Financial modeling with Lévy processes

Peter Tankov
Centre de Mathématiques Appliquées
Ecole Polytechnique
Email: peter.tankov@polytechnique.org
Lecture notes

Contents
1 Introduction 2

2 Lévy processes: basic facts 3

2.1 The Poisson process . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Poisson random measures . . . . . . . . . . . . . . . . . . . . 8

3 Path structure of a Lévy process 11

4 Basic stochastic calculus for jump processes 15

4.1 Integrands et integrators . . . . . . . . . . . . . . . . . . . . . 15
4.2 Stochastic integral with respect to a Poisson random measure 18
4.3 Change of variable formula for Lévy-Itô processes . . . . . . . 21

5 Stochastic exponential of a jump process 23

6 Exponential Lévy models 29

7 The Esscher transform and absence of arbitrage in exponen-

tial Lévy models 31
7.1 Measure changes for Lévy processes . . . . . . . . . . . . . . . 31
7.2 One-dimensional models . . . . . . . . . . . . . . . . . . . . . 33
7.3 Multidimensional models . . . . . . . . . . . . . . . . . . . . . 35

1
8 European options in exp-Lévy models 39
8.1 Numerical Fourier inversion . . . . . . . . . . . . . . . . . . . 42

9 Integro-differential equations for exotic options 43

10 Gap options 45
10.1 Single asset gap options . . . . . . . . . . . . . . . . . . . . . 47
10.2 Multi-asset gap options and Lévy copulas . . . . . . . . . . . . 50

11 Implied volatility 58
11.1 Large/small strikes . . . . . . . . . . . . . . . . . . . . . . . . 60
11.2 Short maturity asymptotics . . . . . . . . . . . . . . . . . . . 61
11.3 Flattening of the smile far from maturity . . . . . . . . . . . . 67

12 Hedging in exponential Lévy models 71

12.1 Quadratic hedging in exponential-Lévy models under the mar-
tingale measure . . . . . . . . . . . . . . . . . . . . . . . . . . 74
12.2 Quadratic hedging in exponential Lévy models under the his-
torical measure . . . . . . . . . . . . . . . . . . . . . . . . . . 80

13 Calibration of exp-Lévy models 87

14 Limits and extensions of Lévy processes 91

1 Introduction
Exponential Lévy models generalize the classical Black and Scholes setup
by allowing the stock prices to jump while preserving the independence and
stationarity of returns. There are ample reasons for introducing jumps in
financial modeling. First of all, asset prices do jump, and some risks sim-
ply cannot be handled within continuous-path models. Second, the well-
documented phenomenon of implied volatility smile in option markets shows
that the risk-neutral returns are non-gaussian and leptokurtic. While the
smile itself can be explained within a model with continuous paths, the fact
that it becomes much more pronounced for short maturities is a clear indica-
tion of the presence of jumps. In continuous-path models, the law of returns
for shorter maturities becomes closer to the Gaussian law, whereas in real-
ity and in models with jumps returns actually become less Gaussian as the

2
horizon becomes shorter. Finally, jump processes correspond to genuinely
incomplete markets, whereas all continuous-path models are either complete
or ’completable’ with a small number of additional assets. This fundamental
incompleteness makes it possible to carry out a rigorous analysis of the hedg-
ing error and find ways to improve the hedging performance using additional
instruments such as liquid European options.
A great advantage of exponential Lévy models is their mathematical
tractability, which makes it possible to perform many computations explic-
itly and to present deep results of modern mathematical finance in a simple
manner. This has led to an explosion of the literature on option pricing
and hedging in exponential Lévy models in the late 90s and early 2000s, the
literature which now contains hundreds of research papers and several mono-
graphs. However, some fundamental aspects such as asymptotic behavior of
implied volatility or the computation of hedge ratios have only recently been
given a rigorous treatment.
For background on exponential Lévy models, the reader may refer to
textbooks such as [20, 60] for a more financial perspective or [3, 44] for a
more mathematical perspective.

2 Lévy processes: basic facts

Lévy processes are a class of stochastic processes with discontinuous paths,
which is at the same time simple enough to study and rich enough for appli-
cations, or at least to be used as building blocks of more realistic models.

Definition 1. A stochastic process X is a Lévy process if it is càdlàg, satisfies

X0 = 0 and possesses the following properties:

• Independent increments;

• Stationary increments;

From these properties it follows that

• X is continuous in probability: ∀ε, lims→0 P [|Xs+t − Xt | > ε] = 0.

• At any fixed time, the probability of having a jump is zero: ∀t, P [Xt− =
Xt ] = 1.

3
 Lévy processes are essentially processes with jumps, because it can be
shown that any Lévy process which has a.s.continuous trajectories is a Brow-
nian motion with drift.

Proposition 1. Let X be a continuous Lévy process. Then there exist γ ∈ Rd

and a symmetric positive definite matrix A such that

Xt = γt + Wt ,

where W is a Brownian motion with covariance matrix A.

Proof. This result is a consequence of the Feller-Lévy central limit theorem,

but since it is important for the understanding of Lévy processes, we give
here a short proof (for the one-dimensional case).
It is enough to show that X1 has Gaussian law, the rest will follow from
the stationarity and independence of increments.

Step 1 Let ξnk := X k − X k−1 . The continuity of X implies that

n n

lim P [sup |ξnk | > ε] = 0,

n k

for all ε. Let an = P [|ξn1 | > ε]. Since

P [sup |ξnk | > ε] = 1 − (1 − P [|ξn1 | > ε])n ,

k

we get that lim(1 − an )n = 1, from which it follows that lim n log(1 − an ) = 0.

n n
But n log(1 − an ) ≤ −nan ≤ 0. Therefore,

lim nP [|X 1 | > ε] = 0 (1)

n n

Step 2 Using the independence and stationarity of increments, we can show

that
1
lim nE[cos X 1 − 1] = {log EeiX1 + log Ee−iX1 } := −A; (2)
n n 2
1
lim nE[sin X 1 ] = {log EeiX1 − log Ee−iX1 } := γ. (3)
n n 2i

4
Step 3 The equations (1) and (2) allow to prove that for every function f
such that f (x) = o(|x|2 ) in a neighborhood of 0, limn nE[f (X 1 )] = 0 which
n
implies that ε > 0

lim nE[X 1 1|X 1 |≤ε ] = γ, (4)

n n n

lim nE[X 21 1|X 1 |≤ε ] = A, (5)

n n n

lim nE[|X 1 |3 1|X 1 |≤ε ] = 0. (6)

n n n

(7)

Step 4 Assembling together the different equations, we finally get

iuX 1
log E[eiuX1 ] = n log E[e n 1X 1 ≤ε ] + o(1)
n

u2
E[X 21 1X 1 ≤ε ] + o(1/n)} + o(1)
= n log{1 + iuE[X 1 1X 1 ≤ε ] −
n n 2 n n

Au2 Au2
= iuγ − + o(1) −−−→ iuγ −
2 n→∞ 2
where o(1) denotes a quantity which tends to 0 as n → ∞.
The second fundamental example of Lévy process is the Poisson process.

2.1 The Poisson process

Definition 2. Let (τi )i≥1 be
Pa sequence of exponential random variables with
parameter λ and let Tn = ni=1 τi . Then the process
X
Nt = 1t≥Tn (8)
n≥1

is called the Poisson process with parameter (or intensity) λ

Proposition 2 (Properties of the Poisson process).

1. For all t ≥ 0, the sum in (8) is finite a.s.

2. The trajectories of N are piecewise constant with jumps of size 1 only.

3. The trajectories are càdlàg.

5
 4. ∀t > 0, Nt− = Nt with probability 1.

5. ∀t > 0, Nt follows the Poisson law with parameter λt:

(λt)n
P [Nt = n] = e−λt
n!

6. The characteristic function of the Poisson process is

E[eiuNt ] = exp{λt(eiu − 1)}.

7. The Poisson process is a Lévy process

The Poisson process counts the events with exponential interarrival times.
In a more general setting, one speaks of a counting process.
Definition 3. Let (Tn ) be a sequence of times with Tn → ∞ a.s. Then the
process X
Nt = 1t≥Tn
n≥1

is called a counting process.

In other words, a counting process is an increasing piecewise constant
process with jumps of size 1 only and almost surely finite.
The first step towards the characterization of Lévy processes is to char-
acterize Lévy processes which are counting processes.
Proposition 3. Let (Nt ) be a Lévy process and a counting process. Then
(Nt ) is a Poisson process.
Proof. The proof uses the characterization of the exponential distribution by
its memoryless property: if a random variable T satisfies

P [T > t + s|T > t] = P [T > s]

for all t, s > 0 then T has exponential distribution.

Let T1 be the first jump time of the process N . The independence and
stationarity of increments give us:

P [T1 > t + s|T1 > t] = P [Nt+s = 0|Nt = 0]

= P [Nt+s − Nt = 0|Nt = 0] = P [Ns = 0] = P [T1 > s],

6
which means that the first jump time T1 has exponential distribution.
Now, it suffices to show that the process (XT1 +t − XT1 )t≥0 is independent
from T1 and has the same law as (Xt )t≥0 . Let f (t) := E[eiuXt ]. Then using
once again the independence and stationarity of increments we get that f (t+
iuX
s) = f (t)f (s) and Mt := ef (t)t is a martingale. Let T1n := n ∧ T1 . Then by
Doob’s optional sampling theorem,

f (T1n + t) ivT1n
 
iu(XT n +t −XT n )+ivT1n n
E[e 1 1 ]=E n
e = E[eiuXt ]E[eivT1 ].
f (T1 )

The proof is finished with an application of the dominated convergence the-

orem.

Compound Poisson process The Poisson process itself cannot be used

to model asset prices because the condition that the jump size is always equal
to 1 is too restrictive, but it can be used as building block to construct richer
models.

Definition 4 (Compound Poisson process). The compound Poisson process

with jump intensity λ and jump size distribution µ is a stochastic process
(Xt )t≥0 defined by
XNt
Xt = Yi ,
i=1

where {Yi }i≥1 is a sequence of independent random variables with law µ and
N is a Poisson process with intensity λ independent from {Yi }i≥1 .

In other words, a compound Poisson process is a piecewise constant pro-

cess which jumps at jump times of a standard Poisson process and whose
jump sizes are i.i.d. random variables with a given law.

Proposition 4 (Properties of the compound Poisson process). Let (Xt )t≥0 be

a compound Poisson process with jump intensity λ and jump size distribution
µ. Then X is a piecewise constant Lévy process and its characteristic function
is given by  Z 
E[eiuXt ] = exp tλ (eiux − 1)µ(dx) .
R

7
Example 1 (Merton’s model). The Merton (1976) model is one of the first
applications of jump processes in financial modeling. In this model, to take
into account price discontinuities, one adds Gaussian jumps to the log-price.
Nt
X
St = S0 ert+Xt , Xt = γt + σWt + Yi , Yi ∼ N (µ, δ 2 ) independents.
i=1

The advantage of this choice of jump size distribution is to have a series

representation for the density of the log-price (as well as for the prices of
European options).
n 2
o
∞ (λt)k exp − (x−γt−kµ)
X 2 2
2(σ t+kδ )
pt (x) = e−λt p .
k! 2π(σ 2 t + kδ 2 )
k=0

2.2 Poisson random measures

The notion of the Poisson random measure is central for the theory of Lévy
processes: we shall use it in the next section to give a full characterization
of their path structure.

Definition 5 (Random measure). Let (Ω, P, F) be a probability space and

(E, E) a measurable space. Then M : Ω × E → R is a random measure if

• For every ω ∈ Ω, M (ω, ·) is a measure on E.

• For every A ∈ E, M (·, A) is measurable.

Definition 6 (Poisson random measure). Let (Ω, P, F) be a probability

space, (E, E) a measurable space and µ a measure on (E, E). Then M :
Ω × E → R is a Poisson random measure with intensity µ if

• For all A ∈ E with µ(A) < ∞, M (A) follows the Poisson law with
parameter E[M (A)] = µ(A).

• For any disjoint sets A1 , . . . An , M (A1 ), . . . , M (An ) are independent.

In particular, the Poisson random measure is a positive integer-valued

random measure. It can be constructed as the counting measure of randomly
scattered points, as shown by the following proposition.

8
Proposition 5. Let µ be a σ-finite measure on a measurable subset E of Rd .
Then there exists a Poisson random measure on E with intensity µ.
Proof. Suppose first that µ(E) < ∞. Let {Xi }i≥1 be a sequence of inde-
µ(A)
pendent random variables such that P [Xi ∈ A] = µ(E) , ∀i and ∀A ∈ B(E),
and let M (E) be a Poisson random variable with intensity µ(E) independent
from {Xi }i≥1 . It is then easy to show that the random measure M defined
by
M (E)
X
M (A) := 1A (Xi ), ∀A ∈ B(E),
i=1

is a Poisson random measure on E with intensity µ.

Assume now that µ(E) = ∞, and choose a sequence
S of disjoint measur-
able sets {Ei }i≥1 such that µ(Ei ) < ∞, ∀i and i Ei = E. We construct a
Poisson random measure Mi on each Ei as described above and define
∞
X
M (A) := Mi (A), ∀A ∈ B(E).
i=1

Corollary 1 (Exponential formula). Let M be a Poisson random measure

Ron (E, E) with intensity µ, B ∈ E and let f be a measurable function with
f (x)
B
|e − 1|µ(dx) < ∞. Then
h R i Z 
f (x)M (dx) f (x)
E eB = exp (e − 1)µ(dx) .
B

Definition 7 (Jump measure). Let X be a Rd -valued càdlàg process. The

jump measure of X is a random measure on B([0, ∞) × Rd ) defined by

JX (A) = #{t : ∆Xt 6= 0 and (t, ∆Xt ) ∈ A}.

The jump measure of a set of the form [s, t] × A counts the number of
jumps of X between s and t such that their sizes fall into A. For a counting
process, since the jump size is always equal to 1, the jump measure can be
seen as a random measure on [0, ∞).
Proposition 6. Let X be a Poisson process with intensity λ. Then JX is a
Poisson random measure on [0, ∞) with intensity λ × dt.

9
 Maybe the most important result of the theory of Lévy processes is that
the jump measure of a general Lévy process is also a Poisson random measure.
Exercise 1. Let X and Y be two independent Lévy processes. Use the
definition to show that X + Y is also a Lévy process.
Exercise 2. Show that the memoryless property characterizes the exponen-
tial distribution: if a random variable T satisfies
∀t, s > 0, P [T > t + s|T > t] = P [T > s]
then either T ≡ 0 or T has exponential law.
Exercise 3. Prove that if N is a Poisson process then it is a Lévy process.
Exercise 4. Prove that if N and N 0 are independent Poisson processes with
parameters λ and λ0 then N + N 0 is a Poisson process with parameter λ + λ0 .
Exercise 5. Let X be a compound Poisson process with jump size distribu-
tion µ. Establish that
R
• E[|Xt |] < ∞ if and only if R |x|f (dx) and in this case
Z
E[Xt ] = λt xf (dx).
R
R
• E[|Xt |2 ] < ∞ if and only if x2 f (dx) and in this case
R
Z
Var[Xt ] = λt x2 f (dx).
R
R
• E[eXt ] < ∞ if and only if ex f (dx) and in this case
R
 Z 
Xt x
E[e ] = exp λt (e − 1)f (dx) .
R

Exercise 6. The goal is to show that to construct a Poisson random measure

on R, one needs to take two Poisson processes and make the first one run
towards +∞ and the second one towards −∞.
Let N and N 0 be two Poisson processes with intensity λ, and let M be a
random measure defined by
M (A) = #{t > 0 : t ∈ A, ∆Nt = 1} + #{t > 0 : −t ∈ A, ∆Nt0 = 1}.
Show that M is a Poisson random measure with intensity λ.

10
3 Path structure of a Lévy process
Definition 8 (Lévy measure). Let X be an Rd -valued Lévy process. The
measure ν defined by

ν(A) = E[#{t ∈ [0, 1] : ∆Xt 6= 0 et ∆Xt ∈ A}], A ∈ B(Rd )

is called the Lévy measure of X.

Theorem 1 (Lévy-Itô-decomposition). Let X be an Rd -valued Lévy process

with Lévy measure ν. Then

1. The jump measure JX of X is a Poisson random measure on [0, ∞)×Rd

with intensity dt × ν.
R
2. The Lévy measure ν satisfies Rd (kxk2 ∧ 1)ν(dx) < ∞.

3. There exist γ ∈ Rd and a d-dimensional Brownian motion B with co-

variance matrix A such that

Xt = γt + Bt + Nt + Mt , where (9)
Z
Nt = xJX (ds × dx) and
|x|>1,s∈[0,t]
Z
Mt = x{JX (ds × dx) − ν(dx)ds}
0<|x|≤1,s∈[0,t]
Z
≡ xJ˜X (ds × dx).
0<|x|≤1,s∈[0,t]

The three terms are independent and the convergence in the last term is
almost sure and uniform in t on compacts.

The triple (A, ν, γ) is called the characteristic triple of X.

The proof is based on the following lemma.

Lemma 1. Let (X, Y ) be a 2-dimensional Lévy process such that Y is piece-

wise constant and ∆Xt ∆Yt = 0 for all t a.s. Then X and Y are independent.

11
Proof. In view of the independence and stationarity of increments, it is
eiuXt
enough to show that X1 and Y1 are independent. Let Mt = E[e iuXt ] and
iuY
e t
Nt = E[e iuYt ] . Then M and N are martingales on [0, 1]. From the indepen-

dence and stationarity of increments we deduce that for every Lévy process
Z,
E[eiuZt ] = E[eiuZ1 ]t and E[eiuZ1 ] 6= 0, ∀u.
This means that M is bounded. By Proposition 3, the number of jumps of Y
on [0, 1] is a Poisson random variable. Therefore, N has integrable variation
on this interval. By the martingale property and dominated convergence we
finally get
" n #
X
E[M1 N1 ] − 1 = E (Mi/n − M(i−1)/n )(Ni/n − N(i−1)/n )
"i=1 #
X
→E ∆Mt ∆Nt = 0,
0≤t≤1

which implies E[eiuX1 +ivY1 ] = E[eiuX1 ]E[eivY1 ].

Proof of Theorem 1.

Part 1 Let A ∈ B(Rd ) with 0 ∈ / Ā. Then NtA = #{s ≤ t : ∆Xs ∈

A} is a counting process and a Lévy process, hence, by Proposition 3, a
Poisson process, which means that JX ([t1 , t2 ] × A) follows the Poisson law
with parameter (t2 − t1 )ν(A) and that JX ([t1 , t2 ] × A) is independent from
JX ([s1 , s2 ] × A) if t2 ≤ s1 . Let us now take two disjoint sets A and B. By
Lemma 1, N A and N B are independent, which proves that JX ([s1 , s2 ] × A)
and JX ([t1 , t2 ] × B) are also independent, for all s1 , s2 , t1 , t2 .

