Numerical Analysis of Jump Diffusion Models:

A Partial Differential Equation Approach

Daniel J. Duffy, Datasim

Summary
We discuss a number of numerical methods that approximate the solution of the
Partial Integro Differential Equation (PIDE) that models contingent claims with
jumps. Starting with the Merton jump model for the underlying asset we motivate
how to find the corresponding PIDE that models a derivative quantity on that
asset. We pose the PIDE in a form that becomes amenable to a solution using the
finite difference method (FDM). We discuss a number of schemes and we focus on
one-factor models and we concentrate on improving the Black Scholes for European
options. The models can be extended to multi-factor option pricing problems but
a discussion is outside the scope of this article. The main goal of this article is to
show the relationship between jump models, PIDEs and numerical analysis. We see
this synergy as a new development in computational finance and new techniques
need to be developed and elaborated upon in the coming years.

Background to Jump Diffusion Models

It has been known for some time that Brownian motion does not model well the
behaviour of stocks (Heston 1993, Bates 1991, Mandelbrot 2004). A number of
extensions have been proposed. For example, Heston has produced a closed form
solution for option models with stochastic volatility and the method has been ex-
tended to model problems where the underlying follows discontinuous stochastic
processes. In particular, we can use the Fourier inversion technique to model prob-
lems involving Levy processes (Carr 2002).

The original Black Scholes model and its extensions assume that the probabil-
ity distribution of the stock price at any given future time is lognormal. If this
assumption is not true we shall get biases in the prices produced by the model.

1
If the true distribution is different from the lognormal distribution we shall un-
derprice or overprice call and put options, depending on the distributions’ tails
(Hull 2000). A number of models have been proposed in order to resolve these
shortcomings, for example:

• Model the volatility as a stochastic process (for example, the Heston model)

• Models where the company’s equity is assumed to be an option on its asset

• Models where the stock price may experiences occasional jumps rather than
continuous changes as happened on 19 and 20 October 1987, for example
(Bates 1991). New models must thus provide some support for random jumps

We now discuss how to incorporate jump models into a PIDE formulation. We

note that very little has been done on this area to date although a number of
initial results exist.

Formulation of the Continuous Problem

There is evidence to suggest that the well-loved geometric Brownian motion model
for stock price behaviour does not always work well in practice. In particular, fi-
nancial instruments do not follow a lognormal random walk (see for example, Bates
1991, Wilmott 1998). Jumps can appear at random times and to this end a number
of alternative models have been proposed, for example the jump diffusion model
(see Merton 1976). A particular case is the Poisson process. The Poisson process
is a special case of a so-called counting process. In general, a random process X(t)
is said to be a counting process if X(t) represents the total number of events that
have occurred in the time interval (0, t).
A counting process must satisfy the following conditions:

1. X(t) ≥ 0, X(0) = 0
2. X(t) is integer-valued
3. X(s) < X(t) if s < t
4. X(t) − X(s) equals the number of events that have occurred in the
interval (s, t)

A Poisson process X(t) is a counting process with rate or intensity λ > 0 if

2
 1. X(0) = 0
2. X(t) has independent and stationary increments
3. P [X(t + dt) − X(t) = 1] = λdt + o(dt)
4. P [X(t + dt) − X(t) ≥ 2] = o(dt)
when o(dt) is a function that tends to zero faster then dt, that is
o(dt)
lim = 0 (Hsu 1997)
dt→o dt

In the current context we prefer to define a Poisson process dq as follows:


0, with probability 1 − λ d t
dq =
1, with probability λ d t
(1)
where λ = Poisson arrival intensity

Thus, there is a probability λdt of a jump in the timestep dt. Actually, the Poisson
process models many kinds of arrival patterns and its applications are numerous,
for example queuing systems, inventory control applications and telecommunica-
tions. It also models the behaviour of underlyings in real options modelling:

• Energy prices (Pilipović 1998)

• Oil prices

• Business models (Mun 2002)

