PRICING AND HEDGING SPREAD OPTIONS IN A LOG-NORMAL MODEL

REN

E CARMONA AND VALDO DURRLEMAN

ABSTRACT. This paper deals with the pricing of spread options on the difference between correlated
log-normal underlying assets. We introduce a new pricing paradigm based on a set of precise lower
bounds. We also derive closed form formulae for the Greeks and other sensitivities of the prices. In
doing so we prove that the price of a spread option is a decreasing function of the correlation parameter,
and we analyze the notion of implied correlation. We use numerical experiments to provide an extensive
analysis of the performance of these new pricing and hedging algorithms, and we compare the results
with those of the existing methods.
1. INTRODUCTION
As Eric Reiner put it at a recent seminar, a multivariate version of the Black-Scholes formula
has yet to emerge. Nevertheless, there is a tremendous number of papers dealing with the issue.
Whether they look at basket options or discrete-time average Asian options, all try to price and hedge
multivariate contingent claims. In order to tackle this problem we start with the simplest case of
bivariate contingent claims, and more precisely of spread options. Spread options already contain
the essence of the difculty in pricing multivariate contingent claims: the linear combination of log-
normal underlying assets. To overcome this obstacle, we introduce a new pricing paradigm. Although
it is original, our work was partly inspired by the article by Rogers and Shi [9] on Asian options. Our
method has several advantages. It gives an extremely good approximation to the price, and computing
the so-called Greeks is straightforward. More importantly it encapsulates the univariate case, that is
to say, it is an extension of the classical results of Black-Scholes and Margrabe. This is certainly a
desirable feature from a mathematical point of view, but also from a risk management point of view.
Indeed, pricing and hedging a single very exotic product is often not the main issue for nancial
engineers, but aggregating different simple products as vanilla options together with basket options
for example is a daily challenge. Pricing and hedging them in a consistent way is the main issue
which our approach helps resolve.
In the next section, we introduce two examples of spread options: one in the equity markets and
the other in the xed income markets. These examples are intended to motivate the forthcoming
computations. The main thrust of the paper is contained in Section 3 where we derive closed form
formulae for our approximation and its error. Section 4 gives a short review of two existing methods
that are closely related to ours. We derive pricing formulae in these models for further comparison
with our methodology. Section 5 applies the results of Section 3 to the case of a spread between two
correlated stocks whose price dynamics are given by geometric Brownian motions, and we derive
closed formformulae for the so-called Greeks. Finally, to illustrate possible extensions of our method,
we show how to include jumps in the dynamics of the underlying assets, and we discuss the notion of
Date: March 16, 2003.
1
2 REN

E CARMONA AND VALDO DURRLEMAN

implied correlation. The accuracy of our approximate price formula, together with the performance
of the subsequent hedging strategies are illustrated by the results of extensive numerical experiments
based on Monte Carlo simulations from the geometric Brownian motion models.
2. TWO EXAMPLES OF SPREAD OPTIONS
In this section, we introduce the notation and the terminology which we use throughout the paper.
In so doing we describe two examples which should be used as a motivation for the need for the kind
of formulae we derive in the sequel.
2.1. Spread Options in the Equity or Energy Markets. We look at the classical setting where be-
sides a riskless bank account with constant interest rate r, our arbitrage-free market model comprises
two assets whose prices at time t are denoted by S
1
(t) and S
2
(t). We assume that their risk-neutral
price dynamics are given by the following stochastic differential equations:
dS
1
(t) = S
1
(t)[(r q
1
)dt +
1
dW
1
(t)]
dS
2
(t) = S
2
(t)[(r q
2
)dt +
2
dW
2
(t)]
where q
1
and q
2
are the instantaneous dividend yields, the volatilities
1
and
2
are positive constants
and W
1
and W
2
are two Brownian motions with correlation . The initial conditions will be denoted
by S
1
(0) = x
1
and S
2
(0) = x
2
. We mainly focus on the pricing of spread options on two stocks
in which case the q
i
s should be interpreted as dividend rates. In the case of a spread option on two
commodity spot prices, the q
i
s should be interpreted as instantaneous convenience yields, while we
should set q
1
= q
2
= r in the case of spread options on futures contracts.
In any case, the price p at time 0 of the spread option with date of maturity T and strike K is given
by the risk-neutral expectation:
(1) p = e
rT
E{(S
2
(T) S
1
(T) K)
+
}.
The special case K = 0 corresponds to an exchange since the pay-off (S
2
(T)S
1
(T))
+
will provide
the holder of the option with the difference S
2
(T) S
1
(T) when S
2
(T) > S
1
(T). So if one would
like to own either one of the indexes at time T, one could simply buy the second one and purchase the
above call spread option with K = 0. This will guarantee that, short of the premium of the option,
we do as well as if we had bought the one ending up being the cheaper of the two in all cases!
The price p of formula (1) can be rewritten in the form:
(2) p = e
rT
E
_
_
x
2
e
(rq
2

2
2
/2)T+
2
W
2
(T)
x
1
e
(rq
1

2
1
/2)T+
1
W
1
(T)
K
_
+
_
which shows that the price p is given by the integral of a function of two variables with respect to a
bivariate Gaussian distribution, namely the joint distribution of W
1
(T) and W
2
(T). Unfortunately,
except in the case of an exchange option (i.e., an option with strike K = 0), the price of the spread
option cannot be given by a formula in closed form.
PRICING AND HEDGING SPREAD OPTIONS 3
2.2. Spread Options in the Fixed-Income Markets. The results obtained in this paper can be ap-
plied in the context of xed-income derivative pricing. Consider a call option on the difference
between the 3-month and the 6-month Libor rates. More precisely if we let = 3 months, we denote
by L(T, ) the 3-month Libor rate spanning the interval [T, T + ], and similarly by L(T, 2) the
6-month Libor rate spanning [T, T + 2]. We consider the option that pays the amount
(L(T, ) 2L(T, 2) k)
+
.
at time T +. The price at time t is given by the conditional expectation with respect to the forward
pricing measure Q
T+
p
t
= B(t, T +)E
Q
T+
Ft
_
(L(T, ) 2L(T, 2) k)
+
_
where B(t, T) denotes the T-zero coupon bond price at time t, and where we use a superscript on the
expectation to emphasize the probability measure with respect to which the expectation is computed,
and a subscript to indicate that the expectation is in fact a conditional expectation with respect to the
-eld used as a subscript. The price p
t
rewrites:
p
t
= B(t, T +)E
Q
T+
Ft
_
(1 +L(T, ) (1 + 2L(T, 2)) k)
+
_
= B(t, T +)E
Q
T+
Ft
_
_
1
B(T, T +)

1
B(T, T + 2)
k
_
+
_
= B(t, T +)E
Q
T+
Ft
_
1
B(T, T +)
_
1
B(T, T +)
B(T, T + 2)
kB(T, T +)
_
+
_
= B(t, T)E
Q
T
Ft
_
_
1
B(T, T +)
B(T, T + 2)
kB(T, T +)
_
+
_
= B(t, T)E
Q
T
Ft
_
_
1
B
T
(T, T +)
B
T
(T, T + 2)
kB
T
(T, T +)
_
+
_
where B
t
(T, T +) = B(t, T +)/B(t, T) is the price at time t of the forward zero coupon bond. Let
us assume from now on that we are in a Gaussian Heath-Jarrow-Morton framework. This is merely
saying that the volatility of the zero coupon B(t, T), say v(t, T), is deterministic i.e. non-random. It
is well-known (see, e.g., [6]) that B
t
(T, T +) and B
t
(T, T + 2) are log-normal martingales under
Q
T
. More precisely, we have:
dB
t
(T, T +)
B
t
(T, T +)
=
t
(T, T +)dW
T
t
for some Q
T
-Brownian motion W
T
provided we set
t
(T, T