Part 2 From the previous part, we deduce that ν(A) < ∞ whenever 0 ∈
/ Ā.
It remains to show that
Z
kxk2 ν(dx) < ∞
kxk≤δ

for some δ > 0. Let

1≥|∆Xs |> Z
X
Xtε = ∆Xs = xJX (ds × dx)
0≤s≤t
ε<|x|≤1,s∈[0,t]

12
and Rtε = Xt − Xtε . Since (Xtε , Rtε ) is a Lévy process, Lemma 1 implies that
Xtε and Rtε are independent. In addition, |E[eiuXt ]| > 0 for all t, u. This
means that
ε ε
E[eiuXt ] = E[eiuRt ]E[eiuXt ].
ε
Therefore, |E[eiuXt ]| is bounded from below by a positive number which does
not depend on ε. By the exponential formula, this is equivalent to
  Z 
iux
exp t (e − 1)ν(dx)  ≥ C > 0,
 
|x|≥ε
R
which gives |x|≥ε (1 − cos(ux))ν(dx) ≤ C̃ < ∞. Since this result is true for
all u, the proof of part 2 is completed.

Part 3 Observe first that the process M is well defined due to the com-
pensation of small jumps and the fact that the Lévy measure integrates kxk2
near zero: introducing the process
Z
ε
Mt = xJ˜X (ds × dx),
ε≤kxk<1,s∈[0,t]

we get that for ε1 < ε2 ,

Z
E[(Mtε1 − Mtε2 )2 ] =t x2 ν(dx)
ε2 ≤kxk<ε1

and so, since the space L is complete, for every t, Mtε converges in L2 as
2

ε → 0. Using Doob’s inequality we show that the convergence is uniform in

t on compact intervals. The process X − N − M is then a continuous Lévy
process independent from N and M in view of Lemma 1. We conclude with
Proposition 1.
Corollary 2 (Lévy-Khintchine representation). Let X be a Lévy process with
characteristic triple (A, ν, γ). Its characteristic function is given by
Au2
  Z 
iuXt iux
E[e ] = exp t iγu − + (e − 1 − iux1|x|≤1 )ν(dx) , (10)
2 R

Proof. Using the previous theorem and the exponential formula, we get
Au2
  Z 
iu(γt+Bt +Nt +Mtε ) iux
E[e ] = exp t iγu − + (e − 1 − iux1|x|≤1 )ν(dx) ,
2 |x|≥ε

and we conclude using the dominated convergence theorem.

13
Example 2 (The variance gamma process). One of the simplest examples
of Lévy processes with infinite intensity of jumps is the gamma process, a
process with stationary independent increments and such that for all t, the
law pt of Xt is the gamma law with parameters λ and ct:
λct ct−1 −λx
pt (x) = x e .
Γ(ct)
The gamma process is an increasing Lévy process whose characteristic func-
tion has a very simple form:

E[eiuXt ] = (1 − iu/λ)−ct .

One can easily show that the Lévy measure of the gamma process has a
density given by

ce−λx
ν(x) = 1x>0 . (11)
x
Starting from the gamma process, we can construct a very popular jump
model: the variance gamma process [50, 48] which is obtained by changing
the time scale of a Brownian motion with drift with a gamma process:

Yt = µXt + σBXt .

Using Yt to model the logarithm of the stock price is usually justified by

saying that the price follows a geometric Brownian motion on a stochastic
time scale given by the gamma process [34]. The variance gamma process
provides another example of a Lévy processes with infinite intensity of jumps,
and its characteristic function is given by
−κt
κσ 2 u2

iuYt
E[e ] = 1 + − iµκu .
2
The parameters have the following intuitive interpretation: σ is the scale
parameter, µ is the parameter of asymmetry (skewness) and κ is responsible
for the kurtosis of the process (thickness of its tails).
Exercise 7. Let X be a Lévy process with characteristic triple (A, ν, γ).
Compute the probability that X will have at least one negative jump of size
bigger than ε > 0 on the interval [0, T ].

14
Exercise 8. Let X be a Lévy process with Lévy measure ν(dx) = λν0 (dx),
R ∞ and satisfies ν0 (R) = ∞. For all n ∈ N, let kn > 0
where ν0 has no atom
be the solution of kn ν0 (dx) = n. For a fixed T , give the law of the random
variable
An = #{t ≤ T : ∆Xt ∈ [kn+1 , kn ]}.
Use this result to suggest a method for estimating λ from an observation of
the trajectory of X on [0, T ], supposing that ν0 is known.

Exercise 9. Let X be a Lévy Rprocess with no diffusion component and a

Lévy measure ν which satisfies R |x|ν(dx) < ∞. Using the Levy-Ito decom-
position, show that the trajectories of X have a.s. finite variation (a function
has finite variation if it can be represented as the difference of two increasing
functions).

Exercise 10. Prove that the Lévy measure of the gamma process is given by
equation (11). Show that the variance gamma process can be represented as
a difference of two independent gamma process, and use this result to deduce
the form of the Lévy measure of the variance gamma process.

4 Basic stochastic calculus for jump processes

4.1 Integrands et integrators
The main application of the stochastic integral in finance is the representation
of self-financing portfolios: in the absence of interest rates, when the price
of the risky asset is a continuous process S, and the quantity of the asset is
denoted by φ, the portfolio value is
Z T
VT = φt dSt
0

We would like this relationship to hold in the presence of jumps as well,

but what are the natural properties to impose on S and φ? The process
S must be right-continuous, since the price jumps arrive inexpectedly. On
the other hand, the hedging strategy φt is based on the observations of the
portfolio manager up to date t; it must therefore be left-continuous. The
following example illustrates this: suppose that the asset price is given by
St = λt − Nt , where Nt is a Poisson process with intensity λ, and let T

15
be the time of the first jump of N . If one could use the (càdlàg) strategy
φt = 1[0,T ) (t) amounting to sell the asset just before the jump, there would
clearly be an arbitrage opportunity, since
Z t
Vt = φt dSt = λt ∧ T.
0

On the other hand, with the càglàd strategy φt = 1[0,T ] (t),

Z t
Vt = φt dSt = λt ∧ T − Nt∧T ,
0

which has zero expectation. It is therefore natural to consider adapted and

left-continuous integrands.
The simplest (and the only one which can be realized in practice) form
of a portfolio strategy is such where the portfolio is only rebalanced a finite
number of times. We define a simple predictable process by
n
X
φt = φ0 1t=0 + φi 1(Ti ,Ti+1 ] (t), (12)
i=0

where T0 = 0, (Ti )i≥0 is a sequence of stopping times, and for each i, φi is

FTi -measurable and bounded. The space of simple predictable processes will
be denoted by S.
For simple predictable processes, the stochastic integral is defined by
Z t n
X
φs dSs := φi (STi+1 ∧t − STi ∧t ) (13)
0 i=0

For a general adapted left-continuous process, the stochastic integral can

then be defined as the continuous extension of the integral for simple pre-
dictable processes, using the topology of uniform convergence on compacts
in probability (ucp).
The sequence of processes (X n ) is said to converge ucp to the process X if
for every t, (X n −X)∗t converges to 0 in probability, where Zt∗ := sup0≤s≤t |Zs |.
We denote by Sucp the space S endowed with the topology of ucp convergence
and by Lucp and Ducp the space of adapted and, respectively, left or right
continuous processes, with the same topology. It is then possible to show that
the space Sucp is dense in Lucp , and to associate the ucp topology with a metric

16
on Ducp , for which this space will be complete. To extend the stochastic
integration operator defined by (13) from Sucp to Lucp , this operator must
be continuous as a mapping from Sucp to Ducp . Whether or not this is true,
depends on the integrator S, and we shall limit ourselves to the integrators
for which this property holds.
Definition 9. The process S ∈ D is a semimartingale if the stochastic
integration operator defined by (13) is a continuous operator from Sucp to
Ducp .
Every adapted càdlàg process of finite variation on compacts is a semi-
martingale. This follows from
Z t 
 
sup  φs dSs  ≤ VarT0 (S) sup φt ,

0≤t≤T 0 0≤t≤T

where VarT0 (S) denotes the total variation of S on [0, T ]. A square integrable
càdlàg martingale is a semimartingale. For a simple predictable
R· process φ
of the form (12), and a square integrable martingale M , 0 φt dMt is also a
martingale and
Z T 2
E φt dMt ≤ sup φ2t E[MT2 ].
0 0≤t≤T,ω∈Ω
n
Suppose now that (φ ) is a sequence of simple predictable processes such that
φn → 0 ucp. Then, using the Chebyshev’s inequality and Doob’s inequality,
we get that
Z · ∗  Z · ∗ 
P n
φt dMt > ε ≤ P φt 1|φnt |≤C dMt > ε + P [(φn )∗ > C]
n
0 0
4C 2 E[MT2 ]
≤ + P [(φn )∗ > C] → 0,
ε2
because the first term can be made arbitrarily small by choosing C sufficiently
small, and the second term can be made small by choosing n sufficiently large.
Since the terms γt and Nt in the Lévy-Itô decomposition have finite vari-
ation and the terms Bt and Mt are square integrable martingales, every Lévy
process is a semimartingale.
A deep result of the general theory of processes [53] is that every semi-
martingale is the sum of a finite variation process and a local martingale.
The notion of local martingale extends that of the martingale: the process
(Xt ) is a local martingale if there exists a sequence of stopping times {Ti }i≥1
such that Ti → ∞ when i → ∞ and for each i, (XTi ∧t ) is a martingale.

17
4.2 Stochastic integral with respect to a Poisson ran-
dom measure
In the Lévy-Itô decomposition, we have already encountered integrals of de-
terministic functions with respect to Poisson random measures and compen-
sated Poisson random measures. In this section, our goal is to extend this
notion of integral to stochastic integrands.
Let M be a Poisson random measure of [0, T ] × R with intensity µ. µ is
supposed to be σ-finite: there exists a sequence Un ↑ R with µ([0, t]×Un ) < ∞
for all t. Typically, M will be the jump measure of a Lévy process. We would
like to define the integral of M or its compensated version with respect to a
predictable function φ : Ω × [0, T ] × R → R, that is, a function which satisfies
(i) For all t, (ω, x) 7→ φ(ω, t, x) is Ft × B(R)-measurable.
(ii) For all (ω, x) t 7→ φ(ω, t, x) is left-continuous.
The stochastic integral of φ with respect to M will be defined in two different
settings:
Case 1: φ satisfies
Z TZ
|φ(t, y)|M (dt × dy) < ∞ p.s.
0 R
P
In this case, the stochastic integral of φ with respect to M = δ(Ti ,yi ) is
defined as the absolutely convergent sum
Z tZ X
φ(t, y)M (dt × dy) := φ(Ti , yi ).
0 R i:Ti ≤t

Case 2: φ is square integrable, that is, it satisfies

Z TZ
E φ2 (t, y)µ(dt × dy) < ∞
0 R

In this case, the construction is more involved, since we need to use the L2
theory and continuous extension once again. We define simple predictable
functions φ : Ω × [0, T ] × R → R via
m
X n X
X m
φ(t, y) = φ0j 1t=0 1Aj (y) + φij 1(Ti ,Ti+1 ] (t)1Aj (y),
j=1 i=1 j=1

18
where T0 = 0, (Ti )i≥1 is a sequence of stopping times; for all j, Aj ∈ B(R)
is such that µ([0, t] × Aj ) < ∞ for all t; and for all i and j, φij is bounded
and FTi -measurable. The stochastic integral of a simple predictable function
with respect to M is defined by
Z tZ n,m
X X
φ(t, y)M (dt × dy) := φ(Ti , yi ) ≡ φij M ((Ti ∧ t, Ti+1 ∧ t] × Aj )
0 R i:Ti ≤t i,j=1

In a similar fashion, we can define the integral with respect to the compen-
sated measure M̃ = M − µ:
Z tZ
Xt = φ(t, y)M̃ (dt × dy)
0 R
n,m
X
:= φij {M ((Ti ∧ t, Ti+1 ∧ t] × Aj ) − µ((Ti ∧ t, Ti+1 ∧ t] × Aj )}
i,j=1

This process is a martingale and satisfies the “isometry relation”:

Z TZ
2
E[XT ] = E φ2 (t, y)µ(dt × dy).
0 R

This isometry allows to extend the notion of stochastic integral with respect
to a compensated Poisson random measure to square integrable predictable
functions. Next, the localization procedure can be used to extend the defi-
nition to all functions φ adapted and left-continuous in t and measurable in
y, such that the process
Z tZ
At := φ2 (s, y)µ(ds × dy)
0 R

is locally integrable.
The stochastic integral with respect to a Poisson random measure is more
general than that with respect to a Poisson process: if S is a piecewise
constant Lévy process,
Z T X Z TZ
φt dSt = φt ∆St = φt yJS (dt × dy),
0 0 R

that is, the integral with respect to a process can be written as the integral
of a specific function with respect to the jump measure of the process.

19
 The stochastic integral with respect to a Poisson random measure allows
us to define a new class of processes, which extends the notion of the Lévy
process, while still preserving an easy-to-understand mathematical structure:
many authors call this class Lévy-Itô processes. Recall that a Lévy process
satisfies (with a little change of notation)
Z tZ Z tZ
Xt = µt + σWt + xM (ds × dx) + xM̃ (ds × dx),
0 |x|>1 0 |x|≤1

where M is a Poisson random measure with intensity dt × ν. For a Lévy-Itô

process, the coefficients can be non-constant and even random:
Z t Z t
Xt = µs ds + σs dWs
0 0
Z tZ Z tZ
+ γs (x)M (ds × dx) + γs (x)M̃ (ds × dx), (14)
0 |x|>1 0 |x|≤1

where µ and σ are adapted locally bounded processes and γt (x) is an adapted
random function, left-continuous in t, measurable in x, such that the process:
Z
γt2 (x)ν(dx)
|x|≤1

is locally bounded.
The class of Lévy-Itô processes enjoys better stability properties than
that of Lévy processes: if (Xt ) is a Lévy-Itô process then for every function
f ∈ C 2 , (fR(Xt )) is also a Lévy-Itô process.
When |x|>1 |γt (x)|ν(dx) is also locally bounded, the process X can be
decomposed onto a ’martingale part’ and a ’drift part’:
Z t Z Z t Z tZ
Xt = (µs + γt (x)ν(dx))ds + σs dWs + γs (x)M̃ (ds × dx),
0 |x|>1 0 0 R

and in the purely martingale case, often found in applications,

Z t Z tZ
Xt = σs dWs + γs (x)M̃ (ds × dx), (15)
0 0 R

and the isometry relation holds:

Z T  Z T Z 
2 2
E[XT ] = E σt dt + E γt2 (x)ν(dx)dt . (16)
0 0 R

20
4.3 Change of variable formula for Lévy-Itô processes
In the absence of jumps, the change of variable formula (Itô formula) for a
function f ∈ C 2 takes the form
Z T
1 T 00
Z
0
f (XT ) = f (X0 ) + f (Xt )dXt + f (Xt )σt2 dt.
0 2 0
When the process has a finite number of jumps on [0, T ], one can write
Xt := Xtc + s≤t ∆Xs and apply the same formula between the jump times:
P

Z T Z T
c 1
X
0
f (XT ) = f (X0 )+ f (Xt )dXt + f 00 (Xt )σt2 dt+ {f (Xt )−f (Xt− )}.
0 2 0 t≤T :∆X 6=0 t

When the number of jumps is infinite, the latter sum may diverge, but we
still have
Z T
1 T 00
Z
0
f (XT ) = f (X0 ) + f (Xt− )dXt + f (Xt )σt2 dt
0 2 0
X
+ {f (Xt ) − f (Xt− ) − f 0 (Xt− )∆Xt }. (17)
t≤T :∆Xt 6=0

To make the decomposition (14) appear and show that the class of Lévy-Itô
processes is stable with respect to transformations with C 2 functions, we
rewrite the above expression as follows:
Z T
1
f (XT ) = f (X0 ) + {µt f 0 (Xt ) + σt2 f 00 (Xt )
0 2
Z
+ (f (Xt + γt (x)) − f (Xt ) − γt (x)f 0 (Xt ))ν(dx)}dt
|x|≤1
Z T Z T Z
0
+ f (Xt )σt dWt + (f (Xt− + γt (x)) − f (Xt− ))M̃ (dt × dx)
0 0 |x|≤1
Z T Z
+ (f (Xt− + γt (x)) − f (Xt− ))M (dt × dx)
0 |x|>1

Exercise 11. Show that for a simple predictable

R· process φ of the form (12),
and a square integrable martingale M , 0 φt dMt is also a martingale and
Z T 2
E φt dMt ≤ sup φ2t E[MT2 ].
0 0≤t≤T,ω∈Ω

21
Hint: Use Doob’s optional sampling theorem. Let (Xt ) be an (Ft )-martingale
and let S and T be bounded stopping times with S ≤ T a.s. Then,

E[XT |FS ] = XS , p.s.

Exercise 12. The quadratic variation or ’square bracket’ of a semimartingale

can be defined by
Z t
2 2
[X]t := Xt − X0 − 2 Xs− dXs .
0

• Compute the quadratic variation for a general Lévy-Itô process, a Lévy

process and a Poisson process.

• Show that if X is a Lévy-Itô process such that [X]t ≡ t and X is a

martingale then X is the standard Brownian motion.

Exercise 13. Let X be a Lévy-Itô process of the form (14), whose coefficients
µ, σ and γ are deterministic and bounded. Applying the change of variable
formula (17) to the function f (x) = eiux , show that the characteristic function
of XT is given by a generalized version of the Lévy-Khintchine formula.

Exercise 14. Let X be a Lévy-Itô process of the form (14) such that

σt2
Z
µt + + (eγt (x) − 1 − γt (x)1|x|≤1 )ν(dx) = 0
2 R

a.s. for all t. Using the change of variable formula (17), show that eXt can
be written in the form (15)R with coefficients to be defined.
γt (x)
Assuming that σt and R (e − 1)2 ν(dx) are a.s bounded by a constant
C, use the isometry relation (16) and Gronwall’s lemma to show that (eXt )
is a square integrable martingale.

Gronwall’s lemma: Let φ be a positive locally bounded function on R+

such that Z t
φ(t) ≤ a + b φ(s)ds
0

for all t and two constants a and b ≥ 0. Then φ(t) ≤ aebt .

