The Journal of Computational Finance (4961) Volume 13/Number 1, Fall 2009

Saddlepoint methods for option pricing

Peter Carr
Bloomberg LP, 731 Lexington Avenue, New York, NY 10022, USA;
email: [email protected]
Dilip Madan
Robert H. Smith School of Business, University of Maryland, College Park, MD 20742, USA;
email: [email protected]
A single saddlepoint approximation for call prices seen as complementary
probabilities that log price exceeds log strike by an independent exponential
under the share measure is developed using a non-Gaussian base. The
suggested base is that of a Gaussian random variable less an exponential
with parameter . It is suggested that be chosen to match the volatility
under the share measure. The method is implemented and observed to be
exact for the BlackScholes model. Six other models with closed forms for
the cumulant generating function are also investigated.
1 INTRODUCTION
The Fourier transform of the density for the logarithm of the stock price has seen
numerous nancial applications. For a theoretical perspective we cite as examples
Dufe et al (2000) and Bakshi and Madan (2000). This transform has become a
standard calibration engine following the methods of Carr and Madan (1999) who
invoked the fast Fourier transform for its speed. Direct Fourier inversion has also
been used and we cite Heston (1993), Bates (1996) and Scott (1997).
Though adequate for near-money options, and this sufces for most calibration
exercises, the method is known to break down for deep out-of-the-money options
where it often gives rise to negative prices. Rogers and Zane (1999) suggest the use
of classical saddlepoint methods to compute option prices and employ in particular
the Lugannini and Rice (1980) approximation as developed in Daniels (1987) and
extensively studied in Jensen (1995). They consider mainly a Gaussian base density
but suggest that we may follow Wood et al (1993) and Butler (2007), for non-
Gaussian bases. Working with a Gaussian base these methods are also used by Xiong
et al (2005) for a variety of models with stochastic rates and volatilities.
These applications of classical saddlepoint methods are used to compute the
probability that the stock is in-the-money for the risk-neutral probability and the
reweighted probability when the stock is itself taken as a numeraire. Thus two
saddlepoint computations are involved in constructing one call option price.
49

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
50 P. Carr and D. Madan
We recognize following Madan et al (2008) that the call price is itself quite
generally a complementary probability itself. Hence one should be able to apply
saddlepoint methods directly in one step to determine the call price. However, even
in the classical BlackScholes case, this density is not Gaussian and the use of a
Gaussian base is not exact. It turns out that in the BlackScholes case the density
reected in call prices as a complementary probability is the density of a Gaussian
variable less an independent exponential. This leads us to select for a non-Gaussian
base the convolution of a normal random variable with a negative exponential.
With this altered base we observe that a single saddlepoint computation does yield
exact BlackScholes prices on adopting the Wood et al (1993) generalization of the
LuganniniRice (denoted as LR in tables and gures) approximation.
We then adopt this base more generally for a host of well known option pricing
models. In the present paper we study six models. They are the Carr, Geman, Madan
and Yor (CGMY) model, Carr et al (2002), its spectrally negative form CGYSN
studied in Eberlein and Madan (2009), the VGSSD Sato process of Carr et al (2007),
the Merton (1976) jump diffusion model with stochastic volatility (SVJ) as studied in
Bakshi et al (1997), the Heston stochastic volatility (HSV) model, Heston (1993) and
its generalization to Lvy processes as developed in the VGSA model of Carr et al
(2003). For all these models we demonstrate that the single saddlepoint method with
the non-Gaussian base given by the density of a Gaussian variate less an independent
exponential provides effective call prices that work well for deep out-of-the-money
options where the more traditional Fourier methods break down into negative prices.
The outline for the rest of the paper is as follows. Section 2 describes call prices as
a complementary probability. Section 3 summarizes the generalized LuganniniRice
approximation of Wood et al (1993). In Section 4 we analyze the BlackScholes case
and show that the base given by the convolution of a Gaussian variate with a negative
exponential would be exact in this case. Computations for the BlackScholes case
are conducted in Section 5 and it is shown that the new base is exact. This base is
then adopted more generally for the other models and results are presented for the
six models discussed above in Section 6. Section 7 concludes.
2 CALL PRICES AS PROBABILITIES
Recently Madan et al (2008) have shown that quite generally the call price is the
probability that the stock price was equal to the strike for the last time before the
maturity. Equivalently it is the probability that after the maturity the stock stays
below the strike forever. Here we conduct a much simpler analysis at the level of
random variables as opposed to processes, and enquire into the nature of the random
variable represented by the call price seen as a probability. We begin by writing the
call price as a function of the log strike k. That is c(k), which is given by:
c(k) =e
rt
_

