This action might not be possible to undo. Are you sure you want to continue?

# Journal of Computational and Applied Mathematics 218 (2008) 270–280

**www.elsevier.com/locate/cam
**

Numerical pricing of options using high-order compact ﬁnite

difference schemes

D.Y. Tangman

,1

, A. Gopaul, M. Bhuruth

∗

Department of Mathematics, University of Mauritius, Reduit, Mauritius

Received 29 August 2006; received in revised form 11 January 2007

Abstract

We consider high-order compact (HOC) schemes for quasilinear parabolic partial differential equations to discretise the Black–

Scholes PDE for the numerical pricing of European and American options. We show that for the heat equation with smooth initial

conditions, the HOC schemes attain clear fourth-order convergence but fail if non-smooth payoff conditions are used. To restore

the fourth-order convergence, we use a grid stretching that concentrates grid nodes at the strike price for European options. For

an American option, an efﬁcient procedure is also described to compute the option price, Greeks and the optimal exercise curve.

Comparisons with a fourth-order non-compact scheme are also done. However, fourth-order convergence is not experienced with

this strategy. To improve the convergence rate for American options, we discuss the use of a front-ﬁxing transformation with the

HOC scheme. We also show that the HOC scheme with grid stretching along the asset price dimension gives accurate numerical

solutions for European options under stochastic volatility.

© 2007 Elsevier B.V. All rights reserved.

MSC: 35A35; 35A40; 65N99

Keywords: European options; American options; High-order compact scheme; Grid stretching; Front ﬁxing

1. Introduction

The Crank–Nicolson scheme is a popular technique used in the numerical pricing of ﬁnancial contracts in a

Black–Scholes framework. Arecent work by McCartin and Labadie [8] has focused on the use of the Crandall–Douglas

scheme for pricing vanilla options. However, it is well known that the kink at the strike price in the payoff function of

various options, causes a lower order rate of convergence for high-order schemes. Recently Oosterlee et al. [9] used a

grid stretching transformation described in [15] in combination with a fourth-order spatial discretisation based on a ﬁve-

point stencil and fourth-order backward differencing formula (BDF4) for time discretisation, to obtain a fourth-order

accurate solution for European options. The non-compact scheme gives rise to a system which has a pentadiagonal

structure and the time evolution is performed over ﬁve time levels requiring option values for four initial time steps.

This brings some complications since only payoff values are available.

∗

Corresponding author.

E-mail addresses: tangmany@mx.uom.ac.mu (D.Y. Tangman), a.gopaul@uom.ac.mu (A. Gopaul), mbhuruth@uom.ac.mu (M. Bhuruth).

1

The research of D. Y. Tangman was supported by a Postgraduate Research Scholarship from the University of Mauritius.

0377-0427/$ - see front matter © 2007 Elsevier B.V. All rights reserved.

doi:10.1016/j.cam.2007.01.035

D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280 271

Our contribution in this paper is the use of high-order compact schemes [6,13] for the pricing of options under the

standard Black–Scholes model and Heston’s stochastic volatility model. The high-order compact scheme (HOC) is

deﬁned over two time levels similar to the Crank–Nicolson scheme and thus it requires only one initialisation step. An

additional advantage is that the method described here leads to tridiagonal linear systems, thus allowing the use of fast

tridiagonal solvers. Applying a similar grid stretching transformation as used in [9,15] in combination with the high-

order compact discretisation leads to an efﬁcient pricing algorithm for the European option. For the American option

pricing problem, we use a time dependent grid stretching transformation proposed in [9] but the linear complementarity

problem is solved using an efﬁcient procedure developed in [14] for locating the free-boundary value. We show that

this gives an optimal exercise curve comparable to [1]; see also [3] while the curve computed by the method described

in [9] is less accurate. However, the grid stretching strategy fails to produce a high-order convergence rate. We therefore

consider the use of a front-ﬁxing transformation for which the solution on a ﬁxed domain is smooth thus not requiring

any grid stretching transformation. We show that the convergence rate strongly depends on an accurate computation of

the free boundary. For European options under the stochastic volatility model of Heston [5], we show that numerical

solutions having high accuracy can be obtained with coarse grids.

An outline of this paper is as follows. In Section 2, we review the numerical pricing of European and American

options in a Black–Scholes setup. In Section 3 we study the use of high-order compact discretisations [6,13] for the heat

and Black–Scholes equations. In Section 4, we discuss the applications of the scheme for pricing American options

and in Section 5 we extend the HOC scheme to stochastic volatility European call option problems.

2. Options pricing in the Black–Scholes framework

We assume that the stock price {S

t

, t ∈ [0, T ]} satisﬁes the stochastic differential equation

dS

t

=(r −o)S

t

dt +oS

t

dW

t

, (1)

where r is the risk-free rate, o is the continuous dividend, o is the volatility parameter, T is the maturity date and

{W

t

}

0t T

is a standard Brownian motion. Under the assumptions of a frictionless market, the value of a European

option at time t is given by

V(S

t

, t ) =e

−(r−o)(T −t )

E

Q

[V(S

T

, T ) | I

t

], (2)

where V(S

T

, T ) is the payoff function, I

t

is the ﬁltration generated by the stock price process, and Qis the equivalent

martingale measure. From the Feynman-Kac Theorem [15], it follows that (2) is equivalent to the Black–Scholes PDE

jV

jt

=LV, (3)

with initial and boundary conditions

V(S, 0) =max(S −E, 0),

V(0, t) =0,

and

V(S, t) =Se

−ot

−Ee

−rt

as S → ∞,

for an European call option, where t =T −t and the spatial operator Lis

L=

1

2

o

2

S

2

j

2

jS

2

+(r −o)S

j

jS

−r.

In contrast, the American call problem is posed [17] as a linear complementarity problem (LCP) of the form

V

t

LV,

V(S, 0) =1(S),

V(S, t)V(S, 0),

272 D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280

and

(V

t

=LV) ∧ (V(S, t) =V(S, 0)), (4)

where 1(S) =max(S −E, 0) for a call option and 1(S) =max(E −S, 0) for a put. It is common practice to change

variables

x =log

_

S

E

_

, t =

1

2

o

2

(T −t ), k =

2r

o

2

, k

o

=

2(r −o)

o

2

,

ˆ : =

1

2

(k

o

−1),

ˆ

[ =

1

4

(k

o

−1)

2

+k,

and

u(x, t) =

1

E

(e

ˆ :x+

ˆ

[t

V(S, t )),

to transform the Black–Scholes PDE (3) into the standard heat equation [10]

ju

jt

=

j

2

u

jx

2

. (5)

We next describe high-order ﬁnite difference schemes for Eqs. (3) and (5).

3. High-order compact discretisations

For a ﬁnite difference discretisation of the spatial operator L, we need to truncate the inﬁnite S-domain (0, ∞) to

O

S

=(0,

ˆ

S

max

). This gives the computational domain

O

S

=

_

S

i

∈ R

+

: S

i

=

ˆ

S

min

+iS, i =0, 1, . . . , m, S =

ˆ

S

max

−

ˆ

S

min

m

_

and

O

t

=

_

t

j

∈ R

+

: t

j

=jt, j =0, 1, . . . , n, t =

T

n

_

.

Let V

j

i

be an approximation to the solution V(S

i

, t

j

) of an option. Also let V

j

=[V

j

1

, . . . , V

j

m−1

]

T

denote the vector

of unknowns at the interior grid points of O

S

. Then the HOC scheme to discretise both the heat equation (5) and

Black–Scholes PDE (3) can be described as follows for a general quasilinear parabolic PDE.