22
5 Stochastic exponential of a jump process
Proposition 7 (Stochastic exponential). Let (Xt )t≥0 be a Lévy-Itô process
with volatility coefficient σ. There exists a unique cadlag process (Z)t≥0 such
that

dZt = Zt− dXt Z0 = 1. (18)

Z is given by:
1
Rt
σs2 ds
Y
Zt = eXt − 2 0 (1 + ∆Xs )e−∆Xs . (19)
0≤s≤t

Proof. Let
Y
Vt = (1 + ∆Xs )e−∆Xs .
0≤s≤t;∆Xs 6=0

The first step is to show that this process exists and is of finite variation. We
decompose V into a product of two terms: Vt = Vt0 Vt00 , where
Y Y
Vt0 = (1 + ∆Xs )e−∆Xs and Vt00 = (1 + ∆Xs )e−∆Xs .
0≤s≤t 0≤s≤t
|∆Xs |≤1/2 |∆Xs |>1/2

V 00 for every t is a product of finite number of factors, so it is clearly of finite

variation and there are no existence problems. V 0 is positive and we can
consider its logarithm.
X
ln Vt0 = (ln(1 + ∆Xs ) − ∆Xs ).
0≤s≤t;|∆Xs |≤1/2

Note that each term of this sum satisfies

0 > ln(1 + ∆Xs ) − ∆Xs > −∆Xs2 .

Therefore, the series is decreasing and bounded from below by − 0≤s≤t ∆Xs2 ,
P
which is finite for every Lévy-Itô process. Hence, (ln Vt0 ) exists and is a de-
creasing process. This entails that (Vt ) exists and has trajectories of finite
variation.

23
 The second step is to apply the Itô formula to the function Zt ≡ f (t, Xt , Vt ) ≡
Rt 2
Xt − 12
0 s ds V . This yields (in differential form)
σ
e t

σt2 Xt − 1 R t σs2 ds 1 t 2
R 1 t 2
R
dZt = − e 2 0 Vt dt + eXt− − 2 0 σs ds Vt− dXt + eXt− − 2 0 σs ds dVt
2
σt2 Xt − 1 R t σs2 ds 1 t 2
R 1 t 2
R
+ e 2 0 Vt dt + eXt − 2 0 σs ds Vt − eXt− − 2 0 σs ds Vt−
2
1 t 2 1 t 2
R R
− eXt− − 2 0 σs ds Vt− ∆Xt − eXt− − 2 0 σs ds ∆Vt .

Now observe that since Vt is a pure jump process,

dVt ≡ ∆Vt = Vt− (e∆Xt (1 + ∆Xt ) − 1).

Substituting this into the above equality and making all the cancellations
yields the Equation (18).
(1) (2)
To understand why the solution is unique, observe that if (Zt ) and (Zt )
(1) (2)
satisfy the Equation (18), then their difference Z̃t = Zt − Zt satisfies the
same equation with initial condition Z̃0 = 0. From the form of this equation,
it is clear that if the solution is equal to zero at some point, it will remain
zero.
Z is called the stochastic exponential or the Doléans-Dade exponential of
X and is denoted by Z = E(X).

Relation between ordinary and stochastic exponential It is clear

from the above results that the ordinary exponential and the stochastic ex-
ponential of a Lévy process are two different notions: they do not correspond
to the same stochastic process. In fact, contrary to the ordinary exponential
exp(Xt ), which is obviously a positive process, the stochastic exponential
Z = E(X) is not necessarily positive. It is easy to see that the stochastic
exponential is always nonnegative if all jumps of X are greater than −1, or,
equivalently, ν((−∞, −1]) = 0.
It is therefore natural to ask, which of the two processes is more suitable
for modeling price dynamics. The following result, due to Goll and Kallsen
[35], shows that the two approaches are equivalent: if Z > 0 is the stochastic
exponential of a Lévy process it is also the ordinary exponential of another
Lévy process and vice versa. Therefore, the two operations, although they
produce different objects when applied to the same Lévy process, end up by
giving us the same class of positive processes.

24
Proposition 8 (Relation between ordinary and stochastic exponentials).
1. Let (Xt )t≥0 be a real valued Lévy process with Lévy triplet (σ 2 , ν, γ) and
Z = E(X) its stochastic exponential. If Z > 0 a.s. then there exists
another Lévy process (Lt )t≥0 with triplet (σL2 , νL , γL ) such that Zt = eLt
where
σ2t X
Lt = ln Zt = Xt − + ln(1 + ∆Xs ) − ∆Xs . (20)
2 0≤s≤t
σL = σ,
Z
νL (A) = ν({x : ln(1 + x) ∈ A}) = 1A (ln(1 + x))ν(dx), (21)
σ2
Z

γL = γ − + ν(dx) ln(1 + x)1[−1,1] (ln(1 + x)) − x1[−1,1] (x) .
2

2. Let (Lt )t≥0 be a real valued Lévy process with Lévy triplet (σL2 , νL , γL )
and St = exp Lt its exponential. Then there exists a Lévy process
(Xt )t≥0 such that St is the stochastic exponential of X: S = E(X)
where
σ2t X 
e∆Ls − 1 − ∆Ls .

X t = Lt + + (22)
2 0≤s≤t

The Lévy triplet (σ 2 , ν, γ) of X is given by:

σ = σL ,
Z
ν(A) = νL ({x : e − 1 ∈ A}) = 1A (ex − 1)νL (dx),
x
(23)
σL2
Z
+ νL (dx) (ex − 1)1[−1,1] (ex − 1) − x1[−1,1] (x) .

γ = γL +
2

Proof. 1. The condition Z > 0 a.s. is equivalent to ∆Xs > −1 for all s
a.s., so taking the logarithm Pis justified
 here. In the proof
of Proposition 7
we have seen that the sum 0≤s≤t ln(1 + ∆Xs ) − ∆Xs converges and is a
finite variation process. Then it is clear that L is a Lévy process and that
σL = σ. Moreover, ∆Ls = ln(1 + ∆Xs ) for all s. This entails that
Z
JL ([0, t] × A) = 1A (ln(1 + x))JX (ds dx)
[0,t]×R

25
 R
and also νL (A) = 1A (ln(1 + x))ν(dx). It remains to compute γL . Substi-
tuting the Lévy-Itô decomposition for (Lt ) and (Xt ) into (20), we obtain

σ2t
Z Z
γL t − γt + + ˜
xJL (ds dx) + xJL (ds dx)
2 s∈[0,t],|x|≤1 s∈[0,t],|x|>1
Z Z
− ˜
xJX (ds dx) − xJX (ds dx)
s∈[0,t],|x|≤1 s∈[0,t],|x|>1
X
− ln(1 + ∆Xs ) − ∆Xs = 0.
0≤s≤t

Observing that
Z
x(JL (ds dx) − JX (ds dx))
s∈[0,t],|x|≤1
X

= ∆Xs 1[−1,1] (∆Xs ) − ln(1 + ∆Xs )1[−1,1] (ln(1 + ∆Xs ))
0≤s≤t

converges, we can split the above expression into jump part and drift part,
both of which must be equal to zero. For the drift part we obtain:
Z 1
σ2
γL − γ + − {xνL (dx) − xν(dx)} = 0,
2 −1

which yields the correct formula for γL after a change of variable.

2. The jumps of St are given by ∆St = St− (exp(∆Lt ) − 1). If X is a Lévy
process such that S = E(X) then since dSt = St− dXt then ∆St = St− ∆Xt
so ∆Xt = exp(∆Lt ) − 1 so ν is given by (23). In particular ∆Xt > −1 a.s.
and it is easily verified that ln E(X) is a Lévy process with characteristics
matching those of L only if X has characteristics given by (23). Conversely
if X is a Lévy process with characteristics given by (23), using (19) we can
verify as above that E(X) = exp Lt .
In view of this result and given that the formulas involving the ordinary
exponential are usually more tractable, the latter is more commonly used for
asset price modeling. However, in some situations, the stochastic exponential
may be a better choice, as shown by the following example.

26
CPPI strategy in the presence of jumps The CPPI (constant propor-
tion portfolio insurance) is a portfolio insurance strategy which allows (in
theory) to keep the portfolio value above a fixed level, while still preserving
some upside potential in case of a favorable market evolution.
To fix the ideas, suppose that the portfolio manager has promised to
the investor a guaraneed capital of N at maturity T (N can be greater or
smaller than the initial investment). To achieve this, the portfolio value Vt
must remain at each date t above the floor Bt , which is equal to the price
of the zero-coupon bond with notional N and maturity T . The difference
Ct = Vt − Bt is called cushion and the CPPI strategy uses the following
algorithm:
• At each date t, if Vt > Bt , invest mCt in the risky asset, where m > 1 is
called the multiplier, and the rest into zero-coupon bonds with maturity
T.

• If Vt ≤ Bt , invest all wealth in zero-coupon bonds with maturity T .

If the price of the risky asset is a continuous process, the portfolio value
remains above the floor, and the dynamics of the cushion is given by
dCt dSt
=m + (1 − m)rdt,
Ct St
where r is the interest rate. In the Black-Scholes model,
dSt
= µdt + σdWt ,
St
this equation can be solved explicitly, and we obtain the final portfolio value
m2 σ 2 T
 
−rT
VT = N + (V0 − N e ) exp rT + m(µ − r)T + mσWT − ,
2
whose expectation is

E[VT ] = N + (V0 − N e−rT ) exp (rT + m(µ − r)T ) .

If µ > r, there seems to be a paradox: there is no risk and the expected return
can be made arbitrarily large by choosing m big enough. This paradox is
easily solved if the price trajectories are allowed to jump, since in this case,
if the multiplier increases, the loss probability increases as well.

27
 Let
dSt
= rdt + dZt ,
St−
where Z is a Lévy process and let τ = inf{t : Vt ≤ Bt } be the first date when
the portfolio passes below the floor (it is possible that τ = ∞). Before τ , the
cushion satisfies
dCt
= mdZt + rdt,
Ct−
and the discounted cushion Ct∗ := Ct
ert
is therefore given by
Ct∗ = E(mZ)t , t < τ.
After τ , the entire portfolio is invested into the risk-free asset, which means
that the discounted cushion remains constant. Therefore,
Ct∗ = E(mZ)t∧τ .
The loss occurs if at some date t ≤ T , Ct∗ ≤ 0, which can happen if and only
if Z has a jump in the interval [0, T ] whose size is less than −1/m. We then
get (see exercise 7)
!
Z −1/m
P [∃t ∈ [0, T ] : Vt ≤ Bt ] = 1 − exp −T ν(dx) .
−∞

See [23] for details on portfolio insurance in the presence of jumps.

Exercise 15. Show that if X and Y are two Lévy processes and Y has no
diffusion component then
X
E(X)E(Y ) = E(X + Y + ∆X∆Y )

Exercise 16 (Martingale property of the stochastic exponential). Let X be

a Lévy process and a martingale. Show that E(X) is a martingale as well.

Hint: Represent X as the sum of a compound Poisson process and a Lévy

process X 0 such that |∆X 0 | < 1. Use Proposition 8 and the previous exercise.
Exercise 17. Use the previous exercise to show that for every Lévy process
X with E[|Xt |] < ∞,
E[E(X)t ] = eE[Xt ] , t > 0.

28
6 Exponential Lévy models
The Black-Scholes model
dSt
= µdt + σdWt
St
2
can be equivalently rewritten in the exponential form St = S0 e(µ−σ /2)t+σWt .
This gives us two possibilities to construct an exponential Lévy model start-
ing from a (one-dimensional) Lévy process X: using the stochastic differential
equation:
dStsde
sde
= rdt + dXt , (24)
St−
or using the ordinary exponential
Stexp = S0exp ert+Xt , (25)
where we explicitly included the interest rate r (assumed constant) in the
formulas, to simplify notation later on. The subsctipts sde for stochastic
differential equation and exp for exponential, used here to emphasize the
fact that S sde and S exp are different processes, will be omitted later on when
there is no ambiguity. Sometimes it will be convenient to discount the price
processes with the numéraire B(t, T ) = e−r(T −t) for some fixed maturity T .
St
In this case Ŝt := B(t,T )
= er(T −t) St and the equations become

dŜt
= dXt (26)
Ŝt−
or Ŝt = Ŝ0 eXt , (27)

Examples of exponential Lévy models Parametric exponential Lévy

models fall into two categories. In the first category, called jump-diffusion
models, the “normal” evolution of prices is given by a diffusion process,
punctuated by jumps at random intervals. Here the jumps represent rare
events — crashes and large drawdowns. Such an evolution can be represented
by a Lévy process with a nonzero Gaussian component and a jump part with
finitely many jumps:
Nt
X
Xt = γt + σWt + Yi , (28)
i=1

29
where (Yi ) are i.i.d. and N is a Poisson process.
In the Merton model (see Example 1), which is the first model of this
type, suggested in the literature, jumps in the log-price X are assumed to
have a Gaussian distribution: Yi ∼ N (µ, δ 2 ).
In the Kou model [42], jump sizes are distributed according to an asym-
metric Laplace law with a density of the form

ν0 (dx) = [pλ+ e−λ+ x 1x>0 + (1 − p)λ− e−λ− |x| 1x<0 ]dx (29)

with λ+ > 0, λ− > 0 governing the decay of the tails for the distribution of
positive and negative jump sizes and p ∈ [0, 1] representing the probability
of an upward jump. The probability distribution of returns in this model has
semi-heavy (exponential) tails.
The second category consists of models with an infinite number of jumps
in every interval, called infinite activity or infinite intensity models. In these
models, one does not need to introduce a Brownian component since the
dynamics of jumps is already rich enough to generate nontrivial small time
behavior [14].
The variance gamma process [15, 48] (see Example 2) is obtained by
time-changing a Brownian motion with a gamma subordinator and has the
characteristic exponent of the form:

1 u2 σ 2 κ
ψ(u) = − log(1 + − iθκu). (30)
κ 2
The density of the Lévy measure of the variance gamma process is given by
c −λ− |x| c
ν(x) = e 1x<0 + e−λ+ x 1x>0 , (31)
|x| x
√2 2 √2 2
θ +2σ /κ θ θ +2σ /κ
where c = 1/κ, λ+ = σ2
− σ2 and λ− = σ2
+ σθ2 .
To define the tempered stable process, introduced by Koponen [41] and
also known under the name of CGMY model [14], one specifies directly the
Lévy density:
c− c+
ν(x) = 1+α
e−λ− |x| 1x<0 + 1+α+ e−λ+ x 1x>0 (32)
|x| − x

with α+ < 2 and α− < 2.

30
7 The Esscher transform and absence of ar-
bitrage in exponential Lévy models
7.1 Measure changes for Lévy processes
To find out whether a given exponential Lévy model is suitable for financial
modeling, one needs to ensure that it does not contain arbitrage opportu-
nities, a property which, by the fundamental theorem of asset pricing, is
guaranteed by the existence of an equivalent martingale measure. The no ar-
bitrage equivalences for exponential Lévy models were studied in [37, 19, 66]
in the one-dimensional unconstrained case and more recently in [40] in the
multidimensional case with convex constraints on trading strategies. In this
section, we start by reviewing the one-dimensional result, and then provide a
multidimensional result (Theorem 3) which is valid in the unconstrained case
only but is more explicit than the one in [40] and clarifies the link between
the geometric properties of the Lévy measure and the arbitrage opportunities
in the model.
In the Black-Scholes model, the unique equivalent martingale measure
could be obtained by changing the drift of the Brownian motion. In models
with jumps, if the Gaussian component is absent, this is no longer possible,
but a much greater variety of equivalent measures can be obtained by altering
the distribution of jumps. The following proposition describes the possible
measure changes under which a Lévy process remains a Lévy process.
Proposition 9 (see Sato [59], Theorems 33.1 and 33.2). Let (X,P) be a
Lévy process on Rd with characteristic triplet (A, ν, γ); choose η ∈ Rd and
φ : Rd → R with
Z
(eφ(x)/2 − 1)2 ν(dx) < ∞. (33)
Rd
and define Z tZ
Ut := η.X + c
(eφ(x) − 1)J˜X (ds dx),
0 Rd
where X c denotes the continuous martingale (Brownian motion) part of X,
and J˜X is the compensated jump measure of X.
Then E(U )t is a positive martingale such that the probability measure P0
defined by
dP0 |Ft
= E(U )t , (34)
dP|Ft

31
is equivalent to P and under P0 , X is a Lévy process with characteristic triplet
(A, ν 0 , γ 0 ) where ν 0 = eφ ν and
Z
0
γ =γ+ x(ν 0 − ν)(dx) + Aη. (35)
|x|≤1

A useful example, which will be the basis of our construction of an equiv-

alent martingale measure is provided by the Esscher transform. Let X be a
Lévy Rprocess on Rd with characteristic triplet (A, ν, γ), and let θ ∈ Rd be such
that |x|>1 eθ.x ν(dx) < ∞. Applying a measure transformation of Proposi-
tion 9 with η = θ and φ(x) = θ.x, we obtain an equivalent probability under
which X is a Lévy process with Lévy measure ν̃(dx) R= eθ.x ν(dx) and third
component of the characteristic triplet γ̃ = γ + Aθ + |x|≤1 x(eθ.x − 1)ν(dx).
Using Proposition 9, the Radon-Nikodym derivative corresponding to this
measure change is found to be

dP0 |Ft eθ.Xt

= = exp(θ.Xt − κ(θ)t), (36)
dP|Ft E[eθ.Xt ]

where κ(θ) := ln E[exp(θ.X1 )] = ψ(−iθ).

Although the two definitions of an exponential Lévy model, via the ordi-
nary exponential (27) or via the stochastic exponential (26), are equivalent,
the set of Lévy processes that lead to arbitrage-free models of the form (27)
does not necessarily coincide with the set that yields arbitrage-free models
of the form (26). In particular, we shall see that the no-arbitrage conditions
for multidimensional stochastic and ordinary exponentials are considerably
different. It will be more convenient to find these conditions for models of
type (26) first and then deduce the conditions for ordinary exponentials us-
ing the transformation Xt := ln E(Y )t . In the multidimensional case, this
transformation must be applied to each component.
In an exponential Lévy model of type (26), the absence of arbitrage is
tantamount to the existence of a probability Q equivalent to P such that E(X)
is a Q-martingale. We will see that when this is the case, it is always possible
to find a martingale probability Q ∼ P under which X remains a Lévy
process, which means that X itself must be a Q-martingale (cf. Proposition
8.23 in[20]).

