* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download 7. Equivalent Martingale Measures So far we have considered
Survey
Document related concepts
Transcript
7. Equivalent Martingale Measures So far we have considered derivative asset pricing exploiting PDEs implied by arbitragefree portfolios. Another approach is to change the probability measure to another probability measure implied by arbitrage-free markets such that under that the (risk-free return discounted) prices become martingales. 1 As for background, consider pricing an European call option. The aim is to find the fair price for the option given the available information. To price the option (Ct), we use the best prediction of the end value in the light of available information, such that (1) Ct = Et [ρ max(ST − K, 0)], where Et is the conditional expectation given information up to time t, and ρ is a discount factor. 2 The no arbitrage theory implies that if the option is replicable, then the discount factor will be the riskfree rate, and the probability measure with respect to which the expectation must be calculated is such that the discounted price process (2) S̃t = e−r tSt, where r is the riskfree return, is martingale. 3 To illustrate the situation, consider the following single period discrete world. Example 7.1: Suppose we have a call option C on stock S and a bank account B. Let the exercise price of the option be K, and assume that there are two possible end values S1 > K > S2 of the stock. So p ³ ³³ ³ ³³ PP S0³ PP P 1 − pPP S1 = u S0, with u > 1 S2 = d S0, with d < 1, where p is the probability that the price goes up to S = u S0 , and S0 is the current price of the stock. Then the option with initial cost C has the end value max{S1 − K, 0}. To replicate this with the stock and bank account with (riskfree) interest rate, r, we may construct the following strategy: 4 Buy one share of the stock by financing it with cash and S2 /(1+r) borrowed at rate r from the bank (after one period the repayment is accordinly S2 ). The value of the initial position is then (3) S0 − 1 S2 . 1+r The end value of the position is according to the stock price as Stock value Loan repayment Total payoff S = S1 S1 −S2 S1 − S2 S = S2 S2 −S2 0 5 We observe that in the case of S = S2 the payoff is 0, the same as with the call options, and in the case S = S1 the total payoff is S1 − S2 = a(S1 − K), where a = (S1 − S2 )/(S1 − K). Thus in all, the payoff of the strategy is exactly the same as the payoff of a call options. This implies that in the absence of arbitrage the cost of the investment must be the same in both cases. That is, buying a call options must have the same value as the other strategy based on one stock and bank loan. So aC = S0 − 1 S2 1+r or C = (S0 − 1 1 S2 )/a = p∗ (S1 − K), 1+r 1+r where p∗ = 1+r−d . u−d 6 We observe that 0 < p∗ < 1 (provided that 1 + r < u Exercise: What is the rationale that this also holds?), so that p∗ can be considered as a conditional probability given the initial price S0 of the stock (the probability depends on u and d which are dependent on S0 , i.e., how much the price should go up to reach the given value S1 , or to decrease to go down to the other possible given value S2 ). These probabilities are called risk neutral, hedging or martingale probabilities or probability measures. The last name is because they make the discounted price process S̃ a martingale. 