Consider the equation

j

2

u

jS

2

=f (S, u, u

t

, u

S

), (6)

for which the Crandall discretisation is given by

u

j+1/2

i−1

−2u

j+1/2

i

+u

j+1/2

i+1

S

2

=

1

12

_

f

j+1/2

i−1

+10f

j+1/2

i

+f

j+1/2

i+1

_

,

D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280 273

where

u

j+1/2

i

=

1

2

(u

j+1

i

+u

j

i

),

f

j+1/2

i+p

=f

⎛

⎝

S

i+p

, u

j+1/2

i+p

,

u

j+1

i+p

−u

j

i+p

t

, (u

S

)

j+1/2

i+p

⎞

⎠

, p =−1, 0, 1,

(u

S

)

j+1/2

i+1

=

1

2S

_

u

j+1/2

i−1

−4u

j+1/2

i

+3u

j+1/2

i+1

_

,

(u

S

)

j+1/2

i−1

=

1

2S

_

−3u

j+1/2

i−1

+4u

j+1/2

i

−u

j+1/2

i+1

_

,

(u

S

)

j+1/2

i

=

−1

2S

_

u

j+1/2

i−1

−u

j+1/2

i+1

_

+

S

20

_

f

j+1/2

i−1

−f

j+1/2

i+1

_

,

will be the fourth accurate for a ﬁxed ratio z = t/(S)

2

. The tedious process of computing the tridiagonal matrix

coefﬁcients can be simpliﬁed using a symbolic software package such as Mathematica.

Another way of obtaining the HOC scheme is to use the method developed in [13,11] where (6) is approximated

using central difference discretisation as

o

2

S

u

i

−

h

2

12

j

4

u

jS

4

+O(h

4

) =f (S

i

, u

i

, (u

t

)

i

, o

S

u

i

) −

c(S)h

2

6

j

3

u

jS

3

+O(h

4

), (7)

with c(S) being the coefﬁcient of u

S

in f. The leading truncation error is only second-order in Eq. (7) and to obtain

HOC schemes, compact approximations are derived for j

3

u/jS

3

and j

4

u/jS

4

by repeatedly differentiating (6) with

respect to S giving

j

3

u

jS

3

=f

1

(S, u, u

S

, u

St

, u

SS

)

and

j

4

u

jS

4

=f

2

(S, u, u

S

, u

SS

, u

SSt

, u

SSS

). (8)

Using (8) to replace the high-order derivative terms in (7) gives a fourth-order approximation and because the highest

derivative is now of order two, central differencing can be used to obtain compact stencils. For the Black–Scholes PDE,

a semidiscretisation of the spatial terms results in the matrix form

(I +A)

ju

jt

=Bu, (9)

where A represents the terms involving u

St

and u

SSt

while B comprises of diffusion, convection and source terms.

Spotz and Carey [13] used a Crank–Nicolson scheme for the time discretisation giving

_

I +A −

tB

2

_

u

k+1

=

_

I +A +

tB

2

_

u

k

. (10)

To obtain the high-order non-compact (HONC) scheme used in [9], we need to consider the discretisation of the

spatial and time derivatives separately. For approximating a kth-order derivative using p grid points, the one line

Mathematica code [2]

CoefficientList[Normal[Series[S ∧ q ∗ Log[S] ∧ k, {S, 1, p −1}]/h ∧ k], S]

gives the ﬁnite difference coefﬁcients where q is the number of intervals before the point of approximation. Thus to

obtain a central fourth-order approximation for the second-order spatial derivative, the triplet (k, p, q) is chosen to be

(2, 5, 2). Since we normally have only one boundary condition at each end of the computational domain, one-sided

274 D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280

approximations for the ﬁrst and last nodes in the stencil need to be used. Also, these approximations need to be fourth-

order accurate so as not to deteriorate the accuracy of the high-order scheme. At S

m−1

, we use (1, 5, 3) and (2, 6, 4) to

obtain approximations

jV

m−1

jS

=

−V

m−4

+6V

m−3

−18V

m−2

+10V

m−1

+3V

m

12S

+O(S

4

)

and

j

2

V

m−1

jS

2

=

V

m−5

−6V

m−4

+14V

m−3

−4V

m−2

−15V

m−1

+10V

m

12S

2

+O(S

4

),

for the ﬁrst and second-order spatial derivatives, respectively. Similarly for the one-sided approximation at S

1

, we use

(1, 5, 1) and (2, 6, 1). After the spatial discretisation has been performed, a matrix A is obtained and only an ODE

system

jV

jt

=AV,

remains to be discretised. Similarly, backward differencing formulas (BDF) can be generated by using one-sided

approximations for the time derivative. BDF-4 is obtained with (1, 5, 4) as

3V

j−3

−16V

j−2

+36V

j−1

−48V

j

+25V

j+1

12t

=AV

j+1

(11)

and (1, 4, 3) gives the BDF-3 formula. Since (11) requires four initialisation steps, Oosterlee et al. [9] used two

Crank–Nicolson steps and one BDF-3 step. For comparison purposes, we will use the same approach.

For the ﬁrst numerical example, we ﬁnd the solution of the heat equation (5) with the smooth initial condition

u(x, 0) =sin(¬x).

The analytical solution is then given by

u(x, t) =e

−¬

2

t

sin(¬x).

Applying the HOC discretisation of both Jain (HOCJ) and Spotz (HOCS) to the heat equation (5) results in the linear

system

Au

j+1

=Bu

j

,

where j represents the time index and the matrices A and B are deﬁned as

A =tridiag

_

1

12

−

z

2

,

10

12

+z,

1

12

−

z

2

_

, B =tridiag

_

1

12

+

z

2

,

10

12

−z,

1

12

+

z

2

_

.

For the HONCscheme, we have a pentadiagonal linear systembased on a ﬁve-time level scheme. The results of Table 1

show that the HOC schemes clearly achieve a fourth-order convergence rate while the popular Crank–Nicolson scheme

is only second-order accurate. The HONC scheme seems to be only third-order accurate. This is because we have used

Table 1

Inﬁnity norm error and convergence rates of Crank–Nicolson (CN), HOCJ, HOCS and HONC schemes for the heat equation with smooth initial

conditions

Grid nodes CN error Order HOCJ/HOCS error Order HONC error Order

20 5.51e −4 – 1.32e −5 – 2.56e −5 –

40 1.40e −4 1.9743 8.26e −7 4.0000 3.25e −6 2.9763

80 3.52e −5 1.9936 5.16e −8 4.0000 4.09e −7 2.9911

D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280 275

Table 2

Inﬁnity norm error and convergence rates of Crank–Nicolson (CN), HOCJ, HOCS and HONC schemes for a European options with ﬁnancial

parameters T =0.25, E =1, o =0.2, r =0.05

Grid nodes CN error Order HOCJ Order HOCS Order HONC Order

20 0.0062 – 0.0039 – 0.0039 – 0.0043 –

40 0.0013 2.2365 8.47e −4 2.2121 8.47e −4 2.2165 8.73e −4 2.2988

80 3.15e −4 2.0630 2.06e −4 2.0386 2.06e −4 2.0412 2.08e −4 2.0698

lower order ﬁrst step approximations to get the initial values for BDF-4. We also tested the BDF-4 scheme with four

exact initial steps and clear fourth-order convergence was obtained.