These quantities can exhibit peaks and spikes in their respective prices; for exam-
ple, in one case the unit price of natural gas in the Netherlands jumped from 30
Euro to more than 1500 Euro in one day during a period of short supply. Fortu-
nately, the price dropped again shortly afterwards. With shares however, the price
plummets as was witnessed in October 1987. The modified stochastic differential
equation for an underlying with jumps is:

3
 dS
S
= µd t + σ dz + (η − 1) d q
where
S = Underlying stock price
µ = Drift rate
σ = Volatility (2)
dz = Increment of Gauss-Wiener process
dq = Poisson process
η − 1 = Impulse function producing a jump from S to Sη
K = E(η − 1), expected relative jump size

In other words, the arrival of a jump is random and this is part of the stochastic
differential equation for S. We thus have two sources of uncertainty. First, the term
σdz corresponds to the usual Brownian motion while the term dq corresponds to
exceptional (and infrequent) events. Two special cases of equation (2) are geomet-
ric Brownian motion and the case of pure jump diffusion

dS
=η−1
S

In this case the path followed by S is continuous most of the time while finite
negative or positive jumps will appear at discrete points in time. We assume that
Brownian motion and jumps are independent. Based on the SDE (2) the resulting
PIDE for a contingent claim V(S,t) that depends on S is given by (Briani 2004,
Merton 1976):

Z ∞
∂V 1 2 2 ∂2V
∂τ
= 2
σ S ∂S 2 + (r − λK)S ∂V
∂S
− rV + (λ V (Sη)g(η)dη − λV )
0
where
τ = T − t = time to expiry
(3)
η = jump amplitude
and Z ∞
g(η) ≥ 0, g(η)dη = 1
0

We rewrite this equation in the more tractable form:

4
 ∂V 2
∂τ
= 12 σ 2 S 2 ∂∂SV2 + (τ − λK)S ∂V
∂S
− (r + λ)V
Z ∞ (4)
+λ V (Sη)g(η)dη
0

We must augment this PIDE by defining the usual payoff terminal conditions
and the boundary conditions when S is zero and infinity. One choice is the linear-
ity boundary conditions at both end points, namely the second-order derivatives
with respect to S are zero (see Duffy 2004 on the numerical approximation).

Moving to the Numerical Analysis: Some Chal-

lenges
In general, an exact or closed solution to the problem (4) is impossible in general,
due mainly to the integral term. For this reason, we must approximate it using
approximate methods. In this article we concentrate on the application of the
finite difference method. We reduce the scope even further by focusing on one-
factor models. Two-factor and three-factor PIDE models are also being studied
in the financial literature but a discussion of these is outside the scope of this
article. Looking at equation (4) we see that it consists of a ’classical’ Black Scholes
part (convection-diffusion equation) and an integral part (that corresponds to the
’memory’ in the system as it were and this is the term that models jumps). It is
intuitively obvious that the PIDE (4) will be harder to solve numerically than the
corresponding Black Scholes PDE. To be more precise, we list a number of hurdles
to be crossed:
• The integral term in the PIDE is on a semi-infinite interval. How are we
going to approximate the integral term? In some cases the integrand may
also have singularities. In general, we speak of the non-local nature of the
integral term
• The PDE part of the PIDE is also defined on an infinite interval. It too
must be truncated, but how? Furthermore, we are now confronted by two
truncated intervals and how should they be chosen so as not to destroy
accuracy?
• Problem (4) may not have a solution in the classical sense and we then may
need to resort to so-called viscosity solutions (Briani 2004, Crandall 1992).
For example, the solution of (4) may not be continuously differentiable and
we may have to find weaker solutions

5
 • Can we apply standard finite difference schemes (Euler, Crank Nicolson, fit-
ting) in combination with numerical integration techniques to produce stable
and accurate approximations? We may still experience spurious oscillation
problems that Crank Nicolson is famous for (Duffy 2004a, Duffy 1980)

• How can we avoid producing a dense system of equations when we approxi-

mate the integral term on a bounded interval?

