Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
3. Monte Carlo Simulations Math6911 S08, HM Zhu References 1. Chapters 4 and 8, “Numerical Methods in Finance” 2. Chapters 17.6-17.7, “Options, Futures and Other Derivatives” 3. George S. Fishman, Monte Carlo: concepts, algorithms, and applications, Springer, New York, QA 298.F57, 1995 2 Math6911, S08, HM ZHU 3. Monte Carlo Simulations 3.1 Overview Math6911 S08, HM Zhu Monte Carlo Simulation • Monte Carlo simulation, a quite different approach from binomial tree, is based on statistical sampling and analyzing the outputs gives the estimate of a quantity of interest Math6911, S08, HM ZHU Monte Carlo Simulation • Typically, estimate an expected value with respect to an underlying probability distribution – eg. an option price may be evaluated by computing the expected payoff w.r.t. risk-neutral probability measure • Evaluate a portfolio policy by simulating a large number of scenarios • Interested in not only mean values, but also what happens on the tail of probability distribution, such as Value-at-risk Math6911, S08, HM ZHU Monte Carlo Simulation in Option Pricing • In option pricing, Monte Carlo simulations uses the risk-neutral valuation result • More specifically, sample the paths to obtain the expected payoff in a risk-neutral world and then discount this payoff at the risk-neutral rate Math6911, S08, HM ZHU Learning Objectives • Understanding the stochastic processes of the underlying asset price. • How to generate random samplings with a given probability distribution • Choose an optimal number of simulation experiments • Reliability of the estimate, e.g. a confidence interval • Estimate error • Improve quality of the estimates: variance reduction techniques, quasi-Monte Carlo simulations. Math6911, S08, HM ZHU 3. Monte Carlo Simulation 3.2 Modeling Asset Price Movement Math6911 S08, HM Zhu Randomness in stock market 9 Math6911, S08, HM ZHU Outlines 1. 2. 3. 4. 5. Markov process Wiener process Generalized Wiener process Ito’s process, Ito’s lemma Asset price models 10 Math6911, S08, HM ZHU References 1. Chapter 12, “Options, Futures, and Other Derivatives” 2. Appendix A for Introduction to MATLAB, Appendix B for review of probability theory, “Numerical Methods in Finance: A MATLAB Introduction” 11 Math6911, S08, HM ZHU Markov process • A particular type of stochastic process whose future probability depends only on the present value • A stochastic process x(t) is called a Markov process if for every n and t1 < t 2 < < t n , we have P(x(t n ) ≤ xn x(t ) for all t ≤ t n −1 ) = P (x(t n ) ≤ xn x(t n −1 )) • The Markov property implies that the probability distribution of the price at any particular future time is not dependent on the particular path followed by the price in the past 12 Math6911, S08, HM ZHU Modeling of asset prices • Really about modeling the arrival of new info that affects the price. • Under the assumptions, The past history is fully reflected in the present asset price, which does not hold any further information Markets respond immediately to any new information about an asset • The anticipated changes in the asset price are a Markov process 13 Math6911, S08, HM ZHU Weak-Form Market Efficiency • This asserts that it is impossible to produce consistently superior returns with a trading rule based on the past history of stock prices. In other words technical analysis does not work. • A Markov process for stock prices is clearly consistent with weak-form market efficiency Math6911, S08, HM ZHU Example of a Discrete Time Continuous Variable Model • A stock price is currently at $40. • Over one year, the change in stock price has a distribution Ν(0,10) where Ν(µ,σ) is a normal distribution with mean µ and standard deviation σ. Math6911, S08, HM ZHU Questions • What is the probability distribution of the change