32
7.2 One-dimensional models
We start with the one-dimensional case. In the sequel, cc(A) denotes the
smallest convex cone containing A and ri(A) denotes the relative interior of
the set A, that is, the interior of A in the smallest linear subspace containing
A. In particular, ri({0}) = {0}.
Theorem 2 (Absence of arbitrage in models based on stochastic exponen-
tials, one-dimensional case).
Let (X, P) be a real-valued Lévy process on [0, T ] with characteristic triplet
2
(σ , ν, γ). The following statements are equivalent:
1. There exists a probability Q equivalent to P such that (X, Q) is a Lévy
process and a martingale.
2. Either X ≡ 0 or (X, P) is not a.s. monotone.
3. One of the following conditions is satisfied:
(i) σ > 0.
R
(ii) σ = 0 and |x|≤1 |x|ν(dx) = ∞.
R
(iii) σ = 0, |x|≤1 |x|ν(dx) < ∞ and −b ∈ ri(cc(supp ν)), where b =
R
γ − |x|≤1 xν(dx) is the drift of X.
Condition 2. implies that if an exponential Lévy model admits an arbi-
trage, it can be realized by a buy-and-hold strategy (if X is increasing) or a
corresponding short sale (if X is decreasing).
R It is easy to see that condition (iii) above is satisfied if and only if σ = 0,
|x|≤1
|x|ν(dx) < ∞ and one of the following is true:
• ν((−∞, 0)) > 0 and ν((0, ∞)) > 0.
• ν((−∞, 0)) > 0 and b > 0.
• ν((0, ∞)) > 0 and b < 0.
• The trivial case of a constant process: ν = 0 and b = 0.
In other words, when a finite-variation Lévy process has one-sided jumps, it
is arbitrage-free if the jumps and the drift point in opposite directions.
Before proceeding with the proof of theorem 2, we will show that for
one-dimensional exponential Lévy models of the form (27), the no-arbitrage
conditions are actually the same as for stochastic exponentials.

33
Corollary 3 (Absence of arbitrage in models based on ordinary exponential,
one-dimensional case). Let (X, P) be a real-valued Lévy process on [0, T ] with
characteristic triplet (σ 2 , ν, γ). The following statements are equivalent:

1. There exists a probability Q equivalent to P such that (X, Q) is a Lévy

process and eX is a martingale.

2. Either X ≡ 0 or (X, P) is not a.s. monotone.

3. One of the following conditions is satisfied:

(i) σ > 0.
R
(ii) σ = 0 and |x|≤1 |x|ν(dx) = ∞.
R
(iii) σ = 0, |x|≤1 |x|ν(dx) < ∞ and −b ∈ ri(cc(supp ν)).

Proof. It suffices to show that ln E(X) is monotone if and only if X is mono-

tone. From [20, Proposition 8.22] it is easy to see that ln E(X) is a finite
variation process if and only if X is a finite variation process. In the finite-
variation case, the stochastic exponential has a simple form:
Y
E(X)t = ebt (1 + ∆Xs ),
s≤t

and it is readily seen that the monotonicity properties of X and log E(X) are
the same.
Proof of theorem 2. We exclude the trivial case X ≡ 0 a.s. which clearly does
not constitute an arbitrage opportunity (every probability is a martingale
measure).
The equivalence 2 ⇐⇒ 3 follows from [20, Proposition 3.10].
3 ⇒ 1. Define a probability P̃ equivalent to P by
Z · Z 
dP̃|FT −x2 ˜
=E (e − 1)JX (ds dx)
dP|FT 0 R T

2
Under P̃, X has characteristic triplet (σ 2 , ν̃, γ̃) with ν̃ = e−x ν and γ̃ =
2
γ + |x|≤1 x(e−x − 1)ν(dx). It is easy to see that E P̃ [eλXt ] < ∞ for all λ ∈ R
R

and all t > 0.

34
 Suppose that the convex function λ 7→ E P̃ [eλX1 ] has a finite minimizer
∗
λ∗ . Then, using the dominated convergence theorem, E P̃ [X1 eλ X1 ] = 0 which
implies that X is a Q-martingale with
∗
dQ|Ft eλ Xt
= (Essher transform)
dP̃|Ft E[eλ∗ Xt ]

To show the existence of a finite minimizer λ∗ , it is sufficient to prove that

E P̃ [eλX1 ] → ∞ as λ → ∞, or, equivalently, that the function

σ2 2
Z
2
f (λ) = log E [e ] = λ + γ̃λ + (eλx − 1 − λx1|x|≤1 )e−x ν(dx).
P̃ λX1
2 R

goes to infinity as λ → ∞. In case (i), f 00 (λ) ≥ σ 2 which means that

f (λ) → ∞ as λ → ∞. In case (ii),
Z Z
0 −x2 2
f (λ) = γ̃ + xe ν(dx) + x(eλx − 1)e−x ν(dx),
|x|>1 R

and it is not difficult to check that limλ→+∞ f 0 (λ) = +∞ and limλ→−∞ f 0 (λ) =
−∞ which means that f (λ) → ∞ as λ → ∞. In case (iii),
Z
2
f (λ) = b + xeλx e−x ν(dx),
0
R

and it is easy to see, by examining one by one the different mutually exclusive
cases listed after the statement of the Theorem, that in each of these cases f 0
is bounded from below on R and therefore once again, f (λ) → ∞ as λ → ∞.
1 ⇒ 2. It is clear that a process cannot be a martingale under one
probability and a.s. monotone under an equivalent probability, unless it is
constant.

7.3 Multidimensional models

In the multidimensional case, the no arbitrage conditions for ordinary and
stochastic exponentials are different. We start with the simpler case of
stochastic exponentials.
Let (X, P) be an Rd -valued Lévy process on [0, T ] with characteristic
triplet (A, ν, γ). To describe the no-arbitrage conditions, we need to separate
the finite and infinite variation components of X. We therefore introduce the

35
linear subspace L ⊆ Rd containing all vectors w ∈ Rd such that w.X is a
finite variation process. From proposition 3.8 and theorem 4.1 in [20], it
follows that
Z
d
L = N (A) ∩ {w ∈ R : |w.x|ν(dx) < ∞},
|x|≤1

where N (A) := {w ∈ Rd : Aw = 0}. Further, denote by X L the projection

of X on L. X L is a finite variation Lévy process with triplet (0, ν L , γ L ), and
we denote its drift by
Z
L L
b := γ − xν L (dx).
L∩{x:|x|≤1}

Theorem 3 (Absence of arbitrage in models based on stochastic exponential,

multidimensional case). Let (X, P) be an Rd -valued Lévy process on [0, T ]
with characteristic triplet (A, ν, γ). The following statements are equivalent:
1. There exists a probability Q equivalent to P such that (X, Q) is a Lévy
process and (X i ) is a Q-martingale for all i.
2. For every w ∈ Rd , the process w.X satisfies one of the equivalent con-
ditions 2. or 3. of Theorem 2.
3. −bL ∈ ri(cc(supp ν L )).
Let us comment on the equivalent conditions of the above theorem.
To understand condition 2., assume that for some w ∈ Rd , the process
w.X does not satisfy the equivalent conditions of Theorem 2, meaning that it
is either strictly increasing or strictly decreasing. Consider a portfolio where
the relative proportions of different assets are kept constant and equal to wi .
The proportions may, of course, change when the underlying assets jump, but
it is assumed that they are readjusted to their constant values immediately
after the jump. Such a strategy is called a fixed-mix strategy. The discounted
value V̂t of such a portfolio satisfies the SDE dV̂t = V̂t− w.dXt , and therefore
either this portfolio constitutes an arbitrage strategy or an arbitrage strategy
can be obtained by shorting this portfolio. Condition 2 thus implies that a
multidimensional exponential Lévy model is arbitrage-free if and only if there
are no fixed-mix arbitrage strategies.
The third condition is a concise characterization of arbitrage-free expo-
nential Lévy models in terms of their characteristic triplets. This condition

36
is always satisfied if the process X has no finite-variation components: in
this case L = {0} and condition 3. reduces to 0 ∈ {0}. If the process is
of finite variation, this condition reduces to −b ∈ ri(cc(supp ν)), that is, the
drift and the finite variation jumps must point in opposite directions.
Proof of theorem 3. 1 ⇒ 2 is readily obtained by an application of Theorem
2 to the process w.X.
2 ⇒ 1. By an argument similar to the one in the proof of Theorem 2,
we can suppose without loss of generality that for all λ ∈ Rd , E[eλ.X1 ] < ∞.
The function f : λ 7→ E[eλ.X1 ] < ∞ is then a proper convex differentiable
∗
function on Rd and if λ∗ is a minimizer of this function, E[X1i eλ .X1 ] = 0 for
all i = 1, . . . , d and we can define an equivalent martingale measure Q using
the Esscher transform ∗
dQ|Ft eλ .X1
:= .
dP|Ft E[eλ∗ .X1 ]
Suppose that w.X is a Lévy process satisfying conditions 2 or 3 of Theorem
2. Then it follows from the proof of this theorem that w.X is either constant
or limλ→∞ E[eλw.X1 ] = ∞. Hence, the function f is constant along every
recession direction, which implies that f attains its minimum (Theorem 27.1
in [55]).
2 ⇒ 3. Suppose −bL ∈ / ri(cc(supp ν L )). Then −bL can be weakly sepa-
rated from cc(supp ν L ) by a hyperplane contained in L, passing through the
origin, and which does not contain −bL or cc(supp ν L ) completely (theorems
11.3 and 11.7 in [55]). This means that there exists w ∈ L such that

bL .w ≥ 0 and x.w ≥ 0, ∀x ∈ supp ν L

with either bL .w > 0 or x.w > 0 for some x ∈ supp ν L . In this case,
ri(cc(supp ν w )) is either {0} or (0, ∞), where the measure ν w is defined
by ν w (A) := ν L ({x ∈ L : w.x ∈ A}). If bL .w > 0, this implies that
−bL .w ∈/ ri(cc(supp ν w )). If bL .w = 0 then necessarily x.w > 0 for some
x ∈ supp ν L which means that in this case ri(cc(supp ν w )) = (0, ∞) and once
again −bL .w ∈ / ri(cc(supp ν w )). In both cases, we have obtained a contradic-
tion with 2.
3 ⇒ 2. Assume that −bL ∈ ri(cc(supp ν L )) and let w ∈ Rd . If w ∈ / L
than w.X has infinite variation and the claim is shown. Assume that w ∈ L
and let Rw := ri(cc(supp ν w )). Rw can be equal to R, half-axis or a single
point {0}. If Rw = R, there is nothing to prove. In the two other cases,

37
−b.w ∈/ Rw means that w weakly separates −bL from cc(supp ν L ) in such
a way that either bL .w > 0 or x.w > 0 for some x ∈ supp ν L , which is a
contradiction with 3.

Case of models based on ordinary exponentials In multidimensional

models of type (27), contrary to the one-dimensional case, the no-arbitrage
conditions are not the same as in models of type (26), as the following ex-
ample illustrates. Let N be a standard Poisson process with intensity λ and
define
1
Xt1 = Nt − λ(e − 1)t; St1 = S01 eXt .
2
Xt2 = −Nt − λ(e−1 − 1)t; St2 = S02 eXt .

The linear combination X 1 + X 2 is nonconstant and monotone, however the

model is arbitrage-free since S 1 and S 2 are easily seen to be martingales.
To check whether a model of type (27) based on an Rd -valued Lévy process
X is arbitrage-free, one should construct the equivalent model of type (26)
by computing Yti = ln E(X i )t for i = 1, . . . , d, and then check the conditions
of Theorem 3 for the process Y . The following remarks can facilitate this
task in some commonly encountered cases:
• The space L of finite variation components is invariant under the map-
ping ln E; therefore, if the process X does not have finite variation
components, the model is arbitrage-free.

• If the Lévy measure ν X of X has full support then the Lévy measure
ν Y of Y satisfies cc(supp ν Y ) = Rd , which implies that the model is
arbitrage-free.

• If an orthant is contained in the support of ν X , this orthant will also

be contained in cc(supp ν Y ).
Exercise 18. Let (X, P ) be a Lévy process with characteristic triple (A, ν, γ),
and let Q be a probability measure defined by
dQ eθXt
|Ft =
dP E[eθXt ]

for θ ∈ R such that E[eθXt ] < ∞. Compute the characteristic triple of X

under Q.

38
Exercise 19.
• Let X be a Lévy process and let f : R → (0, ∞). What condition must
be imposed on the function f for the sum
X
f (∆Xt )
t∈[0,1]:∆Xt 6=0

to converge a.s.?
1
• Let (X, P ) be a Lévy process with Lévy measure |x|1+α and let (X, Q)
be a Lévy process with Lévy measure |x|1+α0 (with α > 0 and α0 > 0).
1

Use the previous question to show that P ∼ Q implies α = α0 . Check

your result using Proposition 9.

8 European options in exp-Lévy models

Given the results of section 7, in any “reasonable” exponential Lévy model
we can assume that there exists a probability measure Q equivalent to P
such that the discounted prices of all assets are Q-martingales. In practice,
this measure is usually found by calibrating the exponential Lévy model to
market quoted prices of European options [7, 21], and the first step in using
the model is therefore to obtain fast pricing algorithms for European calls
and puts.
Prices of European options in exponential Lévy models can be computed
directly from the characteristic function of X which is explicitly known from
the Lévy-Khintchine formula. This idea was first introduced to finance by
Carr and Madan [15] (for European calls and puts) and later extended and
generalized by many authors including [54, 47, 46, 27]. The result given below
is a slight generalization of the one in [27], allowing both discontinuous pay-
off functions and Lévy processes without a bounded density, such as variance
gamma.
We start with a one-dimensional risk-neutral exponential Lévy model
in the form (27). Under the risk-neutral probability, the process eX must
therefore be a martingale, a condition which can be expressed in terms of
the characteristic triplet of X:
Z
A
γ + + (ey − 1 − y1|y|≤1 )ν(dy) = 0.
2 R

39
We consider a European option with pay-off G(ST ) = G(ŜT ) at time T and
denote by g its log-payoff function: G(ex ) ≡ g(x). As above, we denote by
Φt the characteristic function of Xt .

Proposition 10. Suppose that there exists R 6= 0 such that

g(x)e−Rx has finite variation on R, (37)

−Rx 1
g(x)e ∈ L (R), (38)
|ΦT −t (u − iR)|
Z
E[eRXT −t ] < ∞ and du < ∞. (39)
R 1 + |u|

Then the price at time t of the European option with pay-off function G
satisfies

P (t, St ) := e−r(T −t) E[G(ST )|Ft ]

e−r(T −t)
Z
= ĝ(u + iR)ΦT −t (−u − iR)ŜtR−iu du, (40)
2π R

where Z
ĝ(u) := eiux g(x)dx.
R

Proof. By integration by parts for Stieltjes integrals,

Z Z
ix(u+iR) i
ĝ(u + iR) = g(x)e dx = eix(u+iR) dg(x). (41)
R u + iR R
This implies in particular that
C
|ĝ(u + iR)| ≤ , u ∈ R. (42)
|u + iR|

Suppose that R > 0 (the case R < 0 can be treated in a similar manner) and
consider the function Z ∞
Rx
f (x) = e p(dz),
x

where p denotes the distribution of XT −t . From the assumption (39) it follows

Z
eRx p(dx) < ∞
R

40
and therefore limx→∞ f (x) = 0. Clearly also limx→−∞ f (x) = 0. By integra-
tion by parts,
Z N
1 N Rx
Z
1
f (x)dx = e p(dx) + (f (N ) − f (−N ).
−N R −N R

This shows that f ∈ L1 (R) and it follows that

ΦT −t (−u − iR)
Z
e−iux f (x)dx = .
R R − iu
From condition (39) it follows that f can be recovered by Fourier inversion
(cf. [57, Theorem 9.11]):
ΦT −t (−u − iR)
Z
1
f (x) = eiux du. (43)
2π R R − iu
Let us now turn to the proof of (40). From (41), (43) and Fubini’s theorem,
Z
1
ĝ(u + iR)ΦT −t (−u − iR)ŜtR−iu du (44)
2π R
eiux ΦT −t (−u − iR)e(R−iu) log Ŝt
Z Z
1 Rx
= dg(x)e du (45)
2π R R R − iu
Z Z Z ∞
−R(x−log Ŝt )
= dg(x)e f (x − log Ŝt ) = dg(x) p(dz) (46)
R R x−log Ŝt
Z
= g(x + log Ŝt )p(dx) = E Q [G(ŜT )|Ft ] = E Q [G(ST )|Ft ]. (47)
R

Example 3. The digital option has pay-off G(ST ) = 1ST ≥K . In this case for
all R > 0 conditions (37) and (38) are satisfied and

K iu−R
ĝ(u + iR) = .
R − iu
Example 4. The European call option has pay-off G(ST ) = (ST −K)+ . There-
fore, conditions (37) and (38) are satisfied for all R > 1,

K iu+1−R
ĝ(u + iR) = .
(R − iu)(R − 1 − iu)

41
and the price of a call option can be written as an inverse Fourier transform:
e−r(T −t) K iu+1−R ŜtR−iu ΦT −t (−u − iR)
Z
C(t, St ) = du
2π R (R − iu)(R − 1 − iu)
Z kf (iu+1−R)
St e ΦT −t (−u − iR)
= du (48)
2π R (R − iu)(R − 1 − iu)
where k f is the log forward moneyness defined by k f = ln(K/St ) − r(T − t).
This property allows to compute call option prices for many values of k f in
a single computation using the FFT algorithm as explained below.

8.1 Numerical Fourier inversion

In this paragraph we describe an efficient implementation of Formula 48 using
the Fast Fourier Transform (FFT). Let
ΦT −t (−u − iR)
ζ(u) =
(R − iu)(R − 1 − iu)
Option prices can be computed by evaluating numerically the inverse Fourier
transform of ζ:
Z +∞
kf (1−R) 1 f
f
C(k ) = St e e−ivk ζ(v)dv (49)
2π −∞
This integral can be efficiently computed using the Fast Fourier transform, an
algoritm due to Cooley and Tukey [26] which allows to compute F0 , . . . , FN −1 ,
given by,
N
X −1
Fn = fk e−2πink/N , n = 0 . . . N − 1,
k=0

using only O(N ln N ) operations.