) = v(t, T) v(t, T

). Obviously, we
have a similar formula for the dynamics of B
t
(T, T +2). We can therefore infer the distributions of
4 REN

E CARMONA AND VALDO DURRLEMAN

the forward bonds under Q
T
since:
B
T
(T, T +) =
B(t, T +)
B(t, T)
exp
_

1
2
_
T
t

2
s
(T, T +)ds +
_
T
t

s
(T, T +)dW
T
s
_
B
T
(T, T +)
B
T
(T, T + 2)
=
B(t, T +)
B(t, T + 2)
exp
_

1
2
_
T
t

2
s
(T, T +)
2
s
(T, T + 2)ds
+
_
T
t

s
(T, T +)
s
(T, T + 2)dW
T
s
_
.
So if we restrict ourselves to k < 0, the computation of the price reduces to the problem of the pricing
of a spread option in the equity markets. This is indeed the case if we use the formulae

2
1
=
1
T t
_
T
t
(
s
(T, T +)
s
(T, T + 2))
2
ds

2
2
=
1
T t
_
T
t

2
s
(T, T +)ds
=
1
T t
_
T
t

s
(T, T +) (
s
(T, T +)
s
(T, T + 2)) ds.
for the volatilities and the correlation. The case k > 0 can equivalently be treated by means of
symmetry arguments as we derive below.
3. MAIN RESULT
The above pricing problems boil down to computing the following expectation:
(3) = (, , , , , ) = E
_
_
e
X
1

2
/2
e
X
2

2
/2

_
+
_
where , , , and are real constants and X
1
and X
2
are jointly Gaussian N(0, 1) random
variables with correlation . The purpose of this section is to compute this expectation. It can be
expressed as a double integral, but in general, its value cannot be given in closed form. Our goal is to
provide a very good approximation for its value, in such a way to retain all the advantages of closed
form formulae such as the Black Scholes formula.
Without any loss of generality we assume that
= 1 or =
which is equivalent to

2
2 +
2
= 0.
Indeed, is explicitly computable with the Black-Scholes formula whenever the above assumption
does not hold.
PRICING AND HEDGING SPREAD OPTIONS 5
3.1. Symmetries of the Problem. We rst study the symmetries properties of (3). The simple for-
mula:
(4) E{X
+
} = E{(X)
+
} +E{X}
becomes very useful when E{X} can be computed explicitly. In the present situation, we have:
E
_
_
e
X
1

2
/2
e
X
2

2
/2

_
+
_
= E
_
_
e
X
2

2
/2
e
X
1

2
/2
+
_
+
_
+
or equivalently,
(, , , , , ) = (, , , , , ) +
which can make some proofs simpler by reducing the parameter domain for which properties of the
expectation need to be derived.
Formulae derived from (4) will be called parity formulae, and arguments based on parity formulae
will be called parity arguments. This terminology is motivated by the standard call/put parity relating
the price of a European call option to the price of the European put option with the same strike and
expiry. Next, we introduce two independent N(0, 1) random variables Z
1
and Z
2
such that:
X
1
=
_
1
2
Z
1
+Z
2
X
2
= Z
2
and we denote by the unique number [0, ] such that = cos . With these notations, the
expectation giving can be rewritten as:
(5) = E
_
_
e
[sin Z
1
+cos Z
2
]
2
/2
e
Z
2

2
/2

_
+
_
.
Function (3) also enjoys symmetry properties as given by the following proposition.
Proposition 1. The following symmetry relations hold
(, , , , , ) =
_
, , , , ,

_
=
_
, , , , ,

_
where =
_

2
2 +
2
.
Proof. The assumption = or = 1 allows us to divide by and therefore to dene the new
correlations. Let us let

=
_
1
2
and let us write as follows.
(, , , , , ) = E
_
_
e

Z
1
+Z
2

2
/2
e
Z
2

2
/2

_
+
_
= E
_
e
Z
2

2
/2
_
e

Z
1
+()Z
2

2
/2+
2
/2
e
Z
2
+
2
/2
_
+
_
= E
_
_
e

Z
1
+()Z
2
(
2
2+
2
)/2
e
Z
2

2
/2

_
+
_
=
_
,
_

2
2 +
2
, , , ,

_

2
2 +
2
_
6 REN

E CARMONA AND VALDO DURRLEMAN

From the second to the third line, we used the Girsanovs transform. The second symmetry relation
is obtained exactly in the same way.
It is interesting to remark that the above symmetry relations give directly Margrabes formula [4].
Indeed:
(, , , , 0, ) =
_
, , 0, , ,

_
= E
_
_
e
U
2
/2

_
+
_
where U N(0, 1), and this last expectation is given by the Black-Scholes formula.
Symmetry arguments also allow us to price options on the sum of two assets:
E
_
_
e
X
1

2
/2
+e
X
2

2
/2

_
+
_
.
Indeed, if 0, the computation is straightforward. On the contrary if < 0, a symmetry argument
transforms the computation into that of a spread option.
We conclude this section by explaining how parity and symmetries interplay. Let us denote by n,
s and t the following involution functions
n(, , , , , ) = (, , , , , )
s(, , , , , ) =
_
,
_

2
2 +
2
, , , ,

_

2
2 +
2
_
t(, , , , , ) =
_
, , ,
_

2
2 +
2
, ,

_

2
2 +
2
_
Then one easily checks that:
(6) n s = t n.
3.2. The Approximation. We now derive our approximate pricing formula from a set of lower
bounds for the price.
A Preliminary Remark. Recall that the Black-Scholes formula for valuing call options is given by:
C
BS
= Se
qT
(d
0
+

T) Ke
rT
(d
0
)
where
d
0
=
ln(Se
qT
/Ke
rT
)

T
2
.
Here and throughout the paper, we use the notation (x) and (x) for the density and the cumulative
distribution function of the standard Gaussian N(0, 1) distribution, i.e.,
(x) =
1

2
e
x
2
/2
and (x) =
1

2
_
x

e
u
2
/2
du.
PRICING AND HEDGING SPREAD OPTIONS 7
C
BS
can be seen as a function of d
0
, the other parameters remaining xed:
C
BS
= C
BS
(d
0
)
One can then check (this will also be a simple consequence of Proposition 2 below) that the value d
0
is precisely that value of d at which the function C
BS
(d) is attains its maximum, so that:
C
BS
= C
BS
(d
0
) = sup
dR
C
BS
(d).
With this remark in mind, the delta (i.e., the derivative of C
BS
with respect to S) is very easily
computable, and:
=
C
BS
S
+
C
BS
d
d
0
S
= e
qT
(d
0
+

T)
As the reader can check, the brute force differentiation of the Black-Scholes formula requires more
than a single line of computation. Our approximation for the value (3) is based on this idea. We would
like to have an approximation that looks like a Black-Scholes formula. We introduce some degrees
of freedom (like the parameter d in the one-dimensional case) and the values of the parameters we
choose will be those that maximize the value of the approximation. Unlike the one-dimensional case,
we cannot recover the exact value of (3) (except in some very special cases) but the approximation will
turn out to be excellent. But like the one-dimensional case, we can easily compute the sensitivities of
(3) with respect to its various parameters.
The next result is elementary. We state it as a proposition for future reference.
Proposition 2. Let X be a real valued random variable on a probability space (, H, P), and let
A H be any family of events such that {X 0} A. Then,
E{X
+
} = sup
AA
E{X1
A
}
Proof. Let A A,
E{X
+
} =
_
A
X dP =
_
A{X0}
X dP +
_
A{X<0}
X dP