of the stock price over 2 years? – ½ years? – ¼ years? – ∆t years? • Taking limits we have defined a continuous variable, continuous time process Math6911, S08, HM ZHU Variances & Standard Deviations • In Markov processes, changes of the variable in successive periods of time are independent. This means that: – Means are additive – variances are additive – Standard deviations are not additive Math6911, S08, HM ZHU Variances & Standard Deviations (continued) • In our example it is correct to say that the variance is 100 per year. • It is strictly speaking not correct to say that the standard deviation is 10 per year. Math6911, S08, HM ZHU A Wiener Process (See pages 265-67, Hull) • • • Consider a variable z follows a particular Markov process with a mean change of 0 and a variance rate of 1.0 per year. The change in a small interval of time ∆t is ∆z The variable follows a Wiener process if 1.∆ z = ε ∆t where ε is N(0,1) 2. The values of ∆z for any 2 different (nonoverlapping) periods of time are independent Math6911, S08, HM ZHU Properties of a Wiener Process ∆z over a small time interval ∆t: Mean of ∆z is 0 Variance of ∆z is ∆t Standard deviation of ∆z is ∆t ∆z over a long period of time, T: Mean of [z (T ) – z (0)] is 0 Variance of [z (T ) – z (0)] is T Standard deviation of [z (T ) – z (0)] is Math6911, S08, HM ZHU T Taking Limits . . . • What does an expression involving dz and dt mean? • It should be interpreted as meaning that the corresponding expression involving ∆z and ∆t is true in the limit as ∆t tends to zero • In this respect, stochastic calculus is analogous to ordinary calculus Math6911, S08, HM ZHU Generalized Wiener Processes (See page 267-69, Hull) • A Wiener process has a drift rate (i.e. average change per unit time) of 0 and a variance rate of 1 • In a generalized Wiener process the drift rate and the variance rate can be set equal to any chosen constants • The variable x follows a generalized Wiener process with a drift rate of a and a variance rate of b2 if dx=a dt + b dz Math6911, S08, HM ZHU Generalized Wiener Processes (continued) ∆x = a ∆t + b ε ∆t • Mean change in x in time ∆t is a ∆t • Variance of change in x in time ∆t is b2 ∆t • Standard deviation of change in x in time ∆t is b ∆t Math6911, S08, HM ZHU Generalized Wiener Processes (continued) ∆x = a T + b ε T • Mean change in x in time T is aT • Variance of change in x in time T is b2T • Standard deviation of change in x in time T is b T Math6911, S08, HM ZHU Example • A stock price starts at 40 and at the end of one year, it has a probability distribution of N(40,10) • If we assume the stochastic process is Markov with no drift then the process is dS = 10dz • If the stock price were expected to grow by $8 on average during the year, so that the year-end distribution is N (48,10), the process would be dS = 8dt + 10dz • What the probability distribution of the stock price at the end of six months? Math6911, S08, HM ZHU Itô Process (See pages 269, Hull) • In an Itô process the drift rate and the variance rate are functions of time dx=a(x,t) dt+b(x,t) dz • The discrete time version, i.e., ∆x over a small time interval (t, t+∆t) ∆x = a ( x, t ) ∆t + b ( x, t )ε ∆t is only true in the limit as ∆t tends to zero Math6911, S08, HM ZHU Why a Generalized Wiener Process is not Appropriate for Stocks • For a stock price we can conjecture that its expected percentage change in a short period of time remains constant, not its expected absolute change in a short period of time • We can also conjecture that our uncertainty as to the size of future stock price movements is proportional to the level of the stock price Math6911, S08, HM ZHU A Simple Model for Stock Prices (See pages 269-71, Hull) dS = µ S dt + σ S dz Itô’s process where µ is the expected return σ is the volatility. The discrete time equivalent is ∆ S = µS ∆ t + σS ε ∆ t Math6911, S08, HM ZHU Example: Simulate asset prices • Consider a stock that pays no-dividends, has an expected return 15% per annum with continuous compounding and a volatility of 30% per annum. Observe today’s price $1 per share and with ∆t = 1/250 yr. Simulate a path for the stock price. G iven S i -1, w e com pute ∆ S = µ S i −1∆ t + σ S i −1ε T hen S i = S i −1 + ∆t ∆S 29 Math6911, S08, HM ZHU Example: Simulate stock prices Stock Price S at start of period 1.0000 0.9751 0.9870 1.0099 1.0108 1.0165 1.0140 1.0037 1.0180 0.9927 Random sample for ε ∆S during a time period dt -1.3457 0.6132 1.1928 0.0114 0.2667 -0.1590 -0.5652 0.7189 -1.3431 -1.3913 -0.0249 0.0119 0.0229 0.0008 0.0057 -0.0025 -0.0103 0.0143 -0.0253 -0.0256 30 Math6911, S08, HM ZHU Example: Simulate stock prices 31 Math6911, S08, HM ZHU Itô’s Lemma (See pages 273-274, Hull) • If we know the stochastic process followed by x, Itô’s lemma tells us the stochastic process followed by some function G (x, t ) • Since a derivative security is a function of the price of the underlying and time, Itô’s lemma plays an important part in the analysis of derivative securities Math6911, S08, HM ZHU Taylor Series Expansion A Taylor’s series expansion of G(x, t) gives ∂G ∂G ∂ 2G 2 x ∆G = ∆x + ∆t + ½ ∆ ∂x 2 ∂x ∂t ∂ 2G ∂ 2G 2 ∆t + … + ∆x ∆t + ½ 2 ∂x∂t ∂t Math6911, S08, HM ZHU Ignoring Terms of Higher Order Than ∆t and ∆x In ordinary calculus we have ∂G ∂G ∆G = ∆x + ∆t ∂x ∂t In stochastic calculus this becomes ∂G ∂G ∂ 2G 2 ∆G = ∆x + ∆t + ½ ∆x 2 ∂x ∂t ∂x because ∆x has a component which is of order ∆t Math6911, S08, HM ZHU Substituting for ∆x Suppose dx = a( x, t )dt + b( x, t )dz so that ∆x = a ∆t + b ε ∆t Then ignoring terms of higher order than ∆t ∂G ∂G ∂G 2 2 ∆G = ∆x + ∆t + ½ 2 b ε ∆t ∂x ∂t ∂x 2 Math6911, S08, HM ZHU The ε2∆t Term Since ε ≈ N (0,1), E (ε ) = 0 E (ε 2 ) − [E (ε )]2 = 1 E (ε ) = 1 2 It follows that E (ε 2 ∆t ) = ∆t The variance of ∆t is proportional to ∆t 2 and can be ignored. Hence 1 ∂ 2G 2 ∂G ∂G b ∆t ∆G = ∆x + ∆t + 2 2∂x ∂x ∂t Math6911, S08, HM ZHU Taking Limits Taking limits Substituti ng We obtain Math6911, S08, HM ZHU ∂ 2G 2 ∂G ∂G dt + ½ 2 b dt dx + dG = ∂x ∂t ∂x dx = a dt + b dz ⎛ ∂G ∂G ∂ 2G 2 ⎞ ∂G b dz a+ dG = ⎜⎜ + ½ 2 b ⎟⎟ dt + ∂x ∂x ∂t ⎠ ⎝ ∂x This is Ito' s Lemma Application of Ito’s Lemma to a Stock Price Process The stock price process is d S = µS dt + σS d z For a function G of S and t 2 ⎛ ∂G ∂G ∂G ∂ G 2 2⎞ σS dz dG = ⎜⎜ µS + + ½ 2 σ S ⎟⎟ dt + ∂S ∂t ∂S ⎠ ⎝ ∂S Math6911, S08, HM ZHU Example If G = ln S, then dG = ? ⎛ σ2 ⎞ dG = ⎜ µ − ⎟ dt + σ dz 2 ⎠ ⎝ Generalized Wiener process Math6911, S08, HM ZHU A generalized Wiener process for stock price The discrete version of this is ( ln S ( t + ∆ t ) − ln S ( t ) = µ − σ 2 ) / 2 ∆ t + σε ∆t or µ −σ ( S (t + ∆ t ) = S (t ) e 2 ) / 2 ∆t +σ ε ∆t S( t ) follows a lognorm al distribution. It is often referred to as geom etric Brownian m otio n. 40 Math6911, S08, HM ZHU Lognormal Distribution T h e co rre sp o n d in g d e n sity fu n ctio n fo r S ( t ) is 2 2 ⎛ ⎛ ⎞ ⎞ ⎛ ⎞ σ ⎛ ⎞ ⎜ − ⎜ lo g ⎜ x ⎟−⎜µ − ⎟t ⎟ ⎟ S 2 ⎠ ⎠ ⎟ 0 ⎠ ⎝ ⎜ ⎝ ⎝ e xp ⎜ ⎟ 2 2 σ t ⎜ ⎟ ⎜ ⎟ ⎝ ⎠ f (x ) = x σ 2π t fo r x > 0 w ith f ( x ) = 0 fo r x ≤ 0 . It ca n b e p ro v e n th a t E (S ( t ) ) = S 0 e µ t ( E S (t ) 2 ) 2 µ +σ ) t ( =S e 2 0 2 ( v a r ( S ( t ) ) = S 02 e 2 µ t e σ Math6911, S08, HM ZHU 2 t −1 ) Example: Simulate stock prices 42 Math6911, S08, HM ZHU Models for stock price in risk neutral world (r: risk-free interest rate) In a risk neutral world, two process for the underlying asset is: dS = σ S dz + µˆ Sdt or (1) S ( t + ∆t ) − S ( t ) = µˆ S ( t ) ∆t + σ S ( t ) ε ∆t ( ) d lnS = r − σ 2 / 2 dt + σ dz or (2) ( ) lnS(t + ∆t ) − lnS(t ) = r − σ 2 / 2 ∆t + σε Math6911, S08, HM ZHU ∆t 43 Which one is more accurate? • It is usually more accurate to work with InS because it follows a