k
(e
x
e
k
)f (x) dx
The Journal of Computational Finance Volume 13/Number 1, Fall 2009

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
Saddlepoint methods for option pricing 51
where f (x) is the density for x, the logarithm of the stock price, and r is the risk-free
interest rate, assumed to be constant. We divide this call price by the spot price S
0
and note that as:
S
0
=e
rt
_

e
x
f (x) dx
we get that:
c(k) =
c(k)
S
0
=
_

k
(e
x
e
k
)f (x) dx
_

e
x
f (x) dx
(1)
It is clear that this relativized call price (1) is unity at a zero strike and decreases
to zero as the log strike tends to innity and hence this function is a complementary
probability. The negative of its derivative is therefore a density and computation
yields that:
c

(k) =
_

k
e
k
f (x) dx
_

e
x
f (x) dx
Ignoring the normalizing constant we consider the unnormalized function:
g(k) =
_

k
e
k
f (x) dx
that tends to zero as the strike goes to positive or negative innity. Computing the
Fourier transform of g we observe that:
_

e
iuk
g(k) dk =
_

e
iuk
e
k
(1 F(k)) dk
=
_

e
(1+iuk)
1 +iu
f (k) dk
=
(u i)
1 +iu
where (u) is the characteristic function for the logarithm of the stock price. On
incorporating the normalization we observe the transform:
(u) =
(u i)
(i)(1 +iu)
(2)
The transform (u) is the characteristic function of a density whose complemen-
tary distribution function is the normalized call price taken at the log strike. We
also observe that (2) is the product of two characteristic functions, the rst being
(u i)/(i) and the second is (1 +iu)
1
which is the characteristic function
Research Paper www.thejournalofcomputationalnance.com

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
52 P. Carr and D. Madan
of a negative exponential random variable. The rst characteristic function is easily
seen to be the density of the log of the stock, tilted by the stock price itself, or
the exponential of this log stock price. Prices of claims paid out in US dollars can
equivalently be seen as prices of claims paid out in shares, but now valued using this
share price tilted measure. This measure is also called the share measure as we are
transforming payouts from US dollars to shares when employing such a measure.
Hence, by virtue of being a product of characteristic functions, the derivative of
the call price normalized by the forward price is the density of:
Z =X Y
where X is the logarithm of the stock under the share measure and Y is an
independent exponential. The normalized call price is then the probability that
Z > log(K) or equivalently that:
X > ln K +Y
the probability that under the share measure the logarithm of the stock exceeds log
strike by an independent exponential variate.
We can obtain this result without transform methods on noting that the normalized
call price is:
C(K)
S
0
=
E[(S K)
+
]
E[S]
=

E
__
1
K
S
_
+
_
where

E is the share measure. Now we dene:
K
S
=e
y
and let f (y) be the density for y under the share measure with F(y) the correspond-
ing distribution function then:
C(K)
S
0
=
_

0
(1 e
y
)f (y) dy
=
_

0
(1 F(y))e
y
dy
But as e
y
is the density of a positive exponential we have here the probability
that:
ln S ln K > Y
or the probability that the log of the stock exceeds log strike by an independent
exponential.
The Journal of Computational Finance Volume 13/Number 1, Fall 2009