• How to compare one method with another one? For example, is Crank Nicol-
son really better than implicit Euler even though the former method produces
spurious oscillations?

• Finally, having truncated the domain of integration, how do we define bound-

ary conditions at the boundaries of the truncated domain? (Briani 2004)
We wish to address the above issues in this article.

Numerical Analysis of PIDE for Jump Processes and PIDE

We now give an overview of a number of finite difference schemes that approximate
the solution of (4). In fact, we must approximate the solutions of the following
initial boundary value problem associated with (4):
Z ∞
∂u
= Lu + λ u(x + y, t)g(y)dy (5)
∂t −∞

where L is the convection-diffusion part of the PIDE. This PIDE is written in

a slightly different form than (4). Since we are using the ”engineer’s” time we need
to augment the PIDE by an initial condition:

u(x, 0) = ψ(x) for −∞ < x < ∞ (6)

The corresponding boundary conditions are always an issue in these kinds of prob-
lems and we use the following ones (there are other alternatives but they are not
the most pressing issues at this stage):

∂2u
∂x2
= 0 as x → ∞ and x → −∞ (7)

The situation is complicated by the fact that the unknown solution appears in

6
the differential equation and the integral term. In general, we then must construct
two meshes, namely one for the PDE and one for the integral term. These meshes
do not necessarily have to coincide but things become messy in this case because
we have to use some kind of interpolation when we construct the discrete systems
of equations. It is easier to use the same mesh for both the differential and integral
terms.

Another potential problem is that the discrete system of equations can result in a
dense matrix. In the pure PDE we get a band matrix of some kind (for example, a
tridiagonal system) but again the integral term confuses things. We shall see how
to avoid this problem.

Finally, the PDE is defined on a semi-infinite interval and we must truncate this
to a finite interval. Furthermore, the integral term in equation (5) is on an infinite
interval and this must be truncated. This issue is discussed in La Chioma 2003
and we give the main results here. In general, the procedure is to choose two fi-
nite values A and B such that the difference between the infinite and truncated
integrals is less than a given tolerance:

Z ∞ Z B
| f (x)dx − f (x)dx | < 
−∞ A

We approximate the truncated integral by some kind of Newton-Cotes integra-

tion method:

B N
B − AX
Z
f (x)dx ≈ wj f (xj )
A N j=0

In the current problem we have a specific integrand (kernel), namely a probability

density function of the form

y 2
Γδ (y) = √1 exp(− 2δ 2)
2πδ

This function goes to zero very quickly and we only look at this when:

7
 Γδ (y) ≥  ⇔
q √ q p
− −2 δ log( 2π) ≤ y ≤ −2 δ 2 log( δ 2π)
2

In the above context we then choose the limits of integration as follows:

( q √
A = + −2 δ 2 log(δ 2π)
B = −A

We now propose the modified form of equation (5)

Z B
∂u
= Lu + λ u(x + y, t)Γδ (y)dy (8)
∂t A

We have now replaced the integral on an infinite interval by one on a finite interval.
Furthermore, we have also truncated the domain corresponding to the PDE part
of the initial boundary value problem. We are now ready to look at candidate finite
difference solutions.

Comparison of the different Numerical Methods

In the last few years there has been a rapid growth in the interest in PIDE for
financial derivatives. We discuss a number of methods and we give references to
them for the benefit of the reader. We first introduce some notation in order to
promote the understandability of what is to follow:

Lh,k = fully discrete approximate to L

I h = discrete approximation to the integral term I(u)

We do not further specify the exact kinds of approximations for the integral, but
they could be:

• Trapezoidal integration rule

• The tanh integration rule (Duffy 2004)

In the time dimension, we apply exclusively one-step methods and then we can
choose between fully explicit, fully implicit and Crank Nicolson variants (as usual).
In the following discussion we suppress the dependence of the discrete solution on
the index for the S direction. This makes the schemes a bit more readable.