7 This is seen as follows: We easily find that (4) S0 = 1 (p∗ S1 + (1 − p∗)S2 ) = p∗ S̃1 + (1 − p∗ )S̃2 , 1+r where S̃ = S/(1 + r) is the discounted price process. That is, the conditional expectations of the discounted price process S̃ given the information I0 = {S0 } is (5) E∗ [S̃|I0 ] = p∗ S̃1 + (1 − p∗ )S̃2 = S̃0 = S0 , so that S̃ is martingale with respect to the risk neutral probability measure, as stated above. 8 We observe that the option value does not at all depend on the true probabilities p and q = 1 − p! However, if we write the above martingale equaton as (q ∗ = 1 − p∗ ) (6) p∗ q∗ E [S̃|I0 ] = p S̃1 + q S̃2 = pS̃1 + q S̃2, p q ∗ ∗ ∗ where p∗ /p and q ∗/q can be considered kinds of likelihood ratios or odds ratios, judged by the markets for the events that the stock price will be S1 and S2 , respectively. So the market expected value of the future stock price is a kind of likelihood weighted value of the possible future outcomes. 9 Translations of Probabilities Probability Measure As an illustration, consider the probability density f (z) of a standard normal distribution, 1 − 1 z2 (7) f (z) = √ e 2 . 2π Then probability of that the random variable Z is near a specific value z̄ is ³ (8) P 1 1 z̄ − ∆ < Z < z̄ + ∆ 2 2 Z ´ z̄+ 21 ∆ = z̄− 21 ∆ 1 2 1 √ e− 2 z dz, 2π which is a real number (between zero and one). Thus, the probability associates a real number (in this case between zero and one) to intervals on real line, or more generally to (Borel) sets. Such functions are called measures in mathematics or measure functions. 10 Because ∆ is small Z z̄+ 1 ∆ 2 Z 1 1 − 1 z2 1 − 1 z̄ 2 z̄+ 2 ∆ √ e 2 dz ≈ √ e 2 dz 1 1 z̄− 2 ∆ z̄− 2 ∆ 2π 2π 1 − 1 z̄ 2 = √ e 2 ∆. 2π (9) For infinitesimal ∆, denoted as dz, we designate the associated measure by symbol dP (z), or simply by dP . Thus, in the above case we have (10) 1 − 1 z2 dP (z) = √ e 2 dz. 2π 11 Generally, if P is a probability measure, we have (11) Z ∞ −∞ dP = 1. With these notations, e.g., (12) E[X] = Z ∞ −∞ x dP (x). So the expected value is mathematically an integral with respect to probability measure. dP is called sometimes the density of the probability measure P . 12 Changing Probability Measure Martingale model is a central tool for modeling fair prices of derivative securities. However, generally, if St is a risky asset, then given information up to time point t, we have (13) Et[St+h] > (1 + rf )St, (h > 0), because investors want some compensation for the risk, where Et is the conditional expectation, and rf is the risk-free rate. 13 However, we observed in the PDE approach that under the arbitrage-free pricing the riskfree rate should be a proper discounting factor in pricing risky derivative assets. More importantly, the fundamental theorem of asset pricing establishes the equivalence of the absence of arbitrage opportunity and existence of martingale measure in (the stochastic model of) financial markets. 14 A probability measure, P̃ , is a martingale measure for the discounted price process S̃t = e−rtSt, if S̃t is martingale under P̃ , i.e. (14) EP̃ t [S̃s ] = S̃t, whenever s ≥ t, where r is the riskfree rate, and EP̃ t indicates that the conditional expectation is taken with respect to probability measure P̃ . This theory implies that the no-arbitrage price of a contingent claim with underlying security St and (random) payoff X at maturity T is obtained by (15) Ct = EP̃ t h e−rτ X i , where τ = T − t, and P̃ is the martingale measure for the discounted price process S̃t. We say that S̃t is P̃ -martingale. 