See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220243788

A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete

Path-Dependent Options

Article in Operations Research · October 2005

DOI: 10.1287/opre.1050.0219 · Source: DBLP

CITATIONS READS
88 2,103

2 authors:

Mark Broadie Yusaku Yamamoto

Columbia University The University of Electro-Communications
89 PUBLICATIONS 7,976 CITATIONS 121 PUBLICATIONS 688 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mark Broadie on 30 June 2015.

The user has requested enhancement of the downloaded file.

OPERATIONS RESEARCH informs ®
Vol. 53, No. 5, September–October 2005, pp. 764–779
issn 0030-364X
eissn 1526-5463
05
5305
0764 doi 10.1287/opre.1050.0219
© 2005 INFORMS

A Double-Exponential Fast Gauss

Transform Algorithm for Pricing Discrete
Path-Dependent Options
M. Broadie
Graduate School of Business, Columbia University, 3022 Broadway, New York, New York 10027-6902,
[email protected]
Y. Yamamoto
Department of Computational Science and Engineering, Nagoya University, Nagoya 464-8603, Japan,
[email protected]

This paper develops algorithms for the pricing of discretely sampled barrier, lookback, and hindsight options and discretely
exercisable American options. Under the Black-Scholes framework, the pricing of these options can be reduced to evaluation
of a series of convolutions of the Gaussian distribution and a known function. We compute these convolutions efficiently
using the double-exponential integration formula and the fast Gauss transform. The resulting algorithms have computational
complexity of OnN , where the number of monitoring/exercise dates is n and the number of sample points at each date
is N , and our results show the error decreases exponentially with N . We also extend the approach and provide results for
Merton’s lognormal jump-diffusion model.
Subject classifications: finance: asset pricing.
Area of review: Financial Engineering.
History: Received January 2003; revision received March 2004; accepted September 2004.

1. Introduction the extremum or the average is taken over a finite set of

There are many traded options whose payoff depends on time points called monitoring dates. Similarly, for some
the maximum, the minimum, or the average of the under- American-style options, early exercise is only allowed on
lying asset price during the whole or part of the life of a finite set of time points called exercise dates, and such
the option. An example of such path-dependent options is options are known as Bermudan options. For these dis-
the barrier option, which has a payoff equal to that of the crete options, it has long been recognized that the formulas
European option except that the option is nullified or acti- developed under the assumption of continuous monitoring/
vated if the underlying asset price reaches a barrier during exercise give only a poor approximation. If the number
the life of the option. Other examples include the lookback of monitoring dates is not too small, corrections to these
option, whose payoff depends on the difference between formulas proposed by Broadie et al. (1997, 1999) work
the asset price at maturity and the maximum or the min- well and give satisfactory results for the standard barrier
imum of the asset price; and the hindsight option, whose and lookback options. However, when the number of mon-
payoff depends on the difference between the maximum or itoring dates is small or if one needs to price a more
the minimum and a fixed strike price. American options exotic form of path-dependent options, such as the look-
are also sometimes called quasi path-dependent options back option with American features, one has to resort to
because the decision of early exercise is made based on the numerical methods.
path of the asset price. In principle, the price of these discrete options can be
Many of these options have closed-form solutions or calculated by simple forward or backward recursion over
analytical approximation formulas under the Black-Scholes time. For example, in the case of the discrete knockout
(Black and Scholes 1973) model. See Goldman et al. option, which is a barrier option that is nullified if the
(1979) and Conze and Viswanathan (1991) for look- underlying asset price reaches a barrier at one of the moni-
back options and Rubinstein and Reiner (2000) and Rich toring dates, one needs a probability density function (pdf)
(1995) for barrier options. These formulas are based on the Pi Si  at each date i such that Pi Si dSi represents the
assumption that the extremum or the average is taken over probability that the option is still alive at time ti and the
the entire continuous path of the underlying asset, or early asset price at ti is between Si and Si + dSi . Of course, all
exercise is possible at any moment during the life of the pricing computations are done with risk-neutral pdfs (see
option. However, for many traded path-dependent options, Duffie 2001 for more detail). Strictly speaking, it would
764
Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 765

be more appropriate to call Pi Si s “subprobability” den- a discontinuity in its derivative (in the case of Bermudan
sity functions (consistent with a risk-neutral distribution) options). The trapezoidal rule and other lower-order inte-
because they do not integrate to one. However, hereafter gration rules are known to be inefficient for either of the
we simply call them pdfs for brevity. The pdf Pi Si  can cases.
be calculated by multiplying Pi−1 Si−1  with the transition Another development along this direction is the tridiag-
pdf (tpdf) pSi
Si−1 , integrating over Si−1 and setting onal probability algorithm for barrier and hindsight options
to zero any probability outside the barrier. In the case of proposed by Tse et al. (2001). The method uses Gaussian
a Bermudan option, one has to compute the continuation quadrature, which can attain far more accuracy than the
value at time ti−1 . This is calculated by multiplying the trapezoidal rule with the same number of sample points,
option value at time ti with the tpdf pSi
Si−1  and inte- and can compute the option price to a specified accuracy.
grating over Si . In both cases, the key computation in one The main disadvantage of this method is that the sample
step of the recursion is the integration of a product of a points of the Gaussian quadrature are not equally spaced
known function and the tpdf. In particular, under the Black- and the FFT can no longer be used to compute the convo-
Scholes model, the tpdf becomes a Gaussian probability lution. This results in ON 2  computational work for each
density function by adopting the log asset price as a vari- time step. Gaussian quadrature is also used in the approach
able and the above integration reduces to a convolution of of Sullivan (2000).
a given function and the Gaussian density. In this paper, we propose a new algorithm for pricing
Many numerical methods for option pricing are based discrete path-dependent and quasi path-dependent options.
on this idea. For instance, binomial and trinomial methods Our algorithm is also based on the idea of computing the
(see Cox et al. 1979, Boyle 1988) correspond to dividing convolution by numerical integration, but attains faster con-
the time between the two monitoring/exercise dates into vergence and less computational work for each time step at
smaller time steps so that the width of the tpdf for one time the same time by adopting two techniques from numerical
step is small enough that the convolution can be approx- analysis, namely, the double-exponential (DE) integration
imated by a sum of two or three terms. However, such formula (Takahashi and Mori 1974, Mori 1985) and the
an approach needs many time steps even when the num- fast Gauss transform (FGT) (Greengard and Strain 1991).
ber of monitoring/exercise dates is small. Even then the The DE formula computes an integral over a finite inter-
generic binomial and trinomial algorithms require special val or a half-infinite line by converting it to an integral
modifications to price barrier and lookback options (see over the entire real axis. Its convergence is very fast and
Ahn et al. 1999, Boyle and Lau 1994, Cheuk and Vorst the error decreases faster than any negative power of N .
1996, Ritchken 1995). Finite-difference approaches to this Typically, the error decreases as e−cN , where c is some
problem are explored in Boyle and Tian (1998, 1999) and constant, which means that one needs only Od sample
Zvan et al. (2000). Numerical methods for pricing dis- points to attain d digits accuracy. However, as in the case of
crete barrier and lookback options under alternative (i.e., the Gaussian quadrature, its sample points are not equally
non Black-Scholes) processes are considered in Boyle and spaced and the FFT cannot be used. We therefore use the
Tian (1998, 1999), Davydov and Linetsky (2001), and Duan FGT, which is a fast algorithm to compute the convolu-
et al. (2003). tion of a given function and a Gaussian in ON  work.
An efficient approach in the Black-Scholes model in In addition to being asymptotically faster than the FFT, it
the case of discretely monitored/exercisable options is to has a marked advantage that the sample points need not
compute the convolution by numerical integration without be equally spaced. By combining these two techniques, we
using intermediate time steps. The convolution method pro- can construct a fast and accurate algorithm for the pricing
posed by Reiner (2000) adopts this latter approach. The of barrier, lookback, hindsight, and Bermudan options.
method uses equally spaced grid points to discretize the log In this paper, we assume that the underlying asset fol-
asset price at each date and computes the convolution by lows geometric Brownian motion (Black-Scholes model) or
the trapezoidal rule or a quadrature based on cubic fitting. geometric Brownian motion with lognormal jumps (Merton
The convolution is computed via the fast Fourier transform model). However, it is in principle possible to extend our
(FFT), so when the number of points at each date is N , algorithm to other asset price models if the analytical
the computational work for one convolution is ON log N , expression of the transition probability density function is
which is far less than the work of ON 2  that would be given. See, for example, Broadie and Yamamoto (2003)
required for direct evaluation. However, the convergence of for an idea for applying our approach to Kou’s jump-
this method is not very fast. The error due to numerical diffusion model (2002). It is also possible to extend the
integration decreases only as N −c , where c is a constant algorithm to options on more than one asset by using the
and c = 2 for the trapezoidal rule. This means that to attain multidimensional version of the DE formula and the FGT.
d digits accuracy, one needs Oed/c  sample points. This Our algorithm can also be used to price true American
is because for the integral appearing in the convolution, options by letting the time interval t between monitoring
either the integration region is a half-infinite line (in the dates approach zero. However, in this case, the fast con-
case of barrier and lookback options) or the integrand has vergence property of our algorithm is masked by the large
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
766 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