To approximate option prices, we truncate and discretize the integral (49)
as follows:
Z ∞ Z A/2
1 −ivkf 1 f
e ζ(v)dv = e−ivk ζ(v)dv + εT
2π −∞ 2π −A/2
N −1
A X f
= wm ζ(vm )e−ik vm + εT + εD , (50)
2πN m=0

42
where εT is the truncation error, εD is the discretization error, vm = −A/2 +
m∆, ∆ = A/(N − 1) is the discretization step and wm are weights, cor-
responding to the chosen integration rule (for instance, for the trapezoidal
rule w0 = wN −1 = 1/2 and all other weights are equal to 1). Now, setting
knf = k0f + N
2πn
∆
we see that the sum in the last term becomes a discrete Fourier
transform:
N −1
A iknf A/2 X f
e wm e−ik0 vm f (km
f
)e−2πinm/N
2πN m=0

The FFT algorithm allows to compute option prices for the log strikes knf =
k0f + N
2πn
∆
. The log strikes are thus equidistant with the step d satisfying

2π
d∆ = .
N

Typically, k0f is chosen so that the grid is centered around the money, ∆ is
fixed to keep the discretization error low, and N is adjusted to keep the trun-
cation error low and have a sufficiently small step between strikes (increasing
N reduces the truncation error and the distance between consecutive strikes
at the same time). The option prices for the strikes not on the grid must
be computed by interpolation. It should be noted that the FFT method
should only be used when option prices for many strikes must be computed
simultaneously (such as for calibration). If only a small number of strikes
is needed, integration methods with variable step size usually have a better
performance.
.

9 Integro-differential equations for exotic op-

tions
In exponential Lévy models, the price of some options whose pay-off de-
pends on the trajectory of the underlying asset price, such as barriers, can
be expressed as the solution of an equation similar to the Black-Scholes PDE

∂C 1 2 2 ∂ 2 C ∂C
+ σ S 2
= rC − rS .
∂t 2 ∂S ∂S

43
 Let X be a Lévy process such that eX is a martingale under Q (risk-
neutral probability) and the price of the underlying asset is given by St =
S0 ert+Xt . Then the price at time t of the European option

Pt = e−r(T −t) E[(ST − K)+ |Ft ]

can be expressed as deterministic function of t and St : Pt = P (t, St ) and in

addition e−rt P (t, St ) is a martingale.
In a similar manner, for the up-and-out option we get

PtB = e−r(T −t) E[(ST − K)+ 1max0≤s≤T Ss <B |Ft ].

The process e−rt PtB is therefore a martingale. Define now P B (t, S) as the
deterministic function given by

P B (t, S) = e−r(T −t) E[g(ST ∧τt )|St = S],

where g(S) = (S − K)+ 1S<B and τt = inf{s ≥ t : Xs ≥ B}. Then, PtB =

P B (t, St ) if the barrier has not been reached yet.
Under the assumption that eX is a martingale, the dynamics of S is
Z
dSt = rSt dt + St σdWt + St− (ex − 1)J˜X (dt × dx),
R

where JX is the compensated jump measure of X and σ is the volatility

of its diffusion component. Let the function P̃ (t, St ) be sufficiently regular.
Applying the Itô formula to this function

1 2 2 ∂ 2 P̃

∂ P̃ ∂ P̃
dP̃ (t, St ) = + rS + St σ
∂t ∂S 2 ∂S 2
Z ! 
x x ∂ P̃
+ P̃ (t, St e ) − P̃ (t, St ) − St (e − 1) ν(dx)
R ∂S
Z
∂ P̃
+ σSt dWt + (P̃ (t, St ex ) − P̃ (t, St ))J˜X (dt × dx).
∂S R

We can then state the following result (results of this type are known as
verification theorems):

44
Proposition 11. Let P (t, S) be a function which is differentiable with respect
to t once and with respect to S twice, such that the derivative ∂P
∂S
is bounded,
and assume that P satisfies the equation

∂P 1 2 2 ∂ 2 P
Z  
x x ∂P ∂P
+ S σ + P (t, Se ) − P (t, S) − S(e − 1) ν(dx) = rP −rS
∂t 2 ∂S 2 R ∂S ∂S

with the terminal condition P (T, S) = (S − K)+ . Then the price at time t
of a European call option with maturity T and strike K is given by P (t, St ).

The price of a European option in a model with jumps therefore solves an

equation which is similar to the Black-Scholes PDE, but contains a correction
term depending on the Lévy measure. Equations of this type are called
integro-differential equations. It remains to prove that the above equation
admits smooth solutions, and for European options it is a relatively easy
task [24]. One could also prove formally that the price of an up-and-out
option with barrier B solves the same equation with the additional condition
P B (t, S) = 0 for S ≥ B but in this case the problem of regularity is really
difficult and it may become necessary to relax the notion of solution [25].

10 Gap options
The gap options are a class of exotic equity derivatives offering protection
against rapid downside market moves (gaps). These options have zero delta,
allowing to make bets on large downside moves of the underlying without in-
troducing additional sensitivity to small fluctuations, just as volatility deriva-
tives allow to make bets on volatility without going short or long delta. The
market for gap options is relatively new, and they are known under many
different names: gap options, crash notes, gap notes, daily cliquets, gap risk
swaps etc. The gap risk often arises in the context of constant proportion
portfolio insurance (CPPI) strategies [23] and other leveraged products such
as the leveraged credit-linked notes. The sellers of gap options (who can
be seen as the buyers of the protection against gap risk) are typically ma-
jor banks who want to get off their books the risk associated to CPPI or
other leveraged products. The buyers of gap options and the sellers of the
protection are usually hedge funds looking for extra returns.
The pay-off of a gap option is linked to the occurrence of a gap event, that
is, a 1-day downside move of sufficient size in the underlying. The following

45
single-name gap option was commercialized by a big international bank in
2007 under the name of gap risk swap:
Example 5 (Single-name gap option).
• The protection seller pays the notional amount N to the protection
buyer at inception and receives Libor + spread monthly until maturity
or the first occurrence of the gap event, whichever comes first, plus the
notional at maturity if no gap event occurs.

• The gap event is defined as a downside move of over 10% in the DJ

Euro Stoxx 50 index within 1 day (close to close).

• If a gap event occurs between dates t − 1 and t, the protection seller

immediately receives the reduced notional N (1−10∗(0.9−R))+ , where
R = SSt−1
t
is the index performance at gap, after which the product
terminates.
The gap options are therefore similar to equity default swaps, with a very
important difference, that in EDS, the price change from the inception date
of the contract to a given date is monitored, whereas in gap options, only
1-day moves are taken into account.
The pay-off of a multi-name gap option depends of the total number of
gap events occurring in a basket of underlyings during a reference period.
We are grateful to Zareer Dadachanji from Credit Suisse for the following
example.
Example 6 (Multiname gap option).
• As before, the protection seller pays the notional amount N to the
protection buyer and receives Libor + spread monthly until maturity.
If no gap event occurs, the protection seller receives the full notional
amount at the maturity of the contract.

• A gap event is defined as a downside move of over 20% during one

business day in any underlying from a basket of 10 names.

• If a gap event occurs, the protection seller receives at maturity a re-

duced notional amount kN , where the reduction factor k is determined
from the number M of gap events using the following table:
M 0 1 2 3 ≥4
k 1 1 1 0.5 0

46
 The gap options are designed to capture stock jumps, and clearly cannot
be priced within a diffusion model with continuous paths, since any such
model will largely underestimate the gap risk. For instance, for a stock with
a 25% volatility, the probability of having an 10% gap on any one day during
one year is 3 × 10−8 , and the probability of a 20% gap is entirely negligible.

10.1 Single asset gap options

Suppose that the time to maturity T of a gap option is subdivided onto N
periods of length ∆ (e.g. days): T = N ∆. The return of the k-th period will
Sk∆
be denoted by Rk∆ = S(k−1)∆ . For the analytic treatment, we formalize the
single-asset gap option as follows.

Definition 10 (Gap option). Let α denote the return level which triggers
the gap event and k ∗ be the time of first gap expressed in the units of ∆:
k ∗ := inf{k : Rk∆ ≤ α}. The gap option is an option which pays to its holder
the amount f (Rk∆∗ ) at time ∆k ∗ , if k ∗ ≤ N and nothing otherwise.

Supposing that the interest rate is deterministic and equal to r, it is easy

to see that the pay-off structure of example 5 can be expressed as a linear
combination of pay-offs of definition 10.
We first treat the case where the log-returns are independent and station-
ary.

Proposition 12. Let the log-returns (Rk∆ )Nk=1 be i.i.d. and denote the distri-
bution of log R1∆ by p∆ (dx). Then the price of a gap option as of definition
10 is given by
R N
−rT ∞
Z β 1−e β
p∆ (dx)
G∆ = e−r∆ f (ex )p∆ (dx) R∞ , (51)
−∞ 1− e−r∆ β p∆ (dx)

with β := log α < 0.

47
Proof.
∗
G∆ = E e−∆k r f (Rk∆∗ )1k∗ ≤N
 

N
X
= P[k ∗ = n]E[f (Rn∆ )|k ∗ = n]e−∆nr
n=1
XN n−1
Y
−∆nr
= P[Rn∆ ≤ α]E[f (Rn∆ )|Rn∆ ≤ α]e P[Rl∆ > α]
n=1 l=1
R N
∞
Z β 1 − e−rT β p∆ (dx)
= e−r∆ f (ex )p∆ (dx) R∞ .
−∞ 1 − e−r∆ β p∆ (dx)

Numerical evaluation of prices Formula (51) allows to compute gap

option prices by Fourier inversion. For this, weR need to be able to evaluate
x
the cumulative distribution function F∆ (x) := −∞ p∆ (dξ) and the integral
Z β
f (ex )p∆ (dx). (52)
−∞

Let
R φ∆ be the characteristic function of p∆ , and suppose that p∆ satisfies
R |φ∆ (u)|
|x|p∆ (dx) < ∞ and R 1+|u| du < ∞. Let F 0 be the CDF and φ0 the
characteristic function of a Gaussian random variable with zero mean and
standard deviation σ 0 > 0. Then by Lemma 1 in [23],

φ0 (u) − φ∆ (u)
Z
0 1
F∆ (x) = F (x) + e−iux du. (53)
2π R iu
The Gaussian random variable is only needed to obtain an integrable expres-
sion in the right hand side and can be replaced by any other well-behaved
random variable.
The integral (52) is nothing but the price of a European option with
payoff function f and maturity ∆. For arbitrary f it can be evaluated using
the Fourier transform method as described in section 8. For the numerical
evaluation, the integrals must be truncated to a finite interval [−L, L]. Since
∆ is small, the characteristic function φ∆ (u) decays slowly at infinity, which
means that L must be sufficiently large (for example, in a jump-diffusion

48
model with volatility σ, L ∼ σ√C∆ , where C is a constant which depends on
the desired precision, such as C = 5 — see [67]). The computation of the
integrals will therefore be rather costly. For this reason, we do not recommend
to use the exact formula, and propose an approximation, which is based on
an expansion of G∆ around the value ∆ = 0. In other words, instead of
using a numerical method whose computational complexity increases when
∆ is small, we suggest an explicit formula whose precision improves when
∆ → 0.

Approximate pricing formula Suppose that St = S0 eXt , where X is a

Lévy process. This means that p∆ as defined above is the distribution of X∆ .
−4
R β Since r∆ ∼ 10 and the probability of having a gap on a given day
−∞ ∆
p (dx) is also extremely small, with very high precision,
Rβ
β
1 − e−rT −N −∞ p∆ (dx)
Z
x
G∆ ≈ f (e )p∆ (dx) Rβ . (54)
−∞ r∆ + −∞ p∆ (dx)

Our second approximation is less trivial. From [58], we know that for all
Lévy processes and under very mild hypotheses on the function f , we have
Z β Z β
g(x)p∆ (dx) ∼ ∆ g(x)ν(dx),
−∞ −∞

as ∆ → 0, where ν is the Lévy measure of X. Consequently, when ∆ is

nonzero but small, we can replace the integrals with respect to the density
with the integrals with respect to the Lévy measure in formula (54), obtaining
an approximate but explicit expression for the gap option price:
Rβ
β
1 − e−rT −T −∞ ν(dx)
Z
G∆ ≈ G0 = lim G∆ = f (ex )ν(dx) Rβ . (55)
∆→0 −∞ r + −∞ ν(dx)

This approximation is obtained by making the time interval at which returns

are monitored (a priori, one day), go to zero. In other words, G0 is the zero-
order term of the Taylor expansion of G∆ around the point ∆ = 0. The error
of this approximation will therefore decay proportionally to ∆ when ∆ → 0.

A modified gap option For a better understanding of the risks of a gap

option, it is convenient to interpret the pricing formula (55) as an exact price

49
of a modified gap option rather than the true price of the original option.
From now on, we define the single-asset gap option as follows.

Definition 11 (Modified gap option). Let τ = inf{t : ∆Xt ≤ β} be the

time of the first jump of X smaller than
 β. The gap option as a product
which pays to its holder the amount f SSττ− = f (e∆Xτ ) if τ ≤ T and zero
otherwise.

The price of this product is given by

G = E Q [e−rτ f (e∆Xτ )1τ ≤T ]

which is easily seen to be equal to G0 :

Proposition 13. Suppose that the underlying follows an exponential Lévy

model: St = S0 eXt , where X is a Lévy process with Lévy measure ν. Then the
price of the gap option as of definition 11, or, equivalently, the approximate
price of the gap option as of definition 10 is given by
Rβ
β
1 − e−rT −T −∞ ν(dx)
Z
G= f (ex )ν(dx) Rβ
−∞ r + −∞ ν(dx)

with β := log α.

The gap option then arises as a pure jump risk product, which is only
sensitive to negative jumps larger than β in absolute value, but not to small
fluctuations of the underlying. In particular, it has zero delta.

10.2 Multi-asset gap options and Lévy copulas

As explained earlier in this section, a multiname (basket) gap option is a
product where one monitors the total number of gap events in a basket
of underlyings over the lifetime of the option [0, T ]. A gap event is de-
fined as a negative return of size less than α between consecutive closing
prices (close-to-close) in any of the underlyings of the basket. The pay-off
of the product at date T is determined by the total number of gap events in
the basket over the reference period. To compute the price of a multiname
gap option, we suppose that M underlying assets S 1 , . . . , S M follow an M -
i
dimensional exponential Lévy model, that is, Sti = S0i eXt for i = 1, . . . , M ,

50
where (X 1 , . . . , X M ) is an M -dimensional Lévy process with Lévy measure
ν. We make the same simplifying hypothesis as in definition 11, that is, we
define a gap event as a negative jump smaller than a given value β in any
of the assets, rather than a negative daily return. From now on, we define a
multiname gap option as follows.
Definition 12. For a given β < 0, let
M
X
Nt = #{(s, i) : s ≤ t, 1 ≤ i ≤ M and ∆Xsi ≤ β} (56)
i=1

be the process counting the total number of gap events in the basket before
time t. The multiname gap option is a product which pays to its holder the
amount f (NT ) at time T .
The pay-off function f for a typical multiname gap option is given in
example 6. Notice that the single-name gap option stops at the first gap
event, whereas in the multiname case the gap events are counted up to the
maturity of the product.
The biggest difficulty in the multidimensional case, is that now we have
to model simultaneous jumps in the prices of different underlyings. The
multidimensional Lévy measures can be conveniently described using their
tail integrals. The tail integral U describes the intensity of simultaneous
jumps in all components smaller than the components of a given vector.
Given an M -dimensional Lévy measure ν, we define the tail integral of ν by
U (z1 , . . . , zM ) = ν({x ∈ RM : x1 ≤ z1 , . . . , xM ≤ zM }), z1 , . . . , zM < 0.
(57)
The tail integral can also be defined for positive z (see [38]), but we do not
introduce this here since we are only interested in jumps smaller than a given
negative value.
To describe the intensity of simultaneous jumps of a subset of the com-
ponents of X, we define the marginal tail integral: for m ≤ M and 1 ≤ i1 <
· · · < im ≤ M , the (i1 , . . . , im )-marginal tail integral of ν is defined by
Ui1 ,...,im (z1 , . . . , zm ) = ν({x ∈ RM : xi1 ≤ z1 , . . . , xim ≤ zm }), z1 , . . . , zm < 0.
(58)
The process N counting the total number of gap events in the basket is
clearly a piecewise constant increasing integer-valued process which moves

51
only by jumps of integer size. The jump sizes can vary from 1 (in case of a
gap event affecting a single component) to M (simultaneous gap event in all
components). The following lemma describes the structure of this process
via the tail integrals of ν.

Lemma 2. The process N counting the total number of gap events is a Lévy
process with integer jump sizes 1, . . . , M occurring with intensities λ1 , . . . , λM
given by
M
X X
λm = (−1)k−m k
Cm Ui1 ,...,ik (β, . . . , β), 1 ≤ m ≤ M, (59)
k=m 1≤i1 <···<ik ≤M

k
where Cm denotes the binomial coefficient and the second sum is taken over
all possible sets of k integer indices satisfying 1 ≤ i1 < · · · < ik ≤ M .

Proof. Since X is a process with stationary and independent increments, it

follows from formula (56) that N has stationary and independent increments
as well. A jump of size m in N occurs if and only if exactly m components
of X jump by an amount smaller or equal to β. Therefore,
X
λm = ν ({xi ≤ β ∀i ∈ {i1 , . . . , im }; xi > β ∀i ∈
/ {i1 , . . . , im }})
1≤i1 <···<im ≤M
(60)

The expression under the sum sign can be written as

ν ({xi ≤ β ∀i ∈ {i1 , . . . , im }; xi > β ∀i ∈

/ {i1 , . . . , im }})
= ν ({xi ≤ β ∀i ∈ {i1 , . . . , im }})
M
X −m X
+ (−1)p ν ({xi ≤ β ∀i ∈ {i1 , . . . , im } ∪ {j1 , . . . , jp }})
p=1 1≤j1 <···<jp ≤M
{j1 ,...,jp }∩{i1 ,...,im }=∅

M
X −m X
= Ui1 ,...,im (β, . . . , β) + (−1)p Ui1 ,...,im ,j1 ,...,jp (β, . . . , β)
p=1 1≤j1 <···<jp ≤M
{j1 ,...,jp }∩{i1 ,...,im }=∅

Combining this equation with (60) and gathering the terms with identical
tail integrals, one obtains (59).