15 Thus a martingale measure can be viewed as a representation of the market’s current opinion on the evolution of values of underlying assets and the prices of all derivatives contingent to them. Consequently, the knowledge of the martingale measure is all that is needed, in principle, to value whatever derivative securities by the formula of the form (15). 16 Then given a stock price process St with probability measure P , the goal is to find the martingale measure P̃ . This can be accomplished if there is an invertible function (one-to-one) ξ(z) such that (16) dP̃ (z) = ξ(z) dP (z). Now recalling from standard calculus; if G(z) = R g(z) dz then g(z) = dG(z)/dz = G0(z), i.e., g(z) is the (mathematical) derivative of G(z). Alternatively, we can write dG(z) = g(z)dz, and hence Z (17) G(z) = Z dG(z) = g(z)dz. 17 In the same manner, because (18) P̃ (z) = Z z −∞ dP̃ (t) = Z z −∞ ξ(t) dP (t), so that we can adopt notation (19) dP̃ (z) = ξ(z), dP (z) and call ξ as a derivative of P̃ with respect to P . In mathematical measure theory this is known as the Radon-Nikodym derivative. 18 The existence of ξ is guaranteed if P and P̃ satisfy (20) P̃ (dz) > 0 if and only if P (dz) > 0. Actually this condition guarantees besides the existence of ξ (known as Radon-Nikodym Theorem), also the equivalence of P̃ and P in the sense (21) dP̃ (z) = ξ(z) dP (z) and (22) dP (z) = ξ(z)−1dP̃ (z). In this sense P̃ and P are equivalent probability measures. 19 Example 7.2: (GBM) Consider the geometric Brownian motion (23) dSt = µSt dt + σSt dWt , so that (24) µ Zt = log (St/S0 ) = t≥0 ¶ 1 2 µ − σ t + σ Wt 2 and (25) Zt ∼ N (µ∗ t, σ 2 t), where µ∗ = µ − σ 2 /2. Thus (26) dP (z) = √ 1 2π σ 2 t e ∗ 2 t) − (z−µ 2σ 2 t dz. 20 Let (27) µ Z̃t = ¶ rf − 1 2 σ t + σ W̃t, 2 where rf is a riskfree rate, and W̃t is another Wiener process (the next example shows the relationship between W and W̃ ). Then (28) dP̃ (z̃) = √ 1 2π σ 2t ∗ 2 t) − (z−r 2σ 2 t e dz̃, where r∗ = rf − 21 σ 2. Because the density function of the normal distribution is always positive, trivially (29) P (dz) > 0 ⇐⇒ P̃ (dz) > 0. 21 The transformation between these two probability measures is (30) − ξ(z) = e 2 2 (µ∗ −r∗ )z− 1 (µ∗ −r∗ )t 2 2 σ , so that (31) dP̃ (z) = ξ(z) dP (z). 22 The Girsanov Theorem The Radon-Nikodym theorem gives the conditions under which the derivative ξ exist (i.e., that we can move to another probability measure without losing any information, which means that those events that have positive probability have also positive probability under the other measure). The Girsanov Theorem provides the conditions under which the Radon-Nikodym derivative exists for cases where Zt is a continuous stochastic process. 23 For the purpose, let {It}, t ∈ [0, T ] be a family of information sets (T < ∞). Define Rt (32) ξt = e Rt 2 1 0 Xu dWu − 2 0 Xu du , t ∈ [0, T ], where Xt is an It-measurable process (that is once It is given, the value of Xt is known), and Wt is a Wiener process with probability measure P . 24 It is assumed that Xt does not increase too fast, so that (33) · Rt E e 0 Xu du ¸ < ∞, t ∈ [0, T ], called Novikov condition (after a Russian mathematician). Using Itô (34) dξt = ξtXt dWt, from which we immediately see that ξt is a martingale, because Wt is a Wiener process and there is no drift component in (34). 25 This is also easy to see formally. Obviously, from (32) (35) ξ0 = 1. Thus, ξt = ξ0 + (36) = 1+ ·Z (37) E t Z t 0 Z t 0 ¸ ξsXs dWs ξsXs dWs. Z u ξs Xs dWs | Iu = 0 ξs Xs dWs , u < t, 0 Rt i.e., 0 ξsXs dWs is a martingale, and hence Z u (38) E[ξt|Iu] = 1 + ξsXs dWs = ξu, 0 implying that ξt is a martingale. 