropriate value for δ, by ensuring Var(S T S t ) is the same for each process considered. For general β, this is a lot more difficult than MacBeth and Merville s approach, and requires numerical solution. Whilst E(S T S t ) is easy to compute as seen in the proof of Theorem 3.2, we also saw that the variance is not easy to calculate in general. Compared to Beckers method, MacBeth and Merville s method is certainly easier to implement for any β value. Although Beckers approach might be more appropriate, given our interest in S T for the purposes of valuing options, this alignment requires numerical solution for β 1, and has introduced complications that are perhaps unnecessary. On this basis, I have elected to use the alignment procedure of MacBeth & Merville. This is evident in the program cevs.f seen in Appendix D.3 that I use to simulate
84 66 CHAPTER 3. THE CEV MODEL share price. In this function, I call the volatility σ as a parameter, and then use the relationship in Equation (3.20) to determine δ for use in the Euler Equation (3.19) CEV Share Price Series For the purpose of pricing European options, analysis of share price evolution can be restricted to the distribution of possible share price values at the exercise date of the option. This distribution, and the risk-neutral pricing technology of Cox and Ross, enables the estimation of the present value of the option exercise payoff. In order to obtain the distribution of a future share price, either an appropriate form of the Kolmogorov Equations must be solved, or the distribution estimated using simulation. In order to do either of these things, a decision must be made about how to model the evolution of share price based on properties of the past and present share price series. Hence examining time series of CEV share prices will be beneficial in the option pricing setting in that it will indicate features of an appropriate model, which may then be applied to option pricing. Figures 3.2 and 3.3 contain time series {S t } for CEV processes with S t = $5, τ = 1 year, µ = 0.1, σ = 0.3, n = 250 subintervals, and the same Brownian motion increments, {B t }. The processes have β values of 2 (GBM) and -1 respectively. The figures each show three graphs, the first of which is simply the share price series. Since the series have been generated using the same Brownian motion increments, we would expect them to be similar, and in this case it is very difficult to see any difference at all. Comparison of the final price in each series shows that the CEV process has fallen further than the GBM realisation, but aside from this, differences are difficult to detect. The second graph in each figure shows the daily return series {R t }, where R t = ln S t+ t ln S t
85 3.2. THE CEV SHARE PRICE SOLUTION 67 for each process 2, where for daily returns, t = 1 years is approximately 250 one trading day. Superimposed on the daily returns is an estimate of the mean level of the daily returns, calculated using a Lowess filter with a smoothing window of 30 days. Since the filter is based on a symmetric window about the individual points, as before, 15 estimates have been removed from either end so that all estimates of the mean level are based on actual daily returns and not back- or forecasted values. The estimated standard deviation of the daily returns has been calculated using the Lowess filter and the estimate of the mean for mean-correction, and has been restricted in the same way. Lines which are two standard deviations from the mean have superimposed on the daily return series. Whilst in the CEV case the standard deviation of the daily returns is a function of the share price level, in the Black-Scholes case it is a constant. This suggests that the mean ± two standard deviation limits in the second graph in Figure 3.2 should be approximately equidistant over the entire series {R t }, since the series is GBM. Conversely, a CEV process with parameter β = 1 has an inverse relationship between share price level and volatility, and so we should see a widening of the bounds when share price is low, and an opposite effect when the share price is high. Due to the way I have aligned the processes, the classification of the share price as low or high depends on the initial share price S t. That this should be so, follows by examining the volatility for the general process, with t < u < T : ( ) dsu Var S u = δ 2 Su β 2 du using S u = σ 2 S 2 β t = σ 2 ( St S u Su β 2 du ) 2 β du 2 In general, when dividends are paid on the stock, the daily returns should be calculated where d t is the dividend paid at t. R t = ln(s t+ t + d t ) ln S t
86 68 CHAPTER 3. THE CEV MODEL and since β < 2, the relationship between share price and the variance of its instantaneous rate of return is an inverse one. Furthermore, when S u < S t, the volatility will be greater than σ, and when S u > S t, the volatility will be less than σ. Differences between the daily return series for the CEV process in Figure 3.3 and the same series for the GBM process in Figure 3.2 are easier to see than differences between the raw share price series. In particular the series are very different near the end of the graph, which, on close inspection, is one region of the share price series where a difference is apparent. The third graph in each respective figure does show clearly the difference between the two processes. In the case of the geometric Brownian motion, there does not seem to be any clear relationship between share price level and the standard deviation of the daily returns. The standard deviation should be constant, a fact which is not entirely supported by the graph. This is not overly alarming since we are dealing with a single realisation of a process, and examining a property that holds in continuous time using non-parametric smoothing technology. Comparison of this scatterplot to the third graph of the CEV process shows highly evident differences. Here we see a decreasing relationship between share price level, and the standard deviation of its daily returns. The relationship does not appear to be linear, but it is certainly decreasing, in confirmation of what we know to be true for a CEV process with β < 2.
87 3.2. THE CEV SHARE PRICE SOLUTION 69 Share Price Time - years Log-returns Time - years Log-return Standard Deviation Share Price Figure 3.2: Firstly, a typical realisation of GBM, with S t = $5, τ = 1 year, µ = 0.10, σ = 0.3 and n = 250 subintervals; secondly, the daily returns for this series with estimated mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against the estimated standard deviation of the daily returns.
88 70 CHAPTER 3. THE CEV MODEL Share Price Time - years Log-returns Time - years Log-return Standard Deviation Share Price Figure 3.3: Firstly, a typical realisation of share price, with S t = $5, τ = 1 year, µ = 0.10, σ = 0.3 and n = 250 subintervals and CEV parameter β = 1; secondly, the daily returns for this series with estimated mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against the estimated standard deviation of the daily returns.
89 3.2. THE CEV SHARE PRICE SOLUTION Graphical Examination of S T The theoretical distribution for CEV share prices at T, given an initial starting value at t, is given by Equation (3.11), with an alternative form in Equation (3.14). Both these (improper) density functions present problems in terms of computation. In the first case, the Bessel function I ν (z) is not internally computed in SPLUS, and in the second, there is an infinite summation 3. An alternative is due to Schroder (1989), who discusses computation of the CEV option price. Whilst not the focus of his paper, Schroder represents the transition survivor function of S T given S t in terms of a non-central Chisquared random variable. Johnson, Kotz & Balakrishnan (1995) introduce the non-central Chi-squared distribution on page 433 as in the following result. Result 3.3 (The Non-central χ 2 Distribution). If U 1, U 2,..., U ν are independent unit normal random variables, and δ 1, δ 2,..., δ ν then: are constants ν (U j + δ j ) 2 (3.21) j=1 has a non-central χ 2 distribution, with ν degrees of freedom, and non-centrality parameter: λ = ν δj 2. j=1 The density function of such a variable is given as: p(y; ν, λ) = exp( 1 (y + λ)) ν j=0 (y) 1 2 ν+j 1 λ j Γ( 1 2 ν + j)22j j!. (3.22) 3 It does appear that the terms in the summation should converge very quickly to zero, due to the presence of both n! and Γ(ν + n + 1) in the denominator. Together these terms are like n! 2 and will quickly dominate the term in the numerator, suggesting that the summation should converge after few terms.
90 72 CHAPTER 3. THE CEV MODEL Given the definition of Johnson et al., it appears ν Z + is required, however the moment generating function for a non-central Chi-squared variable is defined for all ν > 0. This moment generating function: ( ) M X (t) = (1 2t) 1 λt 2 ν exp 1 2t uniquely defines the non-central Chi-squared variable, and so non-integer degrees of freedom are permitted. This is noted by Johnson et al. Note also that λ must be positive. Using the series expansion of the modified Bessel function shown in Equation (3.13) the density function p(y; ν, λ) can be manipulated to give: p(y; ν, λ) = 1 2 (y/λ) 1 4 (ν 2) exp { 1 2 (λ + y)} I 1 2 (ν 2)( λy) (3.23) which is the same as that given by Johnson et al. in their Equation (29.4). Beginning with Equation (3.23), the equivalence of the two forms of the density function is demonstrated below. p(y; ν, λ) = 1 2 (y/λ) 1 4 (ν 2) e 1 2 (λ+y) I 1 2 (ν 2) ( λy) = 1 2 (y/λ) 1 4 (ν 2) e 1 2 (λ+y) ( 1 2 = ( 1 2 ) 1 2 ν y 1 2 (ν 2) e 1 2 (λ+y) j=0 λy) 1 2 (ν 2) j=0 λ j y j 2 2j j!γ( 1 2 ν + j) The last equation is of course the same as Equation (3.22). ( 1 4 ( λy) 2 ) j j!γ( 1 (ν 2) + j + 1) 2 Theorem 3.3 (The CEV Share Price Distribution Function). The transition distribution function for a CEV share price S T given S t, F ST S t (s, τ) = P (S T < s S t ) is given by the equation: 0 s < 0 1 F ST S t (s, τ) = G( x, 2 β ) s = 0 2 Q(2 x;, 2ỹ) s > 0 2 β
91 3.2. THE CEV SHARE PRICE SOLUTION 73 where ỹ = ks 2 β, k and x are defined in Equation (3.12), and Q(x; ν, λ) is the survivor function at x for a non-central Chi-squared variable with ν degrees of freedom and non-centrality parameter λ. Proof. Note firstly that S T 0, so that for s = 0, P (S T < s S t ) = 0. 1 Secondly, P (S T = 0 S t ) is given by Theorem 3.1, and is G( x, ) as required. 2 β I will defer the proof for the final case, where s > 0, until Section 3.5.2, since the proof is made considerably easier by results which follow. Thanks to Schroder s result, and the fact that the non-central Chi-squared cumulative probability function is an internal function in SPLUS, we can again compare the theoretical distribution for share price at a future time T, to an empirical distribution of simulated values. Using the Euler method, I have simulated 5000 share prices τ = 1 year into the future, with β = 1, and additional parameters S t = $5, µ = 0.1, σ = 0.3 and n = 250 subintervals. Of these simulated series, 38 had final values S T = 0, giving an estimate of 38 P (S T = 0) of = This is less than the true probability of , 5000 corresponding to an expected number of bankruptcies of A histogram of the remaining 4962 values is shown in Figure 3.4 along with the density function for a GBM process with the same parameters (plotted as a broken line), and the true density function (plotted as a solid line) conditional on S T > 0, calculated using an approximation to Equation (3.14): f ST S t (s, τ) = (2 β) k 1 2 β e x z 100 n=0 x n+ 1 2 β z n+ 1 β 2 β n!γ(n β ) where the upper limit of the summation has been replaced by 100. In fact, remarkable accuracy can be achieved using only the first 11 terms of the summation, with a maximum difference of only between the two approximations, where both are calculated at 500 values of S T in the range [0, 10.4]. The theoretical curve is a true density function, which is formed by scaling the function f ST S t (s, τ) above as seen in Equation (3.17): f ST S t (s, τ S T > 0) = f S T S t (s, τ) 1 G( x, 1 2 β ).
92 74 CHAPTER 3. THE CEV MODEL Relative Frequency Share Price Figure 3.4: 5000 realisations of S T, a future CEV price with β = 1, with S t = $5, τ = 1 year, µ = 0.1 and σ = 0.3, the (solid) theoretical density for these prices, and the lognormal density function with the same parameters. In Figure 3.4, we see that again there is good agreement between the simulated share prices and theory. Unlike the GBM equivalent seen in Figure 2.3, in this case there are two sources of error, due to sampling error and use of the Euler approximation for the solution to the SDE (3.1). The close fit of the density function to the histogram suggests that aggregate errors in the Euler method of simulating share price under the CEV model have been small. We can also compare the empirical cumulative distribution function to the theoretical one, which, thanks to Schroder, can be computed directly. These functions are shown in Figure 3.5 where the solid line represents the theoretical curve and the broken line the empirical results from the sample shown in Figure 3.4,with S t = $5, τ = 1, µ = 0.1, σ = 0.3 and β = 1. Even using fairly coarse subdivisions for the estimation of the empirical
93 3.2. THE CEV SHARE PRICE SOLUTION 75 P(Share Price < s) Figure 3.5: The empirical cumulative distribution function of the 5000 realisations of S T shown in Figure 3.4, a future CEV price with β = 1, with parameters S t = $5, τ = 1, µ = 0.1 and σ = 0.3, and the theoretical distribution function for these prices. s curve, the empirical and theoretical distributions are virtually indistinguishable. Estimating the sample mean and standard deviation for all 5000 realisations yields estimates S T = , and s(s T ) = , which are both below the theoretical values of and , despite the fact that fewer bankruptcies were observed than expected. These differences are not significant at the 5% level 4. In Theorem 3.2 I derived a form for the variance of a future CEV share price S T, given S t and β. The variance is given in Equation (3.18) and is used to calculate the theoretical standard deviation for the sample above. 4 A two-sided test of H 0 : µ = 5e 0.1 yields a test statistic of , and a p-value of Similarly a two-sided test of H 0 : σ 2 = yields a test statistic of and a p-value of
94 76 CHAPTER 3. THE CEV MODEL Using SPLUS, it is possible to calculate the standard deviation of S T given S t by restricting the infinite summation to a finite number of terms. This computation fails when β is close to two, since the quantile at which the gamma density function must be evaluated grows very large. Increasing the shape parameter, and the upper limit of the summation does not solve the problem since both Γ(n i ) terms, where i = 1, 2, become too large, 2 β even though their ratio is likely to remain moderate in comparison. Figure 3.6 shows the standard deviation function for 2 β 2, and the additional parameters S t = $5, τ = 1 year, µ = 0.1 and σ = 0.3. Due to computation problems the standard deviation for 1.5 < β < 2 has been estimated by linear interpolation using the standard deviation when β = 1.5 found using the approximation, and that for β = 2, given by Equation (2.7). Standard Deviation of Share Price Figure 3.6: The standard deviation of a future CEV share price S T, with S t = $5, τ = 1 year, µ = 0.1 and σ = 0.3, for 2 β 2. beta The relationship shown in the graph is not monotonic, with the variance decreasing as β decreases from 2 to approximately -0.50, after which the
95 3.3. PROPERTIES OF C T - THE EXERCISE PAYOFF 77 variance begins to increase. 3.3 Properties of C T - the Exercise Payoff Many results proved for the GBM case apply to the CEV exercise payoff also. In particular the form for the probability density function of C T given S t, derived in Theorem 2.3 holds, except of course the functions F ST S t (s, τ) and f ST S t (s, τ) are different. In addition we saw that E(C T S t ) is very similar to the price of the option at t, but in future value terms, and with µ present in the solution instead of r. In the case of CEV exercise payoffs I will restrict myself to noting the form of the cumulative distribution function of the payoff. A graphical representation of both the theoretical and simulated exercise payoffs follows from Figure 3.5 by setting F ST S t (s, τ) = 0 for s < K, and repositioning the curve, moving the origin to s = K. The shape of both curves will be identical for s > K. This result is stated and proved in the following Theorem. Theorem 3.4 (The Distribution of C T option will deliver C T = { 0 S T K S T K given S t ). At exercise, the call S T > K (3.24) the cumulative distribution function of which is, given S t : 0 c < 0 2 F CT S t (c, τ) = Q(2 x; 2 β, 2 kk 2 β ) c = 0 2 Q(2 x;, 2 k(k + c) 2 β ) c > 0 2 β (3.25) where x and k are defined in Equation (3.12). Proof. For K > 0: P (C T = 0 S t ) = P (S T < K S t ) = Q(2 x; 2 2 β, 2 kk 2 β )
96 78 CHAPTER 3. THE CEV MODEL from Theorem 3.3. Likewise, for c > 0 P (C T < c S t ) = P (S T K < c S t ) = P (S T < K + c S t ) ( ) 2 = Q 2 x;, 2 k(k + c) 2 β 2 β again using the distribution function of S T given in Theorem The CEV Option Pricing Formula The price of a European option over a stock whose share price evolves according to Equation (3.1), was derived by John Cox and circulated as an unpublished note 5, which was summarised in Cox (1996). In addition, two special cases of the CEV model, the Absolute, and Square Root CEV models, were described in Cox & Ross (1976). The derivation of these models used the risk-neutral pricing method discussed in Section 2.4, as an alternative to the classical method used by Black and Scholes. The risk-neutral method is applicable because again the option price may be described using an equation which is independent of investor risk preferences. The assumptions of the CEV model are: the instantaneous interest rate is known, and constant; share price is a solution to the SDE (3.1), which implies: the parameters δ and β are constant, and known; no dividends are paid on the share over the life of the option; there are no transaction costs, differential taxes, or short-selling restrictions, and it is possible to trade any fraction of the stock or option. 5 John Cox (1975), Notes on option pricing I: constant elasticity of variance diffusions, unpublished note, Stanford University, Graduate School of Business.
97 3.4. THE CEV OPTION PRICING FORMULA 79 The only difference between these assumptions and those of Black and Scholes are those regarding the way share price evolves over time. All other assumptions are the same. Theorem 3.5 (The CEV Option Price PDE). If S t is a solution to the SDE (3.1), then the price of a European call option C t, with time τ to maturity, must satisfy the partial differential equation (PDE): subject to the boundary condition 1 2 δ2 S β 2 C C + rs S2 S + C t rc t = 0 (3.26) C T = (S T K) + where r is the continuously compounding risk-free rate. Proof. The CEV option price PDE follows from Theorem 2.5, with as does the boundary condition. σ(s t, t) = δs β 2 1 t As in the Black-Scholes case, the PDE (3.26) which the option value must satisfy, is independent of µ and is hence independent of investor preferences. Cox and Ross thus infer that the option may be valued assuming any investor preferences, and in particular, assuming risk-neutrality. Why this should be so can be seen by applying the Kolmogorov Equations introduced in Result 3.1 to their solution. Theorem 3.6 (The Risk-Neutral CEV Solution). The correct solution C t, of the PDE (3.26) subject to the boundary condition C T = (S T K) + is given by the expectation equation: C t = C(S t, τ) = e rτ E {(S T K) + S t } where E is an expectation using a risk-neutral transition density function for share price with the replacement µ = r, and r is the continuously compounding risk-free rate.