generalized Wiener process and (2) is true for any ∆t while (1) is true only in the limit as ∆t tends to zero. For all T, ( ) ln S(T ) − ln S(0) = r − σ 2 / 2 T + σε T i.e., r −σ ( S(T ) = S(0)e 2 ) / 2 T +σε T 44 Math6911, S08, HM ZHU 3. Monte Carlo Simulation 3.3 Generate pseudo-random variates Math6911 S08, HM Zhu Generate pseudorandom variates The usual way to generate psedorandom variates, i.e., samples from a given probability distribution: 1. Generate psedurandom numbers, which are variates from the uniform distribution on the interval (0, 1). Denote as U(0, 1) 2. Apply suitable transformation to obtain the desired distribution Math6911, S08, HM ZHU Generate U(0, 1) variables The standard method is linear congruential generators (LCGs): 1. Given an integer number Zi-1, Z i = (aZ i −1 + c )(mod m ) where a, c, and m are chosen parameters 2. Return number Zi which generate a U(0, 1) variable m Comments : − Z0 : the initial number, the seed of the random sequence. eg. rand(' seed' , 0) generates the same sequence each time − LCGs generates rational numbers instead of real ones m -1 i⎫ ⎧ − Periodic. The sequency ⎨U i = ⎬ has a maximum period of m m ⎭i = 0 ⎩ − Improved MATLAB function :rand(' state' , 0) allows much longer periods Math6911, S08, HM ZHU Generate U(0, 1) variables Math6911, S08, HM ZHU Generate U(0, 1) variables Math6911, S08, HM ZHU Generate general distribution F(x) from U(0, 1) Given the distribution function F ( x ) = P{X ≤ x}, we can generate random variates according to F by the following 1. Draw a random number U~U (0 ,1) 2. Return X = F -1 (U ) Proof : Use the monotonicity of F and the fact that U is uniformly distributed : { } P{X ≤ x} = P F −1 (U ) ≤ x = P{U ≤ F ( x )} = F ( x ) Math6911, S08, HM ZHU Question • How to generate samples according to standard normal distribution N(0,1) from U(0, 1) ? - No analytical form of the inverse of the distribution function - Support is not finite Math6911, S08, HM ZHU Generate Samples from Normal Distribution One simple way to obtain a sample from Ν(0,1) is as follows 12 x = ∑U i − 6 i =1 where U i are independent random numbers from U(0,1), and x is the required samples from N(0,1) Note : - It is from Central Limit Theorem. - The approximation is satisfactory for most purposes Math6911, S08, HM ZHU Central Limit Theorem Let X1, X2, X3, ... be a sequence of independent random variables which has the same probability distribution. Consider the sum :Sn = X1 + ... + Xn. Then the expected value of Sn is nμ and its standard deviation is σ n½. Furthermore, informally speaking, the distribution of Sn approaches the normal distribution N(nμ,σn1/2) as n approaches ∞. Or the standardized Sn, i.e., Then the distribution of Zn converges towards the standard normal distribution N(0,1) as n approaches ∞ Math6911, S08, HM ZHU A Note Uniform random variable X: PDF: ⎧ 1 , if x ∈ ( a,b ) ⎪ f ( x) = ⎨b − a ⎪⎩ 0 , otherwise Mean: b+a E [X] = 2 Variance: Var ( X ) Math6911, S08, HM ZHU b − a) ( = 12 2 Generate Samples from N(0,1): Box-Muller algorithm The Box-Muller algorithm can be implemented as follows: 1. Generate two independent uniform samples from U1 ,U 2 ~ U ( 0,1) 2. Set R 2 = −2 log U1 and θ =2π U 2 3. Set X = R cos θ and Y = R sin θ where X and Y are independent samples from N ( 0,1) . Note: - Box and Muller (1958), Ann. Math. Statist., 29:610-611 - Common and more efficient than the previous method. - But it is costly to evaluate trigonometric functions Math6911, S08, HM ZHU Generate Samples from N(0,1): Improved Box-Muller algorithm The polar rejection (improved Box-Muller) method: 1. Generate two independent uniform samples from U1 ,U 2 ~ U ( 0,1) 2. Set V1 = 2U1 -1, V2 = 2U 2 -1, and S = V12 + V22 3. If S>1, return to step 1; otherwise, return -2 ln S -2lnS X = V1 and Y = V2 S S where X and Y are required independent samples