8
A. Implicit and Explicit Methods
A simple approach is to apply the so-called theta-method for both the PDE and
integral terms:

U n+1 −U n
k
= θ1 Lh,k (U n ) + (1 − θ1 )Lh,k (U n+1 )
(9)
h n h n+1
+θ2 I (U ) + (1 − θ2 )I (U )

This system of equations leads to a dense matrix system in general. This can
be remedied by approximating the integral term only at the known time level n.
For example, using Crank Nicolson for the PDE term and integral evaluation at
level n leads to the equation:

U n+1 −U n 1

k
= 2
Lh,k (U n ) + Lh,k (U n+1 ) + I h (un ) (10)

This is an interesting scheme because it is a tridiagonal matrix system and it

has been proved that this scheme is conditionally stable (see Cont 2003). The au-
thors prove stability and accuracy of the scheme (10) using the so-called viscosity
method. This is needed because the famous and sometimes misused Lax Equiva-
lence Principle is no longer valid due to the fact that solutions of the PIDE may be
non-smooth and higher-order derivatives may not exist. The article Cont 2003 is
well worth reading because it discusses the current problem in a thorough manner
using modern mathematical techniques.

B. Implicit-Explicit Runge-Kutta (IMEX) Methods

It is obvious that the coupling between the differential and integral terms in equa-
tion (5) complicates the discovery of suitable numerical schemes and subsequent
analysis. If would be nice if we could split the problem in some way so that we
could solve several simpler sub-problems. To this end, we introduce a method that
splits the PIDE in such a way that one part is implicit in time and the other
part is explicit in time. The method is called the Implicit-Explicit (IMEX) and
the rationale behind it is to split a scheme into its stiff and non-stiff parts (see
Hundsdorfer 2003 for a good introduction to IMEX methods). In the current case
we define the following ’components’ of the PIDE (5):

9
 Z ∞ 
∂u 2
∂t
= σ ∂∂xu2 + µ ∂u
∂x
+ bu + λ u(x + y, t)Γσ (y)dy − u(x, t)
−∞
∂u
∂t
= H(u) + G(u)
Z ∞  (11)
H(u) = µ ∂u
∂x
+λ u(x + y, t)Γσ (y)dy − u(x, t)
−∞

2
G(u) = σ ∂∂xu2 + bu

In this case we use implicit time-stepping for the diffusion term G(u) and ex-
plicit time-stepping for the convection (advection) term H(u). Then one particular
IMEX scheme for this problem becomes:

U n+1 − U n 1
= H(U n ) + θ G(U n ) + (1 − θ)G(U n+1 ), θ≥ (12)
k 2

Here we see that the Euler method is combined with an A-stable theta method
(Hunsdorfer 2003). A generalisation of this method is given in Briani 2004 where
the authors propose a multistep variant of equation (12).

C. Using Operator Splitting (Method of Fractional Steps)

Operator splitting methods are a powerful technique for splitting a problem into
simpler ones. In general, an n-dimensional problem can be split into a series of one-
dimensional problems using this method. In the current context, operator splitting
has been applied to integro-differential equations for the neutron transport problem
(see Yanenko 1971, page 99). We already know that the PIDE (5) can be split into
a differential and integral form:
∂u
= Lu + Iu (13)
∂t

Based in this remark we split the problem into two sub-problems that we write as
(in an intuitive/semi-formal form):
∂u
∂t
= Iu
(14)
∂u
∂t
= Lu

10
Based on Yanenko 1971 we propose the following splitting scheme:
n+1/2 n
(a) U k −U = I h (α U n+1/2 + βU n )
n+1 n+1/2
(b) U −U k
= Lh (α U n+1 + βU n+1/2 ) (15)
where
α ≥ 0, β ≥ 0, α + β = 1

We can choose different values of α and β to give use implicit or explicit schemes.
For example, we could take an explicit scheme for (15)(a) and the exponentially
fitted implicit-in-time scheme for (15)(b):
n+1/2 n
(a) U k −U = I h (U n )
n+1 n+1/2
(b) U −U k
= Lh,k
E (U
n+1
) (16)
where
Lh,k
E ≡ (Duffy) Exporential fitting operator

See Duffy 2004 for a detailed analysis of exponentially fitted finite difference
schemes.