98 80 CHAPTER 3. THE CEV MODEL Proof. Considering C in the form will yield the partial derivatives C S = e rτ P S C(S t, τ) = e rτ P (S t, τ) 2 C S = 2 P 2 e rτ S 2 C τ = re rτ rτ P P + e τ which can be substituted in Equation (3.26) to give the similar PDE for the function P : 1 2 δ2 S β 2 P t S + rs P 2 t S P τ = 0 (3.27) and boundary condition C T = P T = (S T K) +. From the relationship between P and C, the form of P (S t, t) is known, and its partial derivatives can be evaluated: and so: P (S t, t) = E {(S T K) + S t } = P S = 2 P S = 2 P τ = K K K s=k (s K)f S T S t (s, τ)ds (s K) f S ds (s K) 2 f S ds 2 (s K) f τ ds. Using the above partial derivatives, the left hand side of Equation (3.27) can be evaluated to give: K (s K) { 1 2 δ2 S β 2 f t S + rs f 2 t S f } ds (3.28) τ in which a risk-neutral transition density function, f, is being differentiated with respect to the backward variables S t = S and t. The backward Kolmogorov Equation (3.7) states that the transition density for a general CEV process will satisfy the PDE: 1 2 δ2 S β t 2 f 2 S + µs f t + f t S t t = 0.
99 3.4. THE CEV OPTION PRICING FORMULA 81 Thus, the transition density for a process with µ = r will satisfy the backward Kolmogorov Equation with µ replaced by r, and for such a process is a solution of (3.26). C(S t, t) = e rτ E {(S T K) + S t } (3.29) We know the transition density function for the general CEV process, in which we can replace µ by r, and form the risk-neutral density function for S T given S t, where S T subject to the initial condition S t is a random variable that is a solution to the SDE: for this variable, conditional on S t : f S T S t (s, τ). ds t = rs t dt + δs β 2 t db t (3.30) Note also that Equation (3.29) can be rewritten: C(S t, t) = e rτ E {(S T K) + S t } = e rτ (s K)fS T S t (s, τ)ds K = S t. I will denote the density function = e rτ sfs T S t (s, τ)ds Ke rτ fs T S t (s, τ)ds K = e rτ E (S T S t, S T > K)P (S T > K S t ) Ke rτ P (S T > K S t ) K (3.31) where P (.) is a risk-neutral probability whose definition follows from the final equations above, and since, given S T density function for S T given S t is The CEV Solution f S T S t,s T >K(s, τ) = f S T S t (s, τ) P (S T > K S t ). > K, the correct risk-neutral The solution to the CEV partial differential equation has been outlined by various people, with conflicting results. Most of the disagreement can be
100 82 CHAPTER 3. THE CEV MODEL attributed to typographical errors, but I believe one error in particular is genuine. In order to shed more light on the different versions of the CEV solution published, I will attempt to employ consistent notation, and identify and correct inconsistencies. Cox (1996) presents the CEV solution option pricing formula, and the risk-neutral density function from which it was derived. Comparison of this solution with that in Cox & Ross (1976) for the special case β = 1 (the Square Root CEV model) indicates that the density function given in Cox (1996) is incomplete. Ignoring dividends, Cox s risk-neutral density function should read, for s > 0: where: fs T S t (s, τ) = (2 β)k 1 2 β (xz 1 2β ) 1 2 k = 2r δ 2 (2 β)(e r(2 β)τ 1) 1 2 β e x z I 1 2 β (2(xz) 1 2 ) (β < 2) (3.32) x = ks 2 β t e r(2 β)τ z = ks 2 β (3.33) with s = S T, and the missing term e x z is inserted. As we saw earlier, using the power series expansion of the modified Bessel function yields an alternative form for the risk-neutral density function: f S T S t (s, τ) = (2 β)k 1 2 β e x z n=0 x n+ 1 2 β z n+ 1 β 2 β n!γ(n β ). The risk-neutral density function shown in Equation (3.32) can be used to obtain a formula for the option price given by Cox. Theorem 3.7 (The CEV Option Pricing Formula). The solution to Equation (3.26) subject to the boundary condition C T the formula: C t = S t n=0 ( ) e x x n G kk 2 β, n β Ke rτ Γ(n + 1) n=0 = (S T K) + is given by e x x n+ 1 2 β G ( kk 2 β, n + 1 ) Γ(n β )
101 3.4. THE CEV OPTION PRICING FORMULA 83 where g(x, ν) and G(y, ν) are the gamma probability density and survivor functions respectively, with shape parameter ν and unit scale parameter. Proof. From Theorem 3.5 and the results of Cox & Ross (1976) we can use the risk-neutrality argument to give Cox s option pricing formula. Thus, the solution can be obtained by evaluating: C(S t, t) = e rτ (s K)fS T S t (s, τ)ds. K Note that since z = ks 2 β, when s = S T = K then z = kk 2 β. Again drawing parallels to previous results, and in particular Equation (3.15), we can make the change of variable to z: f S T S t (s, τ)ds = e x z n=0 x n+ 1 2 β z n n!γ(n β )dz. Noting x = ks 2 β t e r(2 β)τ, the solution takes the form: ( (z ) ) 1 C(S t, t) = e rτ 2 β K e x z x n+ 1 2 β z n z=kk k 2 β n!γ(n )dz n=0 2 β = e rτ e x z x n (x/k) 1 2 β z n+ 1 2 β Kx n+ 1 2 β z n kk 2 β n!γ(n ) dz n=0 2 β = e rτ e x z x n S t e rτ z n+ 1 2 β Kx n+ 1 2 β z n kk 2 β n!γ(n ) dz 2 β e x x n = S t n! n=0 n=0 kk 2 β Ke rτ e z n+ z 1 2 β Γ(n n=0 )dz 2 β e x x n+ 1 2 β Γ(n β ) kk 2 β e z z n dz (3.34) n! which is the required solution, and so consistent with the conditional density function for S T given S t in Equation (3.32) above. It is not a trivial matter to show directly that this solution satisfies the partial differential equation above, nor that the density function satisfies the forward equation. Despite this, the CEV option prices exhibit limiting qualities that instill confidence in the solution, as does the earlier comparison of CEV share prices to their theoretical distribution.
102 84 CHAPTER 3. THE CEV MODEL Reconciling Various Forms of the CEV Solution Cox s option pricing formula for the CEV model with β < 2, as given in Equation (3.34), is consistently reported in papers by Beckers (1980), Emanuel & MacBeth (1982), MacBeth & Merville (1980) and Schroder (1989). Furthermore the solution that appears in Jarrow & Rudd (1983) on page 154 is also consistent with that of Cox, however, this solution cannot be obtained by integration of the risk-neutral density function also given in this text. The risk-neutral density function given by Jarrow and Rudd is: f S T S t (s, τ) = φe rτ where: n=0 ( g λs φ t, n φ ) g(λ(se rτ ) φ, n + 1) (3.35) λ = 2r δ 2 φ(e φrτ 1) φ = β 2. To identify the function that should appear in Jarrow and Rudd s notes, the risk-neutral density given in Equation (3.32) can be written in Jarrow and Rudd s notation. Comparison of the two density functions, and noting the definitions of k, x, and z, yields immediately the relationships: k = λe φrτ x = λs φ t z = λ(se rτ ) φ. These can be substituted into Cox s risk-neutral density function to give: f S T S t (s, τ) = (2 β)k 1 2 β e x z = φ(λe φrτ ) 1 φ e x z = φλe φrτ s 1 1 φ = φλe φrτ s 1 1 φ n=0 n=0 n=0 n=0 x n+ 1 2 β z n+ 1 β 2 β n!γ(n (λs φ t 2 β ) (λs φ t ) n 1 φ (λ(se rτ ) φ ) n+1+ 1 φ n!γ(n ) φ ) n 1 φ e λs φ t Γ(n ) φ (λ(se rτ ) φ ) n λ(se rτ ) φ e Γ(n + 1) g(λs φ t, n φ )g(λ(se rτ ) φ, n + 1)
103 3.5. COMPUTING THE OPTION PRICE 85 This final equation is clearly different to the density given by Jarrow and Rudd, who have a constant φe φrτ multiplying the sum. Since the riskneutral density given in Equation (3.32) yields the option pricing formula given by Cox, computing e rτ E ((S T K) + S t ) using the risk neutral density function given by Jarrow and Rudd cannot lead to the required formula. 3.5 Computing the Option Price The option pricing formula given by Cox appears formidable, especially compared to the industry standard Black-Scholes formula. Empirical papers that consider the analysis of simulated CEV prices include Beckers (1980) and MacBeth & Merville (1980). These authors produce prices by limiting the infinite summations to a finite number of terms: n 2 C t S t n=n 1 g(x, n + 1)G n 2 Ke rτ g n=n 1 ( ) kk 2 β, n β ( ) x, n G(kK 2 β, n + 1) 2 β (3.36) where x and k are defined in Equation (3.33), and where n 1 and n 2 are chosen to ensure convergence of the solution 6. Alternatively, these authors and others, including Rubinstein (1985), focus their attention on special cases of the CEV class for which closed form solutions or efficient numerical approximations existed. In particular, there is a closed form solution for Absolute CEV prices, where β = 0, and a numerical approximation for the Square Root CEV prices, where β = The Absolute CEV Model In the Absolute CEV model, share price evolves according to the SDE: 6 Beckers uses n 1 = 1 and n 2 = 995. ds t = µs t dt + δdb t (3.37)
104 86 CHAPTER 3. THE CEV MODEL which is an Ornstein-Uhlenbeck process with an absorbing barrier at zero. Cox & Ross (1976) give a form for the risk-neutral transitional density function with β = 0: f S T S t (s, τ β = 0) = 1 2πWt (exp ( (s S ) te rτ ) 2 exp ( (s + S )) te rτ ) 2 2W t 2W t (3.38) where W t = δ2 2r (e2rτ 1). Using the formula in Equation (3.29) and the density function above yields the closed form solution for the option price when β = 0: C abs t = (S t Ke rτ )Φ(d t ) + (S t + Ke rτ )Φ(y t ) + v t (φ(d t ) φ(y t )) (3.39) where φ(x) is the unit normal probability density function at x, Φ(x) is the unit normal cumulative distribution function, and: ( ) 1 1 e 2rτ 2 v t = δ 2r d t = S t Ke rτ v t y t = S t Ke rτ v t. It is relatively simple to obtain the solution from the risk-neutral density function in the usual way, and in this case it is also easy to show by evaluating partial derivatives, that the solution (3.39) is consistent with the PDE (3.26) with β = 0. This solution is as easy to compute as the Black-Scholes equation, and would be an ideal alternative, if the specification of the model were empirically appropriate Computing the General Model Schroder (1989) derives a result that allows easy computation of option prices for any CEV process. His result would have allowed Beckers, MacBeth and Merville, Rubinstein and others to easily examine CEV prices other than those that prior to Schroder s result were easy to compute. This work has been alluded to in Section where it was used to obtain an expression for the distribution function of future share prices. The key to that result,
105 3.5. COMPUTING THE OPTION PRICE 87 and this one, is to note that the integral of interest, in this case Equation (3.29), can be written in terms of the non-central Chi-squared cumulative distribution function. Theorem 3.8. The CEV option price may be written C t = S t 2 p(2z; 2 + 2, 2x)dz 2 β Ke rτ y y 2 p(2x; β, 2z)dz where y = kk 2 β, and p(y; ν, λ) is the probability density function at y of a Chi-squared random variable with ν degrees of freedom and non-centrality parameter λ, and x and k are defined in Equation (3.33). Proof. Recalling Equation (3.32) and noting from the proof to Theorem 3.1 ds = k 1/(2 β) 1 2 β z 1+1/(2 β) dz we obtain a further form for f S T S t (s, τ)ds: fs T S t (s, τ)ds = x 1/(4 2β) z 1/(4 2β) e x z I 1 (2(xz) 1 2 )dz 2 β which we can use to examine the option price: C t = e rτ E ((S T K) + S t ) = e rτ (s K)fS T S t (s, τ)ds K ( (z ) ) 1 = e rτ 2 β K x 1/(4 2β) z 1/(4 2β) e x z I 1 (2(xz) 1 2 )dz z=kk k 2 β 2 β = S t e x z (z/x) 1/(4 2β) I 1/(2 β) (2 xz)dz since: y Ke rτ e x z (x/z) 1/(4 2β) I 1/(2 β) (2 xz)dz (3.40) y e rτ (x/k) 1/(2 β) = e rτ ( S 2 β t e r(2 β)τ ) 1/(2 β) = St. Schroder recognises that the integrands are both non-central Chi-squared density functions, of the form given by Johnson et al. (1995) seen in Result 3.3. Recall the form for p(y; ν, λ) given in Equation (3.23): p(y; ν, λ) = 1 2 (y/λ) 1 4 (ν 2) exp { 1 2 (λ + y)} I 1 2 (ν 2) ( λy).
106 88 CHAPTER 3. THE CEV MODEL The first integrand in Equation (3.40) is: e x z (z/x) 1/(4 2β) I 1/(2 β) (2 xz). Comparison of this with p(y; ν, λ) above suggests that the integrand is almost a non-central Chi-squared density function with y = 2z ν = β λ = 2x except for the factor 1, which is absent. Hence: 2 e x z (z/x) 1/(4 2β) I 1/(2 β) (2 xz) = 2 p(2z; 2 + 2, 2x). (3.41) 2 β The second integrand in the expression for C t above is: e x z (x/z) 1/(4 2β) I 1/(2 β) (2 xz). Again comparison with p(y; ν, λ) indicates that the integrand is almost a non-central Chi-squared density function with y = 2x ν = β λ = 2z except for the scale factor 1 2. Hence: e x z (x/z) 1/(4 2β) I 1/(2 β) (2 xz) = 2 p(2z; 2 + 2, 2x). (3.42) 2 β Combining Equations (3.41) and (3.42) yields the desired result: C t = S t 2 p(2z; 2 + 2, 2x)dz 2 β Ke rτ y y 2 p(2x; β, 2z)dz (3.43) Note that as written, neither of the integrals in Equation (3.43) are noncentral Chi-squared survivor functions at y. The first integral has integration with respect to z rather than the required 2z, whilst the second has integration with respect to half the non-centrality parameter 2z. Before proving Schroder s result, it is necessary to consider the following result.
107 3.5. COMPUTING THE OPTION PRICE 89 Result p(2x; 2ν + 2, 2z)dz = 1 y x 2 p(2z; 2ν, 2y)dz where p(y; ν, λ) is the probability density function at y of a non-central Chisquared random variable with ν degrees of freedom and non-centrality parameter λ. Proof. A proof of this result can be found in Appendix B.3. Theorem 3.9 (Schroder s Option Pricing Formula). The form of the CEV option price given by Schroder (1989) is: C t = S t Q(2y; β, 2x) Ke rτ (1 Q(2x; 2 2 β, 2y)). where y = kk 2 β, x and k are given in Equation (3.33), and Q(y; ν, λ) is the survivor function at y for a non-central Chi-squared random variable with ν degrees of freedom and non-centrality parameter λ. Proof. The proof of this result follows directly from Theorem 3.8 and Result 3.4. Note the form of C t given by Equation (3.43): C t = S t 2 p(2z; 2 + 2, 2x)dz 2 β Ke rτ y y 2 p(2x; β, 2z)dz. Making the substitution w = 2z in the first integral yields: y 2 p(2z; 2 + 2, 2x)dz = 2 β 2y p(w; β, 2x)dw = Q(2y; β, 2x). Similarly, applying Result 3.4 to the second integral and again making the substitution w = 2z yields: y 2 p(2x; 2 + 2, 2z)dz = 1 2 β = 1 x 2x 2 p(2z; 2 2 β, 2y)dz p(w; 2 2 β, 2y)dw = 1 Q(2x; 2 2 β, 2y).