52
 The process N can equivalently be represented as
M
X (m)
Nt = mNt ,
m=1

where N (1) , . . . , N (M ) are independent Poisson processes with intensities λ1 ,

. . . , λM . Since these processes are independent, the expectation of any func-
tional of NT (the price of a gap option) can be computed as
∞ M
! M
X X Y (λi T )ni
E[f (NT )] = e−λT f knk , (61)
n ,...,n =0 k=1 i=1
n i !
1 M

where λ := M
P
i=1 λi . In practice, after a certain number of gap events, the
gap option has zero pay-off and the sum in (61) reduces to a finite number
of terms. In example 6, f (n) ≡ 0 for n ≥ 4 and
n (λ1 T )2
E[f (NT )] = e−λT 1 + λ1 T + + λ2 T (62)
2
(λ1 T )3 λ1 λ2 T 2 λ3 T o
+ + + . (63)
12 2 2
The price of the protection (premium over the risk-free rate received by the
protection seller) is given by the discounted expectation of 1 − f (NT ), that
is,

e−rT E[1 − f (NT )]. (64)

To make computations with the formula (61), one needs to evaluate the
tail integral of ν and all its marginal tail integrals. These objects are de-
termined both by the individual gap intensities of each component and by
the dependence among the components of the multidimensional process. For
modeling purposes, the dependence structure can be separated from the be-
havior of individual components via the notion of Lévy copula [20, 38], which
is parallel to the notion of copula but defined at the level of jumps of a Lévy
process. More precisely we will use the positive Lévy copulas which describe
the one-sided (in this case, only downward) jumps of a Lévy process, as
opposed to general Lévy copulas which are useful when both upward and
downward jumps are of interest.

53
Positive Lévy copulas Let R := (−∞, ∞] denote the extended real line,
d
and for a, b ∈ R let us write a ≤ b if ak ≤ bk , k = 1, . . . , d. In this case,
(a, b] denotes the interval

(a, b] := (a1 , b1 ] × · · · × (ad , bd ].

d
For a function F mapping a subset D ⊂ R into R, the F -volume of (a, b]
is defined by
X
VF ((a, b]) := (−1)N (u) F (u),
u∈{a1 ,b1 }×···×{ad ,bd }

where N (u) := #{k : uk = ak }. In particular, VF ((a, b]) = F (b) − F (a) for

d = 1 and VQ F ((a, b]) = F (b1 , b2 ) + F (a1 , a2 ) − F (a1 , b2 ) − F (b1 , a2 ) for d = 2.
If F (u) = di=1 ui , the F -volume of any interval is equal to its Lebesgue
measure.
A function F : D → R is called d-increasing if VF ((a, b]) ≥ 0 for all
a, b ∈ D such that a ≤ b. The distribution function of a random vector is
one example of a d-increasing function.
A function F : [0, ∞]d → [0, ∞] is called a positive Lévy copula if it
satisfies the following conditions:
1. F (u1 , . . . , ud ) = 0 if ui = 0 for at least one i ∈ {1, . . . , d},

2. F is d-increasing,

3. Fi (u) = u for any i ∈ {1, . . . , d}, u ∈ R, where Fi is the one-dimensional

margin of F , obtained from F by replacing all arguments of F except
the i-th one with ∞:

Fi (u) = F (u1 , . . . , ud )ui =u,uj =∞ ∀j6=i .

The positive Lévy copula has the same properties as ordinary copula but is
defined on a different domain ([0, ∞]d instead of [0, 1]d ). Higher-dimensional
margins of a positive Lévy copula are defined similarly:

Fi1 ,...,im (u1 , . . . , um ) = F (v1 , . . . , vd )vik =uk ,k=1,...,m;vj =∞,j ∈{i

/ 1 ,...,im } .

The Lévy copula links the tail integral to one-dimensional margins; the
following result is a direct corollary of Theorem 3.6 in [38].

54
Proposition 14.

• Let X = (X 1 , . . . , X d ) be a Rd -valued Lévy process, and let the (one-

sided) tail integrals and marginal tail integrals of X be defined by (57)
and (58). Then there exists a positive Lévy copula F such that the tail
integrals of X satisfy

Ui1 ,...,im (x1 , . . . , xm ) = Fi1 ,...,im (Ui1 (x1 ), . . . , Uim (xm )) (65)

for any nonempty index set {i1 , . . . , im } ⊆ {1, . . . , d} and any (x1 , . . . , xm ) ∈
(−∞, 0)m .

• Let F be an M-dimensional positive Lévy copula and Ui , i = 1, . . . , d

tail integrals of real-valued Lévy processes. Then there exists a Rd -
valued Lévy process X whose components have tail integrals Ui , i =
1, . . . , d and whose marginal tail integrals satisfy equation (65) for any
nonempty index set {i1 , . . . , im } ⊆ {1, . . . , d} and any (x1 , . . . , xm ) ∈
(−∞, 0)m .

In terms of the Lévy copula F of X and its marginal tail integrals, formula
(59) can be rewritten as
M
X X
λm = (−1)k−m k
Cm Fi1 ,...,ik (Ui1 (β), . . . , Uik (β))
k=m 1≤i1 <···<ik ≤M

To compute the intensities λi and price the gap option, it is therefore suffi-
cient to know the individual gap intensities Ui (β) (M real numbers), which
can be estimated from 1-dimensional gap option prices or from the prices of
short-term put options and the Lévy copula F . This Lévy copula will typi-
cally be chosen in some suitable parametric family. One convenient choice is
the Clayton family of (positive) Lévy copulas defined by

−θ

−1/θ
F θ (u1 , . . . , uM ) = u−θ
1 + · · · + uM . (66)

The dependence structure in the Clayton family is determined by a single pa-

rameter θ > 0. The limit θ → +∞ corresponds to complete dependence (all
components jump together) and θ → 0 produces independent components.
The Clayton family has the nice property of being margin-stable: if X has

55
Clayton Lévy copula then all lower-dimensional margins also have Clayton
Lévy copula:

−1/θ
Fiθ1 ,...,im (u1 , . . . , um ) = u−θ −θ
1 + · · · + um .

For the Clayton Lévy copula, equation (59) simplifies to

M
X X
λm = (−1)k−m k
Cm (Ui1 (β)−θ + · · · + Uik (β)−θ )−1/θ .
k=m 1≤i1 <···<ik ≤M

This formula can be used directly for baskets of reasonable size (say, less than
20 names). For very large baskets, one can make the simplifying assumption
that all individual stocks have the same gap intensity: Uk (β) = U1 (β) for all
k. In this case, formula (59) reduces to the following simple result:

Proposition 15. Suppose that the prices of M underlyings follow an M -

dimensional exponential Lévy model with Lévy measure ν. If the individual
components of the basket are identically distributed and the dependence struc-
ture is described by the Clayton Lévy copula with parameter θ, the price of a
basket gap option as of definition 12 is given by
∞ M
! M
X X Y (λi T )ni
E[f (NT )] = e−λT f knk ,
n ,...,n =0 k=1 i=1
n i !
1 M

where
M −m
M
X (−1)j CjM −m
λm = U1 (β)Cm (67)
j=0
(m + j)1/θ

Figure 1 shows the behavior of the intensities λ1 , λ2 and λ10 as a function

of the dependence parameter θ in a basket of 10 names, with a single-name
gap probability of 1%. Note that formula (67) implies
(
0, m<M
lim λm =
θ→∞ U1 (β), m = M.
(
0, m>1
lim λm = ,
θ→0 M U1 (β), m = 1.

56
 0.030
lambda1
lambda2
0.025
lambda10

0.020

0.015

0.010

0.005

0.000
0 1 2 3 4 5 6 7 8 9 10
theta

Figure 1: The intensities λi of different jump sizes of the gap counting process
as a function of θ for M = 10 names and a single-name loss probability of
1%.

0.014 0.0020

0.012

0.0015
0.010

0.008
0.0010
0.006

0.004
0.0005

0.002

0.000 0.0000
0 1 2 3 4 5 6 7 8 9 10 0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20

Figure 2: Left: Expected loss of a multi-name gap option in the Credit Suisse
example as a function of the dependence parameter θ. The single-name loss
probability is 1%. Right: zoom of the right graph for small values of θ.

57
in agreement with the behavior observed in Figure 1.
Figure 2 shows the price of the multiname gap option of example 6 com-
puted using the formula (64). The price achieves a maximum for a finite
nonzero value of θ. This happens because for this particular payoff struc-
ture, the protection seller does not lose money if only 1 or 2 gap events occur
during the lifetime of the product, and only starts to pay after 3 or more
gap events. The probability of having 3 or more gap events is very low with
independent components.

11 Implied volatility
Recall the well-known Black-Scholes formula for call option prices:

C BS (t, St , T, K, σ) = St N (d1 ) − Ke−r(T −t) N (d2 ) (68)

St 2 Z u
log( Ke−rτ ) ± τ σ /2 1 z2
with d1,2 = √ and N (u) ≡ √ e− 2 dz,
σ τ 2π −∞
where τ = T − t. If all other parameters are fixed, (68) is an increasing
continuous function of σ, mapping (0, ∞) into ((St − Ke−rτ )+ , St ). The
latter interval is the greatest interval allowed by arbitrage bounds on call
option prices. Therefore, given the market price Ct∗ (T, K) of a call option,
one can always invert (68) and find the value of volatility parameter which,
when substituted into the Black-Scholes formula, gives the correct option
price:

∃! Σt (T, K) > 0 : C BS (t, St , T, K, Σt (T, K)) = Ct∗ (K, T ). (69)

This value is called the (Black-Scholes) implied volatility of the option. For
fixed (T, K), the implied volatility Σt (T, K) is in general a stochastic process
and, for fixed t, its value depends on the characteristics of the option such as
the maturity T and the strike level K: the function Σt : (T, K) → Σt (T, K)
is called the implied volatility surface at date t (see Figure 3). Using the log
moneyness k = log(K/St ) of the option, one can also represent the implied
volatility surface as a function of k and time to maturity: It (τ, k) = Σt (t +
τ, St ek ). From the independence and stationarity of increments of X, it
follows that the definition of implied volatility (69) is equivalent to
I2τ
E[(eXτ − ek−rτ )+ ] = E[(eIWτ − 2 − ek−rτ )+ ].

58
 60

50

40
80
30
60
20
40
10
20 1
1
0.8
0.8
0.6
0.6 1.5
0.4
0.4
1 1.5
T T 0.2
0.2 ïrT
Ke /S
0
1 KeïrT/S
0
0.5 0.5

Figure 3: Left: Profile of the implied volatility surface as a function of time to

maturity and moneyness for the Merton jump-diffusion model with σ = 15%,
δ = 1 and λ = 0.1. Right: Implied volatility surface as a function of time
to maturity and moneyness for the variance gamma model using parameters
taken from [48]. Note the flattening of the skew with maturity.

Since each side depends only on (τ, k) and not on t one concludes that in
exponential Lévy models, the implied volatility for a given log moneyness k
and time to maturity τ does not evolve in time: It (τ, k) = I0 (τ, k) := I(τ, k).
This property is known as the floating smile property.
In exponential Lévy models, the properties of the implied volatility sur-
faces can be characterized in terms of the asymptotic behavior of the surface
for large and small values of strike and maturity. We start with the large
and small strike behavior which was first analyzed by Roger Lee [45]; this
analysis was subsequently extended and made more precise by Benaim and
Friz [32, 31]. Their results, reviewed below, take a particularly simple form
in the case of Lévy processes, because the critical exponents do not depend
on time. Next, we study the short maturity asymptotics, where it turns out
that the behavior of the implied volatility is very different for out of the
money (OTM) and at the money (ATM) options. Below, we present some
original results for the two cases. Finally, the long-maturity asymptotics
were recently studied by Tehranchi [69, 68] and Rogers and Tehranchi [56].
We review their results in the case of Lévy processes, where once again, the
formulation is particularly simple and interesting links to the large deviations
theory and Cramér’s theorem can be made.

59
11.1 Large/small strikes
The limiting slope of time-rescaled implied variance as a function of log-strike
turns out to be related to the critical exponents of the moment generating
function of the log-price process X, defined by

qt∗ = − inf{u : E[euXt ] < ∞}, rt∗ = sup{u : E[euXt ] < ∞}.

It is clear that the interval [−qt∗ , rt∗ ] is nonempty, because E[euXt ] < ∞ at
least for all u ∈ [0, 1] by the martingale condition.

Proposition 16 (Implied volatility asymptotics at extreme strikes [31]). Fix

τ > 0 and suppose that rτ∗ ∈ (0, ∞) and qτ∗ ∈ (0, ∞) and that the moment
generating function blows up in a regularly varying way around its critical ex-
ponents (see [31] for a precise definition). Then the implied volatility I(τ, k)
satisfies

I 2 (τ, −k)τ I 2 (τ, k)τ

∼ ξ(qτ∗ ) and ∼ ξ(rτ∗ − 1), as k → +∞,
|k| k
√
where the function ξ is defined by ξ(x) = 2 − 4( x2 + x − x).

This proposition extends in a natural way to the case of infinite critical

exponents:

I 2 (τ, −k)τ k→+∞ I 2 (τ, k)τ k→+∞

−−−−→ 0 if qτ∗ = ∞ and −−−−→ 0 if p∗τ = ∞.
|k| k

This was shown already in the original work of Roger Lee [45].
For Lévy processes, the exponents q ∗ and r∗ do not depend on t and
are particularly easy to compute, since the moment generating function is
known from the Lévy-Khintchine formula. In particular, the models with
exponential tail decay of the Lévy measure such as variance gamma, normal
inverse Gaussian and Kou satisfy the necessary conditions for the proposition
16 and their critical exponents coincide with the inverse decay lengths: q ∗ =
λ− and r∗ = λ+ . Figure 4 shows that the asymptotic linear slope of the
implied variance as a function of log strike can be observed for values of k
which are not so far from zero, especially for short maturity options.
In Merton model, the tails of the Lévy measure are thinner than expo-
nential and the critical exponents q ∗ and r∗ are infinite. The remark after

60
 0.035
T=1 month
0.030 T=3 months
T=1 year
0.025

0.020

0.015

0.010

0.005

0.000
ï0.8 ï0.6 ï0.4 ï0.2 0.0 0.2 0.4 0.6 0.8

Figure 4: Smile asymptotics: implied variance multiplied by square root of

maturity as function of log-strike in the variance gamma model with param-
eters taken from [48]: σ = 0.1213, θ = −0.1436, κ = 0.1686.

Proposition 16 then only tells us that the limiting slope of the implied vari-
ance is zero, but other results in [32] allow to compute the exact asymptotics:
for the right tail we have
k
I 2 (τ, k)τ ∼ δ × √ , when δ > 0
k→∞ 2 2 log k
and
k
I 2 (τ, k)τ ∼ µ × , when δ = 0,
k→∞ 2 log k
where δ is the standard deviation of the jump size and µ is the mean jump.

11.2 Short maturity asymptotics

The short maturity behavior of implied volatility in exponential Lévy models
is very different from that observed in stochastic / local volatility models with
continuous paths. While in continuous models the implied volatility usually
converges to a finite nonzero value as τ → 0, in models with jumps the
implied volatility of out of the money or in the money options blows up. On
the other hand, the implied volatility of at-the-money options converges to
the volatility of the diffusion component as τ → 0; in particular it converges

61
to zero for pure jump models. This leads to very pronounced smiles for short
maturity options (in agreement with market-quoted smiles). The intuitive
explanation of this effect is that in most continuous models, the stock returns
at short time scales become close to Gaussian; in particular, the skewness
and excess kurtosis converge to zero as τ → 0. By contrast, in models with
jumps, the distribution of stock returns at short time scales shifts further
away from the Gaussian law; the skewness and kurtosis explode as √1τ and τ1
respectively.
The short maturity asymptotics of implied volatility smile in exponential
Lévy models can be computed by comparing the option price asymptotics
in the Black-Scholes model to those in the exponential Lévy model (many
results in this direction can be found in Carr and Wu [17]). To simplify the
developments, we suppose that the interest rate is zero. Then the normalized
Black-Scholes price satisfies
−k 1 √
cBS (τ, k, σ) = N (d1 ) − ek N (d2 ), d1,2 = √ ± σ τ .
σ τ 2
Using the asymptotic expansion of the function N [1], we get, for the ATM
options (k = 0):
√
σ τ
cBS (τ, k, σ) ∼ √ (70)
2π
and for other options
ek/2 3 3/2 − k22
cBS (τ, k, σ) ∼ √ σ τ e 2σ τ , (71)
k 2 2π
where the notation f ∼ g signifies fg → 1 as τ → 0.
In every exponential Lévy model satisfying the martingale condition, we
have [58]
Z
E[(e − e ) ] ∼ τ (ex − ek )+ ν(dx), for k > 0
Xτ k +
(72)
Z
E[(e − e ) ] ∼ τ (ek − ex )+ ν(dx), for k < 0
k Xτ +
(73)

From these estimates, the following universal result can be deduced: it con-
firms the numerical observation of smile explosion in exponential Lévy models
and gives the exact rate at which this explosion takes place.

62
Proposition 17 (Short maturity asymptotics: OTM options). Let X be a
Lévy process with Lévy measure ν satisfying supp ν = R. Then, for a fixed
log moneyness k 6= 0, the implied volatility I(τ, k) in the exponential Lévy
model St = S0 eXt satisfies

2I 2 (τ, k)τ log τ1

lim = 1. (74)
τ →0 k2
√
Proof. Suppose first that k > 0. It is clear that I(τ, k) τ → 0 as τ → 0
because otherwise the option price would not converge to 0. We then have,
from OTM Black-Scholes asymptotics (71):

cBS (τ, k, I(τ, k))

lim k2
= 1,
τ →0 −
C1 I(τ, k)3 τ 3/2 e 2I 2 (τ,k)τ

where C1 > 0 does not depend on τ . Denote the (normalized) call price in
the exponential Lévy model by c(τ, k). Under the full support hypothesis,
c(τ, k) ∼ C2 τ with C2 > 0 which once again does not depend on τ . By
definition of the implied volatility we then have
C2 τ
lim k2
= 1.
τ →0 −
C1 I(τ, k)3 τ 3/2 e 2I 2 (τ,k)τ

Taking the logarithm gives

1 k2
lim {log(C2 /C1 ) + 3 log I(τ, k) + log τ − 2 } = 0.
τ →0 2 2I (τ, k)τ

Now, knowing that I 2 (τ, k)τ → 0, we can multiply all terms by I 2 (τ, k)τ :

3 k2
lim {I 2 τ log(C2 /C1 ) + I 2 τ log(I 2 τ ) − I 2 τ log τ − } = 0.
τ →0 2 2
Since the first two terms disappear in the limit, this completes the proof in
the case k > 0. The case k < 0 can be treated in a similar manner using put
options.
For ATM options, the situation is completely different, from estimate (70)
we will deduce that the implied volatility does not explode but converges to
the volatility of the diffusion component.