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
Saddlepoint methods for option pricing 53
3 GENERALIZED LUGANNINIRICE APPROXIMATIONS
Wood et al (1993) generalized the LuganniniRice approximation to an arbitrary
base density. For the implementation of the approximation to a particular base
one needs access to the cumulant generating function (CGF) of the base random
variable Z:
G(w) =log(E[exp(wZ)])
the cumulative distribution function (cdf), (w), and the probability density function
(pdf), (w). For a more detailed presentation of the method we refer the reader to
Butler (2007, Theorem 16.1.1, pp. 528531).
Suppose we wish to evaluate the probability that X > y for a random variable
with cumulant function:
K(t ) =log(E[exp(t X)])
The LuganniniRice approximation requires that we rst evaluate the Fenchel
transform of the base CGF as:
H() =G(w()) w()
G

(w()) = (3)
We then dene

t , by the saddlepoint equation:
K

t ) =y
The dominant term in the representation of the target density at y in terms of its
cumulant is then equated to the corresponding dominant term with respect to the
base by:
H(

) =K(

t )

t y
There are two solutions for

as H is concave and we take

< G

(0) if y < K

(0)
and

> G

(0) if y > K

(0) otherwise.
We further dene the required second order terms by:
u =

t (K

t ))
1/2
u

= u(G

(w(

)))
1/2
The complementary probability is then estimated by:
P(X > y) =1 (

) + (

)
_
1
u

1
w(

)
_
For a Gaussian base we have the classical LuganniniRice approximation. We also
know that base density will be exact for all models that are a shift and scale transform
of the base. The base is therefore determined up to shift and scale.
Research Paper www.thejournalofcomputationalnance.com

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
54 P. Carr and D. Madan
4 THE BLACKSCHOLES CASE
We would like to choose a base that is exact for the BlackScholes case. This leads
us to ask what the base is in this case. We rst verify independently that the Black
Scholes price of a call option relative to the initial spot price is the probability that
the log of the stock under the share measure exceeds log strike by an independent
exponential.
Under the share measure the logarithm of the nal stock price is given by:
X =ln(S
0
) +
_
r q +

2
2
_
t +

t Z
where Z is a standard normal variate. We therefore seek the probability that:
X > ln K +Y
for an independent exponential Y.
Conditional on the exponential variate this is given by the probability that:
Z >
ln K/S
0

_
r q

+

2
_

t +
Y

t
or:
N
_
ln S
0
/K

t
+
_
r q

+

2
_

t
Y

t
_
where N(x) represents the cumulative distribution function of a standard normal
distribution and n(x) is its corresponding density function.
The unconditional probability is then:
_

0
dy e
y
N
_
ln S
0
/K

t
+
_
r q

+

2
_

t
y

t
_
=e
y
N
_
ln S
0
/K

t
+
_
r q

+

2
_

t
y

t
_

_

0
e
y
n
_
ln S
0
/K

t
+
_
r q

+

2
_

t
y

t
_
1

t
dy
(on integration by parts)
=N
_
ln S
0
/K

t
+
_
r q

+

2
_

t
_

_

0
e
y
n
_
ln S
0
/K

t
+
_
r q

+

2
_

t
y

t
_
1

t
dy
=N
_
ln S
0
/K

t
+
_
r q

+

2
_

t
_

_

0
e
y
1

t
exp
_

1
2
_
y

_
r q

2
_

t
ln S
0
/K

t
_
2
_
dy
The Journal of Computational Finance Volume 13/Number 1, Fall 2009

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
Saddlepoint methods for option pricing 55
We analyze the nal integral as follows:
_

0
e
y
1

t
exp
_

1
2
_
y

_
r q

+

2
_

t
ln S
0
/K

t
_
2
_
dy
=
_

((rq)/+/2)

t +(ln S
0
K)/

t
1

2
exp
_

z
2
2

_

t z +(r q)t
+

2
t
2
+ln S
0
/K
__
dz
by a change of variable:
=
K
S
0
e
rt +qt
2
t /2
_

((rq)/+/2)

t (ln S
0
/K)/

t
1

2
exp
_

z
2
2

t z
_
dz
=
K
S
0
e
rt +qt
_

((rq)//2)

t (ln S
0
/K)/

t
1

2
e
z
2
/2
dz
by another change of variable:
=Ke
rt +qt
N
_
ln S
0
/K

t
+
_
r q

2
_

t
_
Hence on multiplication by S
0
e
qt
we get the traditional BlackScholes formula.
5 THE CHOICE OF A NON-GAUSSIAN BASE
We observe from the analysis of the previous section that the density reected in call
prices for the BlackScholes model is, up to a shift and change of scale, a Gaussian
density less an independent exponential. This leads us to consider for a base model
the one parameter family of a zero mean variable given by:
Z +
1