108 90 CHAPTER 3. THE CEV MODEL Hence the option pricing formula is: as required. C t = S t Q(2y; 2 + 2, 2x) 2 β Ke rτ 2 (1 Q(2x;, 2y)) (3.44) 2 β Recall that Theorem 3.3 gave the form of the cumulative distribution function for the CEV share price S T given S t. The proof for the form of this function in the case s > 0 utilises results now available. Proof of Theorem 3.3 Proof. We are required to prove the proposition that for s > 0, the cumulative distribution function of S T given S t is given by the equation: where F ST S t (s, τ) = Q(2 x; 2 2 β, 2ỹ) k = 2µ δ 2 (2 β)(e µ(2 β)τ 1) x = 2 β ks t e µ(2 β)τ 2 β ỹ = ks and Q(x; ν, λ) is the survivor function at x of a non-central Chi-squared variable with ν degrees of freedom and non-centrality parameter λ. that from Equation (3.31), the option price at t may be written: C t = e rτ E (S T S t, S T > K)P (S T > K S t ) Ke rτ P (S T > K S t ) whilst Schroder gives the option price as: C t = S t Q(2y; β, 2x) Ke rτ (1 Q(2x; 2 2 β, 2y)). Comparing the two formulae it is clear that: P (S T > K S t ) = 1 Q(2x; 2 2 β, 2y) where y = kk 2 β, and so replacing r by µ in x and y, we obtain: P (S T > K S t ) = 1 Q(2 x; 2 2 β, 2 kk 2 β ) Note
109 3.5. COMPUTING THE OPTION PRICE 91 which yields the desired result: where ỹ = ks 2 β. P (S T < s S t ) = 1 (1 Q(2 x; 2 2 β, 2 ks 2 β )) CEV Option Prices = Q(2 x; 2 2 β, 2ỹ) In the same way that a call price series can be computed for geometric Brownian motion using the Black-Scholes formula, the call price series that corresponds to a simulated CEV process can be calculated using Schroder s formula. A difference is that SPLUS cannot evaluate Schroder s formula when τ is close to zero, because both the quantile at which the non-central Chi-squared survivor function must be evaluated, and the non-centrality parameter become infinite. It appears then that both the Q terms will be zero, yielding an option price of 0, not the same as the required (S T K) +, however a limiting argument will undoubtedly yield the correct result. Computation of Schroder s formula becomes increasingly slow as the time to maturity nears zero (as would computation of the infinite sums, since n 2 must be increasingly large to compensate for growth in k). Figure 3.7 displays a realisation of {S t }, a CEV process with β = 1, S t = $5, τ = 1, µ = 0.1, σ = 0.3, and 250 subintervals. In addition, series corresponding to {C t + Ke rτ } and {Ke rτ } are shown. The difference between the two latter curves is of course the option price series {C t }. The graph exhibits the same general qualities shown in the corresponding graphs for the GBM realisations seen in Figures 2.4 and 2.5. A high degree of correlation is seen between the series {S t } and {C t }, where every movement in the share price is mirrored by a movement in the option price. In the particular realisation in Figure 3.7, the option matures in-the-money, and so, near exercise when the discounting effect on the exercise price is minimal, the option price plus (discounted) exercise price converges to the share
110 92 CHAPTER 3. THE CEV MODEL Price Share price Option price + pv(exercise price) pv(exercise price) Time to Maturity - Years Figure 3.7: A realisation of a CEV share price with initial value S t = $5, and parameters β = 1, τ = 1, µ = 0.1, σ = 0.3, and 250 subintervals; also the CEV option prices for this share series added to the present value of the exercise price, with K = $5, and r = 0.06 and time to maturity indicated on the horizontal scale; finally the present value of the exercise price itself. price, or equivalently, the option price converges to the share price less the (discounted) exercise price. Both MacBeth & Merville (1980) and Beckers (1980) have computed and analysed CEV prices. In particular they compare option prices for various CEV classes with Black-Scholes option prices. MacBeth and Merville choose CEV processes with β = 0, -2, and -4, aligning the processes to GBM using the relationship δ = σs 1 β 2 t where σ is equivalent to the Black-Scholes volatility parameter, as described earlier. I have chosen to reproduce part of MacBeth and Merville s Table 1, using SPLUS and Schroder s pricing formula seen in Equation (3.44). These figures
111 3.5. COMPUTING THE OPTION PRICE 93 are shown in Table 3.1 and correspond to those of MacBeth and Merville except in two cases, shown with an asterisk. These values are given as 0.89 and 0.94 by MacBeth and Merville, instead of 0.88 and 0.95 respectively. The correspondence between the two sets of results indicates that Schroder s formula is indeed reproducing C t, correctly and without the need to make decisions regarding the upper and lower limits of summation in Cox s formula. The fact that two values differ in the second decimal place, indicates that at least twice, the values of n 1 and n 2 chosen by MacBeth and Merville have not ensured convergence of the option price seen in Equation (3.36). In pricing a large bundle of options, such differences could amount to significant pricing errors. σ = 0.2 σ = 0.4 τ K β = 0 β = 2 β = 0 β = Table 3.1: A section of MacBeth and Merville s Table 1, of CEV option prices, calculated for the parameters shown, with additional parameters: S t = $50 and r = Beckers compares Square Root and Absolute CEV prices with Black- Scholes prices. In addition to using Equation (3.36) with n 1 = 0 and n 2 = 995 to obtain the CEV prices, Beckers uses an approximation (attributed to John Cox, but not referenced) for the Square Root CEV case, and the exact formula for the Absolute CEV case, given in Equation (3.39).
112 94 CHAPTER 3. THE CEV MODEL Earlier, I described the alignment process used by Beckers, to ensure his CEV prices were comparable to Black-Scholes prices. This involved solving the equation Var(S T S t ) = St 2 e 2µτ (e σ2 1) where the random variable S T follows a CEV process with β < 2, for δ [which features in Var(S T S t )], so that all three share price processes considered have the same variance at T. In contrast, MacBeth and Merville s alignment procedure ensured all processes have the same variance over the interval [t, t + dt]. Using the form for Var(S T S t, β = 1) given in Theorem 3.2, Beckers shows that for a Square Root CEV process, δ is given by the equation: δ 2 = µs te µτ+σ2τ (e σ2 1) e µτ 1 (3.45) where σ is the Black-Scholes volatility parameter 7. I have used this relationship to give the values seen in Table 3.2. Months to maturity σ Table 3.2: Beckers δ values (rounded to 4 d.p.) for the Square Root CEV process, found using Equation (3.45), with additional parameters S t = $40, and µ = log(1.05). Table 3.3 reproduces a section of Beckers Table II, for the Square Root CEV case with S t = $40 and r = log(1.05). In all but one case, the values given by Schroder s formula correspond exactly to those given by Beckers, with the single exception when (σ, K, τ) = (0.2, 45, 4 12 ), which reads in Beckers table.
113 3.5. COMPUTING THE OPTION PRICE 95 Months to Maturity σ K Table 3.3: A section of Beckers Table II, showing Square Root CEV prices, with additional parameters S t = $40, and r = log(1.05). I have been unable to confirm the Absolute CEV prices given by Beckers, and have not attempted to confirm the values of δ he gives in his Appendix B. These values are obtained by solving the equation: Var(S T S t, β = 0) = S 2 t e 2µτ (e σ2 1) for δ, and are reproduced in Table 3.4. These values are of quite a different scale to δ values given by MacBeth and Merville s alignment procedure, which are 8, 12, and 16 corresponding to σ values 0.2, 0.3, and 0.4 respectively, indicating the values used by Beckers may have been reported incorrectly. Note that prices in both Tables 3.1 and 3.3 exhibit the following properties: 7 MacBeth and Merville s method yields δ values of , , and , corresponding to σ = 0.2, 0.3, and 0.4 respectively. These values are similar to those shown in Table 3.2, however use of these values, instead of Beckers, results in option prices up to 15% less, with the difference increasing with time to maturity and σ.
114 96 CHAPTER 3. THE CEV MODEL Months to maturity σ Table 3.4: Beckers δ values (rounded to 4 d.p.) for the Absolute CEV process, with additional parameters S t = $40, and µ = log(1.05). As K increases, C t decreases; As τ increases, C t increases; As σ increases, C t increases. These general trends, observed for the particular values given above, correspond to the expected behaviour of call option prices as discussed in Chapter Behaviour of CEV Prices As mentioned above, CEV prices appear to have similar properties to Black- Scholes prices, with respect to changes in exercise price, time to maturity, or the volatility parameter σ. Of interest is the behaviour of CEV prices when β changes, and in particular whether CEV prices converge to the Black-Scholes price as β nears two. If this convergence were not evident, then this would indicate a deficiency in the CEV option pricing formula. Cox s formula for CEV option prices, and Schroder s method of computing these prices enable us to examine convergence as β nears two from below. At present we cannot examine convergence from above, however a result derived by Emanuel & MacBeth (1982) completes the CEV class, and gives us a pricing formula for the case β > 2. Result 3.5 (The CEV Option Price for β > 2). In the case β > 2, the risk-neutral transition density function, describing the evolution of the share
115 3.5. COMPUTING THE OPTION PRICE 97 price S t to S T is: fs T S t (s, τ β > 2) = (β 2)k 1/(2 β) (xz 1 2β ) β e x z I 1/(β 2) (2 xz) where all terms are defined earlier 8. The option price is given by: 2 C t = S t Q(2x;, 2y) β 2 Ke rτ (1 Q(2y; 2 + 2, 2x)) (β > 2). (3.46) β 2 Proof. A proof of the option price formula can be found in Appendix B.4. Emanuel and MacBeth cite theoretical support for this model 9, however its use obviously conflicts with the motivation behind the original CEV model, including empirical observations regarding share price evolution. Regardless of the motivation behind each model, any option price over a share whose price evolves according to Equation (3.1) can now be computed using Equation (3.44) for β < 2, the Black-Scholes formula when β = 2, and Equation (3.46) when β > 2. Using these formulae, I have produced the graphs in Figures 3.9 to These figures give call option prices for β in the range 2 β 6, for various σ and K, and where S t = $50 and r = These graphs give the relationship between C t and β for out-of-the-money (K = $55), at-the-money (K = $50), and in-the-money (K = $45) calls as σ changes. Both MacBeth and Merville, and Beckers note the relationship between call price and β. MacBeth and Merville restrict their comments to the parameter choices they consider, but Beckers makes a more general comment: As a general rule, it can be inferred that for in-the-money and at-the-money 8 Note that the risk-neutral density that applies in the case β > 2 is very similar to that when β < 2, with differences only in the scale factor (in this case β 2 instead of 2 β), and the degrees of freedom of the modified Bessel function. In fact for β 2, the general risk-neutral density function could be written: f S T S t (s, τ) = 2 β k 1/(2 β) (xz 1 2β ) 1/(4 2β) e x z I 1/ 2 β (2 xz) 1 9 Rubinstein, Mark ((1981), Displaced Diffusion Option Pricing, Manuscript, University of California, Berkeley.
116 98 CHAPTER 3. THE CEV MODEL options the model price increases as the characteristic exponent [β] decreases, whereas exactly the opposite is true for out-of-the-money options. 10 Whilst Beckers uses a different alignment procedure to MacBeth and Merville, and I have produced the graphs in Figures 3.9 to 3.11 using the latter authors alignment procedure, it appears that the relationship Beckers describes does not hold in general. Even if Beckers comments were restricted to the case 0 β < 2, when σ is very large (perhaps too large given empirical observations), even in this range C t is not a monotonic function of β as shown in the graphs. A positive aspect of the functions shown in Figures 3.9 to 3.11 is that the CEV prices appear to converge to the Black-Scholes price, from above and below, when β approaches 2, for the particular combination of parameters used. This observation lends support to the use of the CEV option pricing formulae as alternatives to the Black-Scholes formula. 3.6 Use of the CEV Model In Section 2.5, I noted the volatility smile phenomenon, which provides evidence against the use of the Black-Scholes formula to price share options. Figure 3.8 shows the Black-Scholes implied volatilities 11 for Absolute CEV option prices. Parameters used to obtain the CEV option prices were S t = $5, τ = 1 year, σ = 0.3, r = 0.06 and $4 K $6. These prices were then used as market option prices, to yield the Black-Scholes implied volatilities shown in the graph. In particular I solved the equation: C mt = BS(σ) for σ, where the C mt are Absolute CEV prices, and BS(σ) is the Black- Scholes formula treated as a function with a single parameter σ. The values of σ that satisfy the relationship are shown in the graph. 10 Beckers (1980), page The implied volatility is the value of σ that equates the Black-Scholes model price, to the observed market price of the option, given other (observable) parameters S t, τ, K and r.
117 3.6. USE OF THE CEV MODEL 99 Implied Volatility Strike Price Figure 3.8: Black-Scholes implied volatilities for Absolute CEV option prices with S t = $5, τ = 1 year, σ = 0.3, r = 0.06 and $4 K $6. The curve in Figure 3.8 does not look dissimilar to the implied volatility curve for options on the S&P 500 index on May 5, 1993, given by Hull (1997) in his Figure Suppose a share price was truly following a CEV process, and the market was correctly pricing options over that share. Using the Black-Scholes formula to obtain an estimate for σ, results in the volatility smile relationship shown in the graph. The fact that the relationship shown is similar to those obtained empirically lends indirect support for the use of the CEV model in practice. Empirical research already alluded to has endeavoured to determine whether the CEV model is an appropriate one. In the next Chapter, I attempt to estimate the coefficient β for real share price series, but am not able to use option price information in the analysis. Rubinstein (1985) uses non-parametric methods to compare option pricing models. Included in the models he con- 12 Hull (1997), page 504.
118 100 CHAPTER 3. THE CEV MODEL siders is the Absolute CEV model, whose option price formula is given in Equation (3.39). In order to identify pricing models consistent with his vast database of option prices, Rubinstein considers pairs of market option prices with all characteristics in common except for either the strike price, or the time to maturity. He then calculates implied volatilities for the pair, and compares the pattern of relative values to those theoretically observed for the various models he considers. The Absolute CEV model is consistent with some of the biases he observes, but like the other models considered, it does not explain significantly all the biases observed.
119 3.6. USE OF THE CEV MODEL 101 sigma = 0.4 sigma = 0.6 sigma = 0.8 Call price Call price Call price beta beta beta sigma = 1 sigma = 1.2 sigma = 1.4 Call price Call price Call price beta beta beta sigma = 1.6 sigma = 1.8 sigma = 2 Call price Call price Call price beta beta beta Figure 3.9: Out-of-the-money CEV option prices, with β between -2 and 6, σ between 0.4 and 2, and additional parameters S t = $50, K = $55, τ = 0.5 years, and r = 0.06.
120 102 CHAPTER 3. THE CEV MODEL sigma = 0.4 sigma = 0.6 sigma = 0.8 Call price Call price Call price beta beta beta sigma = 1 sigma = 1.2 sigma = 1.4 Call price Call price Call price beta beta beta sigma = 1.6 sigma = 1.8 sigma = 2 Call price Call price Call price beta beta beta Figure 3.10: At-the-money CEV option prices, with β between -2 and 6, σ between 0.4 and 2, and additional parameters S t = $50, K = $50, τ = 0.5 years, and r = 0.06.
121 3.6. USE OF THE CEV MODEL 103 sigma = 0.4 sigma = 0.6 sigma = 0.8 Call price Call price Call price beta beta beta sigma = 1 sigma = 1.2 sigma = 1.4 Call price Call price Call price beta beta beta sigma = 1.6 sigma = 1.8 sigma = 2 Call price Call price Call price beta beta beta Figure 3.11: In-the-money CEV option prices, with β between -2 and 6, σ between 0.4 and 2, and additional parameters S t = $50, K = $45, τ = 0.5 years, and r = 0.06.
122 104 CHAPTER 3. THE CEV MODEL
123 Chapter 4 Data Analysis 105
124 106 CHAPTER 4. DATA ANALYSIS 4.1 Introduction In order to apply the Black-Scholes model to determine option prices on a stock, a single parameter, σ, must be estimated. A significant part of the empirical literature on option pricing, post Black & Scholes (1972) has tried to identify the best way to estimate this parameter. These studies have shown that whilst the Black-Scholes model does not seem appropriate due to its restrictive set of conditions, the Black-Scholes implied volatility is a better predictor of future share price volatility than an estimate based on past and present share prices. In order to use the CEV model to estimate option prices, not only must δ be estimated, but also the CEV parameter β. This parameter characterises the strength of the inverse relationship between share price level, and its volatility. Estimation of β values in real data that were significantly different from β = 2 would lend support to the use of the CEV model for option pricing, and provide evidence that the Black-Scholes assumptions regarding share price evolution and its volatility are not satisfied. In particular, I focus on the estimation of β, with an aim to identify whether or not the CEV model seems empirically appropriate. Some empirical papers do use particular CEV models to describe underlying asset evolution but do not attempt to chose the best model for their data, either by implying parameters for option and underlying asset price data or by direct estimation from the underlying asset series alone. These papers include Rubinstein (1985), whose work has been discussed in Chapter 3, and Lauterbach & Schultz (1990), who test the Square Root CEV model against the Black-Scholes and other models for the pricing of warrants 1. Other researchers do endeavour to determine the most appropriate CEV model, by estimating β from time series data. Previously, attempts 1 After adjustment for warrant pricing, this empirical research finds that the Square Root CEV formula describes observed warrant prices significantly better than the Black- Scholes model.
125 4.1. INTRODUCTION 107 have been made by Beckers (1980), MacBeth & Merville (1980), Emanuel & MacBeth (1982), Marsh & Rosenfeld (1983), Tucker, Peterson & Scott (1988) and Melino & Turnbull (1991) to estimate β, with varying degrees of success. Whilst their estimation methods vary, a recurrent theme in the results is that the CEV model indeed appears to be an appropriate model for share price evolution. Since modelling the share price has direct implications on modelling option prices, this finding provides evidence for the use of the CEV option pricing model in the marketplace. Before introducing the estimation procedure I have used, I will briefly summarise the methodologies and the findings of the aforementioned papers Summary of Alternative Methods Beckers (1980) Beckers estimates β values for 47 stocks, using almost five years data for each. The SDE (3.1) yields ( ) dst Var S t S t = δ 2 S β 2 t dt. With daily data, we can set dt = 1, and so rewrite the above relationship in the form: ( ln s ( St+1 S t )) S t = ln δ + β 2 2 ln S t where s(.) is the standard deviation operator. This is effectively the regression equation for β used by Beckers. Estimation of the regression parameters will yield estimates for both δ and β. Application of this technique requires some refinement however, as the dependent variable in the regression is not observable. Using the fact that for daily data µ is very small, giving ( ) St+1 E = e µ 0. S t Beckers approximates the standard deviation on the LHS by ln(s t+1 /S t ). This is appropriate if X = ln(s t+1 /S t ) is approximately normal, in which
126 108 CHAPTER 4. DATA ANALYSIS case, the mean value of X will be proportional to its standard deviation. This result is proved by Beckers. Thus, using the single realisation of X instead of the standard deviation of interest, Beckers estimates β using the regression ln ln S t+1 = a + b ln S t + w t (4.1) S t where β = 2b + 2. He reports low R 2 and Durbin-Watson statistics which lead him to conclude that his regression is incomplete. However, using this technique, Beckers finds that 38 of his 47 cases had β estimates significantly less than two, with three stocks yielding β significantly greater than two, leaving only six stocks consistent with GBM. Also, on the basis of his estimates, Beckers finds evidence to reject the hypothesis that the true β values for the stocks are the same, and concludes that in general, different stocks will have different values of β, just as their volatilities and other characteristics differ. Throughout his analysis Beckers restricts the CEV model to β : 0 β < 2. He concludes that only the 33 estimates in this range support the use of the CEV model, rather than utilising evidence from all the estimates which were significantly above or below two. MacBeth and Merville (1980) MacBeth and Merville consider six stocks with options traded on the Chicago Board of Trade Options Exchange. They use one year s daily data to estimate β for each of the stocks. From the SDE (3.1), the Chi-squared random variable (ds t µs t dt) 2 u δ 2 S β t χ 2 1 t dt can be formed given values of µ and β. A sample of daily returns is used to estimate µ, and the constant δ 2 embedded in what they describe as appropriate scaling. A sample of values is obtained by choosing a value of β, and
127 4.1. INTRODUCTION 109 this sample tested for goodness of fit with a χ 2 1 random variable, using the Chi-squared goodness of fit test. Using this method, MacBeth and Merville were unable to find β values that gave consistency between sample values and the χ 2 1 distribution, for any of their share series. They attributed this to non-normality of the db t estimates. Poor estimation of µ may contribute to the inability to find appropriate β values however. In light of this failed attempt, MacBeth and Merville attempt a regression similar to that of Beckers. The regression equation, found by taking the natural logarithm of the equation above, is ln{(ds t µs t dt) 2 )} ln dt = 2 ln δ + β ln S t + ln χ 2 1. This is used to obtain point estimates of β, under the assumption that the (central) Chi-squared random variables are uncorrelated. It is noted that E(ln χ 2 1) 0, but this should not affect estimates for β. A further regression method is used to obtain confidence intervals for β, using the inverse relationship between β and u t, the Chi-squared random variable. The regression methods are said to yield credible but imprecise estimates for β. In fact the confidence intervals are large, with only integer values of β considered. MacBeth and Merville s point and interval estimates are reproduced in Table Examination of this shows that only one confidence interval excludes geometric Brownian motion as a model for share price evolution. A final integer estimate for β is found by computing implied δ values for each stock. To do this they select four option prices on each stock at random, and for a chosen value of β, use a numerical search routine to find a value for δ. This is repeated for other (integer) values until the implied δ s are approximately equal. This method yields ˆβ s that do lie within the confidence limits in Table 4.1.1, but are very different from the previously obtained point estimates. The estimates are shown in the final column of the table.