monitoring date error of Ot. Therefore, the advantage Let pSi
Si−1  denote the transition probability density of
of our algorithm is fully exploited when pricing purely dis- the asset price. Then, from the definition, Pi Si ni=1 satisfy
crete options. the following recursion formula:
In the next section, we first consider barrier options and 
pS1
S0  if S1 > H 
show how the pricing computation can be reduced to a P1 S1  = (4)
series of convolutions of a function with a Gaussian density. 0 otherwise
We also introduce the basic ideas of the double-exponential 

integration formula and the FGT and formulate our algo- 
 pSi
Si−1 Pi−1 Si−1 dSi−1

 H
rithm by combining them. Extensions of the method to Pi Si  = if Si > H  (5)
Merton’s model and to barrier options with Bermudan fea- 


0 otherwise
tures are also discussed. Section 3 deals with the applica-
tion of our method to lookback and hindsight options, along
i = 2n
with extensions. The application to Bermudan options is
investigated in §4. Numerical results that demonstrate the We can therefore start from P1 S1  defined by Equa-
effectiveness of our method are shown at the end of each tion (4), compute P2 S2      Pn Sn  by the forward recur-
section. Section 5 contains concluding remarks. sion formula (5) and, finally, compute the option price by
Equation (3).
2. Barrier Options For our purposes, it is more convenient to change vari-
ables and work with
2.1. The Pricing Problem  
xt = lnSt  − r − q − 21  2 t (6)
In this section, we set up a framework for pricing a
European barrier option with discrete monitoring dates. Let Then, xt evolves according to the stochastic differential
us consider a Black-Scholes economy with a dividend- equation
paying asset St and a money-market account Bt with a con- dxt = dWt  (7)
stant risk-free interest rate r. Then, under the risk-neutral
probability measure, St follows the stochastic differential and the transition probability density function is seen to be
equation pxi
xi−1  = pG xi − xi−1 

dSt = r − qSt dt + St dWt  (1) 1 x − x 2

≡√ exp − i 2 i−1  (8)
2%t 2 t
where q is a constant dividend yield and Wt is a standard which is a Gaussian density. The option-pricing formula (3)
Brownian motion process. Also, the value of the money- and the recursion formulas (4) and (5) become
market account is given by
Q0DOC S0 KH 

   
Bt = exprt (2)
= e−rT Pn xn  exp xn + r −q − 21  2 T −K dxn 
k
We now consider a time horizon 0 T  and n + 1 discrete (9)
time points ti = it i = 0 1     n, where t = T /n, and
and denote Sti by Si . The discrete down-and-out call option 
with maturity T , monitoring dates ti n−1 pG x1 − ln S0  if x1 > h1 
i=1 , barrier H , and P1 x1  = (10)
strike K is defined as an option whose payoff at time T is 0 otherwise
Sn − K+ if Si > H for all i 1
i
n − 1 and zero other- 

wise. Here x+ denotes maxx 0. Other types of barrier 
 pG xi − xi−1 Pi−1 xi−1  dxi−1
 i−1
options such as the down-and-in call and the up-and-out put Pi xi  = (11)
 if xi > hi 
are defined similarly, and the numerical method we develop 

in this section can be applied in a similar way, so we limit 0 otherwise
ourselves to the down-and-out call in this section. i = 2     n
To formulate the option-pricing problem, we introduce a
set of (risk-neutral) probability density functions Pi Si ni=1 respectively. Here,
such that Pi S dS represents the probability that Sj > H  
k = ln K − r − q − 21  2 T and
for 1
∀j
i and S
Si
S + dS. Then, the price Q0DOC   (12)
of the discrete down-and-out call option at time 0 can be hi = ln H − r − q − 21  2 it
written as follows: Thus, we have established that the price of the down-and-

 call price can be computed by a series of convolutions
Q0DOC S0  K H = e−rT Pn Sn Sn − K dSn  (3) of Pi x and the Gaussian density.
K
Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 767

2.2. The Double-Exponential Formula We introduce the following change of variables, the double-
Note that Pi x i = 1     n − 1 is analytical over the exponential transformation
region hi   by construction, so the integral in Equa-  
%
tion (11) is an integral of an analytical function over a half- x = c + exp sinh u  (16)
infinite line. The double-exponential formula proposed by 2
Takahashi and Mori (1974), Mori (1985), and Ooura and which leads to a new expression for the integral
Mori (1991) is particularly efficient for computing such an
integral, and it is known that when the number of sam-
  
%
 
%

%
ple points N is increased, its discretization error decreases I= f c +exp sinhu exp sinhu coshu du
− 2 2 2
faster than any negative power of N . (17)
To derive the formula heuristically, we consider an inte-
gral of a function gu over the entire real axis. We We can now apply the trapezoidal rule with step size h to
assume that gu is analytical in this region, and the this integral, obtaining the approximation
function value and all of its derivatives approach zero   
quickly when x → ±. We then truncate the infinite 
%
+ Ih = h f c + exp sinhjh
integral − gu du at a lower limit u− and an upper j=− 2
+
limit u . These limits are chosen so that gu and its  
higher-order derivatives can be regarded as zero outside % %
· exp sinhjh coshjh (18)
the limits (for more rigorous discussion, see the com- 2 2
plex functional analysis-based approach in Takahashi and
Mori 1974, Mori 1985, Sugihara 1997). We then determine It is known that if the original integral (15) converges, the
step size h, choose integers N + > 0 and N − < 0 so that integrand in Equation (17) decreases at least as fast as the
N − − 1h < u−
N − h and N + h
u+ < N + + 1h, and double-exponential function exp− exp
u
 as u → ±,
apply the trapezoidal rule with step size h to the integral and the sum (18) can be truncated at a modest value of
jh

 N +h without affecting the computed result. To see this, note first
N −h
gu du. Let the value of the integral approximated in
that the integrand in Equation (15) has to decrease at least
this way be Ih . Then, because we can neglect the truncation
faster than 1/x as x →  for the infinite integral to exist.
error, we can evaluate the error in Ih by the Euler-Maclaurin
Therefore, we assume that f x ∼ x−1−0 as x → , where
formula (Evans 1993, Krommer and Ueberhuber 1998) as
0 > 0, and put this into the integrand of Equation (17).
follows:

+ Then, the integrand becomes
Ih − gu du   −1−0  
− % % %
c + exp sinh u exp sinh u cosh u

N +
N +h 2 2 2
gjh− 21 gN − h+gN + h h− gu du   −0
j=N − N −h % %
∼ exp sinh u cosh u
2 2
m
B2k 2k−1 2k−1 +  
= h g N h−g 2k−1 N − h +Em  (13) %0 %
k=1 2k! ∼ exp − exp u exp u
4 4
where Bk are the Bernoulli numbers with B3 = B5 = · · · = 0  
% %0
and B1 = −1/2, B2 = 1/6, B4 = −1/30, g n u is the nth = exp u − exp u  (19)
derivative of gu, and 4 4