Y
where Z is a standard Gaussian variate and Y is a positive exponential with
parameter . To employ such a base in a saddlepoint approximation requires a
knowledge of its CGF, cdf, pdf and the Fenchel transform of the CGF.
The CGF of our suggested base is:
G(w) =
w
2
2
+
w

ln( +w) +ln()

and its rst and second derivatives are:
G

(w) =w +
1

1
+w
G

(w) =1 +
_
1
+w
_
2
Research Paper www.thejournalofcomputationalnance.com

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
56 P. Carr and D. Madan
For the complementary cdf we evaluate directly:
(y) =Pr
_
Z +
1

Y > a
_
=
_

a1/
1

2
e
z
2
/2
_
1 exp
_

_
z
_
a
1

____
dz
=N
_
1

a
_
exp
_

_
a
1

__ _

a1/
1

2
e
z
2
/2z
dz
=N
_
1

a
_
exp
_
a 1 +

2
2
_
N
_
1

a
_
The pdf follows on differentiation as:
(y) =n
_
1

a
_
+ exp
_
a 1 +

2
2
_
N
_
1

a
_
exp
_
a 1 +

2
2
_
n
_
1

a
_
We also need the GaussFenchel transform of the CGF and for this we solve for:
w() = +
c
2
+
_
c
2
4
+1
c =
1

+
The GaussFenchel transform may then be directly determined from Equation (3).
We now develop the saddlepoint method for the BlackScholes model using this
Gauss minus exponential (GME) base. The exact CGF for the random variable
associated with the call price is the CGF for the log of the stock under the share
measure less an independent exponential and this is:
K(x) =

2
t x
2
2
+

2
t
2
x ln(1 +x)
To observe that the suggested base is exact for this CGF we must observe that we
may recover K by a shift and change of scale of G. Consider an arbitrary shift of a
and scale of b to get that:
ax +G(bx) =ax +

2
t x
2
2
+

x ln
_
+

t x
_
+ln()
Now choose =

t to get:
ax +G(bx) =ax +

2
t x
2
2
+x ln(1 +x)
The Journal of Computational Finance Volume 13/Number 1, Fall 2009

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
Saddlepoint methods for option pricing 57
TABLE 1 LuganniniRice saddlepoint prices for BlackScholes.
Strike Maturity LR GME LR Gauss BlackScholes
36.78 0.5 63.7597 62.9690 63.7597
60.65 0.5 40.2569 39.8812 40.2569
81.87 0.5 20.2050 20.1427 20.2050
90.48 0.5 13.4486 13.3367 13.4486
110.52 1 7.08205 6.7907 7.08205
122.14 1 4.00475 3.8300 4.00475
174.87 1 0.365935 0.3524 0.365935
271.82 1 0.0
3
48714 0.0
3
4722 0.0
3
48714
Finally we choose:
a =

2
t
2
1
to get that ax +G(bx) =K(x).
We computed the saddlepoint LuganniniRice approximation for the Black
Scholes model using this GME base for a volatility of 25%, and a sample of
strikes and maturities with the initial spot at 100, an interest rate of 3% and a zero
dividend yield with the choice of at

t . We also computed the LuganniniRice

approximation using a Gaussian base for the same options. Finally we presented the
BlackScholes price computed from the usual BlackScholes formula. We know that
the LuganniniRice approximation with the GME base is exact and the fact that they
match up serves as a check on our new program for the saddlepoint method, where
the classical BlackScholes formula serves as a benchmark. The results are presented
in Table 1.
6 A GENERAL PURPOSE SADDLEPOINT PRICER
We develop a general purpose saddlepoint pricer using the GME base with the
function call:
ww =optionpriceGMESP(pp, kk, rr, qq, xx, tt, uu, model)
that takes as inputs the vectors pp for the initial spot, kk for the strikes, rr for the
interest rates, qq for the dividend yields, xx the parameter values of a prospective
model, tt for the option maturities, uu a vector denoting iscall, and model the name
of the model. The pricer requires that the following three functions are available
for an option price to be computable. These are the CGF, and its rst and second
derivatives. Given that a Gaussian base is not exact for the classical BlackScholes
formula as observed in Section 5 above, we did not write a general purpose pricer
for this base. In fact we rst applied the Gaussian base to the BlackScholes model
and observed the difference from the classical price; this then led to a need to alter
the base. In this section we present results for just the GME base.
Research Paper www.thejournalofcomputationalnance.com