128 110 CHAPTER 4. DATA ANALYSIS Stock Code Confidence Limits Point Estimate Implied Estimate ATT 2, AVON 8, ETKD 1, EXXN 1, IBM 8, XERX 4, Table 4.1: MacBeth and Merville s β estimates for six stocks. The conclusions drawn from these results are that different stocks in general have different β values, and that β is in general less than two. In addition, MacBeth and Merville add that there was evidence in their sample that β changes over time, and that it can be greater than two. The apparent support for the CEV model provides an explanation for why practitioners who use the Black-Scholes model to value call options must constantly adjust the variance rate they input to the model. 2 Comments by Manaster (1980) on the analysis just described appear at the end of the paper by MacBeth & Merville. These comments focus mainly on the estimation method of β and δ. Manaster suggests that if β and δ were estimated jointly, results even more favourable to the [CEV] model could be achieved. He objects to the daily re-estimation of ˆδ and the variance of returns, since the CEV model has a changing variance built in and reestimation of parameters prevents this aspect of the model from being tested. Emanuel and MacBeth (1982) This paper examines the data used by MacBeth & Merville (1980) in addition to a further year s data for the same six stocks. Emanuel and MacBeth employ a method that they themselves admit is not optimal. They use the relationship ( ) dst s S t S t 2 MacBeth & Merville (1980), page 299. = δs β 2 2 t
129 4.1. INTRODUCTION 111 to estimate β, where the standard deviation on the left hand side is given by an at-the-money call s implied volatility, found using the Black-Scholes model. The Black-Scholes implied volatilities are assumed to be reasonable estimates of share price volatility when calculated using observed at-themoney option prices 3 even though the Black-Scholes model may not be appropriate in general. In order to find ˆβ, squared differences between CEV option prices and selected market option prices 4 are minimised over β. Using this estimation technique, four of the six stocks are found to have β estimates greater than two, lending support to the alternative model that they derive. Marsh and Rosenfeld (1983) Marsh and Rosenfeld s paper does not model share prices but rather interest rates, using the CEV model. In order to obtain estimates of β for their two interest rate series, Marsh and Rosenfeld construct a likelihood function for each series using a modification of the density function given by Cox, seen in Equation (3.11), with τ = 1 day. This likelihood function is then calculated for β {0, 1, 2} for different values of δ, where the lognormal density function is used when β = 2. The three different likelihood functions corresponding to the three values of β, have maxima at a specific value of δ. These maxima are then compared, and the model that gives the largest of these selected. This in turn specifies the appropriate estimate of β. Using the interest rate data, of the three models considered, the lognormal (β = 2) model gave the highest likelihood for each of the time series. Tucker, Peterson and Scott (1988) Tucker et al. use the CEV model to describe yet another underlying asset, foreign currency exchange rates. In order to estimate β and δ for six major currencies, the log-likelihood function derived by Christie (1982) is used and 3 These observations are documented in Mayhew (1995) and Beckers (1981). 4 Options with less than 90 days to maturity were omitted from Emanuel and MacBeth s sample.
130 112 CHAPTER 4. DATA ANALYSIS maximised with respect to β and δ jointly using the Newton-Raphson procedure. This method is discussed in the following sections, as it is similar to that which I have used. Five years data is analysed for the exchange rates (to US dollars) of each of the six currencies: the British pound, Canadian dollar, Deutsche mark, French franc, Japanese yen, and the Swiss franc. The data period is split into five year-long blocks, and β and δ estimated for each of the five subperiods. They test the null hypothesis H 0 : β = 2 against the alternative H A : β 2 for each of the six currencies over each of the five subintervals, and find significant evidence to reject on 26 of the 30 occasions. In addition to the support provided for the CEV model, and in particular Emanuel and MacBeth s model, there was evidence to suggest that β was changing over time for the exchange rates. Melino and Turnbull (1991) Melino & Turnbull also use the CEV model to describe the evolution of exchange rates over time. Spot exchange rate data for five currencies were available for a year period. Like Marsh & Rosenfeld, Melino & Turnbull use an adapted form of the CEV density function for the process to construct a likelihood function. This is evaluated at the restricted range of β = 0, 1,..., 2, for a combination of values for the three other unknown 2 parameters (including δ). In addition to this method, they use an approximation like that of Tucker et al. and find that the maximised value of the (quasi) log-likelihood were virtually identical to those obtained from the continuous time model. 5 Once again though, this is only maximised with respect to the other parameters for fixed values of β. The case of lognormality was rejected in four out of the five tests. 5 Melino & Turnbull (1991), page 259.
131 4.2. ESTIMATING β FROM A SHARE PRICE SERIES Estimating β from a share price series The CEV share price evolves according to the familiar SDE (3.1) reproduced below: ds t = µs t dt + δs β 2 t db t the solutions to which are given by Cox (1996), Black & Scholes (1973), and Emanuel & MacBeth (1982) for the cases β < 2, β = 2, and β > 2 respectively. Given a realisation of a such a share price process, the aim of this section is to obtain estimates of β and δ, for use in option pricing. Recall that Cox s and Emanuel and MacBeth s analyses give us the true transition density for the solution to the above SDE for the case β 2: f St+1 S t (s t+1, 1) = 2 β k 1 2 β ( x z 1 2β ) 1 2 where again: 1 2 β e x z I 1 2 β (2( x z) 1 2 ) k = 2µ δ 2 (2 β)(e µ(2 β)τ 1) x = 2 β ks t e µ(2 β)τ 2 β z = ks apply with τ = 1 day. Theoretically, the likelihood function for a realisation from this process could be formed, and maximum likelihood techniques employed to jointly obtain estimates for β and δ, ˆβ and ˆδ respectively. Problems arise not only due to the presence of the modified Bessel function but also the complicated nature of the density as a function of β. In order to numerically solve a system of equations like f β = 0 β= ˆβ,δ=ˆδ f δ = 0 β= ˆβ,δ=ˆδ the derivatives, which are likely to be of the form g(β, δ)f(β, δ) where f(β, δ) is the density function above treated as a function of β and δ only, and g(β, δ) is some other function, must be evaluated over a finely divided range of β,
132 114 CHAPTER 4. DATA ANALYSIS and for each evaluation, approximation of the modified Bessel function must be made. Whilst the modified Bessel function appears to converge quickly, the maximisation process will still be computationally expensive. In addition, the likelihood function should be evaluated for each of the three cases: β < 2, β = 2 and β > 2, and the value of ˆβ that gives the largest value chosen β Estimation Strategy The β estimation method which I introduce in this section uses maximum likelihood technology, but is an approximate method which is not based on the true likelihood function for a series of CEV share prices. Using the true likelihood is possible, but this alternative method is both easier to implement, and to understand. Although the method is approximate, the estimates which it gives will hopefully be close to those which would be obtained using the true likelihood. At the very least, it is hoped that this approximate method will provide estimates which can be used as starting values in an estimation process that does use the CEV transitional density. In order to estimate β, I will use the fact that the Brownian motion increment db t, which appears in the SDE (3.1), is a normal random variable with zero mean and variance dt. This will allow use of the normal density function to construct a likelihood function that is simpler than one using the true transition density for the CEV share price process. Simplification of the estimation process is made by considering the transformation Y t = ln S t. Using Itô s Lemma, ( ) d ln S t = µ 1 2 δ2 S β 2 t dt + δs β 2 2 t db t (4.2) and so integrating both sides from n to n + 1: ln S n+1 ln S n = µ 1 2 δ2 n+1 n S β 2 t dt + δ µ 1 2 δ2 S n β 2 + δs n β 2 2 ɛ n n+1 n S β 2 2 t db t
133 4.2. ESTIMATING β FROM A SHARE PRICE SERIES 115 where ɛ n is a forward N(0, 1) increment that is independent of S n (and hence Y n ), and it is assumed that the integrands are sufficiently smooth to allow the integral approximations to be made, or alternatively, that the time unit of n is sufficiently small. Letting n be measured in days, the rates µ and δ must now also be measured in days. Since Y n = ln S n, we can write the discrete time equivalent of the SDE (3.1), and so conditional on Y n : where µ(y n ) = µ 1 2 δ2 e (β 2)Yn Y n+1 = Y n + µ(y n ) + δe β 2 2 Yn ɛ n (4.3) N ( Y n + µ(y n ), δ 2 e (β 2)Yn) is the expected value of the daily returns Y n+1 Y n, given Y n. Thus, an approximate transition density function for Y n+1 conditional on Y n = y n can be written: { 1 f Yn+1 Y n (y n+1, 1) = β 2 exp 1 ( ) } 2 yn+1 y n µ(y n ) 2πδe 2 yn 2 δe β 2 2 yn (4.4) and so, for a sample of share prices of size N + 1: S 0, S 1,..., S N, the conditional likelihood for the corresponding series {Y n : n = 0,..., N} can be formed. Use of the approximation directly above has an advantage over the use of the true transition density functions for S n+1 given S n, in that it can be used for any CEV process. The true likelihood function would be a product of density functions that have different forms for β < 2, β = 2 and β > 2, and so given an unknown parameter β, there might be difficulty constructing the correct likelihood function, especially when β is close to two. Note that: N 1 L(β, δ s) = f Y0 (y 0 ) f Yn+1 Y n (y n+1, 1) n=0 l(β, δ s ) = ln L(β, δ s ) = ln f Y0 (y 0 ) + N 1 n=0 ln f Yn+1 Y n (y n+1, 1)
134 116 CHAPTER 4. DATA ANALYSIS where L(β, δ s ) is the likelihood function for the parameters β and δ, given a realisation, s, of a share price series. Conditional on the share price sample s, in order to maximise the likelihood function L(β, δ) to find an estimate for β, it is appropriate to maximise the function N 1 1 l(β, δ) = 1 ln f N N Y 0 (y 0 ) + 1 ln f N Yn+1 Y n (y n+1, 1) (4.5) which, for large N, is approximately: ( { N l(β, δ) = ln β 2 exp 1 ( ) }) 2 yn+1 y n µ(y n ) N n=0 2πδe 2 yn 2 δe β 2 2 yn { = 1 N 1 ln δ β 2 y n 1 N 2 2 ln 2π 1 ( ) } 2 rn 2 n=0 = ln δ β 2 2N N 1 n=0 y n 1 2 n=0 δe β 2 2 yn N 1 1 r n 2 ln 2π (4.6) 2N δ 2 e (β 2)yn n=0 where r n = y n+1 y n µ(y n ) are realisations of the random variable R n = Y n+1 Y n µ(y n ). This log-likelihood function is very similar to that derived by Christie (1982), and implemented by Tucker et al. (1988), however there are a few differences. Christie produces the log-likelihood function: l(β, δ) = N 2 ln(2π) N 2 ln δ2 β 2 2 ln S 1 2δ 2 r 2 S S (β 2) (4.7) where N is sample size, and r S = S is the percentage daily return on the S share price S. In the model above, I am dealing with mean corrected daily returns, rather than the percentage returns Christie uses. In addition Christie sees no need for mean-correction, since the parameter µ is of a much smaller order than the variance of returns, but by making the log transformation, I introduce the correction term seen in Equation (4.2). Finally, Christie does not explicitly take into account the unconditional density function for Y 0. The log-likelihood in Equation (4.6) can be constructed from share price data by first computing the natural logarithm of the share prices, then taking
135 4.2. ESTIMATING β FROM A SHARE PRICE SERIES 117 the first difference of the logs to form a series {r n }. Once again I have chosen to estimate the mean of this series, E(Y n+1 Y n ) = µ(y n ), non-parametrically using the Lowess filter. This estimated mean level can then be subtracted from the daily returns to give the series { r n }, which can be combined with {y n } to calculate l(β, δ). Given the equivalence of maximisation of L and l, as discussed above, Maximum Likelihood estimates for β and δ could be found by solving the simultaneous equations: l β = 0 β= ˆβ,δ=ˆδ l δ = 0. β= ˆβ,δ=ˆδ I do not do this, but rather find an estimate for ˆδ from the second equation directly above, and substitute this back in to Equation (4.6), which I then maximise with respect to β. This second maximisation is done numerically, using SPLUS, the details of which follow. The first partial derivative of l(β, δ) with respect to δ is: l δ = 1 δ + 1 N 1 N n=0 r 2 n. (4.8) δ 3 (β 2)yn e Setting Equation (4.8) equal to zero as described above, and solving for ˆδ 2 (β) yields: ˆδ 2 (β) = 1 N N 1 n=0 r 2 n e (β 2)yn (4.9) which may substituted into Equation (4.6). The function obtained by making this substitution is: ( 1 l(β, ˆδ) = 2 ln 1 N N 1 n=0 r 2 n e (β 2)yn ) β 2 2N N 1 n=0 y n 1 2 ln 2π 1 2 (4.10)
136 118 CHAPTER 4. DATA ANALYSIS and maximisation of this with respect to β is equivalent to minimisation of the function: g(β) = ln ( 1 N N 1 n=0 r 2 n e (β 2)yn ) + β 2 N N 1 n=0 y n (4.11) obtained from l(β, ˆδ) by scaling and removal of additive constants. This function can be minimised numerically in SPLUS using the function nlmin as shown in Appendix D.4. An example of a typical likelihood surface is shown in Figure 4.1, and can be seen to have a ridge-like surface. This means that the parameters β and δ are highly interdependent, and that the estimation procedure is probably not all that well defined for finding ˆβ and ˆδ jointly. Ideally, the likelihood surface would have a distinct peak, which would enable the maximisation process to better pin-point the maximum likelihood estimates ˆβ and ˆδ. The scale of δ 2 in the graph has been altered in the estimation process, by a scale factor dt. Hence an estimate of σ 2 is given by where dt is measured in years. ˆδ 2 1 dt Sβ 2 t The share price used to construct the log-likelihood surface shown in Figure 4.1 yields the estimates: ˆβ = ˆδ = ˆσ = Choosing to minimise g(β), as outlined above, restricts the maximisation process to a cross-section of the log-likelihood surface, obtained by setting δ 2 = ˆδ 2 (β) as given in Equation (4.9). The log-likelihood function is shown again in Figure 4.2, along the cross section δ 2 = ˆδ 2 (β). The vertical line β = ˆβ has been superimposed on the graph of l(β, ˆδ), and clearly cuts the curve at its maximum. Hence it appears that the maximisation process is correctly identifying the maximum likelihood estimate of β.
137 4.2. ESTIMATING β FROM A SHARE PRICE SERIES (scaled) delta squared beta Figure 4.1: The log-likelihood surface, l(β, δ), for a simulated series with S t = $5, τ = 3 years, µ = 0.1, σ = 0.3, n = 250 subintervals per year, and β = 0. log-likelihood Figure 4.2: The cross-section of the log-likelihood surface in Figure 4.1, l(β, ˆδ) for a simulated series with S t = $5, τ = 3 years, µ = 0.1, σ = 0.3, n = 250 subintervals per year, and β = 0. In addition, the line β = ˆβ which identifies the maximum. beta
138 120 CHAPTER 4. DATA ANALYSIS The Variance of ˆβ Maximum likelihood theory indicates that the estimates of β and δ have desirable asymptotic properties. As sample size increases, they will be unbiased and normal, with variances obtained via the information matrix: I(β, δ) = E ( 2 l β 2 2 l δ β 2 l δ β 2 l δ 2 ). (4.12) Approximating N 1 l by l as before, and recalling the form of l(β, δ) in Equation (4.6) the partial derivatives can be evaluated and the information matrix found. The first partial derivatives of l are: l β = N l N 1 β = 1 2 n=0 Y n N 1 n=0 l δ = N l N 1 δ = N δ + R n 2 δ 3 e (β 2)Yn yielding second partial derivatives: n=0 2 N 1 l β = 1 R ny 2 n δ 2 e (β 2)Yn n=0 = 1 N 1 ɛ 2 2 nyn 2 n=0 n=0 2 N 1 l δ β = R ny 2 n δ 3 e (β 2)Yn N 1 ɛ 2 = ny n δ n=0 2 l δ = N N 1 2 δ 3 R n 2 2 δ 4 e (β 2)Yn n=0 = N N 1 δ 3 ɛ 2 n 2 δ 2 n=0 R 2 ny n δ 2 e (β 2)Yn
139 4.2. ESTIMATING β FROM A SHARE PRICE SERIES 121 where the ɛ n are approximately N(0, 1), with E (ɛ 2 n) = Var(ɛ n ) = 1. Hence, by the independence of ɛ n and Y n, elements of the information matrix are: ( ) 2 l E = 1 β 2 2 ( ) 2 N 1 l E = δ β n=0 ( ) 2 l E = 2N δ 2 δ 2 which give the information matrix: ( I(β, δ) = This has the determinant: 1 2 I(β, δ) = 2N N 1 1 δ 2 2 N 1 n=0 E(Y 2 n ) E(Y n ) δ E(Y 2 n ) 1 δ E(Yn ) 1 δ E(Yn ) 2N δ 2 n=0 = N δ 2 ( E(Y 2 N ) 1 N E(Y 2 n ) ( E(Y n )) 2 δ 2 ) ( E(YN )) 2 ).. (4.13) The information matrix yields the variances and covariances of the estimates through the relationship ( ) {I(β, δ)} 1 Var(ˆδ) Cov(ˆδ, ˆβ) = Cov(ˆδ, ˆβ) Var( ˆβ) These quantities will all feature the terms in the determinant, in particular E(ln S n ) and E{(ln S n ) 2 }, quantities which do not appear easy to calculate given the density function for S n given S t, where n > t. An alternative to calculating the theoretical variance of the estimates of β and δ is simulation. Share price series with known β and δ can be simulated using the SPLUS function described in Appendix D.3. The likelihood function can then be constructed for these series, and the maximum likelihood estimates obtained using the method described in Appendix D.4. This is done in the following section.