B2m+2 h2m+2 2m+2 When, for example, 0 ∼ 1, this function becomes less than
Em = N + h − N − h g . (14) 10−16 at u = 4, and the infinite sum of Equation (18) can be
2m + 2!
safely truncated at
jh
∼ 4 if double precision arithmetic
with N − h < . < N + h. Equation (13) holds for any fixed is used. Thus, the number of the sample points is N =
positive integer m. Recalling that higher-order derivatives Oh−1 , and we can conclude that the discretization error
of gu can be regarded as zero at u = N − h and u = N + h, decreases faster than any negative power of N .
we know that the error can be approximated by Em , which Takahashi and Mori (1974) show that the double-
is Oh2m+2 . Because m is arbitrary, we can say that the exponential transformation (16) is optimal for a wide class
error of the trapezoidal rule decreases faster than any power of functions in the sense that it can achieve the maximum
of h in this case. accuracy for a given number of points (Takahashi and Mori
The DE formula exploits this fact by converting an inte- 1974, Mori 1985, Sugihara 1997). However, convergence
gral over a half-infinite region into one over the entire real faster than any negative power of N is attained as long as
axis. Let the integral we want to compute be (a) the transformation is analytical, (b) it maps the origi-

 nal integration region onto the entire real axis, and (c) the
I= f x dx (15) integrand in Equation (17) approaches zero fast enough
c
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
768 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

when x → ±. Hence, it is sometimes more appropriate To compute the multiple sums efficiently, the FGT uses the
to modify the transformation by taking into account special following expansion of the Gaussian:
properties of the original integrand f x. We will do this  
for the convolution (11) in §2.4. −xj −yk 2 /4
  
1 1 yk − y0 0
e = √
It is important to note that the integrand in Equation (11) 6=0 0=0
6! 0! 4
is analytical for asset price models other than the Black-   
x −y xj − x0 6
Scholes model. For example, the integrand for Merton’s · h0+6 0√ 0 √  (25)
lognormal jump-diffusion model (Merton 1992) is also ana- 4 4
lytical because the jump size is lognormally distributed and
This formula can be shown easily by using the expansion
the transition probability density pxi
xi−1  can be rep-
of the Gaussian in terms of Hermite functions h0 x:
resented as a sum of Gaussians. Also, the tpdf in Kou’s
jump-diffusion model (2002) is analytical because it can be 2 
y0
represented as a sum of products of an exponential func- e−x−y = h0 x (26)
0=0 0!
tion and the Hh function, both of which are analytical. This  0
indicates that the DE method has potential application to 0 d 2
h0 x = −1 e−x  (27)
these models. We illustrate an example of this in §2.5. dx

and by further expanding the Hermite function in the right-

2.3. The Fast Gauss Transform hand side of Equation (26). It is known that this expansion
To compute the convolution, we first rewrite the DE for- converges very quickly and the double infinite sum over
mula by defining the sample points aj and wj as follows: 0 and 6 can be truncated at a reasonably small integer,
0 = 6 = 0max . It is known that 0max = 8 is sufficient √ to
−8
+ achieve a relative
√ error of 10 when
y k − y 0 / 4
< 1/2

N
IhN = f aj wj  (20) and
xj − x0 / 4
< 1/2 (Greengard and Strain 1991).
j=N − Now we consider a special case where all the target √
  points xj  are in an interval with center x0 and length 4
%
aij = hi + exp sinhjh  (21) and all the source points
√ yk  are in another interval with
2
  center y0 and length 4. Then, the expansion (25) con-
% % verges quickly by truncating the sums over 0 and 6 at
wj = h exp sinhjh coshjh (22)
2 2 0 = 6 = 0max . By substituting this into Equation (24), we
can approximate Gxj  as
where N − and N + are determined so that N + h ∼  
−N − h ∼ 4 and the total number of sample points is 
N   1 1 yk − y0 0
0max 0max
Gxj   qk √
N = N + − N − + 1. Using this formula, we can approximate k=1 6=0 0=0
6! 0! 4
Equation (11) by   
x0 − y0 xj − x0 6
+
· h0+6 √ √

N 4 4
Pi aij  = pG aij − ai−1 i−1
j  Pi−1 aj  wj  0max  6
j  =N −
 1 xj − x0
= √
j = N −      N +  (23) 6=0
6! 4
   
0max
 x0 − y0 1  N
yk − y0 0
Note that we do not need sample points in the region · h0+6 √ q √ 
0=0 4 0! k=1 j 4
xi < hi because pi xi  is always zero there. Apparently, (28)
Equation (23) requires ON 2  computation for each time
step. Moreover, the FFT cannot be used here to reduce the This expression shows that the computation of Gxj  can
computational work because the sample points aij  and be divided into three steps:
ai−1
j   are not equally spaced. We therefore use the FGT, Step 1. Compute
which is a fast algorithm introduced by Greengard and  
Strain (1991) and Greengard and Sun (1998) to compute 1  N
y −y 0
A0 ≡ qk k√ 0 for 0 = 0     0max 
the discrete convolution of a given function with a Gaussian 0! k=1 4
function in ON  work.
Suppose that we want to calculate the sums Step 2. Compute
0max  
  x0 − y0
N
xj −yk 2 B6 ≡ A0 h0+6 √ for 6 = 0     0max 
Gxj  = qk exp −  j = 12M (24) 4
k=1 4 0=0
Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 769

Figure 1. Illustration of the FGT algorithm. By applying the fast Gauss transform with source points
ai−1 i
j  , target points aj , and

qk = Pi−1 ai−1
k wk  (29)
2
4 = 2 t (30)

we can compute the discrete convolution (23) in ON 

work.
Note that we gave only the basic idea of the FGT, and
in the actual algorithm several techniques to further speed
up the computation are employed. For the details of these
techniques, as well as for the error analysis and extensions
to higher dimensions, consult Greengard and Strain (1991),
Greengard and Sun (1998), and Baxter and Roussos (2002).
We conclude this subsection by noting that there are
Note. The
√ source points xi and target points yj lie in intervals of
length 4 centered at x0 and y0 , respectively.
other approaches to computing the discrete convolu-
tion (23) with computational effort less than ON 2 . One
possibility is to use nonuniform FFTs, or variants of the
Step 3. Compute FFT for unequally spaced grids Dutt and Rokhlin (1993)
  and Ware (1998). One of the disadvantages of this approach
0max
 1 xj − x0 6 is that it needs ON log N  work when the number of grid
Gxj  = B6 √ for j = 1     M
6=0
6! 4 points is N , as opposed to ON  work required by the FGT.
In addition, from our preliminary experiments, nonuniform
When 0max is fixed, Steps 1 and 3 require ON  and FFTs were seen to be about 10 times slower than FFTs
OM computational effort, respectively, while Step 2 can for equally spaced grids. We conclude that it is more effi-
be done in a constant time that does not depend either on N cient to use problem-specific convolution methods such as
or M. See Figure 1 for an illustration of the FGT algorithm. the FGT when they are available. However, convolution
In the general
√ case, we divide the space into intervals based on nonuniform FFTs has a marked advantage that it
of length 4 and apply the above method to each of the can deal with a much wider class of transition probability
possible pairs of a source interval and a target interval. Let density functions. We are exploring the possibility of fast
K and J denote the source interval and the target inter- pricing methods based on this approach.
val, respectively, and yK and xJ denote their centers. The
algorithm can be written as follows: 2.4. The DE-FGT Method for Barrier Options
Step 1. Compute The method described in the previous subsection gives a
 0 straightforward approach to computing the price of discrete
1  yk − yK
A0 K ≡ q √ for 0 = 0     0max down-and-out options using the combination of the DE for-
0! yk ∈K k 4 mula and the FGT. However, we can construct a more effi-
cient method by taking into account the properties of the
and for each source interval K. integrand in Equation (11) and by slightly modifying the
Step 2. Compute transformation (16).
  As xi−1 → , the integrand in Equation (11) approaches
 0
max
xJ − yK
B6 J ≡ A0 K h0+6 √ for 6 = 0     0max zero faster than the Gaussian, because pG xi − xi−1  is
K 0=0 4 Gaussian and Pi−1 xi−1  also approaches zero. If we apply
the ordinary double-exponential transformation to this inte-
and for each target interval J . gral, we have another factor that approaches zero dou-
Step 3. Compute ble exponentially. Although this gives rise to an advantage
  that the integrand vanishes extremely fast and the integra-
0max
 1 xj − xJ 6 tion region can be truncated at a smaller upper bound, it
Gxj  = B6 J √ for j = 1     M
6=0
6! 4 also causes a disadvantage that the integrand changes very
rapidly and a finer mesh is necessary in the integration
Here J is the target interval xj belongs to. region. According to our numerical experiments, this dis-
Because each xj and yk belong to only one interval, advantage exceeds the advantage.
the total work for Step 1 and Step 3 is still ON  and Intuitively, it seems more appropriate to construct a
OM, respectively, while Step 2 needs work proportional transformation that works as the DE transformation near
2
to ONint , where Nint is the number of intervals. the lower end of the integral and approaches an identity
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
770 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

