Simple Improvement Method for Upper Bound of American Option

Transcription

1 Simple Improvement Method for Upper Bound of American Option Koichi Matsumoto (joint work with M. Fujii, K. Tsubota) Faculty of Economics Kyushu University k-matsu@en.kyushu-u.ac.jp 6th World Congress of the Bachelier Finance Society June, 2010@Hilton, Toronto, Canada 1/15

2 Introduction Numerical Methods for Pricing American Option 1. Closed-Form Solution: It is difficult to find a closed-form solution. 2. Lattice Methods: When the condition is simple, the lattice methods give good approximated solutions. 3. Monte Carlo Simulation: When the condition is complicated, the Monte Carlo simulation is practical. Monte Carlo simulation Lower Bound: Astopping time gives a lower bound. The least-square method gives a good stopping time. Longstaff and Schwartz (2001) Upper Bound: Amartingale gives an upper bound. Can we find a good martingale? 2/15

3 Setup The saving account is the numeraire. All prices are discounted prices. T N : Fixed Maturity (Ω, F, P, {F k ; k =0, 1,...,T }) : Filtered probability space S k (k =0, 1,...,T ) : Price Process of Risky Asset H k (k =0, 1,...,T ) : Payoff of American Option V k (k =0, 1,...,T ) : Price of American Option Assumption P is a unique equivalent martingale measure. F k is a natural filtration generated by S. WewriteE k [ ] =E [ F k ]. H is an adapted process. Definition 1 A supersolution is a supermartingale X satisfying X k H k, k =0, 1,...,T 1 and the maturity condition, that is, X T = H T. V is a minimum supersolution. Any supersolution is an upper bound process of the American option. 3/15

4 Main Problem Suppose that a supersolution U is given. Note that U 0 is an upper bound. Suppose that the lower bound process L of the continuation value is given. L k E k[v k+1] V k U k, k < T, {z } continuation value L T = H T (= U T ). We want to improve the upper bound U 0 in the Monte Carlo simulation. Chen and Glasserman (2007) proposes an iterative method. 1. Using the supersolution U, a martingale is given by Mk U = P k t=1 (Ut Et 1[Ut]), k =0, 1,...,T. 2. Using the martingale M, a new supersolution (= upper bound process) is given by Uk M = E k[max k t T (H t M t)] + M k, k =0, 1,...,T. The iterative improvement converges to the true price. The calculation of the conditional expectation is necessary at all times and all states for the Doob decomposition. The lower bound process is not used. We want to find a computationally-efficient improvement method using L. 4/15

5 Basic Result Let T k be the set of the stopping times whose values are greater than or equal to k. Theorem 1 Let τ 1,τ 2 T 0 and τ 1 τ 2. Suppose that V satisfies the martingale property in [0,τ 1 ] [τ 1 +1,τ 2 ], that is, V k = E k [V k+1 ], k [0,τ 1 1] [τ 1 +1,τ 2 1]. Martingale Martingale 0 τ 1 τ 1 +1 τ 2 Let Then w(τ 1,τ 2 ) = E [max (H τ1, E τ1 [U τ2 ])]. V 0 w(τ 1,τ 2 ) U 0. }{{} New Upper Bound The problem is to find an appropriate pair of stopping times (τ 1, τ 2 ). 5/15

6 Methods 1, 2 We use the mathematical convention the minimum over the empty set is, min( ) =+. Lemma 1 Let τ1 =min{k 0 H k > L k} T. Then V satisfies the martingale property in [0,τ1 ],thatis,v k = E k[v k+1] fork [0,τ1 1]. Corollary 1 Let w 1 L = w(τ 1,τ 1 ). Then V 0 w 1 L U 0. Corollary 2 Let w 2 L = w(τ 1, (τ 1 +1) T ). Then V 0 w 2 L w 1 L U 0. wl 2 wl 1. wl 2 is a better upper bound than wl 1. When U k = E k[max k t T (H t M t)] + M k, wl 1 = E[max τ 1 t T (H t M t)], wl 2 = E[max H τ, E 1 τ [max 1 (τ 1 +1) T t T (H t M t)] + M τ ]. 1 w 1 L includes no conditional expectation per path. w 2 L requires only one conditional expectation per path. The iterated method requires T conditional expectations per path. The calculations of w 1 L and w 2 L spend much less time than that of the iterative method. The proposed methods are more efficient. 6/15