140 122 CHAPTER 4. DATA ANALYSIS 4.3 Appraisal of the Estimation Technique The estimation of β outlined above relies on the assumption R n N(0, 1) δ2 Sβ 2 holding at least approximately. The maximum likelihood analysis that follows this assumption yields estimates for both β and δ. These estimates may be used to construct a sample which should be approximately unit normal. The values e n = R n ˆδ 2 S ˆβ 2 n (4.14) can be compared graphically to a unit normal probability density function. If the two distributions are close, it is possible that the estimate of β will also be good. This analysis is not as important for simulated series as it is for real data series, since the simulated series are constructed directly from normal Brownian motion increments. Examination of the distribution of the quantity given in Equation (4.14) for selected real series will be done in the following section. The distribution of the e n values for the simulated series, whose log-likelihood function was examined earlier, is shown in Figure 4.3. This histogram has the standard normal density function φ(z) superimposed over the range of the realised e n values. The fit seems good, but there is an outlying value at z = This outlier does not pose a problem however, and a Chi-squared goodness of fit test for the sample yields a very small test statistic, with a p-value of 96.74% for a test of the hypothesis that the distribution comes from a standard normal distribution. This is not at all surprising because of the simulation method for the share series. If the sample of e n values does appear to have a standard normal distribution, the maximum likelihood procedure should indeed be appropriate. If so, the estimates should be asymptotically unbiased, and have variances
141 4.3. APPRAISAL OF THE ESTIMATION TECHNIQUE 123 Relative Frequency Figure 4.3: e n, given by Equation (4.14), for the simulated series examined previously with S t = $5, τ = 3 years, µ = 0.1, σ = 0.3, n = 250 subintervals per year, and β = 0. obtained via the information matrix. As noted above, the theoretical variances look difficult to compute, so I have used simulation to estimate the standard error of estimates of β. For given parameter values, it is possible to generate CEV share series, and then given these series, it is possible to estimate the parameter β. If all parameter values are fixed, and the share price simulation repeated a large number of times, a sample of ˆβ values will form, from which the mean, standard deviation, and approximate distribution can be estimated. Maximum likelihood theory tells us that for long series, with N large, the estimate of β should be approximately normal, and so the standardised ˆβ values: ˆβ β s( ˆβ) should be approximately unit normal, where s( ˆβ) is an estimate of the standard deviation of ˆβ, given by the sample standard deviation of the ˆβ values.
142 124 CHAPTER 4. DATA ANALYSIS Relative Frequency beta = -2 Relative Frequency beta = -1 Relative Frequency beta = beta = 1 beta = 2 beta = 3 Relative Frequency Relative Frequency Relative Frequency beta = 4 beta = 5 beta = 6 Relative Frequency Relative Frequency Relative Frequency Figure 4.4: Estimates of β, resulting from CEV share price simulation, used for the figures in Table 4.2. Superimposed on these are the density function of an N( β, 1) random variable over the range of the estimates, where β is the sample average of the β estimates.
143 4.3. APPRAISAL OF THE ESTIMATION TECHNIQUE 125 I have simulated share series for the nine integer values of β in the range 2 β 6, with S t = $5, µ = 0.1, and σ = 0.3. For each of these β values I have simulated 1000 series 850 days long, and have used the last 750 observations of each series to estimate β. This is intended to reduce the impact of the fact that all share series begin in the same place, if indeed these initial conditions have an effect, and replicate the conditions under which β will be estimated for the real series. After 100 days, the share price will be a random variable whose properties have been discussed in Chapter 3. Recalling that the CEV model allows bankruptcy, it is possible that the 1000 share price series will result in fewer than 1000 estimates of β. In fact, given the initial conditions, the expected number of bankruptcies for some β is large, as is reflected in the observed numbers. This analysis is meant to estimate β values for series similar to those for the real data examined in the following section. Since these series are all approximately three years long, I have estimated β using only the complete simulated series. Estimating β from the shorter series, where bankruptcy is observed, should not improve the estimates, since fewer observations are available than in the full length series, nor should it affect them, since the bankrupted series contain no information not theoretically present in complete series. This same simulation procedure could be used to test any β estimation procedure. Figure 4.4 shows the distribution of the standardised ˆβ values obtained from the simulation. The normal density curves that have been superimposed are drawn for the range of standardised β estimates, and are for normal variables with mean β, and variance 1, where β is the sample mean of the ˆβ values for each case. The normal curves are clearly not a good fit to the histograms. There are many outlying values in the estimates, and many of the histograms appear to have a sharper peak than the normal density functions, indicating positive kurtosis 6. The poor fit is confirmed by Chi-squared goodness of fit tests, with both the mean and standard deviation estimated from the data, where the null 6 Kurtosis is proportional to the fourth central moment of a random variable X, and is
144 126 CHAPTER 4. DATA ANALYSIS hypothesis that the sample values are normal is rejected in every case, with very high significance. Note also that many of the normal density functions are not centred on zero. The biasedness of the estimates is examined in a table which follows. Further results of the simulation are shown in Table 4.2. The true β value, used for simulating the share price series, is given, with the sample mean and standard deviation of the ˆβ values, β and s( ˆβ) respectively, and the number of series that did not result in bankruptcy, n. β β s( ˆβ) n Table 4.2: Summary of the β estimates for simulated series, obtained using the maximum likelihood procedure described above, with S t = $5, µ = 0.1, σ = 0.3 and where 3 years data is used. Comparison of the β and β values in Table 4.2 appears to indicate that the estimates are biased. The normal density curves superimposed on the ˆβ distributions were in fact centred on the respective β β values, which appear to differ from zero. For each of the above values of β it is possible to given by E{(X µ) 4 } σ 4 3 where µ = E(X) and σ 2 = Var(X). The kurtosis is zero for a normal random variable, and so positive (negative) values indicate that a density is more (less) peaked around its centre than a normal density function.
145 4.3. APPRAISAL OF THE ESTIMATION TECHNIQUE 127 test the hypotheses: H 0 :β = i H A :β i where i = 2, 1, The Central Limit Theorem provides the result: β β N(0, 1). s( ˆβ) n Testing these hypotheses for the nine cases of β results in the test statistics and p-values seen in Table 4.3. The data in the table shows that there is True β Test statistic p-value Table 4.3: p-values for hypothesis test H 0 : β = i, i = 2, 1,..., 6 using the simulated data summarised in Table 4.2. a significant bias present in the estimation procedure when β > 2. The estimates for Cox s CEV class have an average which is not significantly different from the true β at the 5% level, but at a higher significance level, the differences will become statistically significant. The figures in Table 4.3 do indicate that the method of estimating β is failing for some CEV processes, and in particular estimates for β that are greater than two may need adjustment to reflect this bias. The sample standard deviation of the β estimates is also given in Table 4.2, and these are shown in Figure 4.5. There does not appear to be much
146 128 CHAPTER 4. DATA ANALYSIS estimated standard deviation of beta estimates Figure 4.5: Estimates of s( ˆβ), the standard deviation of ˆβ, from Table 4.2 against the true β. beta variation in this quantity, but in general it increases as β deviates from 2, particularly for β > 2. Both the bias noted for this latter case and the large standard deviation do not appear to be due to a preponderance of large values, with the histograms in Figure 4.4 showing symmetry in the ˆβ values. In the following section, where I estimate β for 44 stocks, confidence intervals are formed for the true β (given that the model is appropriate and there really is a true β) using the standard deviations of ˆβ listed in Table 4.2. This is an alternative to deriving the theoretical standard deviation using the information matrix, and estimating it for the particular share price series. Linear interpolation of adjacent standard deviations is used for non-integer estimates of β. It certainly appears that the range should be adequate for the true standard error, which when compared to approximate standard errors of MacBeth and Merville of between 1.5 and 2.5 (assuming confidence interval width of around four standard errors) is very small. Tucker et al.
147 4.4. DATA ANALYSIS 129 (1988) also give the standard deviation of β estimates. For the range of β considered here, their standard errors are between 0.30 and They do not state how these standard errors are obtained, although they differ for similar values of ˆβ indicating a dependence on the realised time series data. It is also important that the method is producing sensible estimates. We know that the variance of daily returns should decrease as share price increases for β < 2, and vice versa for β > 2. This should be evident in a plot of the standard deviation of the daily returns against the share price level. Such a graph was seen in Figure 3.3 for a simulated series, and further plots will be produced for the real share series in the following section. 4.4 Data Analysis This section describes an exploratory study of some Australian share price data, provided by Credit Suisse First Boston NZ Limited. The data set contains share price, trade, and dividend information for stocks traded on the Australian Stock Exchange (ASX) which have exchange traded options. Of this data, there are 44 stocks with full information over the period 2 January 1995 to 17 November A list of these stocks can be found in Appendix C. The time period spanned by the data contains a period of extraordinary turbulence in the world markets, experienced in October This period has been omitted from all analysis, as it appears influences were exogenous to the Australian market. Thus the data analysed was for a period 2 January 1995 to 26 September In particular, the analysis focusses on the estimation of β for each of the series, and on the appropriateness of the technique outlined above. The estimated β values for the 44 dividend corrected share series are shown in Table 4.4, with estimates of s( ˆβ) obtained by linear interpolation of the function shown in Figure 4.5. Confidence intervals are formed using these standard deviations, with Type indicating what sort of CEV process
148 130 CHAPTER 4. DATA ANALYSIS has been estimated. Type 1 corresponds to the CEV class of Cox, with β < 2, Type 2 corresponds to GBM, and Type 3 corresponds to the final class, with β > 2. The estimates have not been corrected for the biases identified in the previous section, but this does not have an effect on the classification of the estimates into the respective Type categories. The bias is significant for β > 2, however none of the lower confidence limits given in Table 4.4 are close enough to two for the bias to have an effect. The closest is that of ANZ, at , which is well above two plus the estimated bias when ˆβ 3 of , given in Table 4.3. Table 4.4: β estimates for the 44 ASX share series, with estimates of s( ˆβ), the standard deviation of ˆβ, and the resulting confidence intervals obtained by simulation. Stock Code ˆβ s( ˆβ) ˆβ 2s( ˆβ) ˆβ + 2s( ˆβ) Type AAA AMC ANI ANZ BHP BIL BOR BPC BRY CBA CCL CML CSR FBG FXJ GIO GMF ICI LLC MAY MIM NBH continued on following page
149 4.4. DATA ANALYSIS 131 continued from previous page Stock Code ˆβ s( ˆβ) ˆβ 2s( ˆβ) ˆβ + 2s( ˆβ) Type NCM NCP NDY OSH PAS PBL PDP PLU PNI QNI QRL RIO RSG SEV SGB SRP STO TAH WBC WMC WOW WPL The estimates of β range from for BHP, to for BOR 7. The table clearly shows that there are many stocks in the sample whose share price does not appear to follow GBM. In fact, at an approximate 95% level, 16 stocks have ˆβ significantly less than 2, and 12 have ˆβ significantly greater than 2, leaving 16 cases for which GBM cannot be ruled out as a plausible evolution model. The estimation of β depends on assumptions alluded to in the previous section, in particular, the quantity e n defined in Equation (4.14) is required to be normal. For each of the share series listed in Table 4.4, I have computed the e n values using the share price and daily return data. For each 7 To construct the confidence interval for the estimates of BOR and CSR, a further simulation was run for β = 7, with β = , s( ˆβ) = , and n = 924.
150 132 CHAPTER 4. DATA ANALYSIS of the stocks, I have tested the hypothesis that the sample of e n values has a standard normal distribution. Using the Chi-squared goodness of fit test, with 27 degrees of freedom, I obtain the p-values shown in Table 4.5. Stock Code p-value Stock Code p-value AAA AMC ANI 0 ANZ BHP BIL BOR BPC 0 BRY 0 CBA CCL 0 CML CSR FBG FXJ GIO GMF 0 ICI LLC MAY MIM NBH NCM NCP NDY OSH 0 PAS PBL 0 PDP 0 PLU PNI QNI QRL 0 RIO RSG SEV SGB SRP 0 STO TAH WBC WMC WOW WPL Table 4.5: p-values for the test of normality of e n for the 44 ASX share series, where e n is given by Equation (4.14). These p-values range from 0 to , with 12 being larger than 10%, indicating that for these shares at least, the estimation approximation may be appropriate. It is interesting to note that the p-values are either very small, indicating clear rejection of the hypothesis that the e n are normal, or quite large, where the null hypothesis cannot be rejected at any reasonable significance level.
151 4.4. DATA ANALYSIS 133 Figure 4.6 shows the sample values e n for the stock AMC, which has been randomly selected from the sample of stocks for which the test for normality was rejected. The p-value for this particular stock was , and the graph helps to show why the null hypothesis of normality was rejected. The mean and variance of the sample are and , which are very close to the required values of 0 and 1 respectively. The histogram shows some large positive values, and again, a sharper peak than the normal density function has (positive kurtosis). Removal of the outliers does not improve the Chi-squared statistic, indicating that the main problems are in the centre of the distribution. In particular the shape of the empirical distribution cannot be well described by a normal probability function. Relative Frequency Figure 4.6: Sample values e n, defined in Equation (4.14), for the share price of AMC, with superimposed N(0, 1) density function. The share series that do appear to have the desired normality properties belong to stocks: ANZ, BHP, BOR, CBA, MAY, MIM, NBH, PNI, QNI, RIO, WMC, and WPL. Of these twelve stocks, four belong to each of the
152 134 CHAPTER 4. DATA ANALYSIS three types, with estimates significantly less than 2, greater than 2, or not significantly different. Of these I have chosen BHP, BOR, and MIM for further analysis. Figures 4.7 to 4.9 are of the type seen before, with three graphs featuring in each. In particular these graphs show the share price series, the daily return series with an estimated mean level, and a plot of the estimated standard deviation of the daily returns against the share price level. This final graph is intended only to show the general relationship between share price level and volatility. As in Figures 3.2 and 3.3, superimposed on the daily returns are estimates of the mean level, and lines two estimated standard deviations from this mean. As before, both these averages were determined using the Lowess filter with a smoothing window of 30 days. Figure 4.7 gives the graphs for the BHP share price series. BHP has ˆβ = , and so we should see that the share price is more volatile when the share price is low, and less volatile when it is high. The share price series shows a lot of movement in the share price, with periods of increasing value and decreasing value. The daily returns seem to have a mean that is oscillating slowly about zero, and a standard deviation that is changing over time. In the scatterplot, we see that there appears to be a negative relationship between the share price level and the standard deviation of the daily returns. Figure 4.8 gives the same graphs for the MIM share price series. MIM has ˆβ = , a value that is not significantly different from 2, and hence is a process that could be GBM. The share price series appears similar to that of BHP, with periods of increase and decrease. Once again the daily returns have a mean level that appears to oscillate about zero, and again have a changing standard deviation. The difference between this series and that of BHP, is that the periods of low volatility do not correspond consistently to periods of high share price level, nor do the periods of high volatility correspond to periods of low share price level. This phenomenon is evident
153 4.4. DATA ANALYSIS 135 by comparing the first and second graphs for the series, but is best seen in the third graph. Here, the points seem to be scattered fairly randomly about the graph, with nothing to suggest an increasing or decreasing relationship, consistent with the estimation of β not significantly different from two. The final Figure 4.9 shows the three graphs for the BOR time series. BOR has ˆβ = which indicates that periods of high volatility should correspond to high share price level, and that periods of low volatility should correspond to periods of low share price level. The share series in the first graph does not show the same oscillation as the other two series, but has a period of oscillation to begin with, and then a period of growth in the latter part of the series. Yet again, the daily returns have a mean that appears to oscillate about zero, and a changing standard deviation. The standard deviation of the daily returns is highest at the end of the series, consistent with the size of ˆβ, since this is where the share price is largest. Lower volatility is seen at other times. The third graph indicates that there is a positive relationship between share price level and the standard deviation of the daily returns. This and the relationship seen for the other two series, is probably not linear though, as the specification of the CEV model would suggest. The three share series BHP, MIM and BOR have properties that are consistent with the ˆβ values estimated from them. In addition, the assumptions of the estimation procedure, in particular normality of the e n are satisfied for these series. Thus, it appears that the estimation procedure is giving sensible estimates for these series at least. Further analysis of the ASX share data examined above could focus on the following issues: Can the β estimation procedure be refined to: incorporate the CEV density function and thus construct the exact likelihood function for a CEV process;
154 136 CHAPTER 4. DATA ANALYSIS identify the source of bias, and either eliminate it, or correct for it; determine the asymptotic properties of the estimates? Does ˆβ change significantly over the length of the series? Using the estimates of β and δ, does the CEV model adequately describe market option prices? Does the CEV model provide better forecasts of future share prices than alternative models? Can implied estimates of β and δ be obtained from market option and share prices, in such a way that forecasts of future share price are improved?