D. Splitting in conjunction with Predictor-Corrector Meth-

ods
We now discuss the following problem: can we devise a scheme that has the com-
putational ease of the explicit Euler scheme while at the same time achieving high-
order accuracy? We answer this question by appealing to the predictor-corrector
methods that is used in the approximation of initial value problems (see Conte
1980, page 379) and that we have applied with success to financial engineering ap-
plications, in particular to the numerical solution of deterministic and stochastic
differential equations (Duffy 2004). Let us recall how predictor-corrector works for
the initial value problem:
du
dt
= f (t, u), 0 < t < T
u(0) = A (17)
u(t) = t (u1 (t), . . . , un (t))

11
If we apply the standard trapezoidal rule to (17) we get the scheme:
un+1 − un 1
= {f (tn , un ) + f (tn+1 , un+1 )} (18)
k 2

The only problem with this scheme is that it is nonlinear since the function f(t,
u) is in general a nonlinear function in u. Hence the system (18) cannot be solved
without resorting to some method such as Newton-Raphson, for example. In order
to resolve this problem we define predictor and corrector solutions as follows:
(0)
un+1 = un + kf (tn , un )
(1) (0) (19)
un+1 = un + k2 [f (tn , un ) + f (tn+1 , un+1 )]

This is the essence of predictor-corrector method. In general, we define the it-

erative scheme:
(j) k (j−1)
un+1 = un + [f (tn , un ) + f (tn+1 , un+1 )] j = 1, 2, . . . (20)
2

and the stopping criterion for a given tolerance TOL is given by

(j) (j−1)
kun+1 − un+1 k
(j)
≤ T OL (21)
kun+1 k

in some suitable norm (for example, the max norm).

We now apply the predictor-corrector method to generalise the scheme (16)(a).

We define:
(0)
Un+1 = Un + kI h (Un )

(1) (0)
Un+1 = Un + k2 [I h (Un ) + I h (Un+1 )]
(22)
more generally

(j) (j−1)
Un+1 = Un + k2 [I h (Un ) + I h (Un+1 )] j = 1, 2, . . .

with the same stopping criteria as in inequality (22).

12
A more detailed account of numerical methods for PIDE can be found in Duffy
2005. An interesting discussion of ADI in conjunction with the discrete Fourier
transform as a means of solving PIDEs is given in Andersen 2000.

Conclusions
We have given an overview of a number of numerical methods that approximate
the solution of the partial integro-differential equations (PIDE) that model jump
diffusion models in option pricing. Since there is no closed solution in general we
must resort to approximate methods. Some conclusions on the advantages and
disadvantages of the finite difference methods are:

• Explicit and Implicit: easy to implement as it builds on well-known schemes.

Conditionally stable. First-order accurate.

• IMEX: robust, modern schemes. Ability to handle stiff problems. Second and
higher order accuracy.

• Operator Splitting: reliable and robust. Watch out for the errors induced by
splitting. You may not always get second order accuracy.

• Predictor-Corrector: a good performer as has been proved in many applica-

tions. Applicable to no-linear kernels.

Acknowledgements
I would like to thank Peter Carr for providing a number of useful insights on the
topic of PIDE and their correct formulation, in particular the treatment of bound-
ary conditions.