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
58 P. Carr and D. Madan
Suppose we have the CGF of the log of the stock under some risk-neutral model
given by the function K(x). The CGF of the density embodied in the call price
relativized by the spot price is the function:
K(x +1) ln(1 +x)
Our base is of the form:
x
2
2
+
x

ln( +x) +ln()

We evaluate at a scale b to write:
b
2
x
2
2
+
bx

ln( +bx) +ln()

If we take =b we get:
b
2
x
2
2
+x ln(1 +x)
This choice forces a matching of the cumulant of the exponential in the rescaled
base to the cumulant of the exponential in the target. The means or drifts can always
be matched by shifts and so we focus on the second order terms of K(x +1) and
b
2
x
2
/2. A matching of the volatility is attained by choosing:
=b =[K

t +1)]
1/2
We observe that for BlackScholes this gives the correct solution for =

t . We
could have followed the suggestion of Wood et al (1993) of matching standardized
skewness but then we would not be exact for BlackScholes.
7 RESULTS FOR SOME ASSORTED MODELS
Apart from the Gaussian model reported on in Table 1 we applied our general
purpose GME saddlepoint pricer to models CGMY of Carr et al (2002), the model
VGSSD for the Sato process of Carr et al (2007), the spectrally negative model
CGYSN studied in Eberlein and Madan (2009), and three stochastic volatility
models, HSV model of Heston (1993), the Merton (1976) SVJ as described in Bakshi
et al (1997), and the Lvy process model with stochastic volatility VGSA of Carr
et al (2003).
The parameter values used for the six models were as follows. For CGMY we
used C =2, G=5, M =10 and Y =0.5. For VGSSD we used =0.2, =0.5,
=0.15 and =0.5. For CGYSN we took =0.1, C =0.5, G=5 and Y =0.5.
The Heston parameters were initial volatility 0.2, long-term volatility 0.2, mean
reversion 2, volatility of volatility 0.5 and correlation 0.7. The parameters for SVJ
were initial variance 0.02, jump arrival rate 2.5, mean jump size 0.001, standard
deviation of jump 0.0155, mean reversion 3, long-term variance 0.3, volatility of
The Journal of Computational Finance Volume 13/Number 1, Fall 2009

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
Saddlepoint methods for option pricing 59
volatility 0.5, and correlation 0.7. The parameters for VGSA were initial speed
10, G=20, M =40, mean reversion 3, long-term speed 6, volatility of speed 7.
For all models we used the initial spot at 100, the interest rate at 0.03, a zero
dividend yield, and a half-year maturity. The strikes ranged from 10 dollars to 200
dollars in steps of 10 dollars. Two computations were performed, the rst is the price
using the fast Fourier transform method of Carr and Madan (1999) that is known
to break down for deep out-of-the-money options, often returning negative prices
in such cases, and the motivation for developing the saddlepoint method developed
here. The second is the GME base saddlepoint method. The results are reported in
Table 2 (see page 60). We observe from this table that the saddlepoint pricer matches
the fast Fourier transform for near-money options and it returns positive prices in
the range of deep out-of-the-money options where the fast Fourier transform has
broken down.
8 CONCLUSION
Call prices are observed to be the complementary probabilities that log price exceeds
log strike by an independent exponential, under the share measure obtained on
tilting the log price by its exponential or the normalized nal stock price. It is
therefore appropriate to approximate deep out-of-the-money call prices using a
single saddlepoint approximation for the appropriate complementary probability.
The issue arises with respect to the base distribution to be used in the saddlepoint
approximation. Typically in a standard application one employs a Gaussian base.
For the BlackScholes model of geometric Brownian motion the exact result is that
of the density and cumulant generating function for the random variable formed by
a Gaussian variable less an independent exponential and this density is not a scale
shift transform of a Gaussian base. Hence the use of a Gaussian base will not be exact
for the BlackScholes model. In order to be exact for the BlackScholes model the
suggested base is that of a Gaussian variable less an independent exponential with
parameter . The method is then to employ saddlepoint approximations with such
a base distribution. The parameter of the exponential distribution in the base, that
matches, on rescaling, the cumulant of the exponential in the target cumulant, is
seen to be the volatility of the risk-neutral distribution under the share measure as
given by the square root of second derivative of the unit shifted cumulant taken at the
solution of the saddlepoint equation. The methods are implemented and observed
to be exact for the BlackScholes model. Six other models with closed forms for
the cumulant generating function are also investigated. These are the CGMY model
of Carr et al (2002), the Sato process model constructed from the variance gamma
model as described in Carr et al (2007), the spectrally negative form of CGMY with
a diffusion component studied in Eberlein and Madan (2009), the HSV model of
Heston (1993), the Merton (1976) jump diffusion model and the stochastic volatility
Lvy process, VGSA of Carr et al (2003).
Research Paper www.thejournalofcomputationalnance.com

