IMA 2003 Workshop, March 12-19, 2003

A Robust Option Pricing Problem

Laurent El Ghaoui Department of EECS, UC Berkeley

3

Robust optimization
standard form: min sup f0 (x, u) : u U , fi (x, u) 0, i = 1, . . . , m
x uU

x Rn is the decision variable u is a parameter vector aecting the problem data set U describes uncertainty on u

a semi-innite optimization problem

4

Agenda

option basics option pricing problem worst-case approach

joint work with: A. dAspremont

5

Options
Options are time bombs. Warren Buett, 2003 let x(T ) denote the price of an asset at time T an European call option with maturity T and strike price K is a contract with payo (x(T ) K )+ = sup(x(T ) K, 0)

Options
Options are time bombs. Warren Buett, 2003 let x(T ) denote the price of an asset at time T an European call option with maturity T and strike price K is a contract with payo (x(T ) K )+ = sup(x(T ) K, 0)

what is the price of the option, assuming no arbitrage (free lunch) is possible?
7

Price of option

fundamental result of nance: assuming a zero risk-free interest rate, the no-arbitrage price of the option is E (x(T ) K )+ for some distribution on the asset price at time T

such a is called risk-neutral

8

Forwards

under the risk-neutral distribution , the current asset price is E x(T )

i.e., is a martingale

Basket options

let x(T ) be a n-vector of prices of assets at time T a basket option with weight vector w and strike K is the contract with payo (wT x(T ) K )+ denote the basket option by (w, K ) and its price by C (w, K ) := E (wT x(T ) K )+

(maturity date is implicit here)

10

Option pricing problem

given w0 , w1 , . . . , wm in Rn + K0 , K1 , . . . , Km in R+ p1 , . . . , pm in R+ (basket weights) (strike prices) (observed option prices)

determine the price of the basket option with weight w0 and strike price K0

11

Option pricing problem

given w0 , w1 , . . . , wm in Rn + K0 , K1 , . . . , Km in R+ p1 , . . . , pm in R+ (basket weights) (strike prices) (observed option prices)

determine the price of the basket option with weight w0 and strike price K0

in practice, we are also given the current asset prices themselves, q Rn + (forwards)
12

Challenges & methods

challenges: the risk-neutral measure is not the empirical distribution of assets hence, we may have to rely on option & forward prices only

two approaches: model-based approach arbitrage-based approach

13

Model-based approach
assume a log-normal diusion model for the asset prices ds = Asdt + Bsdw, where s = log x, and w is a multidimensional Brownian motion t the model to observed option prices

14

Model-based approach
assume a log-normal diusion model for the asset prices ds = Asdt + Bsdw, where s = log x, and w is a multidimensional Brownian motion t the model to observed option prices

with some approximations, this problem is an SDP (Aspremont, 2000)

15

Model-based approach
assume a log-normal diusion model for the asset prices ds = Asdt + Bsdw, where s = log x, and w is a multidimensional Brownian motion t the model to observed option prices

pros and cons: very versatile, and easy to solve (albeit only recently) makes a structural assumption about the risk-neutral measure provides a point estimate for the price of basket (w0 , K0 )
16

No-arbitrage approach
nd bounds on the price of basket (w0 , K0 ) by solving semi-innite LP sup / inf s.t.
T E (w0 x K0 )+

T E (wi x K i ) + = pi , i = 1 , . . . , m

the optimization variable is the risk-neutral probability measure

17

No-arbitrage approach
nd bounds on the price of basket (w0 , K0 ) by solving semi-innite LP sup / inf s.t.
T E (w0 x K0 )+

T E (wi x K i ) + = pi , i = 1 , . . . , m

the optimization variable is the risk-neutral probability measure

pros and cons: problem may be dicult approach provides only bounds, but . . . . . . makes no assumptions about market dynamics
18

Upper bound problem

upper bound problem: psup := sup
P

0 (x) (x)dx :

(x)dx = 1, i (x) (x)dx = pi , i = 1, . . . , m,

where = Rn + P is the set of densities with support in

T i (x) := (wi x Ki )+ , i = 0, 1, . . . , m 19

Upper bound problem

upper bound problem: psup := sup
f F

0 (x)f (x)dx :

f (x)dx = 1, i (x)f (x)dx = pi , i = 1, . . . , m,

Lagrangian: L( , , 0 ) =

0 (x) (x)dx + 0 1 + T p

(x)dx

(x) (x)dx

20

Upper bound problem

upper bound problem: psup := sup
f F

0 (x)f (x)dx :

f (x)dx = 1, i (x)f (x)dx = pi , i = 1, . . . , m,

the dual is the robust linear programming problem dsup := infm T p + 0 x , T (x) + 0 0 (x)

s.t.

