Policy iterated lower bounds and linear MC upper bounds for Bermudan style derivatives

Transcription

1 Finance Winterschool 2007, Lunteren NL Policy iterated lower bounds and linear MC upper bounds for Bermudan style derivatives Pricing complex structured products Mohrenstr Berlin Jan. 2007

2 Bermudan callable products Simple example: (Bermudan) callable interest rate swap Finance Winterschool 2007, Lunteren NL Jan

3 Bermudan callable products Exotic example: cancelable snowball swap Snowball swap: Instead of the floating spot rate the holder pays a starting coupon rate I over the first year and in the forthcoming years (K + previous coupon spot rate) +, where the first coupon I and the strike rate K are specified in the contract. Cancelable snowball swap: The holder has the right to cancel this contract. What is the fair value of this cancelable product? Finance Winterschool 2007, Lunteren NL Jan

4 Optimal stopping Mathematical problem: Optimal stopping (calling) of a reward (cash-flow) process Z depending on an underlying (e.g. interest rate) process L Typical difficulties: L is usually high dimensional, for Libor interest rate models, d = 10 and up, so PDE methods do not work in general Z may only be virtually known, e.g. Z i = E F i j i C(L j) for some pay-off function C, rather than simply Z i = C(L i ) Z may be path-dependent Finance Winterschool 2007, Lunteren NL Jan

5 Optimal stopping The standard Bermudan pricing problem Consider an underlying process L in R D, e.g. a system of asset prices or Libor rates and a set of (future) dates T := {T 0, T 1,..., T k } Bermudan derivative: An option to exercise a cashflow C(T τ, L(T τ )) at a future time T τ T, to be decided by the option holder Valuation: If N, with N(0) = 1, is some discounting numeraire and P the with N associated pricing measure, then with Z τ := C(T τ, L(T τ ))/N(T τ ), the t = 0 price of the option is given by the optimal stopping problem V 0 = sup E F 0 Z (τ), τ {0,...,k} where the supremum runs over all stopping indexes τ with respect to {F Ti, 0 i k}, where (F t ) t 0 is the usual filtration generated by L. Finance Winterschool 2007, Lunteren NL Jan

6 Optimal stopping At a future time point t, when the option is not exercised before t, the Bermudan option value is given by with κ(t) := min{m : T m t}. The process V t = N(t) sup E F t Z (τ) τ {κ(t),...,k} Y t := V t N(t), called the Snell-envelope process, is a supermartingale, i.e. E F s Y t Y s Finance Winterschool 2007, Lunteren NL Jan

7 Optimal stopping Canonical Solution by Backward Dynamic Programming Set Y (i) := Y (T i ), F (i) := F Ti. At the last exercise date T k we have, Y (k) = Z (k) and for 0 j < k, Y (j) = max ( ) Z (j), E F j Y (j+1). The first optimal stopping time (index) is then obtained by } = inf {j, i j k : Y (j) Z (j). τ i Nested Monte Carlo simulation of the price Y 0 would thus require N k samples when conditional expectations are computed with N samples Typically, N=10000, k=10 exercise opportunities, give samples!! Finance Winterschool 2007, Lunteren NL Jan

8 Optimal stopping Finance Winterschool 2007, Lunteren NL Jan

9 Contents New approaches: I A path-by-path policy iteration methodology to improve upon standard methods (e.g. Longstaff-Schwartz, Piterbarg, Andersen) II Application to complex exotic structures III A linear Monte Carlo algorithm for price upper bounds via regression estimators of Doob martingale parts Finance Winterschool 2007, Lunteren NL Jan

10 Part I: Iterative methods Iterative construction of the optimal stopping time References: Kolodko, A., Schoenmakers, J. (2006) Iterative construction of the optimal Bermudan stopping time. Finance and Stochastics, 10(1), Finance Winterschool 2007, Lunteren NL Jan