transformation as xi−1 → , because the original integrand 2.5.2. Computation of Delta and Gamma. Our
pG xi − xi−1 Pi−1 xi−1  vanishes rapidly enough. We there- method can also be used to compute the option delta
fore adopt the following transformation: and gamma. To see this, differentiate both sides of Equa-
  tions (10) and (11) with respect to the initial stock price S0 .
%
c
x = ln e + exp −u
1 + u − e   (31) Then, we know that :Pi /:S0 satisfies the same recur-
2 rence as Pi itself and can therefore be computed with our
This is obtained by replacing the eu in the sinh func- DE-FGT method. After obtaining :Pn /:S0 , we can differ-
tion with 1 + u and taking the logarithm. It has the prop- entiate Equation (3) with respect to S0 , put :Pn /:S0 in the
erty that it approaches the DE transformation x = c + right-hand side, and compute the option delta :Q0DOC /:S0 .
e−c exp−%/2 exp−u when u → − and approaches The option gamma can be computed in the same manner.
a linear transformation x = %/2u when u → . The orig- 2.5.3. Application to Merton’s Model. We can extend
inal integral (15) is converted into our method to deal with the lognormal jump-diffusion

  
%
  model introduced by Merton (1992). In this model, the
I= f ln ec + exp 1 + u − e−u  asset price follows the equation
− 2
  √
exp %/21 + u − e−u  %/21 + e−u   
·   du (32) Si = Si−1 exp r − q − 21  2 − ;< t +  tz0
ec + exp %/21 + u − e−u 
NtP t
  
Applying the trapezoidal rule with step size h to this inte- + 4zl + ? − 21 42  (36)
gral, we can find the sample points and the weights of the l=1
modified DE formula as follows:
  where t is the time interval between ti−1 and ti , NtP t
i c % −jh is the number of jumps during this interval, which follows
aj = ln e + exp 1 + jh − e  and (33)
2 a Poisson process with intensity <, and zl l = 0 1   
  are independent and follow the standard normal distribution
exp %/21 + jh − e−jh  %/21 + e−jh 
wj = h    (34) N 0 1. The constants ? and 4 determine the mean and
ec + exp %/21 + jh − e−jh 
the standard deviation of the jumps, respectively, and
By putting these into Equation (23) and computing the dis-
crete convolution using the FGT, we can construct a method ; = e? − 1 (37)
for pricing the discrete down-and-out options efficiently.
We name this the DE-FGT method. In this model, the market becomes incomplete due to the
Our numerical experiments show that numerical integra- existence of jumps, and the standard argument for option
tion based on the transformation (31) requires only half pricing based on the replicating portfolio no longer holds.
as many sample points as that based on (16) to attain the Merton (1992) derives an option-pricing formula under
same level of accuracy. We therefore use this transforma- the assumption that jump risk is diversifiable so that risk-
tion in the subsequent sections whenever an integral over a neutral pricing is appropriate. Bates (1996) and Naik and
half-infinite line is necessary. Lee (1990) derive option-pricing formulas in representative
agent general equilibrium models. The form of their pric-
2.5. Extensions of the Basic Algorithm ing equations are identical to the Merton formula, but with
altered parameter values that account for the market price
2.5.1. Time-Varying Barriers and Double Barriers. of jump risk. The pricing problems in these models are
So far we have assumed, for simplicity, that the barrier therefore equivalent from a computational viewpoint: One
level H is constant through time and the monitoring dates simply substitutes the appropriate “risk-adjusted” parame-
are equally spaced between t = 0 and t = T . However, as ters into the risk-neutral pricing formula.
can be inferred from the derivation of the algorithm, there We can apply the change of variable
is no difficulty in dealing with the case of time-varying  
barriers or nonequally spaced monitoring dates. We can xt = lnSt  − r − q − 21  2 − ;< t (38)
also extend the algorithm to a double-barrier option with a
lower barrier HL and an upper barrier HH . In this case, the to Equation (36) and obtain an equation for xi :
integral appearing in convolution (11) becomes one over a P
finite interval hL  hH , and one can use the finite-interval √ Nt t
  
xi = xi−1 +  tz0 + 4zl + ? − 21 42  (39)
version of the double-exponential transformation l=1
 
h + hL hH − hL % Note that the Poisson probability can be written as
x= H + tanh sinh u (35)
2 2 2
  <tn
instead of (31). Pr NtP t = n = e−<t  (40)
n!
Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 771

and when the number of jumps is n, xi − xi−1 follows a introduce in §4 as a basis. In this method, we use backward
Gaussian distribution with the variance and mean given by recursion and compute the continuation value at time ti−1
as a convolution of the option price Qi xi  at time ti and
n2 =  2 t + n42 and (41) the transition probability density function. Because Qi xi 
  has a discontinuity in the derivatives at the optimal exer-
@n = n ? − 21 42  (42)
cise boundary xi = xc , we divide the integration region into
respectively. We can then write the tpdf of xi as two parts, namely − xc  and xc  , and apply the DE
formula to each of them.
pxi
xi−1  = pM xi − xi−1  To price barrier options such as down-and-out options


<tn 1 with a Bermudan feature using this algorithm, we only have
≡ e−<t √ to replace the lower integral region with hi  xc , where hi
n! 2%n
n=0 is the barrier level given by Equation (12), and use the DE
xi − xi−1 − @n 2 formula (35) for a finite interval.
· exp −  (43)
2n2
2.6. Numerical Results
As can be easily seen, the probability density
We implemented the DE-FGT method for European down-
M
p xi − xi−1  and-out call options under the Black-Scholes model and
Merton’s model and studied its speed and accuracy. All of
has the following expansion, which corresponds to Equa- the experiments, including those in the later sections, were
tion (25) in the Gaussian case (Broadie and Yamamoto done on a 266 MHz Pentium II PC with Red Hat Linux
2003): using gnu C++ compiler, unless otherwise noted.
  2.6.1. Down-and-Out Call Option Under the Black-
M
  1 1 xi − x 0
0max 0max
p xi − xi−1  = √ Scholes Model. We show results for European down-and-
6=0 0=0
6! 0! 2 out call options under the Black-Scholes model in Figure 2.
 0+6 The parameters are S0 = K = 100, T = 02, r = 01, q = 0,
 
−<t <t
n
1 
· e √ and  = 03. We varied the barrier level from H = 91 to
n=0 n! 2%n  n H = 99 in increments of 2 and set the number of monitor-
     ing dates to n = 5, 25, or 50. These are the values used in

x − x + @n xi−1 − x 6
· h0+6 √ √ 
2n 2
(44) Figure 2. European down-and-out call option price
√ under the Black-Scholes model.
where x and x are the centers of intervals of length 2
Execution time (sec.)
containing xi−1 and xi , respectively. Comparing this expres- 1
sion with Equation (25), we know that we can construct 0.001 0.01 0.1 1 10 100
an algorithm similar to the FGT by replacing the Hermite 10 –1
25 150

function with a weighted sum of shifted and scaled Hermite 100

functions. Specifically, we have only to replace the formula 10 –2
to compute B6 (see §2.3) with the following: 50 200 BGK, n= 50
N  0+6 10 –3 BGK, n =25
0max
 jump
−<t <t
n
1  BGK, n =5
B6 = A0 e √
6=0 n=1 n! 2%n  n 10 – 4 DE-FGT, n =5