63
Proposition 18 (Short maturity asymptotics: ATM options).

1. Let X be a Lévy process R without diffusion component and with Lévy

measure ν satisfying |x|≤1 |x|ν(dx) < ∞. Then, the ATM implied
√
volatility I(τ, 0) in the exponential Lévy model St = S0 eXt falls as τ
for short maturities:

I(τ, 0)
lim √ R R
= 1.
τ →0 2πτ max (ex − 1)+ ν(dx), (1 − ex )+ ν(dx)

2. Let X be a Lévy process with characteristic exponent

ψ(u) = iγu − |u|α f (u)

for 1 < α < 2 and some continuous bounded function f satisfying

lim f (u) = c+ , lim f (u) = c− , 0 < c1 , c2 < ∞.

u→+∞ u→−∞

This includes in particular stable and tempered stable processes with

1 < α < 2. Then, the ATM implied volatility I(τ, 0) in the exponential
Lévy model St = S0 eXt falls as τ 1/α−1/2 for short maturities:

I(τ, 0)
lim √ = 1.
τ →0 Cτ 1/α−1/2 2π

1/α 1/α
with C = Γ(1 − 1/α)(c+ + c− ).

3. Let X be a Lévy process withR a diffusion component with volatility σ

and Lévy measure satisfying x2 ν(dx) < ∞. Then the ATM implied
volatility I(τ, 0) in the exponential Lévy model St = S0 eXt converges to
σ as τ → 0.

The short-maturity smile asymptotics are illustrated in Figure 5: the

ATM implied volatility converges to the value of σ and the out of the money
and in the money volatilities eventually become very large as τ approaches
zero.
Proof.

64
 0.55
K=1
0.50 K=1.1

0.45 K=0.9

0.40

0.35

0.30

0.25

0.20

0.15

0.10
0.0 0.5 1.0 1.5

Figure 5: Smile asymptotics: term structure of implied volatility for different

strikes in Merton’s model with parameters σ = 0.1, λ = 2, µ = −0.1 and
δ = 0.5.

1. Let b denote the drift of X. Since X is a finite-variation process, the

Itô-Tanaka formula applied to the function (1 − eXτ )+ does not yield a
local time term, and we obtain

E[(1 − eXτ )+ ]
 Z τ Z τ Z 
Xt Xt +x + Xt +
=E b e 1Xt ≤0 dt + ν(dx){(1 − e ) − (1 − e ) } dt.
0 0 R

By L’Hopital’s rule,

1
lim E[(1 − eXτ )+ ]
τ →0 τ
Z 
Xτ Xτ +x + Xτ +
= b lim E[e 1Xτ ≤0 ]+lim E ν(dx){(1 − e ) − (1 − e ) } .
τ →0 τ →0 R

From Theorem 43.20 in [59], Xtt → b almost surely as t → 0. From

this we deduce that limτ →0 E[eXτ 1Xτ ≤0 ] = 1b≤0 . Using the dominated
convergence for the second term above, we finally obtain
Z
1
lim E[(1 − e ) ] = b1b≤0 + ν(dx)(1 − ex )+ .
Xτ +
τ →0 τ R

65
 Since by the martingale condition,
Z
b + (ex − 1)ν(dx) = 0,
R

this limit can be rewritten as

Z Z
1
lim E[(1 − eXτ )+ ] = max( x +
(e − 1) ν(dx), (1 − ex )+ ν(dx)).
τ →0 τ

Comparing this expression with the Black-Scholes ATM asymptotics,

we obtain the desired result.

2. Let pt denote the density of Xt (which exists and is square integrable

under the hypotheses of this part). The ATM call option price is given
by Z
c(τ, 0) = (ex − 1)+ pt (x)dx.

Let us fix a constant β < −1 and define

Z
c̃(τ, 0) = (ex − 1)+ eβx pt (x)dx.

Then it follows from results in [58] that

Z
|c(τ, 0) − c̃(τ, 0)| = (ex − 1)+ (1 − eβx )pt (dx) = O(τ )

as τ → 0. This means that it is sufficient to study the decay properties

of c̃. This function is a scalar product of the square integrable func-
tion pt by the square integrable function (ex − 1)+ eβx , whose Fourier
transform is given by
Z ∞
1 1 1
eiux (e(1+β)x −eβx )dx = − =− .
0 iu + β iu + β + 1 (u − iβ)(u − iβ − i)

By the Plancherel theorem we then have

α
eτ ψ(u) eiτ γu−τ |u| f (u)
Z Z
1 1
c̃(τ, 0) = − du = − du.
2π (u − iβ)(u − i − iβ) 2π (u − iβ)(u − i − iβ)

66
 On the other hand, direct computation using Cauchy’s integral formula
(for γ 6= 0) or elementary calculus (for γ = 0) shows that

eiγuτ du
Z
= 2π(eγτ β − eγτ (β+1) )1γ<0 = O(τ )
(u − iβ)(u − i − iβ)

as τ → 0. Then, changing the variable of integration, we obtain

1−1/α α −1/α
τ 1/α eiγzτ (1 − e−|z| f (τ z)
Z
)
c̃(τ, 0) = 1/α 1/α
dz + O(τ ).
2π (z − iτ β)(z − iτ (1 + β))

The dominated convergence theorem then yields

Z 0 α Z ∞ α
−1/α 1 1 − e−c− |z| 1 1 − e−c+ |z|
τ c̃(τ, 0) → dz + dz
2π −∞ z2 2π 0 z2

as τ → ∞. This result generalizes the findings of Carr and Wu [17].

Computing the integrals and comparing the result to the Black-Scholes
at the money asymptotics, we obtain the final result.

3. Under the conditions of this part, we can write the characteristic expo-
nent of X as ψ(u) = iγu − f (u)u2 for a continuous bounded function
2
f satisfying limu→∞ f (u) = σ2 . Then, exactly as in the previous part,
the dominated convergence theorem yields
σ 2 u2
1 − e−
Z
c̃(τ, 0) 1 2 σ
√ → du = √ ,
τ 2π R u2 2π
which is equal to the Black-Scholes ATM asymptotics.

11.3 Flattening of the smile far from maturity

As the time to maturity τ goes to infinity, the implied volatility I(τ, k) in
an exponential Lévy model converges to a constant value I(∞) which does
not depend on k (see Figures 5 and 3). As a result, the implied volatility
smile flattens for long maturities, a phenomenon which is also observed in
the options markets, although with a slower rate. This flattening has been
often attributed (e.g. [16]) to the central limit theorem, according to which,

67
 √process with finite variance, the distribution of increments (Xτ −
for a Lévy
E[Xτ )])/ τ becomes approximately Gaussian as τ goes to infinity. However,
contrary to this intuition, the flattening of the smile is not a consequence of
the central limit theorem, but, rather, of a “large deviation” principle which
governs the tail behavior of the sample average of n i.i.d. random variables.
In fact, as observed by Rogers and Tehranchi [56], the implied volatility
flattens even in models where log-returns have infinite variance such as the
finite moment log-stable process of [16].
To understand this, consider a Lévy process X with E[X1 ] < ∞. Since in
a risk-neutral model E[eXt ] = 1, the Jensen inequality implies that E[Xt ] < 0
for all t. Therefore, by the law of large numbers, Xt → −∞ almost surely
as t → ∞, which means that eXt → 0 a.s. The exercise of a long-dated call
option is thus an event with a very small probability. The probability of
such rare events is given by Cramér’s theorem, which is the cornerstone of
the theory of large deviations, rather than by the CLT.
The normalized price of a call option with log-moneyness k can be written
as
c(τ, k) = E(eXτ − ek )+ = P̃[Xτ ≥ k] − ek P[Xτ ≥ k],
where we introduce the new probability P̃ via the Esscher transform:

dP|˜Ft
:= eXt .
dP|Ft

Denote α = E[X1 ] and α̃ = Ẽ[X1 ]. An easy computation using Proposition

9 shows that
σ2
Z
α = − − (ex − x − 1)ν(dx) < 0
2
2 Z R
σ
α̃ = + (xex − ex + 1)ν(dx) > 0.
2 R

To make the probability of a rare event appear, we rewrite the option price
as
   
Xτ − α̃τ k k Xτ − ατ k
c(τ, k) = 1 − P̃ − > α̃ − −e P ≥ −α + .
τ τ τ τ
These probabilities can be estimated with the help of the famous Cramér’s
theorem which gives the exact convergence rate in the law of large numbers.

68
Theorem 4 (Cramér). Let {Xi }i≥1 be an i.i.d. sequence of random variables
with E[Xi ] = 0 for all i. Then for all x ≥ 0
" n #
1 1 X
lim log P Xi ≥ x = −I(x),
n→∞ n n i=1

where I(x) is the Fenchel transform of the log-Laplace transform of X1 :

I(x) = sup(θx − l(θ)), l(θ) = log E[eθX1 ].
θ

Suppose that the Lévy measure ν is such that1

Z Z
xν(dx) < ∞ and xex ν(dx) < ∞ (75)
|x|>1 |x|>1

and define the log-Laplace transforms by

˜l(θ) := log Ẽ[e−θ(X1 −α̃) ] and l(θ) := log E[eθ(X1 −α) ],

and the respective Fenchel transforms by I˜ and I. A direct computation then

shows that
 2 Z 
˜ σ 2 θx x
I(α̃) = I(−α) = sup (θ − θ ) − (e − θe − 1 + θ)ν(dx) ,
θ 2 R

and that the functions I˜ and I are finite and hence, continuous, in the neigh-
borhood of, respectively, α̃ and −α. Hence the sup above can be restricted
to the interval θ ∈ [0, 1], since the function being maximized is concave and
equal to 0 for θ = 0 and θ = 1. Using Cramér’s theorem and the continuity
of I˜ and I, we then obtain
 2 Z 
1 σ 2 θx x
lim log(1 − c(τ, k)) = sup (θ − θ ) − (e − θe − 1 + θ)ν(dx) .
τ →∞ τ θ∈[0,1] 2 R
(76)
Note that this formula is valid for any k, we can even take k to be a function
of τ as long as k = o(τ ) as τ → ∞. Specializing this formula to the Black-
Scholes model, where ν ≡ 0 and the sup can be computed explicitly, we
get
1 σ2
lim log(1 − cBS (τ, k, σ)) = .
τ →∞ τ 8
1
The finite moment log stable process of Carr and Wu [17] satisfies these hypotheses
although the variance of the log-price is infinite in this model.

69
From Equation (76), it follows in particular that the implied volatility sat-
isfies τ I 2 (τ, k) → ∞ as τ → ∞ (otherwise the call option price would not
converge to 1). Since in the Black-Scholes model the option price depends
only on τ σ 2 but not on τ or σ separately, we can write
1 1
lim log(1 − cBS (τ I 2 (τ, k), k, 1)) = ,
τ →∞ τ I 2 (τ, k) 8

and combining this with (76), we obtain the final result:

Proposition 19 ([69]). Let X be a Lévy process with Lévy measure satisfying

(75). Then the implied volatility I(τ, k) in the exponential Lévy model St =
S0 eXt satisfies
 2 Z 
2 σ 2 θx x
lim I (τ, k) = 8 sup (θ − θ ) − (e − θe − 1 + θ)ν(dx) . (77)
τ →∞ θ 2 R

The exact formula (77) for the limiting long-term implied volatility in
an exponential Lévy model is difficult to use in practice: even if for some
models such as variance gamma it yields a closed form expression, it is rather
cumbersome. However, for small jump sizes, Taylor expansion shows that this
expression is not very different from the total variance of the Lévy process:
Z
I (∞, k) ≈ σ + x2 ν(dx).
2 2

The smile flattening in exponential Lévy models has thus little to do with
the so called aggregational normality of stock returns. One may think that
the implied volatility converges to its limiting value faster for Lévy processes
to which the central limit theorem applies. However, the results of Rogers
and Tehranchi [56] suggest otherwise: they give the following upper bound,
valid in exponential Lévy models as soon as E[|Xt | < ∞], for the rate of
convergence of the implied volatility skew to zero:
 I(τ, k2 )2 − I(τ, k1 )2 
 
lim sup sup τ   ≤ 4, 0 < M < ∞.
τ →∞ k1 ,k2 ∈[−M,M ] k2 − k1 

∂I(τ,k)
See also [68] for explicit asymptotics of the the derivative ∂k
as τ → ∞.

70
12 Hedging in exponential Lévy models
Exponential Lévy models generally correspond to incomplete markets, mak-
ing exact replication impossible. Hedging must therefore be interpreted as
approximation of the terminal pay-off with an admissible portfolio. The
usual practice is to minimize the expected squared deviation of the hedg-
ing portfolio from the contingent claim, an approach known as quadratic
hedging. The resulting strategies are often explicitly computable and, more
importantly, they are linear, because the hedging portfolios can be inter-
preted as orthogonal projections of contingent claims onto the closed linear
subspace of hedgeable portfolios. To hedge a book of options written on the
same underlying, a trader can therefore compute the hedge ratio for every
option in the book and then add them up, just like this is typically done
with delta hedging. This greatly reduces the computational cost of hedging
and is an important advantage of quadratic hedging compared to other, e.g.,
utility-based approaches.
To define the criterion to be minimized in a mean square sense, two ap-
proaches are possible. In the first approach [12, 52, 39], the hedging strategy
is supposed to be self-financing, and one minimizes the quadratic hedging er-
ror at maturity, that is, the expected squared difference between the terminal
value of the hedging portfolio and the option’s pay-off:
Z T Z T
2 0 0
inf E[ |VT (φ) − H| ] where VT (φ) = V0 + φt dSt + φt dSt , (78)
V0 ,φ 0 0

where S 0 is the risk-free asset. If the interest rate is constant, we can choose
the zero-coupon bond with maturity T as the risk-free asset: St0 = e−r(T −t)
and after discounting this problem becomes:
Z T
2
inf E[|VT (φ) − H| ], where VT = V0 + ˆ φt dŜt .
V̂0 ,φ 0

In the second approach [30, 29, 61, 64], strategies that are not self-
financing are allowed, but they are required to replicate the option’s pay-off
exactly: VT (φ) = H. In an incomplete market, this means that the option’s
seller will have to continuously inject / withdraw money from the hedging
portfolio. The cumulative amount of funds injected or withdrawn is called
the cost process. It is given by
Ct (φ) = Vt (φ) − Gt (φ),

71
where
Vt (φ) = φ0t St0 + φt St
and G is the gain process given by
Z t Z t
0 0
Gt = φs dSs + φs dSs .
0 0

The discounted cost process is then given by

Z t
0
Ĉt = φt + φt Ŝt − φs dŜs .
0

The risk-minimizing strategy, as introduced by Föllmer and Sondermann [30],

is a strategy which replicates the option’s pay-off, and has the cost process
which varies as little as possible, that is, this strategy minimizes, at each
date t, the residual cost given by
E[(ĈT − Ĉt )2 |Ft ]. (79)
over all admissible continuations of the strategy from date t onwards. The
risk-minimizing strategy always exists in the martingale case (when the dis-
counted stock price is a martingale), but in the general case, it may fail to
exist even in the most simple examples [61]. Motivated by this difficulty,
Föllmer and Schweizer [29] introduced the notion of locally risk minimizing
strategy, which corresponds to finding the extremum of (79) with respect
to suitably defined small perturbations of the strategy, or, in other words,
measuring the riskiness of the cost process locally in time. Local risk mini-
mization is discussed in detail in section 12.2.
The expectations in (79) and (78) are taken with respect to some proba-
bility which we have to specify. To begin, let us assume that we have chosen
a martingale measure Q and the expectations in (78) and (79) are taken with
respect to Q. In particular, Ŝ is a martingale under Q. Assume now that
H ∈ L2 (Ω, F, Q) and Ŝ is also square-integrable. If we consider portfolios of
the form:
Z T
S = {φ caglad predictable and E| φt dŜt |2 < ∞ } (80)
0

then the set A of attainable pay-offs is a closed linear subspace of L2 (Ω, F, Q),
and the quadratic hedging problem becomes an orthogonal projection:
inf E|VT (φ) − H|2 = inf kH − Ak2L2 (Q) . (81)
V̂0 ,φ A∈A

72
The solution is then given by the well-known Galtchouk-Kunita-Watanabe
decomposition [43, 33], which states that any random variable H ∈ L2 (Ω, F, Q)
can be represented as
Z T
H = E[H] + φH H
t dŜt + NT , (82)
0

where (NtH ) is a square integrable martingale orthogonal to Ŝ. The optimal

hedging strategy is then given by φH and the initial cost of the hedging
portfolio is V0 = e−r(T −t) E[H].
Introducing the martingale Ĥt := E[H|Ft ] generated by H, we have
Z t
Ĥt = E[H] + φH H
s dŜs + Nt ,
0

and the orthogonality implies

Z ·
hĤ − φH
s dŜs , Ŝi ≡ 0,
0

which means that the optimal hedge ratio may be expressed more explicitly
using the predictable covariation of the option price and the stock price:

dhĤ, Ŝit
φH
t = . (83)
dhŜ, Ŝit
In the martingale setting, optimizing the global hedging error (78) we
obtain a strategy which is also risk minimizing in the sense of equation (79).
For any strategy φ, we have
" Z T 2  #
E[(ĈT − Ĉt )2 |Ft ] = (Ĥt − V̂t )2 + E H − Ĥt − φs dŜs Ft

t 
"Z
T 2  #
= (Ĥt − V̂t )2 + E[(NT − Nt )2 |Ft ] + E (φs − φHs )dŜs Ft .

t 

To minimize this expression, we clearly need to take φ = φH and choose φ0

such that V̂t = Ĥt for all t. In this case, the discounted cost process is given
by Z t
Ĉt = V̂t − φH H
s dŜs = E[H] + Nt .
0

73
 We shall see in section 12.2 that in the martingale setting, the strategy
H
φ which minimizes the terminal hedging error also coincides with the locally
risk minimizing strategy of Föllmer and Schweizer [29]. Moreover, it is often
easy to compute in terms of option prices. This is no longer true if Ŝ is not
a martingale. However using the risk-neutral second moment of the hedging
error as a criterion for measuring risk is not very natural: Q represents a
pricing rule and not a statistical description of market events, so the profit
and loss (P&L) of a portfolio may have a large variance while its “risk neu-
tral” variance can be small. Nevertheless, to estimate the expected return
of a stock, and therefore, to distinguish it from a martingale, one needs his-
torical stock return observations covering an extended period of time, often
exceeding the lifetime of the option. Option hedging, on the other hand, is a
“local” business, where one tries to cancel out the daily movements of option
prices with the daily movements of the underlying and locally, every stock
behaves like a martingale. Without contributing to this ongoing argument,
we review both approaches in the next two sections.