11 Part I: Iterative construction of the optimal stopping time Improving upon given input stopping policies We consider an input stopping family (policy) (τ i ), which satisfies the consistency conditions: i τ i k, τ k = k, τ i > i τ i = τ i+1, 0 i < k, and the corresponding lower bound process Y for the Snell envelope Y, Y (i) := E F(i) Z (τ i) Y (i) Example input policies: The policy, τ i i. says: exercise immediately! The policy τ i := inf{j i : L(T j ) G R D } process L enters a certain region G exercises when the underlying The policy τ i = inf{j : i j k, max p: j p k E F(j) Z (p) Z (j) } waits until the cashflow is at least equal to the maximum of still-alive Europeans ahead Finance Winterschool 2007, Lunteren NL Jan

12 Part I: Iterative construction of the optimal stopping time Introduce an intermediate process One step improvement: Ỹ (i) := max EF(i) Z (τ p) p: i p k and use Ỹ (i) as a new exercise criterion to define a new exercise policy τ i : = inf{j : i j k, Z (j) Ỹ (j) } = inf{j : i j k, Z (j) max Z (τp) }, p: j p k Then consider the process Ŷ (i) := E F(i) Z ( τ i) as a next approximation of the Snell envelope Key Proposition It holds 0 i k Y (i) Ỹ (i) Ŷ (i) Y (i), 0 i k Finance Winterschool 2007, Lunteren NL Jan

13 Part I: Iterative construction of the optimal stopping time Iterative construction of the optimal stopping time Take an initial family of stopping times (τ (0) i ) satisfying the consistency conditions and set Y 0 (i) := E F(i) Z i τ (0) i k, τ (0) k = k, τ i > i τ i = τ i+1, (τ (0) i is constructed with τ (m) i Then define ), 0 i k. Suppose that for m 0 the pair ( ) (τ (m) i ), (Y m (i) ) being consistent and Y m (i) := E F (m) (τ iz i ), 0 i k. τ (m+1) i := inf{j : i j k, max EF(j) (m) (τ p Z ) Z (j) } p: j p k =: inf{j : i j k, Ỹ m+1 (j) Z (j) }, 0 i k, and set Y m+1 (i) := E F(i) (m+1) (τ Z i ) Finance Winterschool 2007, Lunteren NL Jan

14 Part I: Iterative construction of the optimal stopping time By the key proposition we thus have Y 0 (i) Y m (i) Ỹ m+1 (i) Y m+1 (i) Y (i), 0 m <, 0 i k. and it is shown that for m 1, where τ i τ (m) i τ (m+1) i τ i, is the first optimal stopping time. We so may take limits and it holds, Y (i) := (a.s.) lim m Y m (i) and τ i := (a.s.) lim m τ (m) i, 0 i k, and, Y (i) = (a.s.) lim E F(i) (m) (τ Z i ) = E F(i) Z (τ i ), 0 i k m Finance Winterschool 2007, Lunteren NL Jan

15 Part I: Iterative construction of the optimal stopping time Theorem The constructed limit process Y coincides with the Snell envelope process Y and (τi ) coincides with (τi ); the family of first optimal stopping times. We have Y (i) = Y (i) = E F(i) Z (τ i ), 0 i k. Moreover: It even holds Y m (i) = Y (i) for m k i After k = #exercise dates iterations the Snell Envelope is attained! Finance Winterschool 2007, Lunteren NL Jan

16 Part I: Iterative construction of the optimal stopping time Iteration procedure vs backward dynamic program Exercise date 0 1 k 2 k 1 k 0 Y (0) 0 Y (0) 1 Y (0) k 2 Y (0) k 1 1 Y (1) 0 Y (1) 1 Y (1) k 2 2 Y (2) 0 Y (2) 1 Y k 2 Y k 1 Y k 1 Iteration level k 1 k Y (k 1) 0 Y 1 Y k 2 Y 0 Y 1 Y k 2 Yk 1 Yk 1 Y k Y k Y k Yk Yk Finance Winterschool 2007, Lunteren NL Jan