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
60 P. Carr and D. Madan
TABLE 2 Call prices for different models.
FFT GME LR FFT GME LR FFT GME LR
Strike CGMY CGMY VGSSD VGSSD CGYSN CGYSN
10 88.2148 90.1571 87.7794 90.1508 88.3875 90.1517
20 79.4972 80.3279 79.1716 80.3051 79.4701 80.3086
30 70.2051 70.5178 70.0198 70.4653 70.1370 70.4738
40 60.6231 60.7397 60.4879 60.6352 60.5219 60.6530
50 51.0359 51.0482 50.7400 50.8222 50.7584 50.8565
60 41.7280 41.6104 40.9403 41.0401 40.9802 41.1042
70 32.9847 32.7303 31.2560 31.3247 31.3234 31.4395
80 25.0952 24.7688 21.9359 21.8432 21.9862 22.0009
90 18.3244 18.0184 13.4748 13.1831 13.3894 13.2631
100 12.8429 12.6191 6.6766 6.4127 6.3690 6.2481
110 8.6739 8.5428 2.4189 2.3804 2.0135 1.9884
120 5.6892 5.6278 0.6988 0.7205 0.3593 0.3611
130 3.6601 3.6387 0.1951 0.2057 0.0314 0.0342
140 2.3324 2.3295 0.0551 6.0095E02 9.8120E04 1.7169E03
150 1.4840 1.4879 0.0151 1.8456E02 2.6410E03 4.8645E05
160 0.9482 0.9537 0.0031 5.9918E03 2.6894E03 8.4152E07
170 0.6012 0.6159 0.0040 2.0547E03 2.8654E03 9.5923E09
180 0.3278 0.4018 0.0236 7.4200E04 3.6956E03 7.7157E11
190 0.0129 0.2652 0.0732 2.8116E04 5.7047E03 4.6470E13
200 0.4588 0.1772 0.1700 1.1139E04 9.4178E03 2.2040E15
HSV HSV SVJ SVJ VGSA VGSA
10 88.2213 90.1511 88.9673 90.1609 88.4818 90.1504
20 79.3860 80.3064 79.7445 80.3240 79.4958 80.3036
30 70.1029 70.4684 70.2563 70.4907 70.1374 70.4614
40 60.5096 60.6416 60.5799 60.6637 60.5149 60.6272
50 50.7435 50.8370 50.7925 50.8491 50.7371 50.8088
60 40.9422 41.0710 40.9711 41.0574 40.9123 41.0192
70 31.2459 31.3837 31.1949 31.3139 31.1521 31.2899
80 21.8595 21.8939 21.6007 21.6865 21.6401 21.7127
90 13.1996 13.0323 12.5598 12.4728 12.8156 12.6732
100 6.0528 5.8409 5.0195 4.8485 5.5627 5.3009
110 1.6344 1.5848 0.8444 0.8202 1.2106 1.1812
120 0.2320 0.2385 0.0521 0.0564 0.1390 0.1503
130 2.48E02 2.87E02 5.67E04 3.46E03 1.44E02 1.81E02
140 6.99E04 3.56E03 2.46E03 2.44E04 4.04E04 2.41E03
150 2.23E03 4.82E04 2.67E03 2.02E05 2.35E03 3.57E04
160 2.62E03 7.19E05 2.69E03 1.93E06 2.63E03 5.90E05
170 3.06E03 1.18E05 2.72E03 2.11E07 2.92E03 1.07E05
180 5.01E03 2.10E06 2.85E03 2.59E08 4.21E03 2.13E06
190 9.97E03 4.08E07 3.18E03 3.53E09 7.46E03 4.58E07
200 1.94E02 8.51E08 3.81E03 5.31E10 1.37E02 1.06E07
The Journal of Computational Finance Volume 13/Number 1, Fall 2009