from N ( 0,1) . Note: - Ahrens and Dieter (1988), Comm. ACM 31:1330-1337 Math6911, S08, HM ZHU Question How to generate samples according to normal distribution N(µ, σ) from N(0, 1)? Math6911, S08, HM ZHU To Obtain 2 Correlated Normal Samples To obtain correlated normal samples ε1 and ε 2 , we first 1. generate two independent stardard normal variates x1 , x 2 ~N ( 0,1) 2. and set ε1 = µ1 + x1 ε 2 = µ2 + ρ x1 + x2 1 − ρ 2 where ρ is the coefficient of correlation. Math6911, S08, HM ZHU To Obtain n Correlated Normal Samples To obtain n correlated normal samples ε1 ,… , ε n , we first 1. generate n independent stardard normal variates x1 ,… ,xn ~N ( 0,1) 2. and set ε =µ +U T X T U Here, U = Σ = ⎡⎣ ρij ⎤⎦ is the Cholesky decomposition of the covariance matrix Σ. Math6911, S08, HM ZHU Simulate the correlated asset prices To simulate the payoff of a derivative depending on multi-variables, how can we generate samples of each variables? µˆ −σ ( S ( t + ∆t ) = S ( t ) e i i 2 i ) / 2 ∆t +σ i ε i ∆t i 60 Math6911, S08, HM ZHU 3. Monte Carlo Simulation 3.4 Choose the number of simulations Math6911 S08, HM Zhu How many trials is enough? X i : Sequence of independent samples from the same distribution 1 M X ( M ) ≡ ∑ X i : Sample mean, an unbiased estimator to the M i=1 quantitiy of interest µ = E [ X i ] 2 1 M ⎡⎣ X i − X ( M ) ⎤⎦ : Sample variance S (M ) = ∑ M − 1 i=1 2 σ2 E ⎡⎢( X ( M ) − µ ) ⎤⎥ = Var ⎡⎣ X ( M ) ⎤⎦ = : the expected value of square error ⎣ ⎦ M as a way to quantify the quality of our estimator 2 Note that σ 2 can be estimated by the sample variance S 2 ( M ) Math6911, S08, HM ZHU Number of Simulations and Accuracy The number of simulation trials carried out depends on the accuracy required. M: number of independent trials carried out µ: expected value of the derivative, S(M): standard deviation of the discounted payoff given by the trials. Math6911, S08, HM ZHU Standard Errors in Monte Carlo Simulation The standard error of the estimate of the option price based on sample size M is: σ M ≈ S2 (M ) M Clearly as M increases, our estimate is more accurate Math6911, S08, HM ZHU (1-α) Confidence Interval of Monte Carlo Simulation In general, the confident interval at level (1- α) may be computed by X ( M ) − z1−α / 2 where z1−α / 2 S 2 (M ) < µ < X ( M ) + z1−α / 2 S 2 (M ) M M is the critical number from the standard normal distribution. Note: Uncertainty is inverse proportional to the square root of the number of trials Math6911, S08, HM ZHU 95% Confidence Interval of Monte Carlo Simulation A 95% confidence interval for the price V of the derivative is therefore given by: X ( M ) − 1.96 Math6911, S08, HM ZHU S2 (M ) M < µ < X ( M ) + 1.96 S2 (M ) M Absolute error Suppose we want to controlling the absolute error such that, with probability (1-α ) , X (M ) − µ ≤ β , where β is the maximum acceptable tolerance. Then the number of simulations M must satisfy z1−α / 2 S ( M ) M ≤ β . 2 Math6911, S08, HM ZHU Relative error Suppose we want to controlling the relative error such that, with probability (1-α ) , X (M ) − µ µ ≤γ, where γ is the maximum acceptable tolerance. Then the number of simulations M must satisfy z1−α / 2 S 2 ( M ) M X (M ) Math6911, S08, HM ZHU ≤ γ 1+ γ . 3. Monte Carlo Simulation 3.5 Applications in Option Pricing Math6911 S08, HM Zhu Monte Carlo Simulation When used to value a derivative dependent on a market variable S, this involves the following steps: 1. Simulate 1 path for the stock price in a risk-neutral world 2. Calculate the payoff from the stock option 3. Repeat steps 1 and 2 many times to get many sample payoff 4. Calculate mean payoff 5. Discount mean payoff at risk-free rate to get an estimate of the value of the option Math6911, S08, HM ZHU Example: European call • Consider a European call on a stock with no-dividend payment. S(0) = $50, E = 52, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • Black-Schole Price: $5.1911 • MC Price with 1000 simulations: $5.4445 Confidence Interval [4.8776, 6.0115] • MC Price with 200,000 simulations: $5.1780 Confidence Interval [5.1393, 5.2167] 71 Math6911, S08, HM ZHU Example: Butterfly Spread • A butterfly spread consists of buying a European call option with exercise price K1 and another with exercise price K3 (K1< K3) and by selling two calls with exercise price K2 = (K1 + K3)/2, for the same asset and expiry date • Consider a butterfly spread with S(0) = $50, K1 = $55, K2 = $60, K3 = 65, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • Exact price: $0.6124 • MC price with 100,000 simulations: $0.6095 Confidence Interval [0.6017, 0.6173] 72 Math6911, S08, HM ZHU Example: Arithmetic Average Asian Option • Consider pricing an Asian average rate call option with discrete arithmetic averaging. The option payoff is ⎧1 max ⎨ ⎩N ⎫ − S t K , 0 ( i) ⎬ ∑ i =1 ⎭ where ti = i∆t and ∆t = T . N N • Consider the option on a stock with no-dividend payment. S(0) = $50, K = 50, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. 73 Math6911, S08, HM ZHU Determining Greek Letters For ∆: 1. Make a small change to asset price 2. Carry out the simulation again using the same random number streams 3. Estimate ∆ as the change in the option price divided by the change in the asset price VMC* −VMC ∆S Proceed in a similar manner for other Greek letters 74 Math6911, S08, HM ZHU 3. Monte Carlo Simulation 3.6 Further Comments Math6911 S08, HM Zhu The parameters in the model • Our analysis so far is useless unless we know the parameters µ and σ • The price of an option or derivative is in general independent of µ • However, the asset price volatility σ is critically important to option price. Typically, 20%< σ < 50% • Volatility of stock price can be defined as the standard deviation of the return provided by the stock in one year when the return is expressed using continuous compounding 76 Math6911, S08, HM ZHU Estimate volatility σ from historical data Assume: n + 1: number of observation at fixed time interval Si : Stock price at end of ith (i = 0, 1, ..., n) interval ∆t : Length of observing time interval in years ⎛ Si ⎞ ui = ln ⎜ ⎟ : n number of approximation for the change ⎝ Si −1 ⎠ of ln S over ∆t 77 Math6911, S08, HM ZHU Estimate volatility σ from historical data Compute the standard deviation of the u i ' s : n 1 2 s= ( ui − u ) ∑ i =1 n −1 where the mean of the u i ' s is 1 n u = ∑ i =1 ui n s ∵ s ≈ σ ∆t ∴σ ≈ σˆ = ∆t The standard error of estimate is about σˆ Math6911, S08, HM ZHU 2n 78 Advantages and Limitations • More efficient for options dependent on multiple underlying stochastic variables. As the number of variables increase, Monte Carlo simulations increases linearly while the others increases exponentially • Easily deal with path dependent options and options with complex payoffs / complex stochastic processes • Has an error estimate • It cannot easily deal with American options • Very time-consuming Math6911, S08, HM ZHU Extensions to multiple variables When a derivative depends on several underlying variables we can simulate paths for each of them in a risk-neutral world to calculate the values for the derivative Math6911, S08, HM ZHU Sampling Through the Tree • Instead of sampling from the stochastic process we can sample paths randomly through a binomial or trinomial tree to value a derivative Math6911, S08, HM ZHU 3. Monte Carlo Simulation 3.7 Variance Reduction Techniques Math6911 S08, HM Zhu Variance Reduction Procedures • Usually, a very large value of M is needed to estimate V with reasonable accuracy. • Variance reduction techniques lead to dramatic savings in computational time • • • • • • Math6911, S08, HM ZHU Antithetic variable technique Control variate technique Importance sampling