We now consider the numerical pricing of a European option. The results of Table 2 show that HOC, HONC and

Crank–Nicolson schemes achieve only second-order convergence. To restore the asymptotic fourth-order convergence

of the HOC and HONC scheme, we use a grid stretching transformation [15,9] that concentrates grid nodes at the strike

price. This procedure has the effect of decreasing the error due to the non-smoothness and therefore yields a better

convergence rate. Let y ∈ [0, 1] be the transformed coordinate, then

S =ç(y) =

1

¸

sinh(c

2

y +c

1

(1 −y)) +k, (12)

deﬁnes a stretched coordinate at k which is chosen to be E and where c

1

=sinh

−1

(¸(

ˆ

S

min

−k)), c

2

=sinh

−1

(¸(

ˆ

S

max

−k))

and ¸ is a stretching parameter. The Jacobian J(y) and Hessian H(y) of the transformation are derived by differentiating

(12) with respect to y giving

J(y) =

jç(y)

jy

=

(c

2

−c

1

)

¸

cosh(c

2

y +c

1

(1 −y))

and

H(y) =

j

2

ç(y)

jy

2

=

(c

2

−c

1

)

2

¸

sinh(c

2

y +c

1

(1 −y)).

Using the chain rule, we ﬁnd that

jV

jS

=

1

J(y)

ju

jy

and

j

2

V

jS

2

=

1

J(y)

2

j

2

u

jy

2

−

H(y)

J(y)

3

ju

jy

, (13)

and the Black–Scholes PDE (3) is transformed to

ju

jt

−

1

2

o

2

ç(y)

2

J(y)

2

j

2

u

jy

2

−

_

(r −o)

ç(y)

J(y)

−

1

2

o

2

ç(y)

2

H(y)

J(y)

3

_

ju

jy

+ru =0,

where u(y, t) = V(S, t). From Table 3, we see that the use of the grid stretching (12) with both HOC and HONC

schemes restores the fourth-order convergence for computing the European option price. The Greeks are also computed

very accurately.

4. American options

To price American options, we make use of the time dependent grid stretching transformation [9] rather than the

ﬁxed grid transformation at E. For a call option, this means that we vary k at each time level with k(0) =max(rE/o, E)

276 D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280

Table 3

Inﬁnity norm error and convergence rates of HOC and HONC schemes with a grid stretching for pricing a European option with ﬁnancial parameters

T =0.25, E =15, o =0.3, r =0.05

Grid nodes V error Order A error Order I error Order

HOCJ scheme

20 0.0017 – 0.0113 – 0.0254 –

40 1.03e −4 4.0271 0.0011 3.3340 0.0015 4.0915

80 6.13e −6 4.0703 8.99e −5 3.6358 9.29e −5 4.0048

160 3.77e −7 4.0242 7.16e −6 3.6496 5.85e −6 3.9890

HOCS scheme

20 0.0353 – 0.0129 – 0.0056 –

40 4.62e −4 3.6799 0.0013 3.3207 6.08e −4 3.2075

80 2.77e −5 4.0603 1.09e −4 3.5637 5.42e −5 3.4879

160 1.71e −6 4.0181 7.44e −6 3.8693 3.77e −6 3.8430

HONC scheme

20 0.0077 – 0.0088 – 0.0022 –

40 4.64e −4 4.0422 0.0011 3.0167 3.05e −4 2.8580

80 2.74e −5 4.0851 9.34e −5 3.5373 3.50e −5 3.1241

160 1.32e −6 4.3773 6.30e −6 3.8900 2.69e −6 3.7032

and k(t) equals the critical asset price S

f

(t). Thus, we need a proper evaluation of this free boundary value S

f

(t).

Oosterlee et al. [9] proposed a predictor corrector solution by ﬁrst solving the LCP with the PSORalgorithmto compute

S

j+1

f

=

1

2

(S

i

f

+S

i

f

+1

), (14)

where the index i

f

is chosen such that

V

j+1

i

f

>S

i

f

−E and V

j+1

i

f

+1

=S

i

f

+1

−E.

However, as shown in Fig. 1, (14) gives an oscillatory optimal exercise curve. Also the iterative procedure of the PSOR

algorithm is very slow and is computationally costly. Instead, we use the modiﬁed Thomas algorithm based on row

operations to transform the tridiagonal linear system into an upper bidiagonal one as in [14]. Then, we guess a ﬁrst

approximation to the free boundary location from the condition

V(S, t) <max(S −E, 0), S

f

<S <

ˆ

S

f

, (15)

assuming that V(

ˆ

S

f

, t) = max(

ˆ

S

f

− E, 0). Inequality (15) developed by Han and Wu [4] was based on the strong

maximumprinciple of parabolic equations. This means that, starting with node S

m−1

, we set the option value at the node

above as the payoff value and compute the option value at the node from the transformed bidiagonal linear system. The

critical index i

f

is found once V

j+1

i

f

>V

0

i

f

is satisﬁed. Finally, a more accurate location is obtained from the smooth

pasting condition

jV

jS

=1. (16)

To this end we compute the delta for two points in the continuity region using (13) and extrapolate to obtain the free

boundary location where (16) is satisﬁed.

For a comparison of the free boundary curve, we use the optimal exercise boundary computed in [1]; see also [3] for

an American put option and use the American put-call symmetry

S

f

(t, r =a, o =b) =

E

2

S

f

p

(t, r =b, o =a)

,

for free boundaries to obtain the early exercise curve for a call. Here S

f

p

(t) represents the free boundary value of an

American put. From Fig. 1, we see that both the HOC and HONC schemes give quite similar valuation of the hedging

D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280 277

0 10 20 30

–0.2

0

0.2

0.4

0.6

0.8

1

S

Δ

American Δ

HNOC

HOC

0 10 20 30

–0.05

0

0.05

0.1

0.15

0.2

S

American Γ

0 0.05 0.1 0.15 0.2 0.25

15

16

17

18

19

S

f

(

τ

)

S

f

(

τ

)

HOC

Exact

HOC

0 0.05 0.1 0.15 0.2 0.25

15

16

17

18

19

HNOC

Exact

HNOC

Γ

τ τ

Fig. 1. American call delta, gamma for HOC and HONC schemes and optimal exercise boundary curves.

Table 4

Inﬁnity norm error and convergence rates of HOC and HONC schemes with grid stretching

Nodes HOCS error Order HONC error Order

20 0.0479 – 0.2862 –

40 0.0165 1.5389 0.0222 3.6857

80 0.0042 1.9855 0.0084 1.4017

160 0.0014 1.6119 0.0034 1.2873

parameters but the optimal exercise curve computed in [9] is less accurate than the one using the modiﬁed Thomas

algorithm. For the test case E = 15, o = 0.3, r = 0, o = 0.07, we used 160 spatial nodes for both methods and the

computed critical asset price are 18.5948 for HOC, 18.5525 for HONC with the exact value being 18.6016. However,

we see fromTable 4 that both the HOCand HONCscheme with the grid stretching strategy cannot achieve fourth-order

accuracy. Also, the grid stretching at the free boundary results in a coarse grid at spot prices thus giving less accurate

option prices. The benchmark for American option prices were obtained by using the accurate Binomial method of

Leisen and Reimer [7] with 15 001 steps.