150 250 DE-FGT, n =25
 
Absolute error

x − x + @n 10 –5 75 DE-FGT, n =50

· h0+6 √  (45) Reiner, n = 5
2n Reiner, n = 25
200
10 –6 Reiner, n = 50
where we have truncated the sum over the number of jumps
at Njump . This algorithm enables us to compute the convo- 10 –7
300

lution of pM x and a given function almost as easily as 100

in the Gaussian case. We refer the reader to Broadie and 10 –8
350
Yamamoto (2003) for more detailed derivation of the algo- 250

rithm and implementation issues. 10 –9

400
2.5.4. Barrier Options with Bermudan Features. It
125
is also easy to extend our DE-FGT method to the pricing of 10 –10
barrier options with Bermudan features. To do this, we use 150
300 450 8,000 8,000
16,000
the DE-FGT method for Bermudan options, which we will 10 –11
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
772 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

Table 1. European down-and-out call prices under the Black-Scholes

model.
n=5 n = 25 n = 50 n=
H = 91 6.1872900302 6.0320261243 5.9770686565 5.8077713405
H = 93 5.9997553594 5.6875323983 5.5843399451 5.2768140270
H = 95 5.6711051343 5.0814151587 4.9067890354 4.3975025272
H = 97 5.1672453684 4.1158152250 3.8339777052 3.0595631676
H = 99 4.4891724312 2.8124392982 2.3363868958 1.1707931057

Broadie et al. (1997), Reiner (2000), and Tse et al. (2001). shown in Table 2. In this case, the computation was done
As reference prices against which to compute the errors, we on a 2.0 GHz Xeon PC.
used the values shown in Table 1, which were obtained by It can be seen from the graph that the convergence behav-
Reiner’s convolution method (2000) with the cubic fitting ior of both Reiner’s method and our DE-FGT method is the
quadrature and N = 32000 grid points for each date. These same as that for options with a smaller number of monitor-
values agree with the results of Tse et al. (2001) to at least ing dates, and the DE-FGT method is superior when higher
10 digits after the decimal point for n = 5 and n = 25. For accuracy is needed.
comparison, we also show the prices of the continuously 2.6.2. Up-and-Out Call Option Under the Black-
monitored down-and-out call options (denoted as n = ) Scholes Model. Up-and-out call options, which are nul-
computed using the analytical formula (Kwok 1998). lified when the underlying asset price reaches a barrier
In Figure 2, the vertical axis and the horizontal axis rep- level from below, are known to present difficulties for
resent the error in the calculated option price and the com- most numerical pricing procedures due to the discontinu-
putational time, respectively, both in log scale. The errors ous nature of their payoff. However, our DE-FGT method,
are root mean square errors of five options with different as well as Reiner’s convolution method, does not suffer
barrier levels, and the time is for computing one option from this at all. In fact, all we need to do is to replace the
price. In the computation of Equation (32), we truncated half-infinite integration interval k  in Equation (9) with
the integral at the lower √bound umin = −40 and the upper a finite interval k hn  and use the DE formula based on
bound umax = ln H + 5 T and used N sample points to Equation (35).
approximate the integral. The value of N is also shown in
the graph. For comparison, we also plotted the error and the Figure 3. European down-and-out call option price
computational time of Reiner’s method with up to 16,000 under the Black-Scholes model (T = 10,
points and Broadie et al.’s (BGK) (1999) trinomial lattice 252 monitoring dates).
method with 5,000 time steps. For the former method, we Execution time (sec.)
truncated
√ the integral of Equation (11) at the upper bound 1
of 10 T . For the latter method, the computation was 0.1 1 10 100 1,000
done on a 2.0 GHz Xeon PC, and their computation times 10 –1
were multiplied by 4, for the Xeon machine was found to 250
be about four times faster than the Pentium II machine. 10 –2
As can be seen from the graph, our method converges 400
DE-FGT, n=252
500
much faster than Reiner’s method and attains the same level 10 –3
Reiner, n =252
of accuracy. Compared with the trinomial lattice method,
10 – 4
our method gives much more accurate prices in much 1,000
shorter time. The error of our method decreases almost
Absolute error

500
10 –5
exponentially with N , because as N is incremented by a 2,000
constant, the position of the corresponding point in the
10 –6
graph moves downward by a constant distance. Finally, our 4,000
method is extremely fast, because even the option with 10 –7
50 monitoring dates can be priced within 0.5 CPU second 600

to an accuracy of 10−10 . 10 –8
8,000

As an example of down-and-out call options with longer

times to maturity and larger number of monitoring dates, 10 –9 16,000

we show the results for one-year T = 10 options with

252 monitoring dates in Figure 3. Other parameters are the 10 –10 700
32,000
same as those for options mentioned above. The reference 800

prices computed by Reiner’s method with 64,000 points are 10 –11

Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 773

Table 2. European down-and-out call prices Table 3. European up-and-out call prices
under the Black-Scholes model under the Black-Scholes model.
(T = 10, 252 monitoring dates).
n = 50
n = 252
H = 121 29102779978
H = 91 113121524522 H = 123 33933815021
H = 93 97292574722 H = 125 38446456577
H = 95 78438846454 H = 127 42550087291
H = 97 56306538930 H = 129 46196180375
H = 99 31673854834

reference prices computed by the convolution method with

Results for up-and-call options with S0 = K = 100, N = 64000 are shown in Table 4.
T = 02, r = 01, q = 0, and  = 03 are illustrated in Figure 5 shows the convergence of the DE-FGT method
Figure 4. Here, the barrier level is varied from H = 121 to and the convolution method. In this case, we extended the
H = 129 and the number of monitoring dates was set to n = upper bound of√integral Equation (11) for the convolution
5, 25, or 50. The reference values computed by Reiner’s method to 30  T to attain the same level of accuracy as
method are shown in Table 3. In this case, the computation in the Black-Scholes case. For the DE-FGT method, we
was also done on a 2.0 GHz Xeon PC. It can easily be seen used the same value of umin as √ in the Black-Scholes case
from the graph that both our method and Reiner’s method and increased umax to ln H + 5  T . Again, it is clear from
can compute the prices with the same speed and accuracy the graph that the error of our method decreases exponen-
as for the down-and-out call options. tially with N .
2.6.3. Down-and-Out Call Option Under Merton’s
Model. We also computed European down-and-out call 3. Lookback Options
option prices under Merton’s model. The parameter values
3.1. Reduction to a One-Dimensional Problem
are the same as for the down-and-out call options under
the Black-Scholes model with T = 02, and for the jump Here we formulate the problem of pricing discrete look-
parameters, we used < = 20, ? = 0, and 4 = 03. The back put options using the notation of §2.1. A lookback
put option on an asset is the right to sell the asset at matu-
rity at the highest asset price between initial and maturity
Figure 4. European up-and-out call option price under dates. Lookback call options are defined similarly and can
the Black-Scholes model. be priced in a completely parallel manner, so here we deal
Execution time (sec.) with only put options. The asset price is monitored on a
1 finite number of dates for discrete lookback options. For
0.01 0.1 1 10
details, see Goldman et al. (1979) or Broadie et al. (1999).
10 –1
Let us consider a discrete lookback option with matu-
150 rity T and monitoring dates ti n−1
i=1 . Also, denote the asset
10 –2 price on these monitoring dates by Si n−1
i=1 and let

DE-FGT, n =50
10 –3
125 Reiner, n=50 Mi = max Sk  (46)
1
k
i
200
10 –4 Then, the payoff of this option at maturity can be written
250
as Mn − Sn . Under the risk-neutral probability measure Q,
Absolute error