Stratified sampling Moment matching Using quasi-random sequences Antithetic Variable Technique Consider to estimate V = E ( f (U ) ) where U ∼ N ( 0 ,1) . The standard Monte Carlo estimate: 1 M VMC = f (U i ) with i.i.d. U i ∼ N ( 0 ,1) . ∑ M i =1 The antithetic variate technique: 1 VAV = M We can prove that M f (U i ) +f ( −U i ) i =1 2 ∑ with i.i.d. U i ∼ N ( 0 ,1) ⎛ f (U i ) +f ( −U i ) ⎞ 1 var ⎜ ⎟ = ⎡⎣ var ( f (U i ) ) + cov ( f (U i ) , f ( −U i ) ) ⎤⎦ 2 ⎝ ⎠ 2 1 ≤ var ( f (U i ) ) 2 if f is monotonic. Math6911, S08, HM ZHU Antithetic Variable Technique It involves calculating two values of the derivative. V1: calculated in the usual way V2: calculated by the changing the sign of all the random standard normal samples used for V1 V is average of the V1 and V2 , the final estimate of the value of the derivative Math6911, S08, HM ZHU Antithetic Variable Technique Note: 1. Monte-Carlo works when the simulated variables "spread out" as closely as possible to the true distribution. 2. Antithetic variates relies upon finding samplings that are anticorrelated with the original random variable. 3. Works well when the payoff is monotonic w.r.t. S 4. Further reading: --P.P. Boyle, Option: a Monte Carlo approach, J. of Finanical Economics, 4:323-338 (1977) --Boyle, Broadie and Glasserman, Monte Carlo methods for security pricing J. of Economic Dynamics and Control, 21: 1267-1321 (1997) --N. Madras, Lectures on Monte Carlo Methods, 2002 Math6911, S08, HM ZHU Example: European call • Consider a European call on a stock with no-dividend payment. S(0) = $50, K = 52, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • Black-Schole Price: $5.1911 • MC Price with 200,000 simulations: $5.1780 Confidence Interval [5.1393, 5.2167] • MCAV Price with 200,000 simulations: $5.1837 Confidence Interval [5.1615, 5.2058] (ratio: 1.75) 87 Math6911, S08, HM ZHU Example: Butterfly Spread • Consider a butterfly spread with S(0) = $50, K1 = $55, K2 = $60, K3 = 65, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • Exact price: $0.6124 • MC price with 100,000 simulations: $0.6095 Confidence Interval [0.6017, 0.6173] • MCAV price with 50,000 simulations: $0.6090 Confidence Interval [0.5982, 0.6198] 88 Math6911, S08, HM ZHU Control Variate Technique We can generalize the previous technique to Zθ = VA +θ ( E (VB ) −VB ) for any θ ∈ . Here, VB is "close" to VA with known mean E (VB ) . In this case, var ( Zθ ) = var (VA −θVB ) = var (VA ) − 2θ cov(VA,VB ) +θ 2 var (VB ) We can prove that var ( Zθ ) < var (VA ) if and only if 0 <θ < 2 Math6911, S08, HM ZHU cov(VA,VB ) var (VB ) . 89 Example: European call • Consider a European call on a stock with no-dividend payment. S(0) = $50, K = 52, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • Black-Schole Price: $5.1911 • MC Price with 200,000 simulations: $5.1780 Confidence Interval [5.1393, 5.2167] • MCCV Price with 200,000 simulations: $5.1883 Confidence Interval [5.1712, 5.2054] (ratio: 2.2645) 90 Math6911, S08, HM ZHU Example: Arithmetic Average Asian Option • Consider the option on a stock with no-dividend payment. S(0) = $50, K = 50, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • We could use the sum of the asset prices as a control variate as we know its expected value and Y is correlated to the option itself 1− e ( ) ⎡ ⎤ E ⎢ ∑ S ( t ) ⎥ = ∑ E ⎡⎣ S ( i ∆t ) ⎤⎦ = S ∑ e = S 1− e ⎣ ⎦ • Another choice of the control variate is the payoff of a geometric average option as this is known analytically N i =0 N i i =0 1 ⎧⎪⎛ N ⎫⎪ ⎞ N max ⎨⎜ ∏ S ( ti ) ⎟ − K , 0 ⎬ ⎠ ⎪⎩⎝ i =1 ⎪⎭ Math6911, S08, HM ZHU N 0 i =0 r N +1 ∆t ri ∆t 0 r ∆t 91 Importance Sampling Technique It is a way to distort the probability measure in order to sample from critical region. For example, in evaluating European call option: F : the unconditional probability distribution function for the asset