To improve the convergence order we consider the front-ﬁxing transformation [16]

y =log

_

S

B(t)

_

,

278 D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280

which turns the free boundary value problem (4) into a non-linear parabolic problem posed on a ﬁxed domain y ∈

[0, y

max

] for an American put option as

j

2

V

jy

2

=

2

o

2

_

jV

jt

−

_

r −

o

2

2

+

B

(t)

B(t)

_

jV

jy

+rV

_

,

with initial and boundary conditions as

V(y, 0) =0, V(0, t) =E −B(t),

jV(0, t)

jy

=−B(t), lim

y→∞

V(y, t) =0.

We note that there is no more singularities due to the kink in the payoff function or due to the free boundary so that

no grid stretching strategy is required. But the unknown time dependent convective coefﬁcient reveals the non-linear

nature of the problem and makes it impossible to derive compact scheme based on the approach due to Spotz and Carey

[13]. On the other hand, in either the Crank–Nicolson scheme or Jain HOC formulation with

B(t) =

B

j+1

+B

j

2

, B

(t) =

B

j+1

−B

j

t

,

we obtain a non-linear tridiagonal compact system. Noting that for y <0, the option values equal the transformed

payoff function, we obtain fourth-order approximation for the Neumann boundary condition using two ﬁctitious points

as follows

V

j+1

−2

−8V

j+1

−1

+8V

j+1

1

−V

j+1

2

12y

=−B

j+1

,

so that we can replace V

j+1

0

and V

j+1

1

into the tridiagonal systembut due to the non-linearity of our scheme, we require

Newton’s method for the solution. Writing the non-linear system in the form

F(V

j+1

, B

j+1

) =A(B

j+1

)V

j+1

−f (V

j

, B

j

, B

j+1

),

where V

j+1

= (V

j+1

2

, V

j+1

3

, . . . , V

j+1

m−1

)

T

and A the tridiagonal matrix depending on B

j+1

, the solution is found by

letting Z =(V

j+1

2

, V

j+1

3

, . . . , V

j+1

m−1

, B

j+1

)

T

and using the iterative procedure

Z

k+1

=Z

k

−J

−1

(Z

k

)F(Z

k

),

where J is the Jacobian of F. Finally V

j+1

0

, V

j+1

1

and V

j+1

m

are computed from the discretisation of the boundary

conditions. The results of Table 5 show that although the HOC scheme has smaller error than the Crank–Nicolson

scheme, yet it does not achieve a higher order convergence rate. The reason is that the solution strongly depends

on an accurate computation of the optimal exercise curve. We believe that a highly accurate computation of the free

Table 5

Inﬁnity norm error and convergence rates of Crank–Nicolson (CN), and HOCJ scheme for T =0.5, E =100, o =0.2, r =0.05

Grid nodes CN error Order HOCJ error Order

Front-ﬁxing

20 0.1819 – 0.0490 –

40 0.0461 1.9800 0.0088 2.4772

80 0.0104 2.1453 0.0022 2.0031

160 0.0019 2.4515 0.0005 2.0029

Front-ﬁxing with accurate free boundary from [1]

20 0.1419 – 0.0295 –

40 0.0401 1.8234 0.0060 2.2986

80 0.0113 1.8293 0.0008 2.8575

160 0.0031 1.8612 0.0001 3.3154

D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280 279

Table 6

Inﬁnity norm error and convergence rates of Crank–Nicolson (CN), Spotz HOC scheme without (HOCS) and with grid stretching (HOCSGS) for a

European option under stochastic volatility for T =0.25, E =10, o =0.9, r =0.03, : =5, [ =0.16, S =10, y =0.2 and j =0.1

Grid nodes CN error HOCS error HOCSGS error Exact

10 ×10 0.1013 0.0811 0.0129

20 ×20 0.0246 0.0172 6.41e-4

40 ×40 0.0061 0.0042 8.63e-6

Price 0.8640 0.8660 0.8701 0.8701

boundary will yield a high convergence rate with the HOC scheme. To present numerical evidence for this, we use the

accurate free boundary computation in [1] and cubic splines interpolation to obtain an accurate approximation for the

left boundary condition. We see from Table 5 that a high-order convergence for HOC schemes is achieved. Our current

work involves designing a method that computes a highly accurate free boundary as part of the solution process.

5. Stochastic volatility

If instead of (1), the stock price process S

t

is governed by the stochastic differential equation

dS

t

=(r −o)S

t

dt +

√

y

t

S

t

dW

1

t

,

and

dy

t

=:([ −y

t

) dt +o

√

y

t

dW

2

t

,

where o is now the volatility of the process y

t

, : >0 is the mean reversion time constant and [ >0 represents the mean

reverting value of y

t

. The resulting two dimensional PDE then becomes

jV

jt

=

1

2

yS

2

j

2

V

jS

2

+joyS

j

2

V

jSjy

+

1

2

o

2

y

j

2

V

jy

2

+rS

jV

jS

+:([ −y)

jV

jy

−rV,

where j ∈ [−1, 1] is the correlation factor. For j = 0, the HOC scheme [12] of the form (10) can easily be derived

using the same procedure steps (7)–(9) by repeatedly differentiating along S and y to ﬁnd approximation for the higher

order term in the leading truncation error. If j = 0, the cross-derivative terms can be approximated as

jV

jSjy

=

V

i+1,k+1

−V

i+1,k−1

−V

i−1,k+1

−V

i−1,k−1

4Sy

,

to maintain the compactness of the discretisation. Here k represents the index of discretisation along the y spatial

dimension. From Table 6 we can see clearly the advantage of using HOC scheme with the grid stretching strategy

where greater accuracy are obtained compared to Crank–Nicolson and using HOC scheme without grid stretching.

6. Conclusion

We have shown that for the Europeans options problem, an accurate pricing algorithm can be obtained if instead

of the popular Crank–Nicolson method, higher order schemes are employed. The compactness of the HOC scheme

discussed here ensures that the discretisation has comparable complexity as the Crank–Nicolson discretisation at each

time step. Together with a grid stretching technique, the expected fourth-order convergence is restored for European

options for a ﬁxed mesh size ratio. For American options, we considered two techniques in combination with the

high-order compact discretisation. First a grid stretching strategy coupled with an efﬁcient procedure for locating the

free boundary and secondly, a front-ﬁxing transformation that poses the problem over a ﬁxed domain. Although the

front-ﬁxing technique leads to higher accuracy, a fourth-order convergence rate is not achieved. We presented numerical

evidence that a highly accurate computation of the free-boundary is required to achieve this high-rate of convergence.

The problem of designing a method for the accurate computation of the free boundary as part of the solution process is

currently being studied. We also presented numerical results when the HOC scheme is used for pricing European call

280 D.Y. Tangman et al. / Journal of Computational and Applied Mathematics 218 (2008) 270–280

options under the Heston’s stochastic volatility model. The results indicated that high accuracy can be obtained using

coarse meshes.

Acknowledgements

The authors thank the anonymous referee whose comments improved our work and the presentation of this paper.

References

[1] X. Chen, J. Chadam, R. Stamicar, The optimal exercise boundary for American put options: analytic and numerical approximations, Preprint,

2000.

[2] B. Fornberg, Calculation of weights in ﬁnite difference formulas, SIAM Rev. 40 (3) (1998) 685–691.

[3] J. Goodman, D.N. Ostrov, On the early exercise boundary of the American put option, SIAM J. Appl. Math. 62 (5) (2002) 1823–1835.

[4] H. Han, X. Wu, A fast numerical method for the Black–Scholes equation of American options, SIAM J. Numer. Anal. 41 (6) (2003)

2081–2095.

