Computational Methods for Quant. Finance II
Finite difference and finite element methods
Lecture 10
Sensitivities and Greeks
I Key task in financial engineering: fast and accurate
calculation of sensitivitiesof market models with respect to model parameters.
I Necessary in modelcalibration, risk analysis and in the pricing andhedging of certain derivative contracts.
I Example: variations of option prices with respect to the spot price or with respect to time-to-maturity, the so-called
“Greeks”.
Outline
Option pricing
Sensitivity analysis
Numerical examples
I Consider the process X to model the dynamics of a single underlying, a basket or a underlying and its “background”
volatility drivers in case of stochastic volatility models.
I For notational simplicity assume that r = 0.
I As shown in many instances of this course 1, the fair price of a Europeanstyle contingent claim with payoff g and
underlying X, i.e.,
v(t, x) = E
g(XT) | Xt = x , is the solution of
∂tv+ Av = 0 in J × Rd, v(T, x) = g(x) in Rd. (1)
1provided some smoothness assumptions
I In (1), we denote by A theinfinitesimal generator of X and assume that it has the following form
(Af )(x) = 1
2tr[Q(x)D2f(x)] + b(x)>∇f (x) + c(x)f (x) +
Z
Rd
f(x + z) − f (x) − z>∇f (x) ν(dz),
(2) where Q : Rd → Rd×d, b : Rd → Rd, c : Rd→ R and ν a d-dimensional L´evy measure satisfying
R
Rdmin{1, |z|2}ν(dz) < ∞ andR
|z|>1|zi|ν(dz) < ∞, i= 1, . . . , d.
I We calculate the sensitivitiesof the solution v of (1) with respect to parametersin the infinitesimal generator A and with respect tosolution arguments x and t.
I Let Sη be the set of admissible parameter.
I Write A(η0) for a fixed parameter η0 ∈ Sη to emphasize the dependence of A onη0 and change to time-to-maturity t→ T − t.
I In (1), admit a non-trivial right hand side f (will become important later on).
I Thus, we consider, instead of (1), the equation
∂tu− A(η0)u = f in J × Rd, u(0, x) = u0 in Rd, (3) with u0= g.
I For the numerical implementation we truncate (3) to a bounded domain G ⊂ Rd and impose boundary conditions on
∂G. Typically, G =Qd
k=1(ak, bk), ak, bk∈ R, bk> ak, k = 1, . . . , d.
I Bilinear form a(η0; ·, ·) : V × V → R associated with infinitesimal generator A(η0), η0∈ Sη,
a(η0; u, v) := −hA(η0)u, viV∗,V, u, v ∈ V,
where we consider the abstract setting as given in Chapter 3 with Hilbert spaces V ⊂ H ≡ H∗⊂ V∗.
I Assume that a(η0; ·, ·) is continuous and satisfies a G˚arding inequality for all η0 ∈ Sη.
I For f ∈ L2(J; V∗) and u0 ∈ H the weak formulationof problem (3) is given by:
Find u ∈ L2(J; V) ∩ H1(J; H) such that (∂tu, v)H+ a η0; u, v
= hf, viV∗,V, ∀v ∈ V , (4) u(0, ·) = u0.
Outline
Option pricing
Sensitivity analysis
Numerical examples
For a market model X we distinguish two classes of sensitivities.
1. The sensitivity of the solution u to a variation Sη 3 ηs:= η0+ sδη, s >0,
of an input parameter η0 ∈ Sη. Typical examples are the Greeks Vega (∂σu), Rho (∂ru) and Vomma (∂σσu). Other sensitivities which are not so commonly used in the financial community are the sensitivity of the price with respect to the jump intensity or the order of the process that models the underlying.
2. The sensitivity of the solution u to a variation of arguments t, x. Typical examples are the Greeks Theta (∂tu), Delta (∂xu) and Gamma (∂xxu).
Today, we focus on the first class!
Sensitivity with respect to model parameters
I Let C ⊇ Sη be a Banach space over a domain G ⊂ Rd. (C is the space of parameters or coefficients in A)
I Denote by u(η0) the unique solution to (4). Introduce the derivative of u(η0) with respect to η0 ∈ Sη as the mapping Dη0u(η0) : C → V
e
u(δη) := Dη0u(η0)(δη) := lim
s→0+
1
s u(η0+ sδη) − u(η0)
, δη∈ C.
I Also introduce the derivative of A(η0) with respect to η0 ∈ Sη
A(δη)ϕ := De η0A(η0)(δη)ϕ := lim
s→0+
1
s A(η0+ sδη)ϕ − A(η0)ϕ
, ϕ∈ V, δη ∈ C.
I Assume that eA(δη) ∈ L( eV, eV∗) with eV a real and separable Hilbert space satisfying
V ⊆ V ⊂ H ∼e = H∗⊂ V∗⊆ eV∗.
I Further assume that there exists a real and separable Hilbert space V ⊆ eV such that eAv ∈ V∗, ∀v ∈ V.
Lemma
Let eA(δη) ∈ L( eV, eV∗), ∀δη ∈ C and u(η0) : J → V, η0 ∈ Sη be the unique solution to
∂tu(η0) − A(η0)u(η0) = 0 in J × Rd,
u(η0)(0, ·) = g(x) in Rd. (5) Then, eu(δη) solves
∂tu(δη) − A(ηe 0)eu(δη) = A(δη)u(ηe 0) in J× Rd,
u(δη)(0, ·) = 0e in Rd. (6)
I Associate to the operator eA(δη) the bilinear form ea(δη; ·, ·) : eV × eV → R given by
ea(δη; u, v) = −h eA(δη)u, viVe∗, eV.