155 4.4. DATA ANALYSIS 137 BHP Share Price Time - days Log-returns Time - days Log-return Standard Deviation Share Price Figure 4.7: The BHP share series; in addition, the daily returns for this series with moving mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against standard deviation of the daily returns.
156 138 CHAPTER 4. DATA ANALYSIS MIM Share Price Time - days Log-returns Time - days Log-return Standard Deviation Share Price Figure 4.8: The MIM share series; in addition, the daily returns for this series with moving mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against standard deviation of the daily returns.
157 4.4. DATA ANALYSIS 139 BOR Share Price Time - days Log-returns Time - days Log-return Standard Deviation Share Price Figure 4.9: The BOR share series; in addition, the daily returns for this series with moving mean, and mean ± 2 standard deviation series; thirdly, a plot of the share price level against standard deviation of the daily returns.
158 140 CHAPTER 4. DATA ANALYSIS
159 Chapter 5 Conclusions 141
160 142 CHAPTER 5. CONCLUSIONS The CEV model has not been used extensively in the option pricing literature to describe observed option prices. Reasons for this may be that the option price solution is not an attractive one, and that there is difficulty in estimating the two parameters β and δ. In researching this thesis, I have attempted to gain an understanding of the CEV model, and examine its application to real share series. Furthermore, I have attempted to use this understanding to provide a more detailed analysis of the model than is found in the journal literature. I have described some of the statistical properties of CEV model share and option prices, and for the special case, the GBM and Black-Scholes models. The estimation of β using the method I describe is not altogether satisfactory, since the procedure appears to yield estimates that are biased. However, it does appear to give estimates that are sensible and conform to empirical observations. Moreover, share series are found to be consistent with each of the three CEV classes, corresponding to β < 2, β = 2, and β > 2. This thesis has raised many questions that remain to be answered, but has provided a platform for further research. I look forward to working further on the estimation of β, and analysing market option prices using the CEV model. Specific opportunities for extensions to this analysis are listed at the end of the previous chapter.
161 Appendix A Definitions Throughout this thesis, I have used the following notation consistently. Shown below is a brief description of the symbol, and a page reference to an appropriate definition. Other symbols used in the body of this document are redefined where used. Symbol Description Page β The CEV parameter; 48 ˆβ an estimate of the CEV parameter β; 113 β the sample mean of n estimates of β; 125 B t the value of Brownian motion at time t; 11 C t the value of a European call option (exercisable at T ) 3 at t; δ CEV parameter; 48 ˆδ an estimate of the CEV parameter δ; 113 db t Brownian motion increment over the period [t, t + dt]; 7 ds t share price increment over the period [t, t + dt]; 7 E (.) expectation under risk-neutrality; 51 φ(q) the value of the standard normal probability density 22 function at q; Φ(q) the value of the standard normal cumulative distribution 22 function at q; Φ 2 (x, y; ρ) the bivariate normal cumulative distribution function; 36 f (.) a transition density function under risk-neutrality;
162 144 APPENDIX A. DEFINITIONS Symbol Description Page f Xt+s Y t (x, s) the (transition) density function for the random variable X t+s given the value of another random variable Y t, evaluated at x, and with time interval s elapsed since t; g(x, ν) the gamma probability density function at x with 58 shape parameter ν; G(x, ν) the gamma survivor function at x with shape parameter 58 ν; g t a function that appears in the expected value of Black- 23 Scholes call prices; h t a function that appears in the Black-Scholes formula; 30 h t a function that appears in the density function of a 32 future Black-Scholes price; k CEV variable, featuring in the general transition density 55 function; k risk-neutral CEV variable, with µ = r; 82 K the exercise price of a European call option; 3 L(θ x ) the likelihood function for unknown parameters θ given data x ; 116 l(θ x ) the log-likelihood function for unknown parameters θ given data x ; µ the mean rate of return on a stock; 11 N(µ, σ 2 ) a normal random variable with mean µ and variance σ 2 ; p(x; ν, λ) the density function of a non-central Chi-squared random 71 variable, with ν degrees of freedom, and non- centrality parameter λ; Q(x; ν, λ) the survivor function of a non-central Chi-squared 73 random variable, with ν degrees of freedom, and noncentrality parameter λ; r the continuously compounding risk-free rate; 3 σ 2 the variance of the rate of return on a stock following 11 geometric Brownian motion; S t share price at time t; 3 s(x) the standard deviation of the random variable X; τ the time until exercise of a European call option, T t; 3 t present time, or the index of a time series; T the exercise time of a European call option; 3 116
163 145 Symbol Description Page x CEV variable, featuring in the general transition density 55 function x risk-neutral CEV variable, with µ = r; 82 z CEV variable, featuring in the general transition density 55 function; z risk-neutral CEV variable; 82 Z the standard normal random variable, with zero mean, and unit variance. 11
164 146 APPENDIX A. DEFINITIONS
165 Appendix B Proofs for Selected Results B.1 Result 2.2 Let the share price S t be the solution to the SDE: ds t = µs t dt + σ t S t db t where σ t is a deterministic function of time. Then the price of a call option on a stock with share price S t must satisfy the PDE: 1 2 σ2 t St 2 2 C S + rs C 2 t S + C t r C t = 0 with boundary condition C T = (S t K) +, and the call price C t is given by Black-Scholes equation, with the substitution: σ 2 = σ 2 1 τ T t σ 2 udu. Proof. The form of the PDE follows from Equation (2.18), with σ(s t, t) = σ t. The Black-Scholes function, with can be written: σ 2 = σ 2 1 τ T t C t = C(S t, τ, σ) σ 2 udu (B.1) 147
166 148 APPENDIX B. PROOFS FOR SELECTED RESULTS and this will satisfy the PDE: 1 2 σ2 t St 2 2 C S + rs C 2 t S + C t r C t = 0. (B.2) These partial derivatives are given by: C S = C S σ= σ 2 C S = 2 C 2 S 2 σ= σ (B.3) and C t = C t + C σ= σ σ σ t σ= σ where σ is given by Equation (B.1). Now C σ = S tφ(h t ) τ = τσst 2 2 C σ= σ S 2 since 2 C S 2 = φ(ht) Sσ. In addition: τ ( σ t = 1 T σ t udu 2 2 σ t T t ) σ= σ = σ2 σ 2 t 2τ σ yielding C t = C t ( σ2 σt 2 )St 2 2 C σ= σ S 2 σ= σ (B.4) and so substituting the partial derivatives given in Equations (B.3) and (B.4) into the PDE (B.2) gives: LHS = 1 2 σ2 t St 2 2 C S + rs C 2 t S + C t r C t = 1 2 σ2 t St 2 2 C S + rs C 2 t S + C t ( σ2 σt 2 )St 2 = 1 2 σ2 St 2 2 C S + rs C 2 t S + C t rc t = 0 by Equation (2.19) as required. σ= σ 2 C S 2 rc t σ= σ
167 B.2. RESULT B.2 Result 3.2 G(y, ν 1) = 1 g(y, n + ν) where G(y, ν) and g(y, ν) are the survivor and probability density functions at y for a gamma random variable with shape parameter ν and unit scale parameter. Proof. Using techniques employed by Schroder (1989), the gamma survivor function G(y, ν) may be written in terms of a recurrence relation, found using integration by parts: G(y, ν) = = y y g(x, ν)dx = e x x ν 1 Γ(ν) = e y y ν 1 Γ(ν) e x x ν 1 dx Γ(ν) + + n=0 y y y e x (ν 1)x ν 2 (ν 1)Γ(ν 1) dx e x x ν 2 Γ(ν 1) dx = g(y, ν) + G(y, ν 1). (B.5) Applying this relationship n + 1 times to G(y, n + ν), we obtain: G(y, n + ν) = g(y, n + ν) + G(y, n 1 + ν) = g(y, n + ν) + g(y, n 1 + ν) + G(y, n 2 + ν). = n g(y, k + ν) + G(y, ν 1). k=0 Hence, letting n, we obtain: g(y, n + ν) + G(y, ν 1) = lim G(y, n + ν) n n=0 (B.6) = 1 lim n P (Y n+ν < y) = 1 (B.7)
168 150 APPENDIX B. PROOFS FOR SELECTED RESULTS where Y n+ν is a gamma random variable with parameter n+ν, and the result follows. B.3 Result 3.4 y 2 p(2x; 2ν + 2, 2λ)dλ = 1 x 2 p(2z; 2ν, 2y)dz where p(y; ν, λ) and Q(y; ν, λ) are the probability density and survivor functions at y for a non-central Chi-squared random variable with ν degrees of freedom and non-centrality parameter λ. Proof. This proof follows the method of Schroder (1989). Note firstly that there are two alternative formulae for call price C t given in Equation (3.34): ( ) e x x n G kk 2 β, n β C t = S t Γ(n + 1) n=0 ( Ke rτ e x x n+ 1 2 β G kk 2 β, n + 1 ) Γ(n ) 2 β and Equation (3.43) n=0 C t = S t 2 p(2z; 2 + 2, 2x)dz 2 β Ke rτ y y 2 p(2x; β, 2z)dz. These equations allow us to note directly that: y y 2 p(2z; 2 + 2, 2x)dz = 2 β n=0 2 p(2x; 2 + 2, 2z)dz = 2 β n=0 A result from the proof in Appendix B.2 is: G(y, n + 1) = g(x, n + 1)G(y, n β ) g(x, n )G(y, n + 1). (B.9) 2 β n g(y, i + 1) = i=0 n g(y, i) i=1 (B.8) (B.10) which follows since G(y, 1) = g(y, 1).
169 B.3. RESULT Letting ν = 1 + 1, Equation (B.8) yields: 2 β x 2 p(2z; 2ν, 2y)dz = g(y, n + 1)G(x, n + ν) n=0 and so developing the right hand side of this expression: RHS = = = g(y, n + 1)G(x, n + ν) n=0 ( g(y, n + 1) G(x, ν 1) + n=0 ( g(y, n + 1) 1 n=0 k=n+1 ) n g(x, k + ν) k=0 g(x, k + ν) where the last two equalities follow from Equations (B.6) and (B.7) respectively. Noting that: g(y, i) = i=1 i=0 e y y i Γ(i + 1) = e y i=0 ) y i i! = e y e y = 1 we can expand the final equation above, and change the order of summation to yield: x 2 p(2z; 2ν, 2y)dz = 1 = 1 = 1 n=0 k=n+1 g(y, n + 1) g(x, k + ν) k 1 g(y, n + 1) g(x, k + ν) k=1 n=0 g(x, k + ν + 1) k=0 k g(y, n) Now noting Equation (B.10), the summation in the final equation can be written: g(x, k + ν + 1) k=0 k g(y, n) = n=1 = n=1 g(x, k + ν + 1)G(y, k + ν) k=0 y 2 p(2x; β, 2z)dz
170 152 APPENDIX B. PROOFS FOR SELECTED RESULTS by Equation (B.9) and making the substitution ν = 1. Hence the result 2 β is proved. y 2 p(2x; 2ν + 2, 2λ)dλ = 1 x 2 p(2z; 2ν, 2y)dz (B.11) B.4 Result 3.5 The option price when β > 2 is given by the formula: C t = S t Q(2x; 2 β 2, 2y) Ke rτ (1 Q(2y; β 2, 2x)) where Q(y; ν, λ) is the survivor function at y, for a non-central Chi-squared variable with ν degrees of freedom, and non-centrality parameter λ. Proof. This proof follows the derivation of Schroder s pricing formula for the standard CEV case, where β < 2. The risk-neutral density function given in the first part of this result is: fs T S t (s, τ β > 2) = (β 2)k 1/(2 β) (xz 1 2β ) β e x z I 1/(β 2) (2 xz). Emanuel and MacBeth point out that this function has a singularity at s = and so any integration should be over a range excluding. Hence, they write the call price at t as the share price at t, less the discounted exercise price, plus the present value of the amount saved by letting the option lapse when S T < K, which follows since: S T K + (K S T ) + = S T K + = (S T K) + 1 { 0 K < S T K S T K S T as required. The present value of the saving (K S T ) + is: K e rτ (K s)fs T S t (s, τ)ds s=0
171 B.4. RESULT and so the option price is given by: K C t = S t Ke rτ + e rτ (K s)fs T S t (s, τ)ds. Note that z is unchanged from its previous definition, and so using we can write: s=0 ds = k 1/(2 β) 1 2 β z 1+1/(2 β) dz f S T S t (s, τ)ds = (z/x) β e x z I 1 (2 xz)dz. β 2 Noting that when s = 0 then z =, and when s = K then z = kk 2 β = y, we find in this case the transformation is a negative one, and so the option price is: ( (z C t = S t Ke rτ e rτ y k ) 1 2 β K ) ( z ) 1/(4 2β) e x z I 1 (2 xz)dz. x β 2 In the proof of Theorem 3.8, we saw e rτ (x/k) 1/(2 β) = S t which leads to yet another representation of the formula for C t : C t = S t (1 e x z (x/z) 1/(2β 4) I 1/(β 2) (2 ) xz)dz y Ke (1 rτ e x z (z/x) 1/(2β 4) I 1/(β 2) (2 ) xz)dz. y The form of the non-central Chi-squared probability density function, p(x; ν, λ) is given in Equation (3.23), and comparison of the above integrands with this function yields the formula: C t = S t (1 y 2 p(2x; β 2, 2z)dz ) Ke rτ (1 y 2 p(2z; β 2, 2x)dz ).
172 154 APPENDIX B. PROOFS FOR SELECTED RESULTS Using Result 3.4, and making the simple transformations seen in the proof of Theorem 3.9, yields the option pricing formula given by Schroder: as required. C t = S t Q(2x; 2 β 2, 2y) Ke rτ (1 Q(2y; β 2, 2x))
173 Appendix C Complete List of Shares Table C.1: ASX Company Codes Stock Code AAA AMC ANI ANZ BHP BIL BOR BPC BRY CBA CCL CML CSR FBG FXJ GIO GMF ICI LLC MAY MIM NBH NCM Company Name Acacia Resources Limited Amcor Limited Australian National Industries Limited Australia & New Zealand Banking Group Limited The Broken Hill Proprietory Company Limited Brambles Industries Limited Boral Limited Burns, Philp & Company Limited Brierley Investments Limited Commonwealth Bank of Australia Coca-Cola Amatil Limited Coles Myer Limited CSR Limited Foster s Brewing Group Limited Fairfax (John) Holdings Limited GIO Australia Holdings Limited Goodman Fielder Limited ICI Australia Limited Lend Lease Corporation Limited Mayne Nickless Limited M.I.M. Holdings Limited North Limited Newcrest Mining Limited continued on following page 155
174 156 APPENDIX C. COMPLETE LIST OF SHARES continued from previous page Stock Code Company Name NCP The News Corporation Limited NDY Normandy Mining Limited OSH Oil Search Limited PAS Pasminco Limited PBL Publishing & Broadcasting Limited PDP Pacific Dunlop Limited PLU Plutonic Resources Limited PNI Pioneer International Limited QNI QNI Limited QRL QCT Resources Limited RIO Rio Tinto Limited RSG Resolute Limited SEV Seven Network Limited SGB St. George Bank Limited SRP Southcorp Limited STO Santos Limited TAH Tabcorp Holdings Limited WBC Westpac Banking Corporation WMC WMC Limited WOW Woolworths Limited WPL Woodside Petroleum Limited
175 Appendix D SPLUS Code D.1 GBM Simulation The following program is an appropriate way of simulating a geometric Brownian motion series in SPLUS. It follows from Equation (2.8), and takes arguments S t, τ, µ, σ, and n the number of subintervals in [t, T ]. It returns share with n + 1 elements, which is a realisation of the GBM process over the time period [t, T ]. gbm.f <- function(s,tau,mu,sigma,n){ } dt <- tau/n db <- rnorm(n,0,sqrt(dt)) (1) share <- c(s,s*exp(cumsum((mu - 0.5*sigma^2)*dt + sigma*db))) share I have labelled the key parts of the program, and will explain the function of these. Line (1) generates n independent N(0, t) random variables, where (2) t = τ. These are the realisations of the n Brownian motion incre- n ments needed to construct {S t }. 157
176 158 APPENDIX D. SPLUS CODE Line (2) computes {S t } using the relationship given in Equation (2.8). A sample of N future prices, all at the smae time T, can be obtained using the following program. It is an alternative to setting n = 1 in bs.f above, and repeating N times. gbm.sample.f <- function(n,s,tau,mu,sigma){ } B <- rnorm(n,0,sqrt(tau)) shares <- S*exp((mu - 0.5*sigma^2)*tau + sigma*b) shares Here, N realisations of an N(0, τ) random variable are generated, and are used to construct S T directly. D.2 Inversion of the Black-Scholes Formula In order to invert the Black-Scholes formula with respect to any of its five arguments a numerical method is required. The Newton-Raphson method is ideal in this case, since the first derivative of the Black-Scholes option price C t = S t Φ(h t ) Ke rτ Φ(h t σ τ) with respect to S t, τ, K, r and σ are known. Result D.1 (The Newton-Raphson Algorithm). The Newton-Raphson algorithm states that if a is an approximate root of the function f(x), then is generally a better approximation. a f(a) f (a) Let C mt be the observed market option price at time t, and let C t = C(S t, τ, K, r, σ).