17 Part I: Dual upper bounds Upper approximations of the Snell envelope by Duality The Dual Method Consider a discrete martingale (M j ) j=0,...,k with M 0 = 0 with respect to the filtration ( F (j)). Following Rogers, Haugh and Kogan, we observe that j=0,...,k ] Y 0 = sup E F 0 Z (τ) = sup E F 0 [Z (τ) M τ τ {0,...,k} τ {0,,...,k} ] E F 0 max [Z (j) M j 0 j k Hence the r.h.s. gives an upper bound for the Bermudan price V 0 = Y 0. Finance Winterschool 2007, Lunteren NL Jan

18 Part I: Dual upper bounds Theorem (Davis Karatzas (1994), Rogers (2001), Haugh & Kogan (2001)) Let M be the (unique) Doob-Meyer martingale part of ( Y ) (j) 0 j k, i.e. is an ( F (j)) -martingale which satisfies M Y (j) = Y 0 + M j F j, j = 0,..., k, with M0 := F0 := 0 and F being such that Fj is F (j 1) measurable for j = 1,..., k. Then we have [ ] Y0 = E F 0 max Z (j) Mj. 0 j k Finance Winterschool 2007, Lunteren NL Jan

19 Part I: Converging upper bounds Convergent upper bounds from a convergent sequence of lower bounds From our previously constructed sequence of lower bound processes Y m (i) with Y m (i) Y (i), we deduce by duality a sequence of upper bound processes: Y m (i) up := E F i max i j k (Z(j) j l=i+1 Y m (l) + j l=i+1 Then, by a theorem of (Kolodko & Schoenmakers 2004), 0 m (i) E F i k 1 j=i max E F(l 1) Y m (l) ) =: Y m (i) + m (i). ( ) E F j Y m (j+1) Y m (j), 0. Thus, by letting m on the r.h.s., (a.s.) lim m m (i) = 0, the sequence Y m up converges to the Snell envelope also, i.e., 0 i k. Hence, (a.s.) lim Y m (i) up m = (a.s.) lim m Y m (i) = Y (i), 0 i k. Finance Winterschool 2007, Lunteren NL Jan

20 Part I: Application: Bermudan swaptions A first numerical example: Bermudan swaptions in the LIBOR market model Consider the Libor Market Model with respect to a tenor structure 0 < T 1 < T 2 <... < T n, e.g. in the spot Libor measure P induced by the numeraire B (t) := B m(t)(t) B 1 (0) m(t) 1 i=0 (1 + δ i L i (T i )) with m(t) := min{m : T m t}. The dynamics of the forward Libor L i (t) is given by a system of SDE s dl i = i j=m(t) δ j L i L j γ i γ j 1 + δ j L j dt + L i γ i dw. Here δ i = T i+1 T i are day count fractions, and t γ i (t) = (γ i,1 (t),..., γ i,d (t)) are deterministic volatility vector functions defined in [0, T i ], called factor loadings. Finance Winterschool 2007, Lunteren NL Jan

21 Part I: Application: Bermudan swaptions A (payer) Swaption over a period [T i, T n ], 1 i k. A swaption contract with maturity T i and strike θ with principal $1 gives the right to contract at T i for paying a fixed coupon θ and receiving floating Libor at the settlement dates T i+1,...,t n. So by this definition, its cashflow at maturity is + n 1 S i,n (T i ) := B j+1 (T i )δ j (L j (T i ) θ). j=i A Bermudan Swaption gives the the right to exercise a cashflow C Tτ := S τ,n (T τ ) at an exercise date T τ {T 1,..., T n } to be decided by the option holder. Finance Winterschool 2007, Lunteren NL Jan