price at time T q: the probability of the asset price ≥ K at maturity, known analytically G = F / q: the probability distribution of the asset conditional on the asset price ≥ K, i.e., importance function Instead of sampling from F, we sample from G. Then the value of the option is the average discounted payoff multiplied by q. Math6911, S08, HM ZHU Stratified Sampling Technique Sampling representative values rather than random values from a probability distribution is usually more accurate. It involves dividing the distribution into intervals and sampling from each interval (stratum) according to the distribution: m E [ X ] = ∑ p j E ⎡⎣ X Y = y j ⎤⎦ j =1 where P {Y = y j } = p j for j = 1, ,m, is known. Note: 1. For each stratum, we sample X conditioned on the even Y = y j 2. The representative values are typically mean or median for each interval Math6911, S08, HM ZHU Moment Matching Moment matching involves adjusting the samples taken from a standard normal distribution so that the first, second, or possibly higher moments are matched. For example, we want to sample from N(0,1). Suppose that the samples are ε i (1 ≤ i ≤ n). To match the first two moments, we adjustify the samples by ε = * i εi − m s where m and s are the mean and standard deviation of samples. The adjusted samples ε i* has the correct mean 0 and standard deviation 1. We then use the adjusted samples for calculation. Math6911, S08, HM ZHU Quasi-Random Sequences • Also called a low-discrepancy sequence is a sequence of representative samples from a probability distribution • At each stage of the simulation, the sampled points are roughly evenly distributed throughout the probability space 1 • The standard error of the estimate is proportional to M Math6911, S08, HM ZHU Further reading • R. Brotherton-Ratcliffe, “Monte Carlo Motoring”, Risk, 1994:53-58. • W.H.Press, et al, “Numerical Reciptes in C”, Cambridge Univ. Press, 1992 • I. M. Sobol, “USSR Computational Mathematics and Mathematical Physics”, 7, 4, (1967): 86-112; P. Jaeckel, “Monte Carlo Methods in finance”, Wiley, Chichester, 2002 • P. Glasserman, “Monte Carlo Methods in Financial Engineering”, Springer-Verlag, New York, 2004 Math6911, S08, HM ZHU Further reading • www.mcqmc.org • www.mat.sbg.ac.at/~schmidw/links.html Math6911, S08, HM ZHU 3. Monte Carlo Simulation Appendix Math6911 S08, HM Zhu Asian Options (Chapter 22) • Payoff related to average stock price • Average Price options pay: – Call: max(Save – K, 0) – Put: max(K – Save , 0) • Average Strike options pay: – Call: max(ST – Save , 0) – Put: max(Save – ST , 0) 99 Math6911, S08, HM ZHU Asian Options (Chapter 22) • For arithmetic averaging: S ave 1 = N N ∑ S ( t ), i =1 i where T = N ∆t • For geometric averaging: S ave = N S ( t1 ) S ( t2 ) Math6911, S08, HM ZHU S (tN ) 10 0 Asian Options (geometric averaging) • Closed form (Kemna & Vorst,1990, J. of Banking and Finance, 14:113-129) because the geometric average of the underlying prices follows a lognormal distribution as well. where N(x) is the cumulative normal distribution function and Math6911, S08, HM ZHU 10 1 Asian Options (arithmetic averaging) • No analytic solution • Can be valued by assuming (as an approximation) that the average stock price is lognormally distributed • Turnbull and Wakeman, 1991, J. of Financial and Quantitative Finance, 26:377389 • Levy,1992, J. of International Money and Finance, 14:474-491 Math6911, S08, HM ZHU 10 2 Example: Asian Put with arithmetric average • Consider an Asian Put on a stock with non-dividend payment. S(0) = $50, X = 50, T = 5 months, and annual risk-free interest rate r = 10% and a volatility σ= 40% per annum. • MC Price with 100,000 simulations: $3.9806 Confidence Interval [3.9195, 4.0227] • MCAV Price with 100,000 simulations: $3.9581 Confidence Interval [3.9854, 3.9309] Math6911, S08, HM ZHU 10 3