21

Dual gives a hedging strategy

let 0 , be feasible for the dual: dsup := infm T p + 0 x , T (x) + 0 0 (x) ()

s.t.

strategy: invest i in basket (wi , Ki ), 0 in cash

22

Dual gives a hedging strategy

let 0 , be feasible for the dual: dsup := infm T p + 0 x , T (x) + 0 0 (x) ()

s.t.

strategy: invest i in basket (wi , Ki ), 0 in cash price of strategy: T p + 0 taking expectations in (), get T p + 0 E 0 (x) = psup (proves weak duality)
23

A special case
we are given option prices on the m = n individual assets, as well as forward prices min / sup s.t.
T E (w0 x K0 )+

E (xi Ki )+ = pi , i = 1, . . . , n E x = q

we may further relax the problem by ignoring the forward price information
24

Special case : upper bound

no-arbitrage: problem is feasible i 0 p q p + K upper bound is given by dsup = max
T w0 p+ i

0j n+1

w0,i min(qi pi , j Ki ) j K0 ,

where 0 = 0 j := (qj pj )/Kj 1 = n+1 upper bound is attained, hence dsup = psup

25

Ignoring forward prices

ignore the constraints E x = q obtain formula by maximizing dsup wrt q :
T T dsup = psup = w0 p + (w0 K K0 )+

26

Ignoring forward prices

ignore the constraints E x = q obtain formula by maximizing dsup wrt q :
T T dsup = psup = w0 p + (w0 K K0 )+

makes sense: concave in p (the RHS of our primal LP) convex in (w0 , K0 ) interpolates pi at w0 = i-th unit vector of Rn , and K0 = Ki
27

Sketch of proof
weak duality follows from homogeneity and convexity of x x+ :
T E (w0 x K0 )+

= =

T T E (w0 (x K ) + (w0 K K0 ))+ T T K K0 )+ E (x K )+ + (w0 w0 T T w0 p + (w0 K K0 )+

(w0 0 )

28

Sketch of proof
weak duality follows from homogeneity and convexity of x x+ :
T E (w0 x K0 )+

= =

T T E (w0 (x K ) + (w0 K K0 ))+ T T K K0 )+ E (x K )+ + (w0 w0 T T w0 p + (w0 K K0 )+

(w0 0 )

T strong duality: choose x = p + K with pty 1 if w0 K K0 , otherwise take a limit of feasible distributions 1 p + K with probability , x= 0 with probability 1 . 29

Lower bound
dual problem reduces to a nite LP (with O(n) variables and constraints)

30

Lower bound
dual problem reduces to a nite LP (with O(n) variables and constraints) if we ignore forward prices pinf = dinf =
i : Ki wi K0

pi wi K0 Kj wj pi wi min(1, ) K0 + wj Kj K0 Ki wi

j : Kj wj <K0

max

i : Ki wi <K0

in which case, direct proof of perfect duality

31

General case: integral transform approach

the option price function C (w, K ) = E (wT x K )+ is an integral transform of measure , with kernel the payo function (wT x K )+

32

General case: integral transform approach

the option price function C (w, K ) = E (wT x K )+ (1)

is an integral transform of measure , with kernel the payo function (wT x K )+

make C the variable, and solve interpolation problem sup / inf C (w0 , K0 ) :
C C

C (wi , Ki ) = pi , i = 1, . . . , m, C of the form (1)

33

LP relaxation
conditions under which C is of form C (w, K ) = E (wT x K )+ for some measure exist, but seem hard to check semi-innite LP relaxation: supC / inf C s.t. C (w0 , K0 ) C (w, K ) convex in (K, w) C (w, K ) homogeneous of degree 1 1 C (w, K )/ K 0 and C (w, K ) nondecreasing in w C (wi , Ki ) = pi , i = 1, . . . , m

34

Finite LP formulation
optimal C of semi-innite LP is piecewise ane hence the semi-innite LP can be reduced exactly to a nite LP sup / inf subject to p0 gi,j 0, 1 gi,n+1 0, gi , (wi , Ki ) = pi , gi , (wj , Kj ) (wi , Ki ) pj pi , i, j = 0, ..., m + n + 1 j = 1, ..., n i = 0, ..., m + n + 1,

i = 0, ..., m + n + 1,

where the variables gi are subgradients of C opt (for upper bound, in special case, LP is exact)
35

Robustness
in practice we have uncertainty: basket weights may change, or not exactly known at present time price information may be noisy (e.g., bid-ask) strike prices also may vary hedging strategies may be implemented with errors

we address the upper bound problem, in the special case

36

Bid-ask spread
bid-ask spread corresponds to a box uncertainty for the vector of observed option prices the worst-case value of upper (lower) bound is attained at pworst = p (or pworst = p) ppp

we may want to introduce correlation in the model . . . ellipsoidal uncertainty on p is as easy

37

Ellipsoidal uncertainty in basket weights

worst-case upper bound under weight uncertainty is psup = max wT p + (wT K K0 )+
wE 2

where E = {w + Ru : we have psup

1} is a given ellipsoid

= max wT p + t(wT K K0 ) : w E , 0 t 1
w,t

t=0,1

max (tK + p)T w + RT (tK + p)

tK0

38

Example
Price bounds on a basket call option 0.05 Simulated price Upper bound (explicit) Lower bound (explicit LP) Upper bound (LP relax.) Lower bound (LP relax.)

0.045

0.04

0.035

0.03 Price

0.025

0.02

0.015

0.01

0.005

0.01

0.02

0.03 Strike K0

0.04

0.05

0.06

0.07

39

Robustness for the general case

i have no idea about this

40

Summary

general problem important in practice special cases yields easy-to-compute bounds developed an LP relaxation for general case relaxation exact in some special cases

41

Further research
imposing smoothness constraints multi-period problem (Bertsimas, 2003) link with model-based approaches

42

Further research
imposing smoothness constraints multi-period problem (Bertsimas, 2003) link with model-based approaches

for more info: dAspremont & El Ghaoui, Static arbitrage bounds for basket option prices, submitted to Oper. Res., 2003
43