[5] S. Heston, A closed-form solution for options with stochastic volatility with applications to bond and currency options, Rev. Financial Studies

6 (1993) 327–343.

[6] M.K. Jain, R.K. Jain, R.K. Mohanty, A fourth-order difference method for the one-dimensional general quasilinear parabolic differential

equation, Numer. Methods Partial Differential Equations 6 (1990) 311–319.

[7] D. Leisen, M. Reimer, Binomial models for option valuation-examining and improving convergence, Appl. Math. Finance 3 (1996) 319–346.

[8] B.J. McCartin, S.M. Labadie, Accurate and efﬁcient pricing of vanilla stock options via the Crandall–Douglas scheme, Appl. Math. Comput.

143 (2003) 39–60.

[9] C.W. Oosterlee, C.C.W. Leentvaar, X. Huang, Accurate American option pricing by grid stretching and high-order ﬁnite differences, Working

papers, DIAM, Delft University of Technology, the Netherlands, 2005.

[10] R. Seydel, Tools for Computational Finance, Springer, Berlin, 2002.

[11] W.F. Spotz, G.F. Carey, High-order compact ﬁnite difference methods, in: V. Andrew, V. Ilin, L. Ridgway Scott (Eds.), ICOSAHOM95,

Proceedings of the Third International Conference on Spectral and High Order Methods, 1996, pp. 397–408.

[12] W.F. Spotz, G.F. Carey, A high-order compact formulation for the 3D Poisson equation, Numer. Methods Partial Differential Equations 12

(1996) 235–243.

[13] W.F. Spotz, G.F. Carey, Extension of high order compact schemes to time dependent problems, Numer. Methods Partial Differential Equations

17 (6) (2001) 657–672.

[14] D.Y. Tangman, A. Gopaul, M. Bhuruth. A fast high-order ﬁnite difference algorithm for pricing American options, J. Comput. Appl. Math.,

submitted.

[15] D. Tavella, C. Randall, Pricing Financial Instruments: The Finite Difference Method, Wiley, New York, 2000.

[16] L. Wu, Y.K. Kwok, A front-ﬁxing ﬁnite difference method for the valuation of American options, J. Financial Eng. 6 (2) (1997) 83–97.

[17] Y.L. Zhu, X. Wu, I.L. Chern, Derivatives Securities and Difference Methods, Springer, New York, 2004.

(2) where V (ST . + (r − )S jS jS 2 In contrast. is the volatility parameter. see also [3] while the curve computed by the method described in [9] is less accurate. t ∈ [0.13] for the pricing of options under the standard Black–Scholes model and Heston’s stochastic volatility model. t) = e−(r− )(T −t) Q E [V (ST . 2. 0) = (S). In Section 2. For European options under the stochastic volatility model of Heston [5].13] for the heat and Black–Scholes equations. we use a time dependent grid stretching transformation proposed in [9] but the linear complementarity problem is solved using an efﬁcient procedure developed in [14] for locating the free-boundary value. T is the maturity date and {Wt }0 t T is a standard Brownian motion. (1) where r is the risk-free rate. ) = 0. In Section 3 we study the use of high-order compact discretisations [6.Y. T ) is the payoff function.D. it follows that (2) is equivalent to the Black–Scholes PDE jV = LV . and Q is the equivalent martingale measure. Options pricing in the Black–Scholes framework We assume that the stock price {St . 0) = max(S − E.15] in combination with the highorder compact discretisation leads to an efﬁcient pricing algorithm for the European option. is the continuous dividend. we discuss the applications of the scheme for pricing American options and in Section 5 we extend the HOC scheme to stochastic volatility European call option problems. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 271 Our contribution in this paper is the use of high-order compact schemes [6. An additional advantage is that the method described here leads to tridiagonal linear systems. T ) | It ]. From the Feynman-Kac Theorem [15]. We therefore consider the use of a front-ﬁxing transformation for which the solution on a ﬁxed domain is smooth thus not requiring any grid stretching transformation. ) = Se− − Ee−r as S → ∞. . (3) for an European call option. we show that numerical solutions having high accuracy can be obtained with coarse grids. ) V (S. where = T − t and the spatial operator L is L= 1 2 2 2 S j2 j − r. the value of a European option at time t is given by V (St . Applying a similar grid stretching transformation as used in [9. We show that this gives an optimal exercise curve comparable to [1]. thus allowing the use of fast tridiagonal solvers. V (S. In Section 4. 0). The high-order compact scheme (HOC) is deﬁned over two time levels similar to the Crank–Nicolson scheme and thus it requires only one initialisation step. and V (S. Under the assumptions of a frictionless market. However. the American call problem is posed [17] as a linear complementarity problem (LCP) of the form V LV . 0). An outline of this paper is as follows. V (S. T ]} satisﬁes the stochastic differential equation dSt = (r − )St dt + St dWt . It is the ﬁltration generated by the stock price process. For the American option pricing problem. We show that the convergence rate strongly depends on an accurate computation of the free boundary. Tangman et al. j with initial and boundary conditions V (S. we review the numerical pricing of European and American options in a Black–Scholes setup. V (0. the grid stretching strategy fails to produce a high-order convergence rate.

. Smax ). 1 ˆ = (k − 1). ) = V (S. (e E to transform the Black–Scholes PDE (3) into the standard heat equation [10] ju j2 u = 2. i = 0. jS 2 for which the Crandall discretisation is given by ui−1 j +1/2 (6) − 2ui j +1/2 S2 + ui+1 j +1/2 = 1 j +1/2 j +1/2 j +1/2 f + 10fi + fi+1 . Then the HOC scheme to discretise both the heat equation (5) and Black–Scholes PDE (3) can be described as follows for a general quasilinear parabolic PDE. (3) and (5). It is common practice to change variables x = log S E .Y. u . . ) = ˆ = 1 (k − 1)2 + k. S = ˆ ˆ Smax − Smin m (5) S S and = j j ∈ R+ : j =j . ∞) to ˆ = (0. 3. 12 i−1 . . . . 2 and u(x. we need to truncate the inﬁnite S-domain (0. Vm−1 ]T denote the vector of unknowns at the interior grid points of S . . j ) of an option. Also let V j = [V1 . . n j j Let Vi be an approximation to the solution V (Si . Consider the equation j2 u = f (S. . 0) for a put. m. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 and (V = LV ) ∧ (V (S. 0)). 0) for a call option and (S) = max(E − S. 1. 1. (4) where (S) = max(S − E. uS ). j = 0. 4 1 ˆ x+ ˆ V (S. High-order compact discretisations For a ﬁnite difference discretisation of the spatial operator L. k= 2r 2 . . = T . t)). Tangman et al.272 D. k = 2(r − ) 2 . . . This gives the computational domain ˆ = Si ∈ R+ : Si = Smin + i S. j jx We next describe high-order ﬁnite difference schemes for Eqs. n. u. = 1 2 2 (T − t). .