10 –5 the price of the option at time 0 is given by

500

10 –6
250 Q0LP S0  = e−rT E0 Mn − Sn  (47)
1,000

10 –7
Table 4. European down-and-out call option prices
2,000
under Merton’s model.
10 –8
300 n=5 n = 25 n = 50
4,000
10 –9
H = 91 8.6304893283 8.2843010923 8.1796345791
H = 93 8.2883832522 7.7161307812 7.5470008678
10 –10 H = 95 7.7707276025 6.8204546460 6.5607004413
350 8,000 H = 97 7.0559324990 5.4877084298 5.0916199042
10 –11
H = 99 6.1639697190 3.7626493142 3.1078183986
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
774 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

Figure 5. European down-and-out call option price we can finally write the option price as
under Merton’s model.
Q0LP S0  = e−qT S0 E0 emn −sn − 1 (54)
Execution time (sec.)
1
0.001 0.01 50 0.1 1 10 100 1,000 This shows that it is sufficient to find the distribution of
10 –1 200 300
mn − sn to compute the option price.
75
To compute Equation (54), we consider the distribution
DE-FGT, n= 5
10 –2 of mi − si i = 0 1     n. Apparently, the distribution
DE-FGT, n= 25
100 250
400 DE-FGT, n= 50 function is zero when mi − si < 0 and there is a finite
10 –3
Reiner, n= 5 probability mass at mi − si = 0. We therefore represent
10 – 4
300 Reiner, n= 25 the distribution of mi − si by two quantities, namely, a
125 Reiner, n= 50
scalar ci , which represents the probability that mi − si = 0,
Absolute error

10 –5
500
and a function gx x > 0, which represents the proba-
350
150 bility density in the region mi − si > 0. Note that the pdf
10 –6 of mi − si can be formally written as ci 4x + gi x, where
4x is the Dirac delta function. At time 0, we have
10 –7 400
600
175

10 –8 c0 = 1 and (55)
450
700 g0 x = 0 (56)
10 –9 500
200 550 800

10 –10
600
by definition. To compute ci and gi x given ci−1 and
16,000 32,000
32,000
gi−1 x, we use the identity
10 –11
mi − si = max0 mi−1 − si 
where Et · is the conditional expectation operator under Q = max0 mi−1 − si−1  + si−1 − si  (57)
given information up to time t.
Equation (47) involves two random variables Sn and Because si−1 − si is an increment of the Brownian motion
Mn at time T , so we need the joint distribution function between time ti−1 and ti and is independent of mi−1 − si−1 ,
of them to compute the expectation value. To reduce the Equation (57) shows that the pdf of mi − si is obtained
dimensionality of the problem, we apply a change of mea-
by computing the convolution of the pdf of mi−1 − si−1
sure (see Babbs 2000, Andreasen 1998) and rewrite Equa-
with that of si−1 − si and collecting all the probability mass
tion (47) as
corresponding to mi − si < 0 to the point mi − si = 0. In
  summary, we have the following recursion formula:
LP −qT  Mn
Q0 S0  = e S0 E0 −1  (48)

Sn 
ḡi x = ci−1 4y + gi−1 yf x − y dy
−
where Et · denotes the conditional expectation operator

under a new measure Q defined by = ci−1 f x + gi−1 yf x − y dy (58)
0
Sn
0
dQ = dQ (49) ci = ḡi x (59)
St e r−qT −t
−

Under this measure, gi x = ḡi x x > 0 (60)

Wt = Wt − t (50) where f x is the probability density function of si−1 − si

given by
becomes a Brownian motion and the stochastic differential    2
equation (1) becomes 1 x + r − q + 21  2 t
f x = √ exp −
2%t 2 2 t
dSt = r − q +  2 St dt + St dWt  (51)  
= pG x + r − q + 21  2 t  (61)
By further introducing the log stock prices
Equations (55), (56), (58), (59), and (60), along with Equa-
tion (54), provide us with a means of computing the
si = lnSi /S0  and (52)
lookback option price by a series of convolutions and
mi = lnMi /S0  (53) integrations.
Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 775

3.2. Application of the DE-FGT Method where

The application of the DE formula and the FGT is now
straightforward. We first rewrite the integral (58) as n2 =  2 t + n42 and (68)
   

 @n = r − q + 21  2 − ;< t + n ? + 21 42  (69)
ḡi x = ci−1 f x + gi−1 y  + bpG x − y   dy   (62)
−b
Then, we have only to replace the function f x given by
where Equation (61) with that given by Equation (67) in the com-
  putation of Equation (58). The resulting convolution can
b = r − q + 21  2 t (63)
be computed efficiently by the DE-FGT method explained
This convolution has the same form as that in Equa- in §2.5.3.
tion (11), and the method described in the previous section
can be applied. After calculating gi x, we can use the DE 3.4. Numerical Results
formula again to compute ci by Equation (59).
3.4.1. Hindsight Call Option Under the Black-Scholes
3.3. Extensions of the Basic Algorithm Model. Figure 6 shows the results of our DE-FGT
method applied to hindsight call options under the Black-
3.3.1. Pricing of Hindsight Options. Using the nota- Scholes model. The parameters are S0 = K = 100, T = 05,
tion we have defined in §3.1, the discrete hindsight call r = 01, q = 0, and  = 03. The number of monitoring
option with strike K is defined as an option whose payoff dates are n = 5, 25, or 50. As reference prices, we used
at T is Mn − K+ . To compute the price Q0HS S0  K of the values given in Tse et al. (2001), which were computed
this option, we use the following relationship between the using the tridiagonal probability algorithm and are claimed
hindsight calls and the lookback puts (see, e.g., Broadie to be accurate up to 10 decimal places. These values, along
et al. 1999): with the computational time for the tridiagonal algorithm,
are shown in Table 5. For comparison, we also included
Q0HS S0  K = Q0LP S0  K + S0 − e−rT K (64) the results obtained with Broadie et al.’s (1999) trinomial
where Q0LP S0  M is the price of a generalized lookback lattice method with 5,000 time steps.
option for which the historical maximum asset price at
time 0 is regarded to be a given value M instead of S0 . We Figure 6. Hindsight call option price under the Black-
can compute the price of such an option easily by changing Scholes model.
the distribution of m0 − s0 from 4x (as given in Equa-
tions (55) and (56)) to 4x − lnM/S0 . Then, by putting 10
Execution time (sec.)
the result into Equation (64), we obtain the hindsight option 0.001 0.01 50 0.1 1 10
price. 20 90
1
3.3.2. Application to Merton’s Model. We next con- DE-FGT, n= 5
sider the pricing of the lookback option under Merton’s 10 –1
70
110 DE-FGT, n= 25
model (1995). Andreasen (1998) shows that under the prob- DE-FGT, n= 50
ability measure Q defined by Equation (49), Si follows a 10 –2
130
new equation: 30 90
40
√ 10 –3
150
Si = Si−1 exp r − q + 21  2 − ;<t +  tz0
Absolute error

NtP t
10 – 4 50
 110 170
+ 4zl +? + 1 2
2
4  (65) BGK, n = 50

l=1 10 –5
190
130
where NtP t follows a Q Poisson process with intensity 60
BGK, n= 25
10 –6
< = <e? (66) 150 BGK, n= 5
210
10 –7
zl
and l = 0 1    are independent and follow the stan-
dard normal distribution N 0 1 under Q . If we define the 10 –8
log stock price by Equation (52), we can show that si−1 − si 170 230
follows the distribution 10 –9 190 250


−< t < tn 1 x + @n 2 70 80
f x = e √ exp −  (67)
n=0 n! 2%n 2n2 10 –10
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
776 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