26 Theorem. (Girsanov) Let Wt be a Wiener process w.r.t probability measure P and w.r.t information sets It, and let Xt be as defined above. Then if the process ξt is a martingale w.r.t information sets It, then W̃t defined by (39) W̃t = Wt − Z t 0 Xu du, t ∈ [0, T ] is a Wiener process w.r.t information sets It and w.r.t probability measure (40) P̃ (A) = EP [1AξT ], where A ∈ IT and 1A is the indicator function of the event A. 27 In heuristic terms: If Wt is a Wiener process with probability measure P , then (41) dW̃t = dWt − Xt dt is a Wiener process with probability measure P̃ , such that dP̃ = ξT dP . Remark 7.1: Generally the theorem gives us a method to find the (equivalent) probability measures with respect to which a drifting process can be turned to a martingale. Remark 7.2: We only change the drift and live the volatility intact. 28 Example 7.3: Consider the general diffusion process (42) dS = a(S, t)dt + b(S, t)dW , where W is the Wiener process w.r.t the probability measure P , (43) dP (w) = √ 1 w2 2πt e− 2t dw, the normal distribution N (0, t). Define a(S, t) b(S, t) and assume that the drift a(S, t) and diffusion b(S, t) (44) Xt = are such that the Novikov condition (33) holds for Xt. Then defining Z (45) W̃ = W − 0 t a(S, u) du b(S, u) is a Wiener process with respect to the probability P̃ given by (40) and (46) dS = b(S, t)dW̃ is a martingale w.r.t the probability measure P̃ . 29 Example 7.4: As in Example 7.2, consider again the geometric Brownian motion (47) dS = µSdt + σSdW , where µ and σ are constants. The probability measure for W is again (43). Let r be the risk-free rate and consider the discounted price series S̃t = e−rt St. (48) Using Ito, (49) dS̃ = (µ − r)S̃dt + σSdW . Now (44) becomes simply (50) Xt = (µ − r)S̃ µ−r = , σ σ S̃ µ−r t σ is Wiener process w..r.t the probability measure P̃ , and (w.r.t. this measure) (51) (52) W̃ = W − dS̃ = σ S̃dW̃ is martingale. P̃ ? 30 In the Girsanov theorem Rt Rt 2 Xu dWu − 12 Xu du 0 ξt = e 0 (53) 1 1 2 = e σ (µ−r)Wt − 2σ2 (µ−r) t. Furthermore, we can consider only the events A = {Wt ≤ w}, w ∈ IR (because here the information sets on the real line are Borel sets that are essentially open intervals). Because Wt ∼ N (0, t) the associated probability measure is (43). Then Z w Z w u2 1 P̃ (A) = EP [1A ξt ] = ξt(u) √ e− 2t du = ξ(u)dP (u), 2π t −∞ −∞ (54) i.e., dP̃ (w) = ξt (w)dP (w) 1 1 2 1 1 2 = e σ (r−µ)w− 2σ2 (r−µ) tdP (w) (55) w2 1 − 2t e = e σ (r−µ)w− 2σ2 (r−µ) t √2π dw t = 1 √ 1 e− 2t (w− 2π t r−µ σ t)2 dw. 31 Denoting w̃ = w − µ−r t, σ we have finally (56) dP̃ (w̃) = √ 1 1 e− 2t w̃ dw̃, 2 2π t which is again the density of the N (0, t) distribution. The end result is that discounted price process (48) is martingale with respect to the probability measure P̃ . The solution of (48) is (57) 1 S̃t = S̃0e− 2 σ 2 t+σdW̃t . In terms of the original process from (48) St = ert S̃, we would have 1 (58) St = S0 e(r− 2 σ or (59) dS = rS dt + σSdW̃ , 2 )t+σ W̃t , i.e., we have essentially replaced the original drift µ with the risk free rate r, and the end result is a process whose discounted price process, S̃t = e−rt St is a martingale. Process (58) is usally called the risk neutral process. 32 Remark 7.3: When pricing options, we calculate the expected values with respect to the distribution of risk neutral process St given in (58). Its distribution is log-normal with density (60) fSt (y) = √ 1 2πt σy e θ̃t ) − (log(y)− 2σ 2 t 2 , y > 0, where (61) θ̃t = log S0 + (r − 1 2 σ )t. 2 Remark 7.4: The original distribution of St is lognormal with density (62) fSt (y) = √ 1 2πt σy e t) − (log(y)−θ 2σ 2 t 2 , y > 0, where (63) θt = log S0 + (µ − 1 2 σ )t. 2 33