_
A{X0}
X dP

_
{X0}
X dP
= E{X
+
}
which proves that sup
AA
E{X1
A
} E{X
+
}. To get the reverse inequality, note that {X 0}
A so that
sup
AA
E{X1
A
} E{X1
{X0}
} = E{X
+
}.
8 REN

E CARMONA AND VALDO DURRLEMAN

Two Different Elementary Lower Bounds. Recall that we are interested in computing E{X
+
} where
X is the following random variable:
X = e
[sin Z
1
+cos Z
2
]
2
/2
e
Z
2

2
/2
.
The main obstruction to a closed formula for this expectation is that the exercise region (i.e. the
set {X 0}) does not have a simple shape. Proposition 2 tells us that any attempt to approximate
the value of this integral by approximating the exercise region by simpler regions will lead to a
lower bound. The two elementary lower bounds that we are about to introduce are based on simple
approximations of the exercise region.
For each R we introduce the random variable:
Y

= sinZ
1
cos Z
2
.
By choosing for A the -eld {Y

} generated by Y

, we get our rst elementary lower bound:

() = sup
A{Y

}
E{X1
A
}.
By choosing A =
_
{Y

d}; d R
_
, we get the second one:

() = sup
dR
E{X1
{Y

d}
}.
The approximation

which we propose is dened as the supremum of the lower bounds

() when
we vary the free parameter . We denote by

a maximizer of

().
(7)

=

(

) = sup
R

() = sup
R
sup
dR
E{X1
{Y

d}
}.
We will also need the approximation

corresponding to

(). Although their values are presumably
different, we will also denote by

a maximizer for that lower bound. The context in which we use

the notation

will always make it clear which one we have in mind, and no confusion will arise.
Proposition 3. We have:

:= sup
R

() = sup
A

{Y

}
_
A
X dP = sup
R
E{E{X|Y

}
+
}
Proof. Since {E{X|Y

} 0} {Y

}, Proposition 2 yields
sup

E{E{X|Y

}
+
} = sup

sup
A{Y

}
_
A
E{X|Y

} dP
By the denition of conditional expectations, we have
sup

sup
A{Y

}
_
A
E{X|Y

} dP = sup

sup
A{Y

}
_
A
X dP
= sup
A

{Y

}
_
A
X dP
=

PRICING AND HEDGING SPREAD OPTIONS 9

The above proposition shows that our approximations are very related to Jensens inequalities. The
function x x
+
being convex, Jensens inequality implies that:
E{X
+
} = E{E{X
+
|Y

}} E{E{X|Y

}
+
} =

().
In fact, we even have:
E{X}
+

E{X
+
}
The rst inequality is just a consequence of the fact that

lim
d
E{X1
{Y

d}
} = 0
and

lim
d+
E{X1
{Y

d}
} = E{X}.
The second one simply comes from the the trivial inclusion
_
{Y

d}; d R
_
{Y

}
and the last one has just been established.
Notice that the lower bounds are much simpler to handle than (3). After all,

() is nothing but a
one-dimensional Gaussian integral.
(8)

() = E{

X()
+
}
with:
(9)

X() = E{X|Y

} = e
cos(+)Y

2
cos
2
(+) /2
e
cos Y

2
cos
2
/2
.
The quantity

() can be computed a bit more explicitly.
Proposition 4.
(10)

() = sup
dR
[(d + cos( +)) (d + cos()) (d)]
We will also denote by d

a maximizer of the above supremum.

Proof. In view of (9),

() can be rewritten as:

() = sup
dR
_
{Y

d}
X dP
= sup
dR
_
{Y

d}
E{X|Y

} dP
= sup
dR
E
__
e
cos(+)Y

(
2
/2) cos
2
(+)
e
cos Y

(
2
/2) cos
2

_
1
{Y

d}
_
10 REN

E CARMONA AND VALDO DURRLEMAN

and

() consists in three terms which can be computed separately. The rst one gives:
E
_
e
cos(+)Y

(
2
/2) cos
2
(+)
1
{Y

d}
_
=
1

2
_
d

e
cos(+)x(
2
/2) cos
2
(+)
e
x
2
/2
dx
=
1

2
_
d+ cos(+)

e
u
2
/2
du
= (d + cos( +))
The other two terms can be computed similarly. This leads to the desired expression for

()
Explicit Computation of

. We are now in a position to explain how to compute

. In order to
maximize the function

(), we need its derivative. A formula for the latter is given in the statement
of the next result.
Proposition 5.

() = sin( +) (d

+ cos( +)) sin (d

+ cos )
Proof. Note that we have already established that

() = E
_

X()1
{Y

}
_
.
We compute the derivative of

() by differentiating both sides of this equation. The derivative of
the function x 1
{x0}
is a Dirac function which is equal to when x = 0 and 0 otherwise.
Consequently, taking derivative under the integral/expectation sign gives:
(11)
d
d
E
_

X()1
{Y

}
_
= E
_
d
d

X()1
{Y

}
_
= E
_
d

X()
d
1
{Y

}
+

X()
d1
{Y

}
d
_
.
The second term in the most right expression is zero since the Lebesgue measure has no atom. For-
mula (11) can easily be justied by convolving the function x 1
{x0}
with an approximate identity,
taking the derivative under the expectation sign, and removing the effect of the regularization by tak-
ing a limit controlled by the Lebesgues dominated convergence theorem. This gives:

() = E
__
sin( +)[ cos( +) +Y

]e
cos(+)Y

(
2
/2) cos
2
(+)
+ sin[ cos +Y

]e
cos Y

(
2
/2) cos
2

_
1
{Y

}
_
= sin( +)
_
d

+ cos(+)

u
e
u
2
/2

2
+ sin
_
d

+ cos

u
e
u
2
/2

2
= sin( +) (d

+ cos( +)) sin (d

+ cos )
which is the desired result.
PRICING AND HEDGING SPREAD OPTIONS 11
So computing explicitly

reduces to computing d

and

. To that end we need a system of two

equations. The equation

) = 0 gives a rst one. In view of Proposition 2, it is clear that the d

that achieves the maximum in (10) has to be a zero of the function:

e
cos(+) x
2
cos
2
(+) /2
e
cos x
2
cos
2
/2
.
This leads to a second equation. Therefore d

and

are necessarily solutions to the following system:

(12)
_
0 = e
cos(

+) d

2
cos
2
(

+) /2
e
cos

2
cos
2

/2

0 = sin(

+) e
cos(

+) d

2
cos
2
(

+) /2
sin

e
cos

2
cos
2

/2
.
From this system we get
e
cos(

+) d

2
cos
2
(

+) /2
=
sin

[ sin(

+) sin

]
e
cos

2
cos
2

/2
=
sin(

+)
[ sin(

+) sin

]
.
In solving for d

in each of the above equations, we see that

is necessarily a solution to:

1
cos
ln
_

sin( +)
[ sin( +) sin]
_

cos
2
(13)
=
1
cos( +)
ln
_

sin
[ sin( +) sin]
_

cos( +)
2
and d

is equal to either the left or the right hand side of equation (13).
We could be more precise and bracket the solution

that corresponds to the maximum. Unfortu-

nately, we would need to distinguish several cases. Since we have not found a nice way of putting the
result, we shall refrain from doing so.
Proposition 6. Let

be the solution of (13) corresponding to the maximum. Let

d

=
1
cos(

)
ln
_
sin(

+)
sin

1
2
( cos(

+) + cos

)
where
= arccos
_

_
,
then
(14)

= (d

+ cos(

+)) (d

+ cos

) (d

)
Special Cases. We now look at which cases our approximations are in fact the true value. We also
explain why in the other cases, our approximation is strictly less than the true value. In view of
Proposition 2, we only need look for the cases where {X 0}

or {X 0}
_
{Y

d}; d R
_
.
Proposition 7.