The author is Daniel J. Duffy ([email protected])

References
Adams, R. A. 1975 Sobolev Spaces Academic Press New York
Andersen, L. and Andreasen, Jesper 2000 Jump-Diffusion Processes: Volatility
Smile Fitting and Numerical Methods for Option Pricing, Review of Operations
Research 4, 231-262
Bates, D. S. 1991 The Crash of ’87: Was it Expected? The Evidence from Options

13
Markets, The Journal of Finance Vol. XLVI, No. 3 July 1991
Briani, M., La Chioma, C. and Natalini, Roberto 2004 Convergence of Numerical
Schemes for Viscosity Solutions to Integro-differential Degenerate Parabolic Prob-
lems arising in Financial Theory Numer. Math. Vol. 98, No. 4, pp. 607-646
Briani, M., Natalini, Roberto and Russo G. 2004 A Implicit-Explicit Numerical
Schemes for Jump–Diffusion Processes, IAC Report 38 (4/2004)
Carr, Peter and Wu, L. 2002 Time-changed Levy processes and option pricing
working paper
La Chioma, C. 2003 Integro-Differential Problems Arising in Pricing Derivatives
in Jump-Diffusion Markets PhD Thesis Rome University
Cont, R. and Voltchkova, E. 2003 A finite difference scheme for option pricing in
jump diffusion and exponential Levy models Internal Report CMAP number 513,
September 2003
Conte, R. and de Boor, C. 1980 Elementary Numerical Analysis An Algorithmic
Approach McGraw-Hill New Auckland
Crandall, M.G., Ishi, H. and Lions, P.L. 1992 User’s Guide to Viscosity Solutions
of Second Order Partial Differential Equations Bulletin of the American Mathe-
matical Society, Volume 27, Number 1, July 1992, pages 1 - 67
Duffy, Daniel J. (1980) Uniformly Convergent Difference Schemes for Problems
with a Small Parameter in the Leading Derivative Ph.D. thesis Trinity College
Dublin
Duffy, Daniel J. 2004 Financial Instrument Pricing in C++ John Wiley and Sons
Chichester
Duffy, Daniel J. 2004a A Critique of the Crank-Nicolson Scheme, Strengths and
Weaknesses for Financial Instrument Pricing Wilmott Magazine July 2004
Duffy, Daniel J. 2005 Finite Difference Methods in Financial Engineering: A Par-
tial Differential Equation Approach John Wiley and Sons Chichester
Heston, S. L. 1993 A Closed-Form Solution for Options with Stochastic Volatility
with Applications to Bond and Currency Options The Review of Financial Studies
1993 Volume 6, number 2, pp.327-343
Hsu, H. (1997) Probability, Random Variables and Random Processes Schaum’s
Outline Series McGraw-Hill New York
Hull, J. 2000 Options, Futures and other Derivative Securities Prentice-Hall En-
glewood Cliffs NJ
Hundsdorfer, W. and Verwer, J. G. 2003 Numerical Solution of Time-Dependent
Advection-Diffusion-Reaction Equations, Springer, Berlin
Larsson, S., Thome, V., and Wahlbin, L. B. 1998 Numerical Solution of Par-
abolic Integro-Differential Equations by the Discontinuous Galerkin Method, Math.
Comp. Vol. 67, No. 221, January 1998, pages 45-71
Mandelbrot, B. and Hudson, R. L. 2004 The (Mis)behaviour of Markets Profile

14
Books London
Merton, R. 1973 Theory of rational option pricing Bell Journal of Economics and
Management Sciences, Vol. 4, pp. 141-183
Merton, R. 1976 Option Pricing when Underlying Stock Returns are Discontinu-
ous, Journal of Financial Economics, 125-144, May 1976
Mun, J. 2002 Real Options Analysis John Wiley and Sons New Jersey
Pilipović, D. 1998 Energy Risk, McGraw-Hill New York
Tavella, Domingo and Randall, Curt 2000 Pricing Financial Instruments,
The Finite Difference Method John Wiley and Sons New York
Wilmott, Paul 1998 Derivatives John Wiley and Sons Chichester UK
Yanenko, I.N. 1971 The Method of Fractional Steps Springer 1971
Zemanian, A.H. 1987 Generalized Integral Transformations Dover New York

15