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.
Saddlepoint methods for option pricing 61
REFERENCES
Bakshi, G., Cao, C., and Chen, Z. (1997). Empirical performance of alternative option pricing
models. Journal of Financial Economics 52, 20032049.
Bakshi, G., and Madan, D. (2000). Spanning and derivative security valuation. Journal of
Financial Economics 55, 205238.
Bates, D. S. (1996). Jumps and stochastic volatility: exchange rate processes implicit in
Deutschemark options. Review of Financial Studies 9, 69107.
Butler, R. (2007). Saddlepoint Approximations with Applications (Cambridge Series in
Statistical and Probabilistic Mathematics). Cambridge University Press, Cambridge.
Carr, P., Geman, H., Madan, D., and Yor, M. (2002). The ne structure of asset returns:
An empirical investigation. Journal of Business 75, 305332.
Carr, P., Geman, H., Madan, D., and Yor, M. (2003). Stochastic volatility for Lvy processes.
Mathematical Finance 13, 345382.
Carr, P., Geman, H., Madan, D., and Yor, M. (2007). Self decomposability and option pricing.
Mathematical Finance 17, 3157.
Carr, P., and Madan, D. (1999). Option valuation using the fast Fourier transform. The Journal
of Computational Finance 2(4), 6173.
Daniels, H. E. (1987). Tail probability approximations. International Statistical Review 55,
3748.
Dufe, D., Pan, J., and Singleton, K. (2000). Transform analysis and asset pricing for afne
jump diffusions. Econometrica 68, 13431376.
Eberlein, E., and Madan, D. (2009). Short positions, rally fears and option markets. Applied
Mathematical Finance, forthcoming.
Heston, S. (1993). A closed-form solution for options with stochastic volatility with applica-
tions to bond and currency options. Review of Financial Studies 6, 327343.
Jensen, J. L. (1995). Saddlepoint Approximations. Oxford University Press, Oxford.
Lugannini, R., and Rice, S. (1980). Saddlepoint approximations for the distribution
of the sum of independent random variables. Advances in Applied Probability 12,
475490.
Madan, D., Roynette, B., and Yor, M. (2008). Option prices as probabilities. Finance Research
Letters 5, 7987.
Merton, R. C. (1976). Option pricing when underlying stock returns are discontinuous.
Journal of Financial Economics 3, 125144.
Rogers, L. C. G., and Zane, O. (1999). Saddlepoint approximations to option prices. Annals
of Applied Probability 9, 493503.
Scott, L. O. (1997). Pricing stock options in jump-diffusion model with stochastic volatility
and interest rates: applications of Fourier inversion methods. Mathematical Finance 7,
413426.
Wood, A. T. A., Booth, J. G., and Butler, R. W. (1993). Saddlepoint approximations
with nonnormal limit distributions. Journal of the American Statistical Association 88,
680686.
Xiong, J., Wong, A., and Salopek, D. (2005). Saddlepoint approximations to option price in
a general equilibrium model. Statistics and Probability Letters 71, 361369.
Research Paper www.thejournalofcomputationalnance.com

2009 Incisive Media. Copying or distributing in print or electronic
forms without written permission of Incisive Media is prohibited.