7 Method 3 Lemma 2 Let τ2 =min{k >τ1 H k > L k} T.ThenV satisfies the martingale property in [τ1 +1,τ2 ],thatis,v k = E k [V k+1] fork [τ1 +1,τ2 1]. Corollary 3 Let wl 3 = w(τ1,τ2 ). Then V 0 w 3 L w 2 L U 0. wl 3 is the best upper bound of the three proposed methods. When U k = E k[max k t T (H t M t)] + M k, «wl 3 = E[max H τ 1, E τ 1 [ max Mt)] + M τ 2 t T(Ht τ 1 ]. We have to calculate τ2. When the lower bound process can be calculated by an analytic formula, the calculation of τ2 is not time-consuming and then the amount of calculation of wl 3 is as much as that of wl. 2 7/15

8 Lower Bound Effect Lemma 3 Let τ a,τ b T 0.Ifτ a τ b,then w(τ a,τ a) w(τ b,τ b), w(τ a, (τ a +1) T ) w(τ b, (τ b +1) T ). Proposition 1 Let L a and L b be lower bound processes. Suppose that L a k L b k, k =0, 1,...,T. L b is a better lower bound process than L a. Then wl 1 a w 1 L b, wl 2 a w 2 L b, w 3 L a w 3 L b. The better a lower bound process is, the greater improvement of upper bound can be expected. 8/15

9 European Option Based Model Let V E be the price process of the European option satisfying V E k = E k[h T ]. M k = V E k V E 0, U k = E k[max(ht Mt)] + Mk. k t T We call this model the European option based model. Proposition 2 Consider the European option based model with L = V E. τ T 0 satisfies τ<τ1,then If U 0 = w(τ,τ) =w(τ,(τ +1) T ). If L is smaller than V E, it fails to improve the upper bound. Proposition 3 In the European option based model, if L = V E,thenwehave U 0 = w 1 L w 2 L = w 3 L. V E is the worst lower bound which may improve the upper bound. We check whether wl 2 = wl 3 generated by V E can improve the upper bound by the numerical analysis. 9/15

10 Simulation Condition The price process is given by the Black Scholes Model, thatis, «S k = S k 1 exp σ2 2 t + σp tξ k, k =1,...,T, H k =max Ke rk t S k, 0, k =0, 1,...,T, where ξ 1,...,ξ T are independent and standard normally distributed. Let L = V E,thatis, L k = KΦ(d(k, T, K, 0)) S k Φ(d(k, T, K,σ 2 )), k =0, 1,...,T 1 where Φ( ) is the standard normal distribution function and 1 d(k, T, K, r) = σ p log K r 12 ««(T k) t S σ2 (T k) t. k S 0 = 100, r =0.04, σ =0.3, t =0.01, T =50, 100, 150. The number of paths for calculating the expectation is 2, 500. The number of paths for calculating the conditional expectation is 500. The antithetic sampling is used. 10/15

11 Better Lower Bound Let L a T = L b T = H T and for k =0, 1,...,T 1,! L a k =max sup E k[h τ ], L b k = sup E k[h τ ] t 0 >k τ T t0,t τ T k+1 where T t0,t is the set of the stopping times whose values are t 0 or T. L a can be calculated by the analytic formula since sup E t0 [H τ] = KΦ(d(t 0, t 1, S t 1, 0)) S t0 Φ(d(t 0, t 1, S t 1,σ 2 )) τ T t1,t + KΦ 2( d(t 0, t 1, S t 1, 0), d(t 0, T, K, 0); t1 t0 T t 0 ) S t0 Φ 2( d(t 0, t 1, St 1,σ 2 ), d(t 0, T, K,σ 2 t1 t0 ); ) T t 0 where Φ 2(, ; ρ) is the standard bivariate normal distribution function. St 1 is a solution of KΦ(d(t 1, T, K, 0)) S t 1 Φ(d(t 1, T, K,σ 2 )) = Ke rt1 t S t 1. L b is used in order to estimate the maximum improvement. Note that L b can be calculated by the lattice tree. 11/15

