Introduction

Finite difference methods

Example

Derivatives and Risk Management

Lecture 13: Finite difference methods

Søren Hesel

Department of Business and Management

Fall 2022

Søren Hesel DRM

 Introduction
Finite difference methods
Example

Outline

1 Introduction

2 Finite difference methods

3 Example

Søren Hesel DRM

 Introduction
Finite difference methods
Example

Introduction

Explicit prices Unrealistic modelling assumptions

Closed form approximations Case by case for assets and models
Simple Numerical techniques Binomial, trinomial trees (relevant for
interest rate models)
Monte Carlo simulation requires distribution assumptions
Finite difference Numerical solution of PDE

We want a numerical technique to be fast, flexible, precise,

converging, easily implementable, easily understandable, ...

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

The fundamental PDE

Single state variable xt with risk-neutral dynamics

dxt = µ(xt , t) dt + σ(xt , t) dzQ

t

Asset price given by a function f (x, t), which for all (x, t) ∈ S × [0, T)
solves the PDE (why?)

∂f ∂f 1 ∂2 f
(x, t) + µ(x, t) (x, t) + σ(x, t)2 2 (x, t) − r(x, t)f (x, t) = 0,
∂t ∂x 2 ∂x
along with the boundary condition f (x, T) = H(x) for x ∈ S.
• We focus on finding a numerical solution, i.e. values of f in a
sample of S × [0, T]
▶ 3 finite difference methods: “explicit”, “implicit”, “Crank-Nicolson”

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

The basic idea: discretization

Discretize in value space

• Approximate the state space S × [0, T] by the lattice (matrix)

{x0 , x1 , . . . , xJ } × {0, ∆t, . . . , N∆t}
▶ choice of x0 ≡ xmin and xJ ≡ xmax depends on the specific problem
▶ let fj,n denote f (xj , n∆t) and similarly for µ, σ, and r
∂f fj+1,n −fj−1,n
• Approximate ∂x (xj , n∆t) by Dx fj,n = 2∆x
• Approximate ∂2 f 2 fj+1,n −2fj,n +fj−1,n
∂x2 (xj , n∆t) by Dx fj,n = (∆x)2

Subject to boundary changes

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

The basic idea: discretization

Discretize in time

• Approximate ∂f
∂t (xj , n∆t) by ...
fj,n −fj,n−1
▶ D−
t fj,n = ... the explicit finite difference method
∆t
fj,n+1 −fj,n
▶ D+
t fj,n = ... the implicit finite difference method
∆t

Hence:
Explicit all values fj−1,n , fj,n , fj+1,n are known - find fj,n−1
Implicit Only fj,n+1 is known - need to find fj−1,n , fj,n and fj+1,n

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

The explicit finite difference method

2
∂f ∂ f
• Replace the derivatives ∂x ∂f
, ∂x2 , and ∂t in the PDE by Dx f , D2x f ,
−
and Dt f
• Rewrite resulting equation as

fj,n−1 = αj,n fj−1,n + βj,n fj,n + γj,n fj+1,n ,

where
! !
1 σ2j,n µj,n σ2j,n
αj,n = ∆t − , βj,n = 1 − ∆t rj,n + ,
2 (∆x)2 ∆x (∆x)2
!
1 σ2j,n µj,n
γj,n = ∆t + .
2 (∆x)2 ∆x

• Backward iterative procedure

• “Explicit” ... not necessary to solve system of equations
• Convergence requires small time steps ⇝ simple but
time-consuming method
Søren Hesel DRM
 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

Convergence of the explicit method

• We approximate both ∂f ∂f
∂t and ∂x
• Since we only use one fj,n−1 , the approximation of ∂f
∂t has to be
”precise”
• Rule of thumb:
∆t 1
σ 2
⩽
(∆x) 2

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

The implicit finite difference method

2
∂f ∂ f
• Replace the derivatives ∂x ∂f
, ∂x2 , and ∂t in the PDE by Dx f , D2x f ,
and D+
t f
• Rewrite resulting equation as
aj,n fj−1,n + bj,n fj,n + cj,n fj+1,n = fj,n+1 ,
where
! !
1 σ2j,n µj,n σ2j,n
aj,n = − ∆t − , bj,n = 1 + ∆t rj,n + ,
2 (∆x)2 ∆x (∆x)2
!
1 σ2j,n µj,n
cj,n = − ∆t +
2 (∆x)2 ∆x
• Backward iterative procedure
• Need to fix f0,n and fJ,n or to add equations like
b0,n f0,n + c0,n f1,n = d0,n+1 , aJ,n fJ−1,n + bJ,n fJ,n = dJ,n+1
• “Implicit” ... necessary to solve system of equations
• Convergence ensured as ∆t, ∆x → 0
Søren Hesel DRM
 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

Implicit system of equations

Set of equations, we consider the boundary equations later.