2 ⎛ j +1/2 = f ⎝Si+p . convection and source terms. uS . p. − ui+1 u fi−1 − fi+1 = 2 S i−1 20 = will be the fourth accurate for a ﬁxed ratio = /( S)2 . Thus to obtain a central fourth-order approximation for the second-order spatial derivative. uSSS ). 6 jS 3 (7) with c(S) being the coefﬁcient of uS in f. ui .D.11] where (6) is approximated using central difference discretisation as 2 S ui (uS )i+1 j +1/2 1 j +1/2 j +1/2 − h 2 j4 u + O(h4 ) = f (Si . − 4ui + 3ui+1 u 2 S i−1 1 j +1/2 j +1/2 j +1/2 j +1/2 (uS )i−1 = . {S. the triplet (k. jS 4 (8) Using (8) to replace the high-order derivative terms in (7) gives a fourth-order approximation and because the highest derivative is now of order two. u. . j +1/2 p = −1. 5. For approximating a kth-order derivative using p grid points. (7) and to obtain HOC schemes. 12 jS 4 S ui ) − c(S)h2 j3 u + O(h4 ). / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 273 where ui j +1/2 1 j +1 j = (ui + ui ). ui+p . Since we normally have only one boundary condition at each end of the computational domain. a semidiscretisation of the spatial terms results in the matrix form (I + A) ju = Bu. q) is chosen to be (2. uSS . the one line Mathematica code [2] CoefficientList[Normal[Series[S ∧ q ∗ Log[S] ∧ k. central differencing can be used to obtain compact stencils. jt (9) where A represents the terms involving uS and uSS while B comprises of diffusion. Another way of obtaining the HOC scheme is to use the method developed in [13. we need to consider the discretisation of the spatial and time derivatives separately. uSS . uSS ) jS 3 and j4 u = f2 (S. The tedious process of computing the tridiagonal matrix coefﬁcients can be simpliﬁed using a symbolic software package such as Mathematica. S] gives the ﬁnite difference coefﬁcients where q is the number of intervals before the point of approximation. (10) To obtain the high-order non-compact (HONC) scheme used in [9]. uS . 2). 1. (u )i . p − 1}]/h ∧ k]. uS . Spotz and Carey [13] used a Crank–Nicolson scheme for the time discretisation giving I +A− B 2 uk+1 = I + A + B 2 uk . 1. one-sided . fi+p j +1/2 ui+p − ui+p j +1 j ⎞ j +1/2 . compact approximations are derived for j3 u/jS 3 and j4 u/jS 4 by repeatedly differentiating (6) with respect to S giving j3 u = f1 (S. Tangman et al. (uS )i+p ⎠ . 0.Y. u. −3ui−1 + 4ui − ui+1 2 S S −1 j +1/2 j +1/2 j +1/2 j +1/2 j +1/2 (uS )i + . For the Black–Scholes PDE. The leading truncation error is only second-order in Eq.

we will use the same approach. BDF-4 is obtained with (1. these approximations need to be fourthorder accurate so as not to deteriorate the accuracy of the high-order scheme.9911 .9763 2. + . 5. we ﬁnd the solution of the heat equation (5) with the smooth initial condition u(x. where j represents the time index and the matrices A and B are deﬁned as A = tridiag 10 1 1 − . 12 2 12 12 2 For the HONC scheme.16e − 8 Order – 4.Y. j remains to be discretised. [9] used two Crank–Nicolson steps and one BDF-3 step. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 approximations for the ﬁrst and last nodes in the stencil need to be used. 1). a matrix A is obtained and only an ODE system jV = AV . we have a pentadiagonal linear system based on a ﬁve-time level scheme. backward differencing formulas (BDF) can be generated by using one-sided approximations for the time derivative. For the ﬁrst numerical example. This is because we have used Table 1 Inﬁnity norm error and convergence rates of Crank–Nicolson (CN). 4.51e − 4 1. 3) gives the BDF-3 formula. HOCS and HONC schemes for the heat equation with smooth initial conditions Grid nodes 20 40 80 CN error 5. The analytical solution is then given by u(x. For comparison purposes. HOCJ. jS 2 12 S 2 for the ﬁrst and second-order spatial derivatives. Oosterlee et al. − . Tangman et al. 5.274 D. At Sm−1 . The results of Table 1 show that the HOC schemes clearly achieve a fourth-order convergence rate while the popular Crank–Nicolson scheme is only second-order accurate. 4) as 3V j −3 − 16V j −2 + 36V j −1 − 48V j + 25V j +1 = AV j +1 12 (11) and (1.52e − 5 Order – 1. Also.9743 1. 6. The HONC scheme seems to be only third-order accurate. we use (1. Since (11) requires four initialisation steps. − . ) = e− 2 sin( x). 1) and (2.0000 HONC error 2. Similarly for the one-sided approximation at S1 .09e − 7 Order – 2.32e − 5 8. After the spatial discretisation has been performed.0000 4.25e − 6 4.9936 HOCJ/HOCS error 1. 6. 4) to obtain approximations −Vm−4 + 6Vm−3 − 18Vm−2 + 10Vm−1 + 3Vm jVm−1 = + O( S 4 ) jS 12 S and j2 Vm−1 Vm−5 − 6Vm−4 + 14Vm−3 − 4Vm−2 − 15Vm−1 + 10Vm = + O( S 4 ). respectively. 0) = sin( x). 3) and (2.56e − 5 3. + . Applying the HOC discretisation of both Jain (HOCJ) and Spotz (HOCS) to the heat equation (5) results in the linear system Auj +1 = Buj .40e − 4 3. 5. Similarly. we use (1. 12 2 12 12 2 B = tridiag 1 10 1 + .26e − 7 5.

73e − 4 2.2121 2. ).0043 8. American options To price American options. We also tested the BDF-4 scheme with four exact initial steps and clear fourth-order convergence was obtained. r = 0. The results of Table 2 show that HOC.0412 HONC 0.25. ) = V (S. The Greeks are also computed very accurately. we make use of the time dependent grid stretching transformation [9] rather than the ﬁxed grid transformation at E.47e − 4 2.2988 2. then S = (y) = 1 sinh(c2 y + c1 (1 − y)) + k. 1] be the transformed coordinate. we see that the use of the grid stretching (12) with both HOC and HONC schemes restores the fourth-order convergence for computing the European option price. we use a grid stretching transformation [15.2365 2. (12) ˆ ˆ deﬁnes a stretched coordinate at k which is chosen to be E and where c1 =sinh−1 ( (Smin −k)). This procedure has the effect of decreasing the error due to the non-smoothness and therefore yields a better convergence rate. jy where u(y.0039 8. Let y ∈ [0.06e − 4 Order – 2. = − 2 jy 2 jS 2 J (y) J (y)3 jy and the Black–Scholes PDE (3) is transformed to ju 1 − jt 2 2 (13) (y)2 j2 u (y) 1 − (r − ) − J (y) 2 J (y)2 jy 2 2 (y)2 H (y) J (y)3 ju + ru = 0. E = 1. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 275 Table 2 Inﬁnity norm error and convergence rates of Crank–Nicolson (CN). HOCJ.47e − 4 2.9] that concentrates grid nodes at the strike price.0039 8.D.0630 HOCJ 0.0698 lower order ﬁrst step approximations to get the initial values for BDF-4. c2 =sinh−1 ( (Smax −k)) and is a stretching parameter. To restore the asymptotic fourth-order convergence of the HOC and HONC scheme.2. = 0.05 Grid nodes 20 40 80 CN error 0. We now consider the numerical pricing of a European option.15e − 4 Order – 2.0386 HOCS 0.Y. jy 2 j (y) (c2 − c1 ) = cosh(c2 y + c1 (1 − y)) jy Using the chain rule. E) . 4. From Table 3.0062 0. HOCS and HONC schemes for a European options with ﬁnancial parameters T = 0. The Jacobian J (y) and Hessian H (y) of the transformation are derived by differentiating (12) with respect to y giving J (y) = and H (y) = j2 (y) (c2 − c1 )2 = sinh(c2 y + c1 (1 − y)). we ﬁnd that 1 ju jV = jS J (y) jy and 1 j2 u H (y) ju j2 V .08e − 4 Order – 2. this means that we vary k at each time level with k(0) = max(rE/ .06e − 4 Order – 2. HONC and Crank–Nicolson schemes achieve only second-order convergence.0013 3. Tangman et al.2165 2. For a call option.