= whenever = 0, or = 0, or = 0, or = +1 or = 1.
12 REN

E CARMONA AND VALDO DURRLEMAN

Proof. When = 0,
{X 0} =
_
e
Z
2

2
/2
0
_
Z
2
= {Y

}
when = 0,
{X 0} =
_
e
[sin Z
1
+cos Z
2
]
2
/2
0
_
{Y

}
when = 0,
{X 0} =
_
e
[sin Z
1
+cos Z
2
]
2
/2
e
Z
2

2
/2
0
_
=
_
sinZ
1
+ ( cos )Z
2
ln(/) +
2
/2
2
/2
_
{Y

}
for = arccos
_

_
, when = 1,
{X 0} =
_
e
Z
2

2
/2
e
Z
2

2
/2
0
_
{Y

}.
In any other case, the exercise region {X 0} will exhibit some curvature and the lower bounds
will therefore be strict lower bounds.
Proposition 8.

= whenever = 0, or = 0, or = 0, or = 1, or
= +1 and ( ) > 0
or
= +1 and ( ) < 0 and
_

( )
e
(+)/2
Proof. By simple inspection of the proof of the last proposition, we see that except in the case = 1,
{X 0} will be either of the form {Y

x} or of the form {Y

x}. The later can be rewritten as

{Y
+
x} which is of the desired form. The only problem can arise when = 1 and when the
function
e
x
2
/2
e
x
2
/2

has two zeroes. This gives the conditions of the proposition.

Numerical evidence suggests that the lower bound

is extremely accurate. We shall illustrate this
claim later in the paper. They are several cases where this approximation is indeed the true price as
shown in the Proposition 8.
3.3. Upper Bound and Control of the Error. In order to control the error of the approximation
provided by

, we derive upper bounds for . In view of Proposition 2, it is easy to get lower bounds
for the expectation, but much more difcult to get good upper bounds. Indeed, trying to simplify the
exercise region can only lead to lower bounds.
PRICING AND HEDGING SPREAD OPTIONS 13
Elementary Upper Bound. Since the standard Gaussian distribution is rotation-invariant, for every
R we can replace the couple (Z
1
, Z
2
) by the couple (cos Z
1
+ sinZ
2
, sinZ
1
cos Z
2
)
without changing the distribution of X.
(15) = E
_
_
e
[sin(+)Z
1
+cos(+)Z
2
]
2
/2
e
[sin Z
1
+cos Z
2
]
2
/2

_
+
_
Recall that we have a nice expression for

() given in Proposition 3. We see that for each R,
we have:

() = E{X
+
} E{E{X|Y

}
+
}
= E{X
+
E{X|Y

}
+
}
E{(X E{X|Y

})
+
}.
We used the elementary inequality (a +b)
+
a
+
+b
+
in the last step. This last expression rewrites
E
__
e
cos(+)Z
2
(
2
/2) cos
2
(+)
_
e
sin(+)Z
1
(
2
/2) sin
2
(+)
1
_
e
cos Z
2
(
2
/2) cos
2

_
e
sin Z
1
(
2
/2) sin
2

1
__
+
_
.
Using again the elementary inequality (a + b)
+
a
+
+ b
+
and the independence of Z
1
and Z
2
, we
get:
E
_
_
e
sin(+)Z
1
(
2
/2) sin
2
(+)
1
_
+
_
+E
_
_
e
sin Z
1
(
2
/2) sin
2

1
_
+
_
=
_

_
| sin( +)|
2
_

| sin( +)|
2
__
+
_

_
| sin|
2
_

| sin|
2
__
.
The one-dimensional Gaussian integral computation is similar to the derivation of the Black-Scholes
formula. This upper bound can also be optimized over . Let be the following normalized Black-
Scholes function
(x) = (|x/2|) (|x/2|) = |(x/2) (x/2)| , x R.
We have the following upper bound.
Proposition 9.
min
_

(0) +( sin);

() +( sin)
_
It remains to explain how we compute

(). This is done in the following proposition.
Proposition 10. For each R, the lower bound

() is given by:

() =
1
2
( ) +
_
[(d
2
+ cos( +)) (d
1
+ cos( +))]
[(d
2
+ cos ) (d
1
+ cos )] [(d
2
) (d
1
)]
_
where, d
1
and d
2
are the zeroes of the function x D(x) dened by:
(16) D(x) = e
cos(+)x(
2
/2) cos
2
(+)
e
cos x(
2
/2) cos
2

.
14 REN

E CARMONA AND VALDO DURRLEMAN

when the latter has two zeroes, with the convention that d
1
= d
2
= when the equation D(x) = 0
has no solution, and d
2
= when the equation D(x) = 0 has exactly one solution d
1
. = 1 is
related to the sign of D between its zeroes.
Proof. Since:

() = E{D(Y

)
+
} = E{D(Y

)1
{D(Y

)>0}
}
the rst task is to properly describe the set E = E() dened by:
E = {x R; D(x) > 0}.
The function x D(x) is continuously differentiable on R and may have 0, 1 or 2 zeroes. Let us
denote by d
1
the unique zero when the function D has exactly one zero, and by d
1
< d
2
the values of
the two zeroes when they exist. Therefore, E can be any of the following six sets.
, (d
1
, d
2
), (d
1
, +), (, d
1
), (, d
1
) (d
2
, +), R.
A third constant is dened by setting = +1 in the rst three cases, and = 1 in the remaining
cases. Without any loss of generality we can restrict ourselves to the case = +1, because the case
= 1 can be derived from the case = 1 and the parity formula (4).
With our convention on the values of d
1
and d
2
, we always have E = (d
1
, d
2
) in the rst three
cases we are concentrating on, i.e., when = 1. Consequently the denition (10), (9) of

() can be
rewritten as:

() = E
__
e
cos(+)Y

(
2
/2) cos
2
(+)
e
cos Y

(
2
/2) cos
2

_
1
{d
1
<Y

<d
2
}
_
.
As

(),

() consists in three terms which can be computed separately. The rst one gives:
E
_
e
cos(+)Z
2
(
2
/2) cos
2
(+)
1
{d
1
<Y

<d
2
}
_
=
1

2
_
d
2
d
1
e
cos(+)x(
2
/2) cos
2
(+)
e
x
2
/2
dx
=
1

2
_
d
2
+ cos(+)
d
1
+ cos(+)
e
u
2
/2
du
= [(d
2
+ cos( +)) (d
1
+ cos( +))]
The other two terms can be computed similarly. This leads to the following expression for

():

() = [(d
2
+ cos( +)) (d
1
+ cos( +))]
[(d
2
+ cos ) (d
1
+ cos )]
[(d
2
) (d
1
)]
which is the desired result when = +1. The case = 1 is obtained similarly, the fact that the set
E is now the complementary set of what it was before accounts for the changes in the formula. The
proof of the proposition is now complete.
PRICING AND HEDGING SPREAD OPTIONS 15
Improvement of the Upper Bound via Symmetry Arguments. The upper bound of Proposition 9 can
be improved. As we can see in this proposition, when = 0, we have

(0)

min
_

(0);

() +( sin)
_

(0)
The upper bound is equal to the lower bound and hence is the true value. The same is true when = 0
or = 1 (i.e., sin = 0). However, in the case = 0, it does not give us that the lower bound is in
fact exact. The previous symmetry relations will allow us to get a better upper bound that will equal
the lower bound also when = 0.
This leads to the following improvement of our upper bound.
Proposition 11.
min
_
min
_