22 Part I: Application: Bermudan swaptions 10 yr. Bermudan swaption: (20 underlying LIBORS) Comparison of Y 1, Y 2, Y 1,up, where τ (0) i i (trivial initial stopping family) θ d Y 1 (SD) Y 2 (SD) Y 1,up (SD) (0.5) (2.4) (0.7) (0.4) (2.4) (0.7) (ITM) (0.4) (2.1) (0.6) (0.4) (2.0) (0.6) (0.4) 381.2(1.6) 382.9(0.8) (0.3) 364.4(1.5) 366.4(0.8) (ATM) (0.3) 343.5(1.3) 345.6(0.7) (0.3) 338.7(1.2) 341.2(0.8) (0.2) 121.0(0.6) 121.3(0.4) (0.2) 113.8(0.5) 114.9(0.4) (OTM) (0.2) 100.7(0.4) 101.5(0.3) (0.2) 96.9(0.4) 97.7(0.3) Finance Winterschool 2007, Lunteren NL Jan

23 Part I: Application: Bermudan swaptions Conclusions from the tables: The computed lower bound Y 2, hence the second iteration, is within 1% or less (relative to the price) of the Dual upper bound Y 1,up Computation times (order of minutes) may be considered low in view of the high-dimensionality of the problem! Finance Winterschool 2007, Lunteren NL Jan

24 Part I: General conclusions Some more general remarks The iterative approach provides a general method for improving upon any given input stopping policy obtained by other means (e.g. Andersen, Longstaff- Schwartz) Computation times may be reduced further by a scenario selection method by Bender, Kolodko, Schoenmakers Finance Winterschool 2007, Lunteren NL Jan

25 Part II: Exotic products Pricing of path-dependent cancellables References: C. Bender, A. Kolodko, and J. Schoenmakers. Iterating cancellable snowballs and related exotics. Risk, pages , September Finance Winterschool 2007, Lunteren NL Jan

26 Part II: Exotic products Consider a path dependent cancelable contract which generates (possibly negative) cash-flows C 1,..., C τ up to a cancelation date τ. The cash-flows of this contract are equivalent to an aggregated cash-flow at cancellation date, B (T τ )Z τ := B (T τ ) τ Z j, j=1 with Z i := C i /B (T i ) being discounted cash-flows with respect to the numeraire B. Product price at time zero: V0 cancel := sup E F 0 Z τ = τ {1,...,k} sup E F 0 τ {1,...,k} τ Z j, j=1 where the supremum is taken over all stopping indices with values in the set {1,..., k}. Finance Winterschool 2007, Lunteren NL Jan

27 Part II: Exotic products Path-dependent callables A path dependent callable contract generates C τ+1,..., C k when called at τ. It is equivalent to the sum of a non-callable and a cancelable one (and vice versa): V0 call := sup E F 0 τ {1,...,k} k = E F 0 Z j + j=1 k j=τ+1 Z j sup E F 0 τ {1,...,k} τ ( Z j ) j=1 Finance Winterschool 2007, Lunteren NL Jan

28 Part II: Snowballs Example: The cancelable snowball swap Pays semi-annually a constant rate I over the first year and in the forthcoming years (Previous coupon + A Libor) +, semi-annually, where A is given in the contract. For this case we take K i := I, i = 0, 1, K i := (K i 1 + A i L i (T i )) +, i = 2,..., n 1, with A 2 := 0.03, A i+1 = A i for even i, A i+1 = A i for odd i. Cancelation is allowed for 2 τ < n, n = 20 (10 years) Effective discounted cashflows at T j : Z j := (L j 1(T j 1 ) K j 1 ) δ j 1, B (T j ) hence aggregated up to cancelation Z τ = τ j=1 Z j. Finance Winterschool 2007, Lunteren NL Jan

29 Part II: Snowballs Take an input policy satisfying construct the new policy and compute the iterated price Iterating the snowball swap i τ i k, τ k = k, τ i > i τ i = τ i+1, 0 i < k, τ i := inf{j i : Z j max p: j p k EF j Z τp } = inf{j i : 0 max p: j p k EF j Ŷ 0 := E F 0 Z τ0, τ p q=j+1 which is generally an improvement of Y 0 due to policy τ. Z q } Finance Winterschool 2007, Lunteren NL Jan