0422 4. (14) gives an oscillatory optimal exercise curve.0077 4.25.32e − 6 Order error Order error Order – 4. Sfp ( .85e − 6 0. jS (16) To this end we compute the delta for two points in the continuity region using (13) and extrapolate to obtain the free boundary location where (16) is satisﬁed.0353 4.8580 3. (15) ˆ ˆ assuming that V (Sf . = a) for free boundaries to obtain the early exercise curve for a call.0703 4.05e − 4 3.77e − 7 0.276 D. a more accurate location is obtained from the smooth f pasting condition jV = 1. = b) = E2 .0015 9. Then.30e − 6 – 3.50e − 5 2.0603 4.09e − 4 7.03e − 4 6.0056 6.0851 4. Here Sfp ( ) represents the free boundary value of an American put.99e − 5 7. (14) where the index if is chosen such that Vif j +1 > Sif − E and Vif +1 = Sif +1 − E. we need a proper evaluation of this free boundary value Sf ( ). [9] proposed a predictor corrector solution by ﬁrst solving the LCP with the PSOR algorithm to compute Sf j +1 1 = 2 (Sif + Sif +1 ).5373 3. 1.0113 0. = 0. The j +1 critical index if is found once Vif > Vi0 is satisﬁed. 1. Oosterlee et al. r = 0. Inequality (15) developed by Han and Wu [4] was based on the strong maximum principle of parabolic equations.5637 3.0129 0.8430 – 2. j +1 However. see also [3] for an American put option and use the American put-call symmetry Sf ( .77e − 6 0. r = a.08e − 4 5.8900 0. we set the option value at the node above as the payoff value and compute the option value at the node from the transformed bidiagonal linear system. From Fig. For a comparison of the free boundary curve.9890 – 3.71e − 6 0. we use the optimal exercise boundary computed in [1].64e − 4 2.0271 4. starting with node Sm−1 . Thus. we guess a ﬁrst approximation to the free boundary location from the condition V (S.44e − 6 0.Y. E = 15.42e − 5 3.3340 3.4879 3. Instead.0011 9.3207 3. This means that. Tangman et al.0254 0. we see that both the HOC and HONC schemes give quite similar valuation of the hedging .6358 3.74e − 5 1. ) = max(Sf − E.77e − 5 1.1241 3.6799 4.2075 3.16e − 6 0. Also the iterative procedure of the PSOR algorithm is very slow and is computationally costly.05 Grid nodes HOCJ scheme 20 40 80 160 HOCS scheme 20 40 80 160 HONC scheme 20 40 80 160 V error 0.3773 0.0181 – 4.0048 3. r = b.6496 – 3. 0).34e − 5 6.0915 4.0088 0.8693 – 3.0013 1.13e − 6 3.3.0011 8. we use the modiﬁed Thomas algorithm based on row operations to transform the tridiagonal linear system into an upper bidiagonal one as in [14]. ˆ S f < S < Sf .0022 3.0242 – 3.62e − 4 2. as shown in Fig. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 Table 3 Inﬁnity norm error and convergence rates of HOC and HONC schemes with a grid stretching for pricing a European option with ﬁnancial parameters T = 0.29e − 5 5.0167 3. Finally.0017 1. 0).69e − 6 – 4. ) < max(S − E.7032 and k( ) equals the critical asset price Sf ( ).

Tangman et al.4 0.0014 Order – 1. = 0. The benchmark for American option prices were obtained by using the accurate Binomial method of Leisen and Reimer [7] with 15 001 steps. the grid stretching at the free boundary results in a coarse grid at spot prices thus giving less accurate option prices. Table 4 Inﬁnity norm error and convergence rates of HOC and HONC schemes with grid stretching Nodes 20 40 80 160 HOCS error 0. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 277 American Δ 0.2 0 10 S 19 18 S f (τ) S f (τ) 17 16 15 HOC 19 18 17 16 15 20 HNOC HOC American Γ 0.1 τ 0.D.5389 1. r = 0.25 0 0.9855 1.2 0 –0.15 0.05 0.0084 0.4017 1.2873 parameters but the optimal exercise curve computed in [9] is less accurate than the one using the modiﬁed Thomas algorithm.2 0.0165 0.6857 1. 1. To improve the convergence order we consider the front-ﬁxing transformation [16] y = log S .2 1 0.6119 HONC error 0.0042 0.6 Exact HNOC 0 0. American call delta.15 0. For the test case E = 15.07. = 0.2 0.1 Γ 0.0034 Order – 3.6016.0479 0. However.1 τ 0. we used 160 spatial nodes for both methods and the computed critical asset price are 18.3. we see from Table 4 that both the HOC and HONC scheme with the grid stretching strategy cannot achieve fourth-order accuracy.5948 for HOC.05 0.05 0 –0.15 0.05 0 10 S HNOC 20 30 30 Exact HOC 0. Also. B( ) .Y.8 Δ 0.25 Fig.2862 0.0222 0.5525 for HONC with the exact value being 18. 18. gamma for HOC and HONC schemes and optimal exercise boundary curves.

0019 – 1. the solution is found by j +1 j +1 Z k+1 = Z k − J −1 (Z k )F (Z k ).1419 0. But the unknown time dependent convective coefﬁcient reveals the non-linear nature of the problem and makes it impossible to derive compact scheme based on the approach due to Spotz and Carey [13].0401 0.8575 3. .2.0008 0.2986 2.8293 1. V (0.0104 0. The reason is that the solution strongly depends on an accurate computation of the optimal exercise curve.0088 0. we obtain fourth-order approximation for the Neumann boundary condition using two ﬁctitious points as follows V−2 − 8V−1 + 8V1 12 y j +1 j +1 j +1 j +1 − V2 j +1 = −B j +1 .4515 0. the option values equal the transformed payoff function. B j . V3 j +1 . 2 B ( )= B j +1 − B j . Grid nodes Front-ﬁxing 20 40 80 160 0.8612 0. . On the other hand.0461 0. . We note that there is no more singularities due to the kink in the payoff function or due to the free boundary so that no grid stretching strategy is required. Vm−1 )T and A the tridiagonal matrix depending on B j +1 . B j +1 ) = A(B j +1 )V j +1 − f (V j . jy with initial and boundary conditions as V (y. V1 conditions. .8234 1. yet it does not achieve a higher order convergence rate. Vm−1 . Finally V0 .0031 2.0001 – 2. ) = E − B( ).0113 0. ) = −B( ). . E = 100.278 D.05 Order HOCJ error Front-ﬁxing with accurate free boundary from [1] 20 40 80 160 0. We believe that a highly accurate computation of the free Table 5 Inﬁnity norm error and convergence rates of Crank–Nicolson (CN).5. so that we can replace V0 and V1 into the tridiagonal system but due to the non-linearity of our scheme.0490 0.0029 CN error Order j +1 j +1 j +1 = 0. Tangman et al. and Vm are computed from the discretisation of the boundary where J is the Jacobian of F .0060 0. and HOCJ scheme for T = 0. V3 j +1 . ) = 0.0022 0.9800 2. . we obtain a non-linear tridiagonal compact system.Y.0295 0. . jy jV (0.0005 – 2. 0) = 0.4772 2. Noting that for y < 0.1453 2. .1819 0. r = 0. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 which turns the free boundary value problem (4) into a non-linear parabolic problem posed on a ﬁxed domain y ∈ [0.3154 . y→∞ lim V (y. in either the Crank–Nicolson scheme or Jain HOC formulation with B( ) = B j +1 + B j . we require Newton’s method for the solution. where V j +1 = (V2 letting Z = (V2 j +1 j +1 j +1 . B j +1 )T and using the iterative procedure . B j +1 ).0031 – 1. Writing the non-linear system in the form F (V j +1 . The results of Table 5 show that although the HOC scheme has smaller error than the Crank–Nicolson scheme. ymax ] for an American put option as 2 j2 V = 2 2 jy 2 jV B( ) − r− + j 2 B( ) jV + rV .