12 Numerical Result (Lower Bound Effect) K =90(OTM) T U 0 wl 3 wl 3 a w 3 L V b (0.002) 3.469(0.002) 3.465(0.002) 3.463(0.002) (0.006) 5.856(0.006) 5.845(0.006) 5.821(0.006) (0.010) 7.612(0.009) 7.584(0.010) 7.542(0.010) K = 100 (ATM) T U 0 wl 3 wl 3 a w 3 L V b (0.004) 7.608(0.004) 7.596(0.004) 7.581(0.004) (0.009) (0.008) (0.009) (0.009) (0.015) (0.013) (0.014) (0.014) K = 110 (ITM) T U 0 wl 3 wl 3 a w 3 L V b (0.006) (0.006) (0.006) (0.006) (0.013) (0.011) (0.012) (0.012) (0.019) (0.016) (0.018) (0.019) U 0 > w 3 L > w 3 L a > w 3 L b > V 0. L L a L b, Lower Bound Effect 2. w 3 L b > V 0. The proposed methods can improve the upper bound efficiently but cannot attain the true price. 12/15

13 Bermudan Max Call Option on five Assets Suppose that the price processes S i for i =1,...,5aregivenbyS0 i = S 0, «Sk i = Sk 1 i exp q σ2 t + σ p «tξk i, k =1,...,T. 2 H k =max`max 1 i 5 Sk i Ke rk t, 0, k =0, 1,...,T. K = 100, q =0.1, σ=0.2, r =0.05, T = t 3. The number of paths for calculating the expectation and the conditional expectation are 250, 000 and 500 respectively. An upper bound process is generated by the single European options. A lower bound process is based on the least square method. The true price V 0 is the point estimate in Broadie and Glasserman (2004). t S 0 U 0 wl 1 wl 2 V 0 1/ (0.015) (0.015) (0.014) / (0.019) (0.020) (0.019) / (0.023) (0.024) (0.023) / (0.014) (0.014) (0.013) / (0.018) (0.018) (0.017) / (0.021) (0.022) (0.021) /15

14 Concluding Remarks We have proposed a simple and computationally tractable improvement method for the upper bound of American options. The method is based on two stopping times. The stopping times are generated from a lower bound process of the continuation value. A better, namely higher lower bound process gives a greater improvement of the upper bound. Our method can be used together with the approximation of lower bound process by the least square method. 14/15

15 References L. Andersen and M. Broadie, Primal-dual simulation algorithm for pricing multidimensional American options, Management Science 50 (2004), pp M. Broadie and M. Cao, Improved lower and upper bound algorithms for pricing American options by simulation, Quantitative Finance 8 (2008), pp M. Broadie and P. Glasserman, A stochastic mesh method for pricing high-dimensional American options, Journal of Computational Finance 7 (2004), pp N. Chen and P. Glasserman, Additive and multiplicative duals for American option pricing, Finance and Stochastics 11 (2007), pp M. Haugh and L. Kogan, Pricing American options: a dual approach, Operations Research 52 (2004), pp A. Kolodko and J. Schoenmakers, Iterative Construction of the Optimal Bermudan Stopping Time, Finance and Stochastics 10 (2006), pp F. Longstaff and E. Schwartz, Valuing American options by simulation: a simple least-squares approach, The Review of Financial Studies 14 (2001), pp M. S. Joshi, A Simple Derivation of and Improvements to Jamshidian s and Rogers Upper Bound Methods for Bermudan Options, Applied Mathematical Finance 14 (2007), pp L. C. G. Rogers, Monte Carlo valuation of American options, Mathematical Finance 12 (2002), pp /15