177 D.2. INVERSION OF THE BLACK-SCHOLES FORMULA 159 The partial derivatives of this function with respect to its arguments can be shown to be 1 : C = Φ(h t ) S t C τ = S tσφ(h t ) 2 + rke rτ Φ(h t σ τ) τ C K = e rτ Φ(h t σ τ) C r = Kτe rτ Φ(h t σ τ) C σ = S t τφ(ht ). These may be used in the Newton-Raphson result to obtain implied parameters of the Black-Scholes function. For example, the value of S that satisfies the equation c = C(S; τ, K, r, σ) where τ, K, r and σ are held constant, can be found in SPLUS using the function bsinvs.f below: bsinv.f<-function(tau,k,r,sigma,c,a=4,n=100,precision=5e-07) { for(i in 1:n){ } ht<-(log(a[i]/k)+(r+0.5*sigma^2)*tau)/sqrt(sigma^2*tau) a[i+1]<-a[i]-(bs.f(a[i],tau,k,r,sigma)-c)/pnorm(ht) (1) diff<-abs(a[i+1]-a[i]) if(diff<precision)return(a[i+1],i) (2) } This function has required arguments: τ, K, r, σ and C m, and optional arguments: a - the initial guess of S, n - the maximum number of iterations 1 Note that C S, C τ, C C C r and σ are always positive, and K is consistent with discussion in Chapter 1. is always negative. This
178 160 APPENDIX D. SPLUS CODE - and precision - the precision required in the estimate of S, set initially to The estimate of S is returned with the number of iterations needed for convergence to within the specified (or default) precision. I have labelled two lines of the program, which have the following function. Line (1) evaluates the Newton-Raphson result, using an initial estimate for the root, a[i], the Black-Scholes function bs.f, which simply calculates the Black-Scholes price with parameters S t = a[i], τ = tau, K = K, r = r, and σ = sigma, and the first partial derivative of the Black-Scholes option price with respect to share price: Φ(h t ) = pnorm(ht), to obtain a better approximation to the root a[i+1]. Line (2) returns the estimate of the root if the process has converged according to the argument precision. In addition, the number of iterations is returned. If the process does not converge within n iterations, the program returns an error message. D.3 CEV Share Price Simulation The following program can be used to simulate the solution to the constant elasticity of variance SDE (3.1). It approximates the solution using the Euler method as described in Section 3.2.2, and has arguments S t, τ, µ, σ = δs β/2 1 t, β and n the number of subintervals in [t, T ]. The share price is initially set at zero for the length of the series, and then replaced if S ti positive. The program returns share with n + 1 elements. is cevs.f<-function(s,tau,mu,sigma,beta,n,db=rnorm(n,0,sqrt(tau/n))) { share <- c(s,rep(0,n)) ds <- NULL
179 D.4. ESTIMATION OF β 161 delta <- sigma*s^(1-beta/2) (1) for ( i in 1:n ) { ds <- mu*share[i]+tau/n + delta*share[i]^(beta/2)*db[i] (2) if (share+ds>0) {share[i+1] <- share[i]+ds} (3) else return(share) } share } This program is fairly similar to that used to generate the GBM price series, however in this case the share prices are calculated using Equation (3.19). Specifically, the three numbered lines have the following functions. Line (1) obtains a value of δ for the given values S t, β and σ. This is equivalent to the alignment procedure of MacBeth & Merville (1980) discussed in Section Line (2) calculates the share price increment ds ti. Line (3) calculates the share price S ti+1 = S ti + ds ti, however if this is negative the simulation process is stopped and the present and future values of the series are set equal to zero. Regardless of bankruptcy, the entire series (of length n + 1) is returned. D.4 Estimation of β The following program can be used to estimate the value of β from a share series. The function shown in Equation (4.11) is constructed and minimised for the given share and dividend series. betaest2.f_function(share,dividends=rep(0,length(share))){ N_length(share)
180 162 APPENDIX D. SPLUS CODE share2_share+dividends r_log(share2[-1])-log(share[-length(share)]) (1) m_length(r) ave_lowess(r,f=30/m) (2) r_r-ave$y r_r[21:(length(r)-20)] (3) share_share[21:(length(share)-21)] (4) assign("share",share,frame=1) (5) assign("r",r,frame=1) (6) result_nlmin(function(z){ length(r)*log(mean(r^2/(share^z)))+z*sum(log(share)) },0)[[1]] (7) } list(beta=result+2,share=share,r=r) (8) The function has required argument {S t }, a series of closing share prices, and {d t }, the series of dividends paid to holders of the share. The two series should be aligned so that element i of each series is for the same day. have labelled eight lines of the program as follows (1), and will explain the function of these. Line (1) constructs the daily returns for the share series, where the return on day i is given by: ( ) Si + d i r i = ln. S i 1 There are m observations in this series where there were N = m + 1 observations in the share series. Line (2) estimates the mean level µ(y n ) of the log return series, since E(ln S n+1 ln S n ) = µ(y n ) is given by Equation (4.2). The Lowess filter is used for this purpose with a smoothing window of 30 observations. I
181 D.4. ESTIMATION OF β 163 Lines (3) and (4) remove 20 observations from the ends of both the share price series, and the mean corrected log-returns. In order to estimate µ(y n ) at the ends, the Lowess filter estimated observations to maintain a symmetric 30 observation window. By removing 20 observations from each end of the series, the effects of this estimation will not affect the estimation of β. Lines (5) and (6) pass the data share and r to the minimisation procedure. Line (7) defines the function which must be minimised, g(β), and orders the minimisation procedure nlmin to return the value of β which minimises g(β). It will begin its search at β = 2 corresponding to z=0. Line (8) returns the estimated value for β, and the share price and mean corrected return series.
182 164 APPENDIX D. SPLUS CODE
183 Bibliography Abramowitz, M. & Stegun, I. A., eds (1968), Handbook of Mathematical Functions, Dover Publications, Inc. Beckers, S. (1980), The constant elasticity of variance model and its implications for option pricing, Journal of Finance 35(3), Beckers, S. (1981), Standard deviations implied in option prices as predictors of future stock price volatility, Journal of Banking and Finance 5, Black, F. (1975), Fact and fantasy in the use of options, Financial Analysts Journal 31(4), 36 41, Black, F. & Scholes, M. (1972), The valuation of option contracts and a test of market efficiency, Journal of Finance 27(2), Black, F. & Scholes, M. (1973), The pricing of options an corporate liabilities, Journal of Political Economy 81(3), Chiras, D. & Manaster, S. (1978), The information content of option prices and a test of market efficiency, Journal of Financial Economics 6, Christie, A. A. (1982), The stochastic behaviour of common stock variances, Journal of Financial Economics 10,
184 166 BIBLIOGRAPHY Cleveland, W. S. (1979), Robust locally weighted regression and smoothing scatterplots, Journal of the American Statistical Association 74, Copeland, T. E. & Weston, J. F. (1988), Financial Theory and Corporate Policy, third edn, Addison-Wesley. Cox, D. & Miller, H. (1965), The Theory of Stochastic Processes, Methuen and Co. Ltd. Cox, J. C. (1996), The constant elasticity of variance option pricing model, The Journal of Portfolio Management Special Issue, Cox, J. C. & Ross, S. A. (1976), The valuation of options for alternative stochastic processes, Journal of Financial Economics 3, Emanuel, D. C. & MacBeth, J. D. (1982), Further results on the constant elasticity of variance call option pricing model, Journal of Financial and Quantitative Analysis 17(4), Feller, W. (1951), Two singular diffusion problems, Annals of Mathematics 54(1), Geske, R. (1979), The valuation of compound options, Journal of Financial Economics 7, Harrison, J. M. & Kreps, D. M. (1979), Martingales and arbitrage in multiperiod securities markets, Journal of Economic Theory 20, Hull, J. C. (1997), Options, Futures, and Other Derivatives, 3rd edn, Prentice-Hall, Inc. Jarrow, R. A. & Rudd, A. (1983), Option Pricing, R. D. Irwin. Johnson, N. L., Kotz, S. & Balakrishnan, N. (1995), Continuous Univariate Distributions, Vol. 2, second edn, John Wiley & Sons, Inc.
185 BIBLIOGRAPHY 167 Kloeden, P. & Platen, E. (1992), Numerical solution of stochastic differential equations, Springer-Verlag. Latané, H. A. & Rendleman, R. J. (1976), Standard deviations of stock price ratios implied in option prices, Journal of Finance 31(2), Lauterbach, B. & Schultz, P. (1990), Pricing warrants: An empirical study of the Black-Scholes model and its alternatives, Journal of Finance 45(4), MacBeth, J. D. & Merville, L. J. (1980), Tests of the Black-Scholes and Cox call option valuation models, Journal of Finance 35(2), Manaster, S. (1980), Discussion, Journal of Finance 35(2), Marsh, T. & Rosenfeld, E. (1983), Stochastic processes for interest rates and equilibrium bond prices, Journal of Finance 38(2), Mayhew, S. (1995), Implied volatility, Financial Analysts Journal 51(4), Melino, A. & Turnbull, S. (1991), The pricing of foreign currency options, Canadian Journal of Economics 24, Merton, R. (1973), Theory of rational option pricing, The Bell Journal of Economics and Management Science 4, Mikosch, T. (1994), An elementary introduction to stochastic differential equations, ISOR Lecture Note Series, Victoria University of Wellington. Rubinstein, M. (1985), Nonparametric tests of alternative option pricing models using all reported trades and quotes on the 30 most active CBOE option classes from August 23, 1976 through August 31, 1978, Journal of Finance 40(2),
186 168 BIBLIOGRAPHY Schroder, M. (1989), Computing the constant elasticity of variance option pricing formula, Journal of Finance 44(1), Tucker, A. L., Peterson, D. R. & Scott, E. (1988), Tests of the Black-Scholes and constant elasticity of variance currency call option valuation models., The Journal of Financial Research 11(3),
The Black-Scholes Formula
FIN-40008 FINANCIAL INSTRUMENTS SPRING 2008 The Black-Scholes Formula These notes examine the Black-Scholes formula for European options. The Black-Scholes formula are complex as they are based on the
Black-Scholes Equation for Option Pricing
Black-Scholes Equation for Option Pricing By Ivan Karmazin, Jiacong Li 1. Introduction In early 1970s, Black, Scholes and Merton achieved a major breakthrough in pricing of European stock options and there
ARBITRAGE-FREE OPTION PRICING MODELS. Denis Bell. University of North Florida
ARBITRAGE-FREE OPTION PRICING MODELS Denis Bell University of North Florida Modelling Stock Prices Example American Express In mathematical finance, it is customary to model a stock price by an (Ito) stochatic
Valuing Stock Options: The Black-Scholes-Merton Model. Chapter 13
Valuing Stock Options: The Black-Scholes-Merton Model Chapter 13 Fundamentals of Futures and Options Markets, 8th Ed, Ch 13, Copyright John C. Hull 2013 1 The Black-Scholes-Merton Random Walk Assumption
The Black-Scholes-Merton Approach to Pricing Options
he Black-Scholes-Merton Approach to Pricing Options Paul J Atzberger Comments should be sent to: atzberg@mathucsbedu Introduction In this article we shall discuss the Black-Scholes-Merton approach to determining
Call Price as a Function of the Stock Price
Call Price as a Function of the Stock Price Intuitively, the call price should be an increasing function of the stock price. This relationship allows one to develop a theory of option pricing, derived
Does Black-Scholes framework for Option Pricing use Constant Volatilities and Interest Rates? New Solution for a New Problem
Does Black-Scholes framework for Option Pricing use Constant Volatilities and Interest Rates? New Solution for a New Problem Gagan Deep Singh Assistant Vice President Genpact Smart Decision Services Financial
Monte Carlo Methods in Finance
Author: Yiyang Yang Advisor: Pr. Xiaolin Li, Pr. Zari Rachev Department of Applied Mathematics and Statistics State University of New York at Stony Brook October 2, 2012 Outline Introduction 1 Introduction
Institute of Actuaries of India Subject CT3 Probability and Mathematical Statistics
Institute of Actuaries of India Subject CT3 Probability and Mathematical Statistics For 2015 Examinations Aim The aim of the Probability and Mathematical Statistics subject is to provide a grounding in
Notes on Black-Scholes Option Pricing Formula
. Notes on Black-Scholes Option Pricing Formula by De-Xing Guan March 2006 These notes are a brief introduction to the Black-Scholes formula, which prices the European call options. The essential reading
On Black-Scholes Equation, Black- Scholes Formula and Binary Option Price
On Black-Scholes Equation, Black- Scholes Formula and Binary Option Price Abstract: Chi Gao 12/15/2013 I. Black-Scholes Equation is derived using two methods: (1) risk-neutral measure; (2) - hedge. II.
Jorge Cruz Lopez - Bus 316: Derivative Securities. Week 11. The Black-Scholes Model: Hull, Ch. 13.
Week 11 The Black-Scholes Model: Hull, Ch. 13. 1 The Black-Scholes Model Objective: To show how the Black-Scholes formula is derived and how it can be used to value options. 2 The Black-Scholes Model 1.
第 9 讲 : 股 票 期 权 定 价 : B-S 模 型 Valuing Stock Options: The Black-Scholes Model
1 第 9 讲 : 股 票 期 权 定 价 : B-S 模 型 Valuing Stock Options: The Black-Scholes Model Outline 有 关 股 价 的 假 设 The B-S Model 隐 性 波 动 性 Implied Volatility 红 利 与 期 权 定 价 Dividends and Option Pricing 美 式 期 权 定 价 American
CS 522 Computational Tools and Methods in Finance Robert Jarrow Lecture 1: Equity Options
CS 5 Computational Tools and Methods in Finance Robert Jarrow Lecture 1: Equity Options 1. Definitions Equity. The common stock of a corporation. Traded on organized exchanges (NYSE, AMEX, NASDAQ). A common
Black-Scholes-Merton approach merits and shortcomings
Black-Scholes-Merton approach merits and shortcomings Emilia Matei 1005056 EC372 Term Paper. Topic 3 1. Introduction The Black-Scholes and Merton method of modelling derivatives prices was first introduced
Mathematical Finance
Mathematical Finance Option Pricing under the Risk-Neutral Measure Cory Barnes Department of Mathematics University of Washington June 11, 2013 Outline 1 Probability Background 2 Black Scholes for European
Black-Scholes Option Pricing Model
Black-Scholes Option Pricing Model Nathan Coelen June 6, 22 1 Introduction Finance is one of the most rapidly changing and fastest growing areas in the corporate business world. Because of this rapid change,
Moreover, under the risk neutral measure, it must be the case that (5) r t = µ t.