8660 HOCSGS error 0. an accurate pricing algorithm can be obtained if instead of the popular Crank–Nicolson method.8701 boundary will yield a high convergence rate with the HOC scheme. Together with a grid stretching technique.k+1 − Vi+1.0042 0. > 0 is the mean reversion time constant and > 0 represents the mean reverting value of yt . S = 10. E = 10.0246 0. a fourth-order convergence rate is not achieved.2 and = 0. a front-ﬁxing transformation that poses the problem over a ﬁxed domain.0172 0. jSjy 4 S y to maintain the compactness of the discretisation.03. the expected fourth-order convergence is restored for European options for a ﬁxed mesh size ratio. To present numerical evidence for this.41e-4 8. 6. For = 0. Stochastic volatility If instead of (1). 5. and dyt = ( − yt ) dt + √ yt dWt2 .8701 Exact 0.1 Grid nodes 10 × 10 20 × 20 40 × 40 Price CN error 0. y = 0. The compactness of the HOC scheme discussed here ensures that the discretisation has comparable complexity as the Crank–Nicolson discretisation at each time step.0129 6.0061 0. The resulting two dimensional PDE then becomes jV 1 j2 V = yS 2 2 + j 2 jS yS 1 j2 V + jSjy 2 2 y jV j2 V jV + ( − y) − rV . From Table 6 we can see clearly the advantage of using HOC scheme with the grid stretching strategy where greater accuracy are obtained compared to Crank–Nicolson and using HOC scheme without grid stretching.k+1 − Vi−1.8640 HOCS error 0. Conclusion We have shown that for the Europeans options problem. Although the front-ﬁxing technique leads to higher accuracy. = 0. the cross-derivative terms can be approximated as jV Vi+1. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 279 Table 6 Inﬁnity norm error and convergence rates of Crank–Nicolson (CN). we use the accurate free boundary computation in [1] and cubic splines interpolation to obtain an accurate approximation for the left boundary condition.9. where is now the volatility of the process yt . We see from Table 5 that a high-order convergence for HOC schemes is achieved. r = 0. we considered two techniques in combination with the high-order compact discretisation. = 0.1013 0. = 5. higher order schemes are employed.16.25. The problem of designing a method for the accurate computation of the free boundary as part of the solution process is currently being studied. Tangman et al.k−1 = . We also presented numerical results when the HOC scheme is used for pricing European call . Spotz HOC scheme without (HOCS) and with grid stretching (HOCSGS) for a European option under stochastic volatility for T = 0.Y.k−1 − Vi−1.D.63e-6 0. We presented numerical evidence that a highly accurate computation of the free-boundary is required to achieve this high-rate of convergence. Here k represents the index of discretisation along the y spatial dimension.0811 0. + rS jS jy jy 2 where ∈ [−1. If = 0. For American options. Our current work involves designing a method that computes a highly accurate free boundary as part of the solution process. the stock price process St is governed by the stochastic differential equation √ dSt = (r − )St dt + yt St dWt1 . 1] is the correlation factor. First a grid stretching strategy coupled with an efﬁcient procedure for locating the free boundary and secondly. the HOC scheme [12] of the form (10) can easily be derived using the same procedure steps (7)–(9) by repeatedly differentiating along S and y to ﬁnd approximation for the higher order term in the leading truncation error.

L. Ridgway Scott (Eds. Huang. Comput. Heston. [15] D. R. Spotz. J. Numer. A closed-form solution for options with stochastic volatility with applications to bond and currency options. Leentvaar. Proceedings of the Third International Conference on Spectral and High Order Methods. C. Reimer. 2000.K. pp. D. Math. Spotz. the Netherlands. Pricing Financial Instruments: The Finite Difference Method. A front-ﬁxing ﬁnite difference method for the valuation of American options. L. 62 (5) (2002) 1823–1835. Y. G. Methods Partial Differential Equations 17 (6) (2001) 657–672. Numer.. [17] Y. A.K. Wu. Math.Y. Extension of high order compact schemes to time dependent problems.F. DIAM. A fourth-order difference method for the one-dimensional general quasilinear parabolic differential equation. Financial Studies 6 (1993) 327–343. 143 (2003) 39–60. SIAM Rev. Math. Andrew. Finance 3 (1996) 319–346.F. [13] W. M. Acknowledgements The authors thank the anonymous referee whose comments improved our work and the presentation of this paper. submitted. [16] L. Tavella. Wu. [12] W. Tools for Computational Finance. G. New York. [4] H. Kwok. A fast high-order ﬁnite difference algorithm for pricing American options. Wiley. 40 (3) (1998) 685–691.F. V.W. C. Binomial models for option valuation-examining and improving convergence. Methods Partial Differential Equations 6 (1990) 311–319. Comput. Appl. Zhu. [3] J. Mohanty. A fast numerical method for the Black–Scholes equation of American options. Appl. Methods Partial Differential Equations 12 (1996) 235–243. Ostrov. Carey. [6] M. J.F.W. On the early exercise boundary of the American put option. Han. [10] R. [9] C. Rev. S. Carey.J.Y. R. R.K. [7] D. Appl. Math. X. High-order compact ﬁnite difference methods. [11] W. Numer. Fornberg. Labadie. I. Chern.N. [5] S. Working papers. X. McCartin. Spotz.L. X. SIAM J. New York. Preprint. [8] B. A high-order compact formulation for the 3D Poisson equation. SIAM J.M. ICOSAHOM95. M. Tangman. Appl.K.). References [1] X. 2002. 6 (2) (1997) 83–97. Chadam. Bhuruth. 41 (6) (2003) 2081–2095. The results indicated that high accuracy can be obtained using coarse meshes. Randall. The optimal exercise boundary for American put options: analytic and numerical approximations. G.F.280 D. Numer. in: V. Derivatives Securities and Difference Methods. / Journal of Computational and Applied Mathematics 218 (2008) 270 – 280 options under the Heston’s stochastic volatility model. 2005. Accurate and efﬁcient pricing of vanilla stock options via the Crandall–Douglas scheme. Wu. Goodman. Jain. Jain. Delft University of Technology. [2] B. Chen. Gopaul. Springer. Financial Eng. .C. 397–408. Oosterlee. Leisen. Seydel. Tangman et al. 1996. Anal.F. Springer. Accurate American option pricing by grid stretching and high-order ﬁnite differences. Ilin. Calculation of weights in ﬁnite difference formulas. 2004. J. Carey. 2000. Stamicar. [14] D. Berlin.