Table 5. Hindsight call option prices under From the numerical point of view, Equation (73) is dif-
the Black-Scholes model. ficult to evaluate accurately because the integrand contains
the max operator and its higher-order derivatives are dis-
n=5 n = 25 n = 50
continuous. For the standard Bermudan option,
Price 14.9413046399 17.6028684623 18.3264598300
Time 0.66s 20.55s 442.48s hi Si  = Si − K+  (74)
(tridiagonal)
and there is a value Sic called the optimal exercise price
Here we again see that our method is extremely fast for each exercise date i i = 0 1     n such that hi Si  >
and attains an accuracy of 10−10 within a fraction of one Ci Si  for S > Sic and hi Si 
Ci Si  for S
Sic . Hence,
second even for the case of n = 50. The lattice method the higher-order derivatives of the integrand are discontin-
gives reasonably accurate results, but requires much longer uous at Si = Sic . This is the main reason why the conver-
computational time. It is not easy to compare the perfor- gence of the high-order multinomial methods (Alford and
mance of our method with that of the tridiagonal algorithm, Webber 2001, Broadie and Yamamoto 2003) is slower for
because the implementation in Tse et al. (2001) is based Bermudan options than for European options. The DE for-
on MATLAB, while ours is based on C++. However, we mula applied directly to Equation (73) will not work well
can point out that while the computational time of the either, for the efficiency of the DE formula hinges on the
tridiagonal probability algorithm increases drastically with assumption that the integrand is analytical.
the number n of monitoring dates, We therefore choose to locate the optimal exercise price
√ our computational time at each exercise date by the bisection method, divide the
seems to increase only as On n when the accuracy of
the result is fixed. This is natural, because the width of the integration region into two at the price, and then apply the
transition probability DE formula for each of the subregions. We move to the log
√ √ density function pxi
xi−1  is pro- asset price defined by Equation (6) and denote the log opti-
portional to t = √T /n, and therefore we need the step
size proportional to T /n to attain the same level of accu- mal exercise price at ti by xic . Then, Equation (73) can be
racy. √
Hence, the computational work√for each convolution rewritten as
is O n and the total work is On n.
xi+1
c

Ci xi  = e−rt pxi+1
xi hi+1 xi+1  dxi+1

−
4. Bermudan Options

+ pxi+1
xi Ci+1 xi+1  dxi+1  (75)
c
4.1. The DE-FGT Method for Bermudan Options xi+1

The rational price of Bermudan options can be calculated Both of the integrals appearing in Equation (75) become
as a discounted expectation value (under the risk-neutral convolutions under the Black-Scholes or Merton’s model
measure) of the payoff under the optimal (adapted) exercise and can be computed efficiently by the DE-FGT methods
strategy, that is, we have described in §§2.4 and 2.5.3, respectively.
To be more specific, assume that we know the value
Q0 S0  = sup e−rC E0 hC SC  (70) c
of xi+1 , that ai+1 i+1
j  and bj  are the sample points of
C c c
the DE formula in the regions xi+1   and − xi+1 ,
where ht St  is the payoff from exercise at time t and C respectively, and that we have the the values of hi+1 xi+1  at
is a stopping time that takes a value on the discrete set points ai+1 i+1
j  and the values of Ci+1 xi+1  at points bj .

ti ni=1 . It is well known that Q0 S0  can be computed by Clearly, these assumptions hold for i = n − 1. Then, one
the backward recursion time step of our method proceeds as follows:
(1) Determine a lower bound xLc and an upper bound xHc
Qi Si  = maxhi Si  Ci Si  (71) for xic . Then, using the fact that the value of Ci xi  for an
arbitrary point xi can be computed from the given values
Ci Si  = e−rt Ei Qi+1 Si+1  (72) of hi+1 xi+1  and Ci+1 xi+1  through Equation (75) and the
DE formula, compute the log optimal exercise price xic by
where Qi and Ci are the option value and the continuation the bisection method.
value at time ti , respectively, and Ei · is the expectation (2) Generate the sample points aij   of the DE formula
value operator given information up to time ti . Equa- in the region xic  . Also, generate sample points bji   of
tion (72) can be written more explicitly using the transition the DE formula in the region − xic .
probability density function pSi+1
Si : (3) Compute the values of Ci xi  at aij   and bji   by

the DE-FGT method.

Ci Si  = e−rt pSi+1
Si  It may appear that the computation of xic by the bisec-
0 tion method requires extra work and lowers the efficiency
· maxhi+1 Si+1  Ci+1 Si+1  dSi+1  (73) of this method. However, we can reduce this work by also
Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 777

using the FGT for Step 1 above. Let m and N denote the Figure 7. Bermudan call option price under the Black-
number of points used in the bisection and the number of Scholes model.
sample points used in the DE formula, respectively. If direct
Execution time (sec.)
evaluation is used in the bisection to compute Ci xi , the 1
computational work is OmN . Instead, we can perform 0.01 0.1 1 10 100
Steps 1 and 2 of the FGT (see §2.3) before the bisection, 10 –1 100
using the sample points ai+1j  and the values of hi+1 xi+1  binomial
there. Note that in these two steps, no information on the 10 –2
150
multinomial-FGT
target points is necessary. We then have the coefficients 25 DE-FGT
B6 J . Similarly, we perform Steps 1 and 2 using the sample 10 –3 1,000
200

points bji+1  and the values of Ci+1 xi+1  there, and obtain 2,000
3,000

another set of coefficients B6 J . Once we have these coef- 10 – 4

Absolute error
50 250
ficients, we can evaluate Ci xi  for each xi in O1 work, 100
using Step 3 of the FGT. Hence, the work for the bisec- 10 –5 200 32,000
tion can be reduced to Om. Note that the computation of 400

B6 J and B6 J was originally done in Step 3 of the above 10 –6
300
algorithm, so they incur no extra work.
It is also possible to reduce m by a judicious choice of 10 –7 350 51,200
the lower bound xLc and the upper bound xHc . For example,
because the optimal exercise boundary of the Bermudan 10 –8 400 102,400
c
call option is a monotonically decreasing function of t, xi+1
c
gives a lower bound on xi . An upper bound can easily be 10 –9
450
obtained from the boundary for an infinite maturity Amer- 600
ican call option. 10 –10
The DE-FGT method we proposed in this subsection has
computational work of Onm + N  when the number of
exercise dates is n. It can also be expected that the compu- DE-FGT method are also shown in the graph. It can be
tational error decreases faster than any negative power of N seen from the graph that the binomial method is accept-
because both of the integrands appearing in Equation (75) able if low accuracies are required, but if higher accuracies
are analytical. We will confirm this in the next subsection. are required, the multinomial-FGT method is much faster.
Our DE-FGT method is even faster and shows exponen-
4.2. Numerical Results tial decrease of the absolute error, attaining an accuracy of
10−10 within one CPU second.
4.2.1. Bermudan Call Options Under the Black-
Scholes Model. We show results for Bermudan call 4.2.2. Bermudan Put Options Under Merton’s
options under the Black-Scholes model in Figure 7. The Model. Finally, we show pricing results for Bermudan put
parameters are S0 = 100, T = 05, r = 003, q = 007, options under Merton’s jump-diffusion model. The param-
 = 02, and n = 10. The strike price K was varied from
eters are S0 = 40, K = 30 to 50 √ (in increments of 5),
90 to 110 in increments of 5. The reference values com-
T = 10, √r = 008, q = 0,  = 005, < = 50, ? = 0,
puted using the multinomial FGT method (Broadie and
and 4 = 005. We show the reference values computed
Yamamoto 2003) are shown in Table 6. The multinomial-
FGT method is a variant of the multinomial method, which using a modified version of the multinomial-FGT method
has a large number of branches and which uses the FGT to (referred to as FGT II in Broadie and Yamamoto 2003)
speed up the computation of the continuation values. The with b = 102400 in Table 7.
number of branches used is 2b + 1, where b = 409600. Here we compare the performance of four numerical
In Figure 7, we show the results of three numerical meth- methods: Amin’s algorithm (1993), FGT I, FGT II (Broadie
ods: the binomial method, the multinomial-FGT method, and Yamamoto 2003), and our DE-FGT method. FGT I is
and our DE-FGT method. The number of time steps for a variant of Amin’s algorithm that uses the FGT to reduce
the binomial method, the value of b for the multinomial- the work of computing the continuation values at each time
FGT method, and the number of sample points N for the step from ON 2  to ON . FGT II is a multinomial-FGT

Table 6. Bermudan call option price under the Black-Scholes

model.
K = 90 K = 95 K = 100 K = 105 K = 110
Price 10.73001013 7.32288562 4.75727741 2.94105489 1.73255637
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
778 Operations Research 53(5), pp. 764–779, © 2005 INFORMS

Table 7. Put option prices under Merton’s model.