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

The Crank-Nicolson finite difference method

• Takes the average of the “explicit method equation”
fj,n+1 − fj,n 1
+ µj,n+1 Dx fj,n+1 + σ2j,n+1 D2x fj,n+1 − rj,n+1 fj,n+1 = 0
∆t 2
and the “implicit method equation”
fj,n+1 − fj,n 1
+ µj,n Dx fj,n + σ2j,n D2x fj,n − rj,n fj,n = 0,
∆t 2
which results in
Aj,n fj−1,n + Bj,n fj,n + Cj,n fj+1,n = −Aj,n+1 fj−1,n+1 + B∗j,n+1 fj,n+1 − Cj,n+1 fj+1,n+1 ,
where
! !
1 σ2j,n µj,n 1 σ2j,n
Aj,n = ∆t − , Bj,n = −1 − ∆t + rj,n ,
4 (∆x)2 ∆x 2 (∆x)2
! !
1 σ2j,n µj,n 1 σ2j,n
Cj,n = ∆t 2
+ , B∗j,n = −1 + ∆t + rj,n .
4 (∆x) ∆x 2 (∆x)2
• Still requires the solution of a system of linear equations
• Converging faster than the other two methods
Søren Hesel DRM
 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

Other issues
• What to do at the boundaries x0 and xJ ?
▶ Handled on a case-by-case basis
▶ Impose conditions on function values or the derivatives; maybe use
non-standard approximations of derivatives
• How to solve a tridiagonal system of equations? Matrix
inversion is inefficient. Better approach:
1 Gauss elimination (involves only the matrix)
2 Forward substitution
3 Backward substitution
• American-style assets? Two alternatives:
1 At each time step compute fj,n for all j as above. Then overwrite
with max(fj,n , Fj,n ).
2 Implicit and Crank-Nicolson: Build early-exercise check into the
backward substitution procedure. Is more efficient!
• Multiple state variables? Two and three can be handled with
some added complexity
Søren Hesel DRM
 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

Boundary conditions

Has to be done case by case basis

• For some, fix the value at the upper and lower boundaries
▶ fJ,n = k ⇒ aJ,n = 0, bJ,n = 1 and dJ,n+1 = k
▶ f0,n = 0 ⇒ b0,n = 1, c0,n = 0 and d0,n+1 = 0
2
∂ f
• Assuming linear model at the upper, ∂x ∂f
2 = 0 and ∂x is

approximated using one-sided backward-looking difference as in

(16.4) then
fJ,n+1 − fJ,n fJ,n − fJ−1,n
+ µJ,n − rJ,n fJ,n =0 ⇒
∆t  ∆x 
∆t ∆t
µJ,n fJ−1,n + 1 + rJ,n ∆t − µJ,n fJ,n =fJ,n+1
∆x ∆x

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

American style derivatives

Exercise region

• There are two regions in S × [0, T]

Continuation region The PDS is fulfilled
Exercise region Exercise is optimal
• Border, x̃t , t ∈ [0, T] between continuation and exercise is
unknown
• Add a boundary condition

f (x̃t ) = F(x̃t , t),

where F(x, t) is exercise payoff

Søren Hesel DRM

 Introduction Discretization of the fundamental PDE
Finite difference methods The three variants
Example Other issues

American style derivatives

Exercising

• Simplest method: Use max (fj,n , F(xj , n∆t))

• When using implicit method:

fJ,n = max {dj,n+1 /bJ,n , F(xJ , n∆t)}

for j = J − 1J − 2, . . . , 1, 0 :
fj,n = max {[dj,n+1 − cj,n fj+1,n ] /bj,n , F(xj , n∆t)}

Søren Hesel DRM

 Introduction
Finite difference methods
Example

Example: The Black-Scholes model

(Tutorial: Hull’s Question 21.18)

Consider a stock price with risk-neutral dynamics

dSt = (r − q)St dt + σSt dzQ
t ,

where S0 = 20, q = 0, r = 0.1, and σ = 0.3.

Use the explicit finite difference to price an American put option with:
4 1
K = 21; T= ; ∆S = 4; ∆t =
12 12

Hints:
• Set Smax = 20 · ∆S = 80
• Boundary conditions: American put ⇝ (21.29)–(21.30):
S = 0 ⇒ P = K; S = Smax ⇒ P ≈ 0
• Also try the implicit finite difference method!
Søren Hesel DRM