I The variational formulationto (6) reads:
Find eu(δη) ∈ L2(J; V) ∩ H1(J; H) such that ∀v ∈ V (∂teu(δη), v)H+ a η0; eu(δη), v
= −ea δη; u(η0), v
, (7) eu(δη)(0) = 0 .
I Problem (7) admits a unique solution eu(δη) ∈ V due to the assumptions on a(η0; ·, ·), eA and u(η0) ∈ V.
Example: BS model
I One-dimensional diffusion process X with inf. generator (r = 0)
ABSf = 1
2σ2∂xxf −1 2σ2∂xf.
I Sensitivity of the price with respect to the volatility σ. The set of admissible parameters Sη is Sη = R+ with η = σ.
I Have
A(δσ)f = δσσe 0∂xxf− δσσ0∂xf ∈ L(V, V∗), with δσ ∈ R = C.
I Bilinear form ea(δσ; ·, ·) appearing in the weak formulation (7) of eu(δσ) is given by
ea(δσ; ϕ, φ) = δσσ0(∂xϕ, ∂xφ) + δσσ0(∂xϕ, φ).
I In this example, eV = V = H01(G).
Example: Tempered stable model
I One-dimensional pure jump process X with tempered stable density k(z) = |z|−1−α(c+e−β+|z|1{z>0}+ c−e−β−|z|1{z>0}) and infinitesimal generator
AJf = Z
R
(f (x + z) − f (x) − zf0(x))k(z)dz.
I Sensitivity of the price with respect to the jump intensity parameter α of X. Have Sη = (0, 2) with η = α and A(δα)f = δαe
Z
R
(f (x + z) − f (x) − zf0(x))ek(z)dz ∈ L( eV, eV∗), where the kernel ek is given by
ek(z) := − ln |z|k(z).
I In this example, eV = eHα/2+ε(G) ⊂ eHα/2(G) = V, ε > 0.
I Consider the finite element discretization of (6), (7). We obtain the matrix form
Find eum+1∈ RN such that for m = 0, . . . , M − 1, (M + θ∆tA)eum+1 = (M − (1 − θ)∆tA
e um
−∆t eA(θum+1+ (1 − θ)um), e
u0 = 0 .
I Here, eA is matrix of the bilinear form ea(δη; ·, ·) in the basis of VN, eAij = ea(δη; bj, bi) for 1 ≤ i, j ≤ N .
Furthermore, um+1, m = 0, . . . , M − 1, is the coefficient vector of the finite element solution uN(tm+1, x) ∈ VN to (4).
Denote by y ← solve(B, x) the output of a generic solver for a linear system Bx = y. Then, we have following algorithm to compute sensitivities with respect to model parameters.
Choose η0 ∈ Sη, δη ∈ C.
Calculate the matrices M, A and eA.
Let u0 be the coefficient vector of u0N in the basis of VN. Set eu0 = 0.
For j = 0, 1, . . . , M − 1,
u1 ← solve M + θ∆tA, (M − (1 − θ)∆tA)u0 Set f := eA(θu1+ (1 − θ)u0).
e
u1 ← solve M + θ∆tA, M − (1 − θ)∆tA)eu0− ∆tf) Set u0:= u1, eu0 := eu1.
Next j
Theorem
Assume u,ue∈ C1(J; V) ∩ C3(J ; V∗). Then, there holds the error bound
keuM − euMNk2+ ∆t
M −1X
m=0
keum+θ− eum+θN k2V
≤ C X
v∈{u,eu}
( (∆t)2RT
0 k¨v(τ )k2∗dτ θ∈ [0, 1]
(∆t)4RT
0 k...
v(τ )k2∗dτ θ= 12 + Ch2(s−r) X
v∈{u,eu}
Z T
0
k ˙v(τ )k2e
Hs−rdτ + Ch2(s−r) max
0≤t≤Tku(t)k2e
Hs.
Hence: the approximated sensitivities converge with the same, optimal rate as the approximated option price!
Outline
Option pricing
Sensitivity analysis
Numerical examples
One-dimensional models
I BS and variance gamma model. European put with K = 100, T = 1.0 and r = 0.01.
I We calculate the Greeks Delta, ∆ = ∂sV, and Gamma, Γ = ∂ssV. For BS, we additionally compute the Vega, V = ∂σV.
I Choose σ = 0.3 and for the variance gamma model ν = 0.04, θ= −0.2.
I As predicted in Theorem 2, all Greeks convergence with the optimal rate as the price V itself.
101 102 103 104 10−7
10−6 10−5 10−4 10−3 10−2 10−1 100 101
s = −2.0
N
Error
Price Delta Gamma Vega
101 102 103 104
10−8 10−6 10−4 10−2 100 102
s = −2.0
N
Error
Price Delta Gamma
Convergence rates of Greeks for a European put in the Black-Scholes (left) and variance gamma model (right).
Multi-variate models
I Heston stochastic volatility model.
I Calculate the sensitivity eu(δρ) with respect to correlation ρ of the Brownian motions that drive the underlying and the volatility.
I The derivative with respect to ρ of AHκ is AeHκ(δρ) = 1
2δρβ(y∂xy+ κy2∂x)
I Corresponding stiffness matrix AeHκ = −1
2βB1⊗ (Bx2 + κMx22).
I European call with K = 100 and T = 0.5. Model parameters:
α= 2.5, β = 0.5, m = 0.025 and ρ0 = −0.4.
102 103 104 105 10−3
10−2 10−1 100 101
s = −1.0
N
Error
Price Sensitivity
Convergence rate of sensitivity eu(δρ) in the Heston stochastic volatility model.