30 Part II: Snowballs Numerical results for typical market data Improved Andersen d Y (0; τ A ) (SD) Ŷ (0; τ A ) (SD) Y up (0; τ A ) (SD) (0.238) (0.318) (0.247) (0.231) (0.389) (0.293) (0.217) (0.460) (0.476) outer and 500 inner paths for Ŷ and outer (with 500 inner) paths for Y up. Improved least-squares regression method (Piterbarg) d Y (0; τ LS ) (SD) Ŷ (0; τ LS ) (SD) Y up (0; τ LS ) (SD) (0.243) (0.632) (0.313) (0.238) (0.466) (0.346) (0.224) (0.469) (0.379) outer and 500 inner paths for Ŷ and outer (with 500 inner) paths for Y up. Finance Winterschool 2007, Lunteren NL Jan

31 Part II: Snowballs Improving an Andersen-like optimization of the LS exercise boundary d Y (0; τ LS-A) (SD) Ŷ (0; τ LS-A) (SD) Y up (0; τ LS-A) (SD) (0.237) (0.349) (0.244) (0.230) (0.345) (0.244) (0.219) (0.456) (0.418) (0.217) (0.445) (0.430) outer and 100 inner paths for Ŷ and 5000 outer (with 500 inner) paths for Y up. τ LS-A, i = inf{j i : Z j H j + Y LS, j } with optimized constants H j. Finance Winterschool 2007, Lunteren NL Jan

32 Part II: Snowballs Message: (i) Price the callable using Pitterbarg s version of Longstaff-Schwartz; (ii) Improve the obtained exercise boundary with an Andersen-like optimization; (iii)compute the Dual upperbound due to the stopping time τ LS-A, 0; (iv)if there is still a significant gap between lower and upper bound, then improve the policy τ LS-A, i by the iteration method. Finance Winterschool 2007, Lunteren NL Jan

33 Part III: Fast upper bounds True upper bounds via non-nested Monte Carlo Joint work with D. Belomestny and C. Bender Finance Winterschool 2007, Lunteren NL Jan

34 Part III: Fast upper bounds For any martingale M Tj, 0 j k with respect to the filtration (F Tj ; 0 j k) starting at M 0 = 0 [ ] Y up 0 (M) := EF 0 max (Z T j M Tj ) 0 j k is an upper bound for the price of the Bermudan option with discounted cashflow Z Tj. Exact Bermudan price is attained at the martingale part M of the Snell envelope, YT j = YT 0 + MT j + FT j, M T 0 = F T 0 = 0 and F T j is F Tj 1 measurable Finance Winterschool 2007, Lunteren NL Jan

35 Part III: Fast upper bounds (I) Assume the underlying process L to be Markovian, and the filtration F to be generated by a d-dimensional Brownian motion W. (II) Assume Y Tj = u(t j, L(T j )) is some approximation of the Snell envelope Y T j, 0 j k, with Doob decomposition Y Tj = Y T0 + M Tj + F Tj, M T0 = F T0 = 0 and F Tj is F Tj 1 measurable. It then holds: Y Tj+1 Y Tj = M Tj+1 M Tj + F Tj+1 F Tj M Tj+1 M Tj = Y Tj+1 E T j [Y Tj+1 ], with M Tj =: Tj 0 H t dw t =: Tj 0 h(t, L(t))dW t, j = 0,..., k. Finance Winterschool 2007, Lunteren NL Jan

36 Part III: Fast upper bounds We are going to estimate h(, ) (hence H) at the finite partition π = {t 0,..., t I } such that t 0 = 0, t I = T, and {T 0,..., T k } π. We may write formally, Y Tj+1 Y Tj H tl (W tl+1 W tl ) + F Tj+1 F Tj. t l π;t j t l <T j+1 By multiplying both sides with (W d t i+1 W d t i ), T j t i < T j+1, and taking F ti - conditional expectations, we get by the F Tj+1 -measurability of F Tj, and so define H d t i 1 [ ] E F t i (Wt d t i+1 t i+1 Wt d i )Y Tj+1, i H π t i := 1 π i E F t i [ ( π W i ) Y Tj+1 ], Tj t i < T j+1, with π i := t i+1 t i and π W d i := W d t i+1 W d t i. Finance Winterschool 2007, Lunteren NL Jan