(0); s(0)
_
+(

); min
_

(); t(0)
_
+(

);
min
_
s(); t( )
_
+||
_

_
_
Improvement of the Upper Bound via Monotonicity Arguments. It turns out that the previous upper
bound is not always decreasing in whereas the true price is. Obviously, the largest decreasing
function below our upper bound will be a better upper bound. The fact that (3) is decreasing with
respect to seems to be known but we did not nd a proof of it in the nancial literature. This is the
topic of the next proposition.
Proposition 12. () is decreasing on [1, 1].
Proof. Let us denote by f the bivariate standard Gaussian density with correlation :
f(, u, v) =
exp
_

1
2
u
2
2uv+v
2
1
2
_
2
_
1
2
We will use the following identity:
f

=

2
f
uv
.
Let us rewrite using f.
=
_
R
2
_
e
u
2
/2
e
v
2
/2

_
+
f(, u, v) dudv
=
_
R
2
(e
x
e
y
)
+
f
_
,
x ln +
2
/2

,
y ln +
2
/2

_
dxdy

16 REN

E CARMONA AND VALDO DURRLEMAN

Since f is C
1
((1, 1) R R), we can use the Lebesgues differentiation theorem to compute the
derivative of with respect to .

=
_
R
2
(e
x
e
y
)
+
f

_
,
x ln +
2
/2

,
y ln +
2
/2

_
dxdy

=
_
R
2
(e
x
e
y
)
+

2
f
uv
_
,
x ln +
2
/2

,
y ln +
2
/2

_
dxdy

=

2

_
R
2
(e
x
e
y
)
+
f
_
,
x ln +
2
/2

,
y ln +
2
/2

_
dxdy

=

2

_
R
2
_
e
u
2
/2
e
v
2
/2

_
+
f(, u, v) dudv
To show that this last expression is always negative, notice that it is the limit as h and k go to zero of
1
hk
_
R
2
(g( +h, +k) g(, +k) g( +h, ) +g(, )) f(, u, v) dudv
where g(, ) =
_
e
u
2
/2
e
v
2
/2

_
+
. Note that for any reals x and h > k > 0, we have
1
hk
_
(x +h k)
+
(x +h)
+
(x k)
+
+x
+
_
=
1
hk
_
k1
{x+hk}
(x +h)1
{0x+hk}
+k1
{xk}
+x1
{0xk}
_
=
1
hk
_
k
_
1
{xk}
1
{x+hk}
_
+x1
{0xk}
(x +h)1
{0x+hk}
_

1
hk
_
k
_
1
{xk}
1
{x+hk}
_
+x1
{0xk}
_

1
h
_
1
{xk}
1
{x+hk}
+1
{0xk}
_
=
1
h
_
1
{x0}
1
{xkh}
_
0.
The fact that h > k was only used at the last line. This implies that the quantity we are studying as
k > h > 0 go to zero is always negative and since the limit exists by the previous argument, we have

0.
Numerical examples show that this improved upper bound is quite accurate (about 10%). However
this upper bound does not catch the extreme accuracy of the lower bound as the numerical experiments
that we ran (see below) suggest.
4. OTHER ANALYTICAL APPROXIMATIONS
There have already existed attempts to give an efcient analytical approximation of . We present
two of them: the Bacheliers model and the Kirks model, both of them have been considered by
markets practitioners.
PRICING AND HEDGING SPREAD OPTIONS 17
4.1. The Bacheliers Model. In most typical applications, all the underlying indexes are modelled
by means of log-normal distributions or at least, exponential transformations of standard distributions.
This is motivated in part by the inherent positivity of asset prices. But the positivity restriction does
not apply to the spreads themselves, since the latter can be negative as differences of positive quanti-
ties. Indeed, computing histograms of historical spread values shows that the marginal distribution of
a spread at a given time extends on both tails, and surprisingly enough, that the Gaussian distribution
can give a reasonable t. This simple remark is the starting point of a series of papers proposing to use
arithmetic Brownian motion (as opposed to the geometric Brownian motion leading to the log-normal
distribution of the indexes) for the dynamics of spreads. In so doing, prices of options can be derived
by computing simple Gaussian integrals leading to simple closed form formulae. It was originally
advocated by Shimko in the early nineties. See [7] for a detailed expos e of this method.
If we approximate the distribution of the spread
S = e
X
1

2
/2
e
X
2

2
/2
by the Gaussian distribution, the least we can ask is that it matches the rst two moments. Therefore:
S N(E{S}, var{S})
and classical computations give
E{S} =
var{S} =
2
_
e

2
1
_
2
_
e

1
_
+
2
_
e

2
1
_
.
Proposition 13. If the value of the spread at maturity is assumed to have the Gaussian distribution,
the Bacheliers approximation
B
is given by:
(17)
B
= ( )
_

B
_
+
B

B
_
where we used the notation:

B
=
_

2
_
e

2
1
_
2 (e

1) +
2
_
e

2
1
_
Proof. Plainly,

B
= E{(S )
+
}
= E{( +
B
)
+
}
for some Gsn(0, 1) random variable . Consequently:

B
=
1

2
_

B
( +
B
u)e
u
2
/2
du
from which we easily get the desired result.
Based on Edgeworth expansions, Jarrow and Rudd [2] improved the Bacheliers model in taking
into account higher order moments (skew and kurtosis). We refer to Mbafeno [5] for the formula
giving an approximate price for a spread option.
18 REN

E CARMONA AND VALDO DURRLEMAN

4.2. The Kirks Model. More recently, Kirk in [3] proposed the formula

K
=
_
_
ln
_

+
_

K
+

K
2
_
_
( +)
_
_
ln
_

+
_

K


K
2
_
_
where

K
=

2
2

+
+
2
_

+
_
2
As the reader can immediately see, the approximation is exact when = 0 and when = 0. However
unlike our lower bound, it cannot be exact when neither = 1 nor when = 0. The numerical
approximation of the price is very good but, as we will see later, the resulting hedging strategy does
not always perform very well.
5. PRICING OPTIONS ON THE SPREAD OF GEOMETRIC BROWNIAN MOTIONS
5.1. Pricing Formula. In this section we apply the results obtained in section 3 to the case of a
option on the difference of two assets. Recall formula (2). This expectation is a particular case of the
expectations computed before provided we set
(18) = x
2
e
q
2
T
=
2

T = x
1
e
q
1
T
=
1

T and = Ke
rT
.
So for each R, the number p() =

() is a lower bound for the price p. Following the discussion
of Section 3 we introduce the approximation p given by the supremum of the lower bounds p() when
we vary the free parameter .
(19) p = sup
R
p()
According to Proposition 6 we have:
Proposition 14. Let

be the solution of (13) corresponding to the maximum. Let

d

=
1
cos(

T
ln
_
x
2
e
q
2
T

2
sin(

+)
x
1
e
q
1
T

1
sin

1
2
(
2
cos(

+) +
1
cos

T
then
(20) p = x
2
e
q
2
T

_
d

+
2
cos(

+)