LECTURE 7: BLACK SCHOLES THEORY 1. Introduction: The Black Scholes Model In 1973 Fisher Black and Myron Scholes ushered in the modern era of derivative securities with a seminal paper 1 on the pricing
The Black-Scholes pricing formulas
The Black-Scholes pricing formulas Moty Katzman September 19, 2014 The Black-Scholes differential equation Aim: Find a formula for the price of European options on stock. Lemma 6.1: Assume that a stock
Option pricing. Vinod Kothari
Option pricing Vinod Kothari Notation we use this Chapter will be as follows: S o : Price of the share at time 0 S T : Price of the share at time T T : time to maturity of the option r : risk free rate
Hedging Illiquid FX Options: An Empirical Analysis of Alternative Hedging Strategies
Hedging Illiquid FX Options: An Empirical Analysis of Alternative Hedging Strategies Drazen Pesjak Supervised by A.A. Tsvetkov 1, D. Posthuma 2 and S.A. Borovkova 3 MSc. Thesis Finance HONOURS TRACK Quantitative
Derivation and Comparative Statics of the Black-Scholes Call and Put Option Pricing Formulas
Derivation and Comparative Statics of the Black-Scholes Call and Put Option Pricing Formulas James R. Garven Latest Revision: 27 April, 2015 Abstract This paper provides an alternative derivation of the
Pricing Barrier Options under Local Volatility
Abstract Pricing Barrier Options under Local Volatility Artur Sepp Mail: [email protected], Web: www.hot.ee/seppar 16 November 2002 We study pricing under the local volatility. Our research is mainly
Four Derivations of the Black Scholes PDE by Fabrice Douglas Rouah www.frouah.com www.volopta.com
Four Derivations of the Black Scholes PDE by Fabrice Douglas Rouah www.frouah.com www.volopta.com In this Note we derive the Black Scholes PDE for an option V, given by @t + 1 + rs @S2 @S We derive the
Invesco Great Wall Fund Management Co. Shenzhen: June 14, 2008
: A Stern School of Business New York University Invesco Great Wall Fund Management Co. Shenzhen: June 14, 2008 Outline 1 2 3 4 5 6 se notes review the principles underlying option pricing and some of
How To Price A Call Option
Now by Itô s formula But Mu f and u g in Ū. Hence τ θ u(x) =E( Mu(X) ds + u(x(τ θ))) 0 τ θ u(x) E( f(x) ds + g(x(τ θ))) = J x (θ). 0 But since u(x) =J x (θ ), we consequently have u(x) =J x (θ ) = min
On Market-Making and Delta-Hedging
On Market-Making and Delta-Hedging 1 Market Makers 2 Market-Making and Bond-Pricing On Market-Making and Delta-Hedging 1 Market Makers 2 Market-Making and Bond-Pricing What to market makers do? Provide
Lecture Notes: Basic Concepts in Option Pricing - The Black and Scholes Model
Brunel University Msc., EC5504, Financial Engineering Prof Menelaos Karanasos Lecture Notes: Basic Concepts in Option Pricing - The Black and Scholes Model Recall that the price of an option is equal to
Pricing of an Exotic Forward Contract
Pricing of an Exotic Forward Contract Jirô Akahori, Yuji Hishida and Maho Nishida Dept. of Mathematical Sciences, Ritsumeikan University 1-1-1 Nojihigashi, Kusatsu, Shiga 525-8577, Japan E-mail: {akahori,
Option Pricing. Chapter 4 Including dividends in the BS model. Stefan Ankirchner. University of Bonn. last update: 6th November 2013
Option Pricing Chapter 4 Including dividends in the BS model Stefan Ankirchner University of Bonn last update: 6th November 2013 Stefan Ankirchner Option Pricing 1 Dividend payments So far: we assumed
Review of Basic Options Concepts and Terminology
Review of Basic Options Concepts and Terminology March 24, 2005 1 Introduction The purchase of an options contract gives the buyer the right to buy call options contract or sell put options contract some
THE BLACK-SCHOLES MODEL AND EXTENSIONS
THE BLAC-SCHOLES MODEL AND EXTENSIONS EVAN TURNER Abstract. This paper will derive the Black-Scholes pricing model of a European option by calculating the expected value of the option. We will assume that
Option Valuation. Chapter 21
Option Valuation Chapter 21 Intrinsic and Time Value intrinsic value of in-the-money options = the payoff that could be obtained from the immediate exercise of the option for a call option: stock price
Merton-Black-Scholes model for option pricing. Peter Denteneer. 22 oktober 2009
Merton-Black-Scholes model for option pricing Instituut{Lorentz voor Theoretische Natuurkunde, LION, Universiteit Leiden 22 oktober 2009 With inspiration from: J. Tinbergen, T.C. Koopmans, E. Majorana,
Introduction to Options. Derivatives
Introduction to Options Econ 422: Investment, Capital & Finance University of Washington Summer 2010 August 18, 2010 Derivatives A derivative is a security whose payoff or value depends on (is derived
Journal of Financial and Economic Practice
Analyzing Investment Data Using Conditional Probabilities: The Implications for Investment Forecasts, Stock Option Pricing, Risk Premia, and CAPM Beta Calculations By Richard R. Joss, Ph.D. Resource Actuary,
A Simulation-Based lntroduction Using Excel
Quantitative Finance A Simulation-Based lntroduction Using Excel Matt Davison University of Western Ontario London, Canada CRC Press Taylor & Francis Croup Boca Raton London New York CRC Press is an imprint
Computational Finance Options
1 Options 1 1 Options Computational Finance Options An option gives the holder of the option the right, but not the obligation to do something. Conversely, if you sell an option, you may be obliged to
Lecture 12: The Black-Scholes Model Steven Skiena. http://www.cs.sunysb.edu/ skiena
Lecture 12: The Black-Scholes Model Steven Skiena Department of Computer Science State University of New York Stony Brook, NY 11794 4400 http://www.cs.sunysb.edu/ skiena The Black-Scholes-Merton Model
Black Scholes Merton Approach To Modelling Financial Derivatives Prices Tomas Sinkariovas 0802869. Words: 3441
Black Scholes Merton Approach To Modelling Financial Derivatives Prices Tomas Sinkariovas 0802869 Words: 3441 1 1. Introduction In this paper I present Black, Scholes (1973) and Merton (1973) (BSM) general
A Genetic Algorithm to Price an European Put Option Using the Geometric Mean Reverting Model
Applied Mathematical Sciences, vol 8, 14, no 143, 715-7135 HIKARI Ltd, wwwm-hikaricom http://dxdoiorg/11988/ams144644 A Genetic Algorithm to Price an European Put Option Using the Geometric Mean Reverting
EXP 481 -- Capital Markets Option Pricing. Options: Definitions. Arbitrage Restrictions on Call Prices. Arbitrage Restrictions on Call Prices 1) C > 0
EXP 481 -- Capital Markets Option Pricing imple arbitrage relations Payoffs to call options Black-choles model Put-Call Parity Implied Volatility Options: Definitions A call option gives the buyer the
Understanding Options and Their Role in Hedging via the Greeks
Understanding Options and Their Role in Hedging via the Greeks Bradley J. Wogsland Department of Physics and Astronomy, University of Tennessee, Knoxville, TN 37996-1200 Options are priced assuming that
Black and Scholes - A Review of Option Pricing Model
CAPM Option Pricing Sven Husmann a, Neda Todorova b a Department of Business Administration, European University Viadrina, Große Scharrnstraße 59, D-15230 Frankfurt (Oder, Germany, Email: [email protected],
BINOMIAL OPTIONS PRICING MODEL. Mark Ioffe. Abstract
BINOMIAL OPTIONS PRICING MODEL Mark Ioffe Abstract Binomial option pricing model is a widespread numerical method of calculating price of American options. In terms of applied mathematics this is simple
Options: Valuation and (No) Arbitrage
Prof. Alex Shapiro Lecture Notes 15 Options: Valuation and (No) Arbitrage I. Readings and Suggested Practice Problems II. Introduction: Objectives and Notation III. No Arbitrage Pricing Bound IV. The Binomial
Jung-Soon Hyun and Young-Hee Kim
J. Korean Math. Soc. 43 (2006), No. 4, pp. 845 858 TWO APPROACHES FOR STOCHASTIC INTEREST RATE OPTION MODEL Jung-Soon Hyun and Young-Hee Kim Abstract. We present two approaches of the stochastic interest
Option Pricing. Chapter 11 Options on Futures. Stefan Ankirchner. University of Bonn. last update: 13/01/2014 at 14:25
Option Pricing Chapter 11 Options on Futures Stefan Ankirchner University of Bonn last update: 13/01/2014 at 14:25 Stefan Ankirchner Option Pricing 1 Agenda Forward contracts Definition Determining forward
The Binomial Option Pricing Model André Farber
1 Solvay Business School Université Libre de Bruxelles The Binomial Option Pricing Model André Farber January 2002 Consider a non-dividend paying stock whose price is initially S 0. Divide time into small
Constant Elasticity of Variance (CEV) Option Pricing Model:Integration and Detailed Derivation
Constant Elasticity of Variance (CEV) Option Pricing Model:Integration and Detailed Derivation Ying-Lin Hsu Department of Applied Mathematics National Chung Hsing University Co-authors: T. I. Lin and C.
Convenient Conventions
C: call value. P : put value. X: strike price. S: stock price. D: dividend. Convenient Conventions c 2015 Prof. Yuh-Dauh Lyuu, National Taiwan University Page 168 Payoff, Mathematically Speaking The payoff
Black-Scholes and the Volatility Surface
IEOR E4707: Financial Engineering: Continuous-Time Models Fall 2009 c 2009 by Martin Haugh Black-Scholes and the Volatility Surface When we studied discrete-time models we used martingale pricing to derive
FIN-40008 FINANCIAL INSTRUMENTS SPRING 2008
FIN-40008 FINANCIAL INSTRUMENTS SPRING 2008 Options These notes consider the way put and call options and the underlying can be combined to create hedges, spreads and combinations. We will consider the
Lecture 6: Option Pricing Using a One-step Binomial Tree. Friday, September 14, 12
Lecture 6: Option Pricing Using a One-step Binomial Tree An over-simplified model with surprisingly general extensions a single time step from 0 to T two types of traded securities: stock S and a bond
Chapter 2: Binomial Methods and the Black-Scholes Formula
Chapter 2: Binomial Methods and the Black-Scholes Formula 2.1 Binomial Trees We consider a financial market consisting of a bond B t = B(t), a stock S t = S(t), and a call-option C t = C(t), where the
Lectures. Sergei Fedotov. 20912 - Introduction to Financial Mathematics. No tutorials in the first week
Lectures Sergei Fedotov 20912 - Introduction to Financial Mathematics No tutorials in the first week Sergei Fedotov (University of Manchester) 20912 2010 1 / 1 Lecture 1 1 Introduction Elementary economics
Lecture. S t = S t δ[s t ].
Lecture In real life the vast majority of all traded options are written on stocks having at least one dividend left before the date of expiration of the option. Thus the study of dividends is important
The Behavior of Bonds and Interest Rates. An Impossible Bond Pricing Model. 780 w Interest Rate Models
780 w Interest Rate Models The Behavior of Bonds and Interest Rates Before discussing how a bond market-maker would delta-hedge, we first need to specify how bonds behave. Suppose we try to model a zero-coupon
LECTURE 15: AMERICAN OPTIONS
LECTURE 15: AMERICAN OPTIONS 1. Introduction All of the options that we have considered thus far have been of the European variety: exercise is permitted only at the termination of the contract. These
LOGNORMAL MODEL FOR STOCK PRICES
LOGNORMAL MODEL FOR STOCK PRICES MICHAEL J. SHARPE MATHEMATICS DEPARTMENT, UCSD 1. INTRODUCTION What follows is a simple but important model that will be the basis for a later study of stock prices as
Lecture 4: The Black-Scholes model
OPTIONS and FUTURES Lecture 4: The Black-Scholes model Philip H. Dybvig Washington University in Saint Louis Black-Scholes option pricing model Lognormal price process Call price Put price Using Black-Scholes
Finite Differences Schemes for Pricing of European and American Options
Finite Differences Schemes for Pricing of European and American Options Margarida Mirador Fernandes IST Technical University of Lisbon Lisbon, Portugal November 009 Abstract Starting with the Black-Scholes
Lecture 11: The Greeks and Risk Management
Lecture 11: The Greeks and Risk Management This lecture studies market risk management from the perspective of an options trader. First, we show how to describe the risk characteristics of derivatives.
EC3070 FINANCIAL DERIVATIVES
BINOMIAL OPTION PRICING MODEL A One-Step Binomial Model The Binomial Option Pricing Model is a simple device that is used for determining the price c τ 0 that should be attributed initially to a call option
Stocks paying discrete dividends: modelling and option pricing
Stocks paying discrete dividends: modelling and option pricing Ralf Korn 1 and L. C. G. Rogers 2 Abstract In the Black-Scholes model, any dividends on stocks are paid continuously, but in reality dividends
Diusion processes. Olivier Scaillet. University of Geneva and Swiss Finance Institute
Diusion processes Olivier Scaillet University of Geneva and Swiss Finance Institute Outline 1 Brownian motion 2 Itô integral 3 Diusion processes 4 Black-Scholes 5 Equity linked life insurance 6 Merton
The Valuation of Currency Options
The Valuation of Currency Options Nahum Biger and John Hull Both Nahum Biger and John Hull are Associate Professors of Finance in the Faculty of Administrative Studies, York University, Canada. Introduction
Binomial lattice model for stock prices
Copyright c 2007 by Karl Sigman Binomial lattice model for stock prices Here we model the price of a stock in discrete time by a Markov chain of the recursive form S n+ S n Y n+, n 0, where the {Y i }
The interest volatility surface
The interest volatility surface David Kohlberg Kandidatuppsats i matematisk statistik Bachelor Thesis in Mathematical Statistics Kandidatuppsats 2011:7 Matematisk statistik Juni 2011 www.math.su.se Matematisk
INTERNATIONAL COMPARISON OF INTEREST RATE GUARANTEES IN LIFE INSURANCE
INTERNATIONAL COMPARISON OF INTEREST RATE GUARANTEES IN LIFE INSURANCE J. DAVID CUMMINS, KRISTIAN R. MILTERSEN, AND SVEIN-ARNE PERSSON Abstract. Interest rate guarantees seem to be included in life insurance
PRICING DIGITAL CALL OPTION IN THE HESTON STOCHASTIC VOLATILITY MODEL
STUDIA UNIV. BABEŞ BOLYAI, MATHEMATICA, Volume XLVIII, Number 3, September 003 PRICING DIGITAL CALL OPTION IN THE HESTON STOCHASTIC VOLATILITY MODEL VASILE L. LAZAR Dedicated to Professor Gheorghe Micula
FINANCIAL ECONOMICS OPTION PRICING
OPTION PRICING Options are contingency contracts that specify payoffs if stock prices reach specified levels. A call option is the right to buy a stock at a specified price, X, called the strike price.
Other variables as arguments besides S. Want those other variables to be observables.
Valuation of options before expiration Need to distinguish between American and European options. Consider European options with time t until expiration. Value now of receiving c T at expiration? (Value
FIN 411 -- Investments Option Pricing. Options: Definitions. Arbitrage Restrictions on Call Prices. Arbitrage Restrictions on Call Prices
FIN 411 -- Investments Option Pricing imple arbitrage relations s to call options Black-choles model Put-Call Parity Implied Volatility Options: Definitions A call option gives the buyer the right, but
Institutional Finance 08: Dynamic Arbitrage to Replicate Non-linear Payoffs. Binomial Option Pricing: Basics (Chapter 10 of McDonald)
Copyright 2003 Pearson Education, Inc. Slide 08-1 Institutional Finance 08: Dynamic Arbitrage to Replicate Non-linear Payoffs Binomial Option Pricing: Basics (Chapter 10 of McDonald) Originally prepared
Oscillatory Reduction in Option Pricing Formula Using Shifted Poisson and Linear Approximation
EPJ Web of Conferences 68, 0 00 06 (2014) DOI: 10.1051/ epjconf/ 20146800006 C Owned by the authors, published by EDP Sciences, 2014 Oscillatory Reduction in Option Pricing Formula Using Shifted Poisson
Option Portfolio Modeling
Value of Option (Total=Intrinsic+Time Euro) Option Portfolio Modeling Harry van Breen www.besttheindex.com E-mail: [email protected] Introduction The goal of this white paper is to provide
Simple formulas to option pricing and hedging in the Black Scholes model
Simple formulas to option pricing and hedging in the Black Scholes model Paolo Pianca Department of Applied Mathematics University Ca Foscari of Venice Dorsoduro 385/E, 3013 Venice, Italy [email protected]
American Options. An Undergraduate Introduction to Financial Mathematics. J. Robert Buchanan. J. Robert Buchanan American Options
American Options An Undergraduate Introduction to Financial Mathematics J. Robert Buchanan 2010 Early Exercise Since American style options give the holder the same rights as European style options plus
where N is the standard normal distribution function,
The Black-Scholes-Merton formula (Hull 13.5 13.8) Assume S t is a geometric Brownian motion w/drift. Want market value at t = 0 of call option. European call option with expiration at time T. Payout at
Lecture 6 Black-Scholes PDE
Lecture 6 Black-Scholes PDE Lecture Notes by Andrzej Palczewski Computational Finance p. 1 Pricing function Let the dynamics of underlining S t be given in the risk-neutral measure Q by If the contingent
Option Pricing. 1 Introduction. Mrinal K. Ghosh
Option Pricing Mrinal K. Ghosh 1 Introduction We first introduce the basic terminology in option pricing. Option: An option is the right, but not the obligation to buy (or sell) an asset under specified
Numerical Methods for Option Pricing
Chapter 9 Numerical Methods for Option Pricing Equation (8.26) provides a way to evaluate option prices. For some simple options, such as the European call and put options, one can integrate (8.26) directly
Vanna-Volga Method for Foreign Exchange Implied Volatility Smile. Copyright Changwei Xiong 2011. January 2011. last update: Nov 27, 2013
Vanna-Volga Method for Foreign Exchange Implied Volatility Smile Copyright Changwei Xiong 011 January 011 last update: Nov 7, 01 TABLE OF CONTENTS TABLE OF CONTENTS...1 1. Trading Strategies of Vanilla
CHAPTER 7: PROPERTIES OF STOCK OPTION PRICES
CHAPER 7: PROPERIES OF SOCK OPION PRICES 7.1 Factors Affecting Option Prices able 7.1 Summary of the Effect on the Price of a Stock Option of Increasing One Variable While Keeping All Other Fixed Variable
American and European. Put Option
American and European Put Option Analytical Finance I Kinda Sumlaji 1 Table of Contents: 1. Introduction... 3 2. Option Style... 4 3. Put Option 4 3.1 Definition 4 3.2 Payoff at Maturity... 4 3.3 Example
Master of Mathematical Finance: Course Descriptions
Master of Mathematical Finance: Course Descriptions CS 522 Data Mining Computer Science This course provides continued exploration of data mining algorithms. More sophisticated algorithms such as support
Valuation of Razorback Executive Stock Options: A Simulation Approach
Valuation of Razorback Executive Stock Options: A Simulation Approach Joe Cheung Charles J. Corrado Department of Accounting & Finance The University of Auckland Private Bag 92019 Auckland, New Zealand.
9 Basics of options, including trading strategies
ECG590I Asset Pricing. Lecture 9: Basics of options, including trading strategies 1 9 Basics of options, including trading strategies Option: The option of buying (call) or selling (put) an asset. European
HPCFinance: New Thinking in Finance. Calculating Variable Annuity Liability Greeks Using Monte Carlo Simulation
HPCFinance: New Thinking in Finance Calculating Variable Annuity Liability Greeks Using Monte Carlo Simulation Dr. Mark Cathcart, Standard Life February 14, 2014 0 / 58 Outline Outline of Presentation
1 Capital Asset Pricing Model (CAPM)
Copyright c 2005 by Karl Sigman 1 Capital Asset Pricing Model (CAPM) We now assume an idealized framework for an open market place, where all the risky assets refer to (say) all the tradeable stocks available
Exam MFE Spring 2007 FINAL ANSWER KEY 1 B 2 A 3 C 4 E 5 D 6 C 7 E 8 C 9 A 10 B 11 D 12 A 13 E 14 E 15 C 16 D 17 B 18 A 19 D
Exam MFE Spring 2007 FINAL ANSWER KEY Question # Answer 1 B 2 A 3 C 4 E 5 D 6 C 7 E 8 C 9 A 10 B 11 D 12 A 13 E 14 E 15 C 16 D 17 B 18 A 19 D **BEGINNING OF EXAMINATION** ACTUARIAL MODELS FINANCIAL ECONOMICS
Exam P - Total 23/23 - 1 -
Exam P Learning Objectives Schools will meet 80% of the learning objectives on this examination if they can show they meet 18.4 of 23 learning objectives outlined in this table. Schools may NOT count a
THE FUNDAMENTAL THEOREM OF ARBITRAGE PRICING
THE FUNDAMENTAL THEOREM OF ARBITRAGE PRICING 1. Introduction The Black-Scholes theory, which is the main subject of this course and its sequel, is based on the Efficient Market Hypothesis, that arbitrages
Valuation of Asian Options
Valuation of Asian Options - with Levy Approximation Master thesis in Economics Jan 2014 Author: Aleksandra Mraovic, Qian Zhang Supervisor: Frederik Lundtofte Department of Economics Abstract Asian options
Alternative Price Processes for Black-Scholes: Empirical Evidence and Theory
Alternative Price Processes for Black-Scholes: Empirical Evidence and Theory Samuel W. Malone April 19, 2002 This work is supported by NSF VIGRE grant number DMS-9983320. Page 1 of 44 1 Introduction This