37 Part III: Fast upper bounds The corresponding approximation of the martingale M is MT π j := Ht π i ( π W i ). t i π;0 t i <T j Theorem: lim E π 0 where π denotes the mesh of π. [ ] max M T π 0 j k j M Tj 2 = 0 Finance Winterschool 2007, Lunteren NL Jan

38 Part III: Fast upper bounds The conditional expectations in the definition of H π are, in fact, functions of L(t i ). Precisely, H π t i = h π (t, L(t i )) = 1 π i E (t i,l(t i )) [ ( π W i ) u(t j+1, L(T j+1 )) ], T j t i < T j+1. which may be computed by regression: Take basis functions ψ(t i, ) = (ψ r (t i, ), r = 1,..., R) and N independent samples (t i, n L(t i )), n = 1,..., N of L(t i ) constructed from the Brownian increments π nw i, n = 1,..., N. Construct the regression matrix where A t i := (A t i A ti ) 1 A t i, A ti = (ψ r (t i, n L(t i ))) n=1,...,n,r=1,...,r Finance Winterschool 2007, Lunteren NL Jan

39 Part III: Fast upper bounds Result: ( ĥ π (t i, x) = ψ(t i, x) A π W i t i π i =: ψ(t i, x) β ti, ) Y Tj+1, T j t i < T j+1 where ( π W i π i Y Tj+1 ) = ( π n W d i π i ny Tj+1 ) n=1,...,n, d=1,...,d nỹt j+1 := u(t j+1, n L(T j+1 )), and β ti is the R D matrix of estimated regression coefficients at time t i., Finance Winterschool 2007, Lunteren NL Jan

40 Part III: Fast upper bounds Ŷ up ( M π ) = 1 Ñ Ñ n=1 True linear Monte Carlo upperbound: max 0 j k z(t j, n L(Tj )) ĥ π (t i, n L(Tj ))( π Wi ), t i π;0 t i <T } j {{} ( ) by doing a new simulation n L(Tj ), π n W i n = 1,..., Ñ. ( ) is always a martingale, so the upper bound is true! Finance Winterschool 2007, Lunteren NL Jan

41 Part III: Numerical example: Max Call on D assets Black-Scholes model: dx d t = (r δ)x d t dt + σx d t dw d t, d = 1,..., D, Pay-off: Z t := z(x t ) := (max(x 1 t,..., X D t ) κ) +. T k = 3yr, k = 9 (ex. dates), κ = 100, r = 0.05, σ = 0.2, δ = 0.1, different D and x 0 D x 0 Lower Bound Upper Bound Upper Bound Upper Bound Y 0 Y up 0 ( M π ) Y up (0) 10 4,200 up Y 10 4, ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ±0.12 Finance Winterschool 2007, Lunteren NL Jan

42 Part I+II+III: Literature D. Belomestny, C. Bender, J. Schoenmakers. True upper bounds for Bermudan products via non-nested Monte Carlo. Working paper C. Bender and J. Schoenmakers. An iterative method for multiple stopping: Convergence and stability. Adv. Appl. Prob., 38(3): , C. Bender, A. Kolodko, and J. Schoenmakers. Iterating cancellable snowballs and related exotics. Risk, pages , September C. Bender, A. Kolodko, and J. Schoenmakers. Enhanced policy iteration for American options via scenario selection. Quantitative Finance, tent. accepted A. Kolodko and J. Schoenmakers. Iterative construction of the optimal Bermudan stopping time. Finance and Stochastics, 10:27-49, J. Schoenmakers. Robust Libor modelling and pricing of derivative products. Chapman & Hall - CRC Press, Boca Raton London New York Singapore, Finance Winterschool 2007, Lunteren NL Jan