T
_
x
1
e
q
1
T

_
d

+
1
sin

T
_
Ke
rT
(d

)
Note that this formula is as close to the Black-Scholes formula as we could hope. We will see
later that they have many features in common.
Proposition 8 takes the form:
Proposition 15. The approximation p is equal to the true price p when K = 0, or x
1
= 0, or x
2
= 0,
or = 1. In particular, p is given by Margrabes formula when K = 0, and by the classical
Black-Scholes formula when x
1
x
2
= 0.
PRICING AND HEDGING SPREAD OPTIONS 19
For the cases of equality when = +1, we refer to Proposition 8.
Proof. There is nothing to be proven but we show how we recover Margrabes formula in the case
K = 0. Since it encompasses the Black-Scholes formula, this will also show how we deal with case
x
1
x
2
= 0. In order to do that, we take the

that we computed in the proof of proposition 7, that is

= + = + arccos
_

_
.
where is now
=
_

2
1
2
1

2
+
2
2
We easily compute that

2
sin(

+) =
1
sin

so that
d

=
1

T
ln
_
x
2
e
q
2
T
x
1
e
q
1
T
_

1
2
(
2
cos(

+) +
1
cos

T.
Furthermore

2
cos(

+)
1
cos

=
hence,
d

+
2
cos(

+)

T =
1

T
ln
_
x
2
e
q
2
T
x
1
e
q
1
T
_
+

T
2
d

+
1
cos

T =
1

T
ln
_
x
2
e
q
2
T
x
1
e
q
1
T
_

T
2
.
Finally, we get
p = x
2
e
q
2
T

_
1

T
ln
_
x
2
e
q
2
T
x
1
e
q
1
T
_
+

T
2
_
x
1
e
q
1
T

_
1

T
ln
_
x
2
e
q
2
T
x
1
e
q
1
T
_

T
2
_
which is Margrabes formula, see [4].
5.2. Hedging and the Computation of the Greeks.
First Order Derivatives. Exactly as in the case with a single asset, our derivation gives as a side effect
a sub-hedge for the option.
Proposition 16. Holding at each time t T

1
= e
q
1
T

_
d

+
1
cos

T
_
and

2
= e
q
2
T

_
d

+
2
cos(

+)

T
_
of the underlying assets will provide a sub-hedge for the option (i.e., such a portfolio gives a sub-
replication of the pay-off of the option.)
20 REN

E CARMONA AND VALDO DURRLEMAN

Proof. We need to compute d p/dx
1
and d p/dx
2
. Note that the price functional depends on x
2
through
both x
2
and

. Therefore
d p
dx
2
=
p
x
2
+
p

x
2
but since we take as the price the optimal lower bound p/ = 0 at

. The remaining term is

computed easily:
p
x
2
=

x
2
E
__
x
2
e
q
2
T
2
cos(

+)W(T)
2
2
cos
2
(

+)T/2
x
1
e
q
1
T
1
cos

W(T)
2
1
cos
2

T/2
Ke
rT
_
1
{W(T)d

}
_
= e
q
2
T
E
_
e

2
cos(

+)W(T)
2
2
cos
2
(

+)T/2
1
{W(T)d

}
_
= e
q
2
T

_
d

+
2
cos(

+)

T
_
The computation is similar for p/x
1
.
In the same manner we can get other partial derivatives of the price with respect to various inter-
esting parameters. These are the so-called Greeks of the nancial literature. We give some of them in
the following proposition.
Proposition 17. Let
1
and
2
denote the sensitivities of the price functional (20) with respect to the
volatilities of each asset, be the sensitivity with respect to their correlation parameter , be the
sensitivity with respect to the strike price K and be that with respect to the maturity time T.

1
= x
1
e
q
1
T

_
d

+
1
cos

T
_
cos

2
= x
2
e
q
2
T

_
d

+
2
cos(

+)

T
_
cos(

+)

T
= x
1
e
q
1
T

_
d

+
1
cos

T
_

1
sin

sin

T
= (d

) e
rT
=

1

1
+
2

2
2T
q
1
x
1

1
q
2
x
2

2
rK
Proof. These formulae are derived exactly in the same way as we did for the deltas. Again, the key
observation is that
p

) = 0
so that we only need differentiate as if

were constant. The formula for comes from the following

homogeneity property:
p(x
1
, x
2
,
1
,
2
, q
1
, q
2
, K, r, T) = p(x
1
e
q
1
T
, x
2
e
q
2
T
,
1

T,
2

T, 0, 0, Ke
rT
, 0, 1).
We can express it using the Greeks we have just computed.
PRICING AND HEDGING SPREAD OPTIONS 21
Second Order Derivatives. The second order derivatives are also of fundamental importance. Let

11
,
12
and
22
be the second derivatives with respect to the current values of the assets. The
simplication that arose for the rst order derivatives does not hold anymore. Indeed,
d
2
p
dx
2
1
=

2
p
x
2
1
+ 2

2
p
x
1

x
1
+

2
p

2
_

x
1
_
2
=

2
p
x
2
1

2
p

2
_

x
1
_
2
+ 2
_

2
p
x
1

+

2
p

x
1
_

x
1
=

2
p
x
2
1

2
p

2
_

x
1
_
2
+ 2
d
dx
1
_
p

x
1
=

2
p
x
2
1

2
p

2
_

x
1
_
2
.
However, we propose to approximate the second order derivative by the corresponding partial deriv-
ative. This is supported by different observations. The rst one is that in case the approximation is
the true price, the corresponding will also be exact since in those case,

does not depend neither

on x
1
nor on x
2
. The second observation is that

does not vary very much with respect to those

parameters. Finally, we have the surprising fact that this approximation on the s exactly balances
the approximation on p so that the Greeks satisfy the Black-Scholes equation!
Instead of explaining how we computed the s, we give them and show that they satisfy the
Black-Scholes equation.
Proposition 18. Suppose

11
=
e
2q
1
T

2
+
1

_
d

+
1
cos

T
_
2

12
=
e
(q
1
+q
2
)T

2
+
1

_
d

+
1
cos

T
_

_
d

+
2
cos(

+)

T
_

22
=
e
2q
2
T

2
+
1

_
d

+
2
cos(

+)

T
_
2
,
then
+
1
2

2
1
x
2
1

11
+
1

2
x
1
x
2

12
+
1
2

2
2
x
2
2

22
+ (r q
1
)x
1

1
+ (r q
2
)x
2

2
r p = 0
Proof. Let us denote
a
1
=
1
cos

T
a
2
=
2
cos(

+)

T,
22 REN

E CARMONA AND VALDO DURRLEMAN

we also let x
1
= x
1
e
q
1
T
and x
2
= x
2
e
q
2
T
. According to Proposition 17 and equation (20),
+
1
2