K = 30 K = 35 K = 40 K = 45 K = 50
Price 2.70636133 4.58094453 6.99527475 9.90671277 13.25586096

method that takes the effect of jumps into account. See and obtain the same order of computational work and con-
Broadie and Yamamoto (2003) for more details about these vergence properties as in the Black-Scholes case.
two methods. The convergence results of the four methods
are shown in Figure 8. The numbers in the graph represent
the number of grid points or sample points used at each
Acknowledgments
time step. Again, the two FGT-based methods, FGT I and This work was partially supported by NSF grants DMS-
FGT II, converge much faster than Amin’s algorithm, but 0074637 and DMS-0410234 and by the Ministry of Educa-
the DE-FGT method shows exponential convergence and is tion, Science, Sports and Culture, Grant-in-Aid for Young
even faster. Scientists, 16760053, 2004 and Grant-in-Aid for the 21st
Century COE “Frontiers of Computational Science.”
5. Conclusion
In this paper, we have shown that under the Black-Scholes References
framework, the price of many path-dependent and quasi Ahn, D. H., S. Figlewski, B. Gao. 1999. Pricing discrete barrier options
path-dependent options—such as the barrier, lookback, with an adaptive mesh model. J. Derivatives 6(4) 33–44.
hindsight, and Bermudan options—can be computed by a Alford, J., N. Webber. 2001. Very high order lattice methods for one factor
series of convolutions of the Gaussian distribution and a models. Working paper, Cass Business School, London, UK.
known function. Using this fact, we proposed a new pricing Amin, K. 1993. Jump diffusion option valuation in discrete time.
J. Finance 48(5) 1833–1863.
algorithm, the DE-FGT method, that computes this convo-
Andreasen, J. 1998. The pricing of discretely sampled Asian and lookback
lution by a combination of the double-exponential integra- options: A change of numeraire approach. J. Computational Finance
tion formula and the FGT. 2(1) 5–30.
The computational work of our method is OnN  when Babbs, S. 2000. Binomial valuation of lookback options. J. Econom.
the number of monitoring/exercise dates is n and the num- Dynamics Control 24(11–12) 1499–1525.
ber of sample points at each date is N . Theoretically, the Bates, D. 1996. Jumps and stochastic volatility: Exchange rate processes
implicit in Deutsche mark options. Rev. Financial Stud. 9 69–107.
errors of our method decrease faster than any negative
Baxter, B., G. Roussos. 2002. A new error estimate of the fast Gauss
power of N , and our experiments on the above four options transform. SIAM J. Sci. Comput. 24(1) 257–259.
show an exponential decrease of the errors. We also extend Black, F., M. Scholes. 1973. The pricing of options and corporate liabili-
the method to Merton’s lognormal jump-diffusion model ties. J. Political Econom. 81 637–654.
Boyle, P. 1988. A lattice framework for option pricing with two state
variables. J. Financial Quant. Anal. 23 1–12.
Figure 8. Bermudan put option price under the Merton
Boyle, P., S. H. Lau. 1994. Bumping up against the barrier with the bino-
jump-diffusion model. mial method. J. Derivatives 1(4) 6–14.
Execution time (sec.) Boyle, P., Y. S. Tian. 1998. An explicit finite difference approach to the
1 pricing of barrier options. Appl. Math. Finance 5(1) 17–43.
0.1 1 10 100 1,000 Boyle, P., Y. S. Tian. 1999. Pricing lookback and barrier options under
the CEV process. J. Financial Quant. Anal. 34(2) 241–264.
10 –1
Broadie, M., Y. Yamamoto. 2003. Application of the fast Gauss transform
100
to option pricing. Management Sci. 49(8) 1071–1088.
100
10 –2 200 Broadie, M., P. Glasserman, S. G. Kou. 1997. A continuity correction for
100 300 discrete barrier options. Math. Finance 7(4) 325–349.
200
10 –3
300 1200 Broadie, M., P. Glasserman, S. G. Kou. 1999. Connecting discrete and
continuous path-dependent options. Finance Stochastics 3(1) 55–82.
Absolute error

200 1200
Cheuk, T., T. Vorst. 1996. Complex barrier options. J. Derivatives 4(1)
10 –4 8–22.
100 200 Amin
300 Conze, A., R. Viswanathan. 1991. Path dependent options: The case of
FGT I lookback options. J. Financial Econom. 7 229–263.
10 –5
400 FGT II Cox, J., S. Ross, M. Rubinstein. 1979. Option pricing: A simplified
DE-FGT approach. J. Financial Econom. 7 229–263.
10 –6 Davydov, D., V. Linetsky. 2001. The valuation and hedging of barrier
400
and lookback options under the CEV process. Management Sci. 47
500 949–965.
10 –7
600 51,200 Duan, J.-C., E. Dudley, G. Gauthier, J.-G. Simonato. 2003. Pricing dis-
cretely monitored barrier options by a Markov chain. J. Derivatives
10 –8 10(4) 9–31.
 Broadie and Yamamoto: A Double-Exponential Fast Gauss Transform Algorithm for Pricing Discrete Path-Dependent Options
Operations Research 53(5), pp. 764–779, © 2005 INFORMS 779

Duffie, D. 2001. Dynamic Asset Pricing Theory, 3rd ed. Princeton Uni- Ooura, T., M. Mori. 1991. The double exponential formula for oscillatory
versity Press, Princeton, NJ. functions over the half infinite interval. J. Computational Appl. Math.
Dutt, A., V. Rokhlin. 1993. Fast Fourier transforms for nonequispaced 38 353–360.
data. SIAM J. Sci. Comput. 14(6) 1368–1393. Reiner, E. 2000. Convolution methods for exotic options. Presentation at
Evans, G. 1993. Practical Numerical Integration. John Wiley and Sons, Columbia University, New York.
New York. Rich, D. R. 1994. The mathematical foundations of barrier option pricing
Goldman, M., H. Sosin, M. A. Gatto. 1979. Path-dependent options: Buy theory. Adv. Futures Options Res. 7 267–371.
at the low, sell at the high. J. Finance 34 1111–1128. Ritchken, P. 1995. On pricing barrier options. J. Derivatives 3(2) 19–28.
Greengard, L., J. Strain. 1991. The fast Gauss transform. SIAM J. Sci.
Rubinstein, M., E. Reiner. 2000. Breaking down the barriers. RISK
Statist. Comput. 12(1) 79–94.
8(September) 28–35.
Greengard, L., X. Sun. 1998. A new version of the fast Gauss transform.
Sugihara, M. 1997. Optimality of the double exponential formula—
Documenta Mathematica, extra volume ICM(III) 575–584.
Functional analysis approach. Numerische Mathematik 75 379–395.
Kou, S. 2002. A jump diffusion model for option pricing. Management
Sci. 48(8) 1086–1101. Sullivan, M. A. 2000. Pricing discretely monitored barrier options. J. Com-
putational Finance 3(4) 35–52.
Krommer, A., C. Ueberhuber. 1998. Computational Integration. SIAM,
Philadelphia, PA. Takahashi, H., M. Mori. 1974. Double exponential formulas for numerical
Kwok, Y. K. 1998. Mathematical Models of Financial Derivatives. integration. Publ. RIMS Kyoto Univ. 9 721–741.
Springer-Verlag, Singapore. Tse, W. M., L. K. Li, K. W. Ng. 2001. Pricing discrete barrier and hind-
Merton, R. 1992. Continuous-Time Finance. Basil Blackwell, London, UK. sight options with the tridiagonal probability algorithm. Management
Sci. 47(3) 383–393.
Mori, M. 1985. Quadrature formulas obtained by variable transformation
and the DE-rule. J. Computational Appl. Math. 12 & 13 119–130. Ware, A. F. 1998. Fast approximate Fourier transforms for irregularly
Naik, V., M. Lee. 1990. General equilibrium pricing of options on the spaced data. SIAM Rev. 40(4) 838–856.
market portfolio with discontinuous returns. Rev. Financial Stud. 3(4) Zvan, R., K. Vetzal, P. Forsyth. 2000. PDE methods for pricing barrier
493–521. options. J. Econom. Dynam. Control 24 1563–1590.

View publication stats