2
1
x
2
1

11
+
1

2
x
1
x
2

12
+
1
2

2
2
x
2
2

22
+ (r q
1
)x
1

1
+ (r q
2
)x
2

2
r p
=
1
2

2
1
x
2
1

11
+
1

2
x
1
x
2

12
+
1
2

2
2
x
2
2

22

1
+
2

2
2T
=
1
2T(
1

1
+
2

2
)
_

2
1
x
2
1
T(d

+a
1
)
2
2
1

2
x
1
x
2
T(d

+a
1
)(d

+a
2
)
+
2
2
x
2
2
T(d

+a
2
)
2
a
2
1
x
2
1
(d

+a
1
)
2
+ 2a
1
a
2
x
1
x
2
(d

+a
2
)(d

+a
1
)
a
2
2
x
2
2
(d

+a
2
)
2
_
=
1
2(
1

1
+
2

2
)
_

2
1
sin
2

x
2
1
(d

+a
1
)
2
+
2
2
sin
2
(

+) x
2
2
(d

+a
2
)
2
2
1

2
(cos cos(

+) cos

) x
1
x
2
(d

+a
1
)(d

+a
2
)
_
=
[
1
sin

x
1
(d

+a
1
)
2
sin(

+) x
2
(d

+a
2
)]
2
2(
1

1
+
2

2
)
= 0
The last equality comes from equation (12).
Since these formulae are the true achievement of our method, we plotted two of them:
1
and in
Figure 1. As Garman pointed out in [1], it is worth noticing that spread options exhibit the unusual
feature of negative vegas. Indeed it is not very common that the price of a derivative product decreases
when the volatility of the underlying increases. Here is an heuristic argument for that: imagine that
the two assets are highly correlated so that the value of the spread is very likely to stay (for instance)
out of the money. The corresponding price is expected to be low. Suppose
1
decreases, the spread
increases its variance, the spread increases its probability of being in the money and therefore the
price of the option increases. The reader can also check that 0. This phenomenon has the same
rationale, as the correlation increases, the variance of the spread decreases.
6. NUMERICAL PERFORMANCE
6.1. Approximation Error. In order to illustrate the accuracy of our approximation, we present the
result of a Monte Carlo analysis in the case of the geometric Brownian motion. We x the values of
the parameters of the marginal dynamical equations according to Table 1, and we vary the values of
the correlation coefcient and of the strike K. Results are reported in Table 2. As we can see the
agreement is excellent. This strongly supports our formula, the price to pay is extremely cheap since
we only need to compute numerically a zero of a given function. This is done very efciently by a
Newton-Raphson method. The computation time is immediate.
We plot the relative error between our approximation and a Monte Carlo simulation against the
volatility parameters and against strike/correlation in Figure 2. On the right panel, we clearly see
the bias of our lower bound when both volatilities are very high. On the contrary, on the left panel,
we observe an increase of the relative error as K becomes large but the peaks are either negative or
PRICING AND HEDGING SPREAD OPTIONS 23
FIGURE 1. Left: Price sensitivity with respect to
1
for different volatilities
1
and

2
; here = 70%. Right: Price sensitivity with respect to the correlation for different
strikes and correlations; here
1
= 20% and
2
= 10%.
positive, which means that we see the error due to Monte-Carlo and not the bias of our approximation.
Asset 1 Asset 2
x 100 110
q 2% 3%
15% 10%
TABLE 1. Model data together with r = 5% and T = 1.
6.2. Comparisons with Bacheliers and Kirks Approximations. Let us now compare our for-
mula with these two other models. As we have already pointed out, the main interesting feature of
our model comes from the easy computation of the Greeks which in turn gives sensible hedging port-
folios. The Kirks formula (and also the Bacheliers formula) also gives rise to two deltas and we can
therefore see how the two hedging strategies perform along a given path (or scenario.)
To this end, we estimate the standard deviation of the tracking error for both models. What we
call tracking error is the difference between the payoff of the option at maturity and the value of
the discretely re-balanced replicating portfolio. In Figure 3, we plot the standard deviation of the
tracking error against the number of re-hedging times in log-scales. Those are typical examples:
in most cases the agreement between Kirks model and the lower bound is very good but in some
cases our lower bound performs much better. In any cases, these two models clearly outperform the
Bacheliers model. The reason is that the Bacheliers model is a one-factor model that cannot capture
the whole structure of the true two-factor model we are dealing with.
24 REN

E CARMONA AND VALDO DURRLEMAN

K
-1 -0.5 0 0.3 0.8 1
-20
29.656
29.653
29.656
28.994
28.995
29.442
28.381
28.381
28.773
28.070
28.069
28.311
27.770
27.770
27.790
27.754
27.754
27.754
-10
21.868
21.867
21.868
20.904
20.907
21.129
19.888
19.892
20.200
19.270
19.271
19.516
18.381
18.382
18.431
18.244
18.244
18.244
0
15.133
15.133
15.133
13.917
13.914
13.917
12.523
12.525
12.523
11.561
11.559
11.561
9.632
9.632
9.632
8.821
8.821
8.821
5
12.244
12.242
12.244
10.956
10.953
11.068
9.445
9.442
9.601
8.367
8.366
8.542
5.967
5.968
6.148
4.454
4.454
4.454
15
7.521
7.521
7.521
6.242
6.243
6.574
4.744
4.743
5.202
3.679
3.678
4.188
1.342
1.344
1.858
0.049
0.049
0.049
25
4.201
4.203
4.201
3.129
3.129
3.680
1.961
1.962
2.718
1.219
1.220
2.062
0.103
0.105
0.926
0.000
0.000
0.000
TABLE 2. The number appearing in italic in the center of each box is the result of
a stratied Monte-Carlo computation with 100,000 trials, the number on top is our
lower bound approximation (20), the improved upper bound appearing at the bottom.
We used bold faces when the lower bound and the upper bound are equal to the true
value. The values of the parameters used for these runs are given in Table 1.
Note also that from these plots it is impossible to see that the hedging strategy induced by our
formula is merely sub-replicating!
7. IMPLIED CORRELATION
7.1. A Tractable Jump-Diffusion Model. We are now going to show that our previous result can t
in the more general case of a jump-diffusion process. The process we are looking at is often referred
as the Mertons jump model in the nancial literature. In the same spirit, we allow for jumps in the
risk-neutral dynamics of S
1
and S
2
.
(21)
dS
i
(t)
S
i
(t)
= (r q
i

0
)dt +
i
dW
i
(t) + (e
J
i
(t)
1)dN
i
(t) + (e
J
0
(t)
1)dN
0
(t)
where N
0
, N
1
and N
2
are three independent Poisson processes with intensity
0
,
1
and
2
. They
are also independent of W
2
and W
1
. (J
i
(t))
t0,i=0,1,2
is a sequence of independent Gaussian random
variables (m
i
, s
2
i
). We will often need the following quantities

i
= e
m
i
+s
2
i
/2
1.
PRICING AND HEDGING SPREAD OPTIONS 25
FIGURE 2. Relative errors between our approximation (20) and Monte-Carlo. Data
are in given in Table 1.
FIGURE 3. Behavior of the tracking error as the number of re-hedging times in-
creases. The model data are x
1
= 100, x
2
= 110,
1
= 10%,
2
= 15% and T = 1.
= 0.9, K = 30 (left) and = 0.6, K = 20 (right).
We introduced three Poisson processes for the sake of generality but we shall consider two special
cases. The rst case is
1
=
2
= 0. This case would be more appropriate in the context of
equity markets where the jumps (generally downwards) account for the sudden moves of the whole
market. In such a case the jump component can be taken as being the same for each stock in a rst
approximation. A second case of interest is
0
= 0. This is more realistic in the case of spark-spread
options in the energy markets where the two components of the spread are gas and electricity future
26 REN

E CARMONA AND VALDO DURRLEMAN

contracts. They come from very different markets and the sudden jumps in the price dynamics have
no reason for being correlated and we can take them independent.
By a simple use of It os formula, we get the integrated price dynamics
S
i
(T) = x
i
exp
_
_
(r q
i

2
i
/2
0

i
)T +
i
W
i
(T) +
N
0
(T)

k=1
J
0
(k) +
N
i
(T)

k=1
J
i
(k)
_
_
Given N
0
(T), N
1
(T) and N
2
(T), S
1
(T) and S
2
(T) will still have the log-normal distribution so
that our lower bound can be computed.
Proposition 19. If we denote by p(x
1
, x
2
,
1
,
2
, ) our lower bound computed with log-normal
distributions, the price of the spread option in Mertons model is
(22) p
jumps
= e
(
1
+
2
+
3
)T

i=0

j=0

k=0
(
1
T)
i
(
2
T)
j
(
3
T)
k
i!j!k!
p ( x
1
, x
2
,
1
,
2
, )
with
x
1
= x
1
e
(
0

0
+
1

1
)T+i(m
1
+s
2
1
/2)+k(m
0
+s
2
0
/2)
x
2
= x
2
e
(
0

0
+
2

2
)T+j(m
2
+s
2
2
/2)+k(m
0
+s
2
0
/2)

1
=
_

2
1
+ (is
2
1
+ks
2
0
)/T

2
=
_

2
2
+ (js
2
2
+ks
2
0
)/T
=

1

2
+ks
2
0
/T
_

2
1
+ (is
2
1
+ks
2
0
)/T
_

2
2
+ (js
2
2
+ks
2
0
)/T
Proof. Formula (22) is obtained by conditioning on the numbers of jumps N
0
(T), N
1
(T) and N
2
(T)
prior to T and by computing simple Gaussian integrals to derive the values of the parameters x
1
, x
2
,

1
,
2
, and to use in (22).
7.2. Implied correlation. As a simple illustration the tractability of our method, we compute the
implied correlation in this model. It is well-known in the one-dimensional case that jumps in the
stock price dynamics give rise to the so-called volatility smile. In other words the volatilities that
need to be plugged in in the Black-Scholes formula to give the observed prices vary when strike or
time-to-maturity change. In our case, we compute the price of spread options for different strikes
and maturities when the underlying model is a jump-diffusion one. Then, we look for the correlation
parameter that would give the same price. Note that we have here to decide with which volatility
parameters our trader is inverting our formula. Since we have assumed that he is pricing spread
options without knowing that the underlying process has jumps, it is not fair to assume that he has
access to the parameters
1
and
2
. However, we can assume that he has observed the stock prices
and that he can estimate the variance of their log-return based on these data. In the case when there
were no jumps, this would lead to an unbiased estimator of the volatility. He will therefore take

i
=
_

2
i
+
0
(s
2
0
+m
2
0
) +
i
(s
2
i
+m
2
i
)
as inputs in his pricing formula.
PRICING AND HEDGING SPREAD OPTIONS 27
We plot the implied volatility in Figure 4 and the implied correlation 5 for a given set of data
(Table 3). It is interesting to note that the two cases we are considering lead to two radically different
correlation smiles.
Independent jumps have the tendency to decrease the implied correlation for out-of-the-money
options. At-the-money options have a relatively constant implied correlation. Note that since the
two jump processes we are adding are independent and independent of the Brownian motions, they
tend to decrease the correlation between the two assets. Let us now look at the term structure of the
implied correlation. It is increasing in maturity for both negative strikes and positive strikes whereas
it is relatively constant for small strikes.
On the other hand, simultaneous jumps tend to increase the implied correlation for out-of-the-
money options. The implied correlation for at-the-money options is extremely at is it should. The
implied correlation also exhibit a very different term structure behavior as it is decreasing for both
negative strikes and positive strikes. Like in the previous case it attens for long maturities.
The volatility and correlation structure is consistent with that found by Mbafeno in [5] in the case
of independent jumps, which is reasonable to assume here as we have already argued. The implied
correlation is computed there, following Shimko, via
=

2
spread

2
1,imp

2
2,imp
2
1,imp

2,imp
where
spread
is the volatility of the spread (computed through the historical variance of the spread)
and
1,imp
,
2,imp
are the implied volatilities of each of the assets. With the data given in Table 3,
such an implied correlation for a spread option with maturity T and strike K would be given by
(K, T) =

2
spread

2
1,imp
(x
2
K, T)
2
2,imp
(x
1
+K, T)
2
1,imp
(x
2
K, T)
2,imp
(x
1
+K, T)
.
As we can see in Figure 6, this clearly underestimates the implied correlation (in a jump-diffusion
world) but has a similar time-to-maturity structure for deep in- or out-of-the-money spread options.
Asset 1 Asset 2
x 100 100
15% 10%
0.2 0.01
m -0.04 -0.03
s
2
0.01 0.01
15.75% 10.05%
TABLE 3. Model data together with = 70%, r = 5% and
0
= 0.
8. CONCLUSION
This paper discussed a new pricing paradigm for spread options. The originality of this article
relies on the introduction of this new algorithm to compute accurate prices for options written on the
difference of two underlying indexes. We compared the performance of our algorithm to some of the
28 REN

E CARMONA AND VALDO DURRLEMAN

FIGURE 4. Implied volatility smile for a call option on the rst asset (left) and on
the second one (right.) The data are given in Table 3.
FIGURE 5. On the left panel: implied correlation smile for a spread option in the
case of perfectly correlated jumps. The data are given in Table 3. On the right panel
is the case where the jumps are independent. In that case
0
= 0.2, m
0
= 0.1 and
s
0
= 0.01.
popular existing methods, both from a numerical and an analytical points of view. The advantages of
our method are clear: its implementation is easy, run time is short, and its results are very accurate.
Moreover, along with the price we can easily compute all the sensitivities, so that hedges and expo-
sures are precisely known. In fact, all the hedging techniques derived from the Black-Scholes for-
mula can be implemented in our framework. Finally, unlike with the other approximations (whether
they are numerical or analytical) we always know on which side of the true price we are.
PRICING AND HEDGING SPREAD OPTIONS 29
FIGURE 6. Comparison of Shimkos implied correlation with our jump-diffusion
implied correlation in case of a positive strike (left) and a negative one (right).
As an added bonus, our formulae lead to efcient computions of implied quantities such as volatili-
ties and correlation so that one is now able to build more complex models (stochastic volatility and/or
correlation, jumps). These improved models will hopefully better t the market reality. We believe
that a numerical study of the implied correlation based on our results can lead to a better understand-
ing of this quantity.
Finally, we would like to mention some possible extensions of our results. Our method can clearly
be generalized to the case of an option on a linear combination of assets (basket, rainbow options
. . .) or any linear combination of prices of a single asset at different times (discrete-time average
Asian option for example.) Such extensions would provide efcient algorithms to compute prices and
hedges for these options.
REFERENCES
[1] M. Garman (1992): Spread the Load, Risk, London.
[2] R. Jarrow and A. Rudd (1982): Approximate option valuation for arbitrary stochastic processes. J. Financial Econom-
ics, 10, 347-369.
[3] E. Kirk (1995): Correlation in the Energy Markets, in Managing Energy Price Risk. London: Risk Publications and
Enron.
[4] W. Margrabe (1978): The value of an option to exchange one asset for another. The Journal of Finance, 33, 177-186.
[5] A. Mbafeno (1997): Co-movement term structure and the valuation of energy spread options, in: M. Dempster and S.
Pliska (eds), Mathematics of Derivative Securities, Cambridge University Press.
[6] M. Musiela and M. Rutkowski (1997): Martingale methods in nancial modelling, Springer-Verlag, Berlin-
Heidelberg.
[7] G. Poitras (1998): Spread Options, Exchange Options, and Arithmetic Brownian Motion. The Journal of Futures
Markets, 18, 487-517.
[8] R. Rebonato (2001): Volatility and Correlation in the pricing of Equity, FX and Interest-Rate Options. John Wiley &
Sons.
[9] L. C. G. Rogers and Z. Shi (1995): The value of an Asian option. Journal of Applied Probability, 32, 1077-1088.
30 REN

E CARMONA AND VALDO DURRLEMAN

DEPARTMENT OF OPERATIONS RESEARCH AND FINANCIAL ENGINGEERING, PRINCETON UNIVERSITY, PRINCE-
TON, NJ 08544, ALSO WITH THE BENDHEIM CENTER FOR FINANCE AND THE APPLIED AND COMPUTATIONAL
MATHEMATICS PROGRAM.
E-mail address: rcarmona@princeton.edu
DEPARTMENT OF OPERATIONS RESEARCH AND FINANCIAL ENGINGEERING, PRINCETON UNIVERSITY, PRINCE-
TON, NJ 08544.
E-mail address: vdurrlem@princeton.edu