12.1 Quadratic hedging in exponential-Lévy models un-

der the martingale measure
Although the quadratic hedging problem is “solved” by the Galtchouk-Kunita-
Watanabe decomposition, from a practical point of view the problem is of
course to compute the risk minimizing hedge φH H
t . Formulas for φt with var-
ious degrees of explicitness and under various assumptions on the driving
process X and on the pay-off G were given in [12, 28, 9, 36] and several other
papers. In particular [24] provide the expressions for hedge ratios in the case
when the hedging portfolio itself contains options. In the case of European
pay-offs and exponential Lévy models, the problem was solved in [39] using
Fourier analysis techniques. Their method, reviewed in section 12.2 covers
the general case as well as the martingale case. In this section, we provide
another Fourier-based result, which is specialized to the martingale setting
but works under different regularity assumptions on the pay-off than in [39],
which include, for instance, digital options.
Proposition 20 (Quadratic hedge in exponential-Lévy models, martingale
case). Let X be a Lévy process with Lévy measure ν, diffusion coefficient σ,
and characteristic function Φ, such that eX is a martingale and assume:
i. The log-payoff function satisfies the conditions (37) and (38).

74
 ii. The integrability condition (39) holds for all t < T .
iii. The Lévy measure of X satisfies
Z
e2(x∨Rx) ν(dx) < ∞. (84)
|x|>1

Then the optimal quadratic hedging for a European option with pay-off G(ST )
at date T in an exponential Lévy model St = S0 ert+Xt amounts to holding a
position in the underlying
Z
1 R−iu−1
φt = ĝ(u + iR)ΦT −t (−u − iR)Ŝt− Υ(R − iu)du (85)
2π R
κ(y + 1) − κ(y) − κ(1)
where Υ(y) = , and κ(z) := log E[ezX1 ], (86)
κ(2) − 2κ(1)
or, equivalently, φt = φ(t, St− ) where:
σ 2 ∂P (t, S) + S1 ν(dz)(ez − 1)[P (t, Sez ) − P (t, S)]
R
∂S
φ(t, S) = R (87)
σ 2 + (ez − 1)2 ν(dz)
with P (t, S) = e−r(T −t) E Q [G(ST )|St = S] the option price at date t when the
underlying is at the level S.
Remark 1. Condition (84), which is the only assumption imposed in addition
to those of Proposition 10, guarantees that both the price process St and the
option pay-off G(ST ) are square integrable.
Proof. By Itô formula, the discounted stock price dynamics is given by
Z T Z TZ
ŜT = Ŝ0 + Ŝt σdWt + Ŝt (ex − 1)J˜X (dt × dx). (88)
0 0 R

To prove the proposition using the formula (83), we now need to obtain
a similar integral representation for the option’s discounted price function
P̂ (t, St ) = er(T −t) P (t, St ).
Let t < T . Applying the Itô formula under the integral sign in (40), we
find
Z Z t
1
P̂ (t, St ) − P̂ (0, S0 ) = duĝ(u + iR) ΦT −s (−u − iR)(R − iu)ŜsR−iu σdWs
2π R 0
Z Z t Z
1
+ duĝ(u + iR) R−iu
ΦT −s (−u − iR)Ŝs− (e(R−iu)z − 1)J˜X (ds × dz).
2π R 0 R
(89)

75
Let us first assume that σ > 0 and study the first term in the right-hand side
of (89), which can be written as
Z Z t
µ(du) Hsu dWs
R 0

where

µ(du) = |ĝ(u + iR)ΦT −t (−u − iR)|du (90)

is a finite positive measure on R and

σĝ(u + iR)ΦT −s (−u − iR)
Hsu = (R − iu)ŜsR−iu
2π|ĝ(u + iR)ΦT −t (−u − iR)|

By the Fubini theorem for stochastic integrals (see [53, page 208]), we can
interchange the two integrals in (90) provided that
Z t
E µ(du)|Hsu |2 ds < ∞ (91)
0

Under the assumption (39) it is easy to check that

ΦT −s (−u − iR)
≤C
|ΦT −t (−u − iR)|
for all s ≤ t ≤ T for some constant C > 0 which does not depend on s and
t. To prove (91) it is then sufficient to check
Z tZ
E |ĝ(u + iR)ΦT −t (−u − iR)||Ŝs2(R−iu) |2 (R − iu)2 dudt < ∞
0 R

which holds because

σ 2 u2
|ΦT −t (−u − iR)| ≤ Ce−(T −t) 2 (92)

Therefore, the first term on the right-hand side of (89) is equal to

Z t Z
σ
σ̃s dWs , σ̃s = duĝ(u + iR)ΦT −s (−u − iR)(R − iu)ŜsR−iu . (93)
0 2π R

This also shows that σ̃s = σSs ∂P ∂S

(s,Ss )
.

76
 Let us now turn to the second term in the right-hand side of (89). Here
we need to apply the Fubini theorem for stochastic integrals with respect to
a compensated Poisson random measure [4, Theorem 5] and the applicability
condition boils down to
Z tZ Z
2(R−iu) 2
E |ĝ(u + iR)ΦT −t (−u − iR)||Ŝs | |e(R−iu)z − 1|2 ν(dz)dudt < ∞
0 R R

If σ > 0, this is once again guaranteed by (92), and when σ = 0,

Z
|e(R−iu)z − 1|2 ν(dz) = ψ(−2iR) − 2<ψ(−u − iR).
R

Since, for some C < ∞,

T −t
|<ψ(−u−iR)ΦT −t (−u−iR)| = |<ψ(−u−iR)e(T −t)<ψ(−u−iR) | ≤ Ce 2
<ψ(−u−iR)
,
the integrability condition is satisfied and we conclude that
Z t Z tZ
P̂ (t, St ) − P̂ (0, S0 ) = σ̃s dWs + γ̃s (z)J˜X (ds × dz) (94)
0 0 R

for all t < T with σ̃ as above and

Z
1 R−iu (R−iu)z
γ̃s (z) = duĝ(u + iR)ΦT −s (−u − iR)Ŝs− (e − 1) (95)
2π R
= P̂ (s, Ss− ez ) − P̂ (s, Ss− ).
The optimal (risk-minimizing) hedge is obtained from formula (83):
R
σ Ŝt σ̃t + Ŝt R ν(dz)(ez − 1)γ̃t (z)
φ̂t = R .
Ŝt2 (σ 2 + R (ez − 1)2 ν(dz))
Substituting the expressions for σ̃ and γ̃ in terms of option prices into the
above expression, we obtain (87) directly. On the other hand, the Fourier
representations (93) and (95) and an application of Fubini’s theorem yield
(85).
As a by-product, the martingale representation (94) also yields the ex-
pression for the residual risk of a hedging strategy:
Z T Z  2 
2 z z
E[(φ) ] = E dt ν(dz) P̂ (t, St− e ) − P̂ (t, St− ) − Ŝt− φt (e − 1)
0 R
"Z 2 #
T 
2 ∂P
+ E Ŝt− φt − (t, St− ) σ 2 dt . (96)
0 ∂S

77
This allows us to examine whether there are any cases where the hedging
error can be reduced to zero, i.e., where one can achieve a perfect hedge for
every option and the market is complete. Hedging error is zero if and only
if, for almost all t, there exists k ∈ R with:
∂P
(σSt , (P (t, St ez ) − P (t, St ))z∈supp ν ) = k(σSt , (St (ez − 1))z∈supp ν )
∂S
This is only true in two (trivial) cases:
• The Lévy process X is a Brownian motion with drift: ν = 0 and we
retrieve the Black-Scholes delta hedge
∂P
φt = ∆BS (t, St ) = (t, St ).
∂S
• The Lévy process X is a Poisson process with drift: σ = 0 and there is
a single possible jump size: ν = δx0 (x). In this case the hedging error
equals
Z T  2 
x0 x0
E dt P̂ (t, St− e ) − P̂ (t, St− ) − Ŝt− φt (e − 1)
0

so by choosing
P (t, St− ex0 ) − P (t, St− )
φt =
St− (ex0 − 1)
we obtain a replication strategy.
In other cases, the market is incomplete (an explicit counter-example may
be constructed using power option with pay-off HT = (ST )α ).

Delta-hedging vs. optimal strategy We see that the optimal strategy

(87) can be represented as a weighted average of the delta hedging ∂P ∂S
and
a certain integral involving the sensitivities of the option price to various
possible jumps. But how far is the optimal strategy from the pure delta
hedging? To answer this question, if option prices are regular (e.g. when
σ > 0) and jumps are small, we can perform a Taylor expansion with respect
to the jump size in equation (87), obtaining
S ∂ 2P
Z
∂P
∆(t, S) = + 2 2
ν(dz)(ez − 1)3 .
∂S 2Σ ∂S

78
1.0 1.0
Delta Delta
0.9 0.9
Optimal Optimal
0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0.0 0.0
0.80 0.85 0.90 0.95 1.00 1.05 1.10 1.15 1.20 0.80 0.85 0.90 0.95 1.00 1.05 1.10 1.15 1.20

Figure 6: Hedge ratios for the optimal strategy of proposition 20 and the
delta hedging strategy as function of stock price S. Left: hedging with stock
in Kou model: the optimal strategy introduces a small asymmetry correction
to delta hedging. Right: variance gamma model close to maturity (2 days):
the optimal strategy is very far from delta hedging.

where Z
2 2
Σ =σ + (ez − 1)2 ν(dz).

Typically in equity markets the jumps are negative and small, therefore
∆(t, S) < ∂P
∂S
and the optimal strategy represents a small (of the order of
third power of jump size) asymmetry correction. This situation is repre-
sented in Figure 6, left graph. On the other hand, for pure-jump processes
such as variance gamma, we cannot perform the Taylor expansion, because
2
the second derivative ∂∂SP2 may not even exist, and the correction may there-
fore be quite large (see Figure 6, right graph).

How big is the hedging error? To answer this question, we simulated

the terminal value of the hedging portfolio and that of the option’s payoff
over 10000 trajectories for different strategies and different parameter sets.
In the first case study, Kou model with parameters estimated from market
data (MSFT) during a calm period was used, and the option to hedge was a
European put with strike K = 90% of the spot price and time to maturity
T = 1 year. The hedging errors are given in Table 1 and the left graph in
Figure 7 shows the P&L histograms. For this parameter set, the optimal

79
 Strategy Root of mean squared error
Delta hedging 0.0133
Optimal quadratic 0.0133
Delta hedging in Black-Scholes model 0.0059
(error due to discrete hedging)
No hedging 0.107
Table 1: Hedging errors for different strategies in Kou model expressed
in percentage of the initial stock price. Model parameters were estimated
from MSFT time series. The “Black-Scholes” strategy corresponds to delta-
hedging in the Black-Scholes model with equivalent volatility.

strategy is very close to delta hedging, and consequently, the hedging error
is the same for delta hedging as for the optimal strategy. On the other hand,
this error is very low, it is only twice as big as what we would get in the
Black and Scholes model with equivalent volatility (this error in the Black-
Scholes model is due to the fact that in the simulations, the portfolio is only
rebalanced once a day and not continuously).
In the second case study, Kou model with unfrequent large negative jumps
(10%) was used, and we wanted once again to hedge an OTM European put
(K = 90%, T = 1). The hedging errors are given in Table 12.1 and the
P&L histograms in Figure 7, right graph. Here we see that first, the optimal
strategy has a much better performance than delta-hedging, and second,
even this performance may not be sufficient, since the residual error is still
of order of 4% of the initial stock price. This means that in this context,
the market is “strongly incomplete” and hedging with stock only does not
allow to make the risk at terminal date sufficiently small. In this case, to
improve the hedging performance, one can include additional liquid assets,
such as options on the same underlying, or variance swaps, into the hedging
portfolio.

12.2 Quadratic hedging in exponential Lévy models

under the historical measure
Throughout this section, to simplify notation, we suppose that the interest
rate is equal to zero; the formulas for the general case can be obtained by
working with discounted claims. Let S be the price process of the underlying,

80
 Strategy Root of mean squared error
Delta-hedging 0.051
Optimal quadratic 0.041
No hedging 0.156
Table 2: Hedging errors for different strategies in Kou model expressed in
percentage of the initial stock price. A parameter set ensuring the presence
of large negative jumps was taken.

100 25
Deltaïhedging Deltaïhedging
90
Optimal hedging Optimal hedging
80 BlackïScholes 20

70

60 15

50

40 10

30

20 5

10

0 0
ï0.05 ï0.04 ï0.03 ï0.02 ï0.01 0.00 0.01 0.02 0.03 0.04 0.05 ï0.20 ï0.15 ï0.10 ï0.05 0.00 0.05 0.10

Figure 7: Histograms of the residual hedging error in Kou model. Left:

parameters estimated from MSFT time series. This graph shows the residual
hedging error in Kou model with the optimal quadratic strategy (solid line),
in Kou model with the delta-hedging strategy (dashed line) and in the Black-
Scholes model with the delta-hedging strategy (dash-dot line). In the latter
case, the error is only due to discrete-time hedging, and this curve was include
to assess the magnitude of the discretization error for other tests. Right:
strong negative jumps.

81
and suppose that it can be written in the form
Z t
St = S0 + Mt + αs dhM is (97)
0

for some square integrable martingale M and some predictable process α.

If
R S is 2xan exponential of a Lévy process X with Lévy measure ν satisfying
|x|>1
e ν(dx) < ∞ and diffusion coefficient σ, which can be written as
Z t Z t Z tZ
St = S0 + γSu du + Su σdWu + Su− (ez − 1)J˜X (du × dz), (98)
0 0 0 R

then the representation (97) holds with

Z t Z tZ
Mt = Su σdWu + Su− (ez − 1)J˜X (du × dz)
Z0 t  Z 0 R 
2 2 z 2
hM it = Su σ + (e − 1) ν(dz) du
0 R
γ
αt = R

St σ + R (ez − 1)2 ν(dz)
2

We then introduce the so-called mean-variance tradeoff process

Z t
Kt := αs2 dhM is .
0

In an exponential Lévy model, the mean-variance tradeoff is deterministic:

γ 2t
Kt = R .
σ 2 + R (ez − 1)2 ν(dz)

Local risk minimization The locally risk minimizing strategy [29, 62]
is a (not necessarily self-financing) trading strategy whose discounted cost
process Ĉ is a martingale orthogonal to M . This strategy is optimal in the
sense that we eliminate all the risk associated to the underlying with hedging,
and the only part of risk that remains in the cost process is the risk which
is orthogonal to the fluctuations of the underlying, and hence, cannot be
hedged with it. If the market is complete, then all risk is explained by the
underlying and the cost process of a locally minimizing strategy becomes

82
constant, that is, the strategy becomes self-financing. As already mentioned,
the locally risk minimizing strategy also has the interpretation of minimizing
the residual risk (79) with respect to suitably defined small perturbations
of the strategy [62]. Since the cost process is nonconstant, the locally risk
minimizing strategy is not a self-financing strategy in general however since
C is a martingale with mean zero this strategy is self-financing on average.
The locally risk minimizing strategy is closely related to an extension
of the Kunita-Watanabe decomposition to semimartingale setting, known as
the Föllmer-Schweizer decomposition [29, 61, 64, 65].
Definition 13. Let H ∈ L2 (P) be a contingent claim. A sum H = H0 +
RT H
0
φu dSu + LH
T is called the Föllmer-Schweizer decomposition of H if H0
is F0 -measurable, φH is an admissible trading strategy and LH is a square
integrable martingale with LH0 = 0, orthogonal to M .

Given a Föllmer-Schweizer decomposition for the claim H, the locally risk

minimizing strategy for H can be constructed by taking φt = φH t for all t,
0 H
R t H is Ct = H0 + Lt for all t, which
and choosing φ such that the cost process
amounts to φ0t = H0 + LHt − φ H
t St − φ dSu .
0 u

Relationship withR · the minimal martingale measure Define a process

Z via Z := E(− 0 αs dMs ) and assume that Z is a strictly positive square
dQ|
integrable martingale. Then we can define a new measure QM by dP|FFt := Zt .
t
By Girsanov-Meyer theorem ([53], theorem 36 in chapter 3), we have that
(i) QM is a martingale measure, that is, S becomes a martingale under Q
and (ii) any square integrable martingale which is orthogonal to M under P
remains a martingale under Q (although it may no longer be orthogonal to
M ). This measure is known as the minimal martingale measure [2, 64].
The minimal martingale measure allows to express the Föllmer-Schweizer
decomposition in a more explicit form. First, compute the process LH :
Z t
H QM H QM
Lt = E [LT |Ft ] = E [H|Ft ] − H0 − φH
u dSu .
0

Since LH
0 = 0, the initial capital for the Föllmer-Schweizer strategy is H0 =
M M
E Q [H]. Let HtM := E Q [H|Ft ]. The orthogonality condition under P then
yields an analogue of formula (83):
dhH M , SiPt
φH
t = .
dhS, SiPt

83
 In models with jumps, the minimal martingale measure does not always
exist as a probability measure (but may turn out to be a signed measure).
In an exponential-Lévy model of the form (98), the density of the minimal
martingale measure simplifies to Z = E(U ) with
 Z tZ 
γ z ˜
Ut = − 2 R z σWt + (e − 1)J(ds × dz) .
σ + R (e − 1)2 ν(dz) 0 R

By Proposition 9, this yields a probability change if

γ(ex − 1)
R < 1 ∀x ∈ supp ν,
σ 2 + R (ez − 1)2 ν(dz)

which imposes a strong restriction on the drift parameter γ. If this condition

is not satisfied, the Föllmer-Schweizer decomposition may still exist, but the
interpretation using the minimal martingale measure is no longer valid, and
the initial capital may turn out to be negative.
The existence of a Föllmer-Schweizer decomposition has been studied by
many authors (see for example [2, 64]), and in particular it was shown that
the decomposition always exists in the case of exponential Lévy models. For
these models, explicit formulas for the coefficients of this decomposition for
European options are given in [39]:

Proposition 21 (Föllmer-Schweizer decomposition for European options in

exponential Lévy models [39]).

• Case of exponential pay-offs. Let z ∈ C with STz ∈ L2 (P). Then the

contingent claim H(z) = STz admits a Föllmer-Schweizer decomposition
with

φ(z)t = Υ(z)eη(z)(T −t) St−

z−1
Z t
η(z)(T −t)
L(z)t = e Stz −e η(z)T
S0z − φ(z)u dSu ,
0

where the coefficients Υ and η are given by

k(z + 1) − k(z) − k(1)

Υ(z) = , η(z) = k(z) − k(1)Υ(z),
k(2) − 2k(1)

and k(z) = log E[ezX1 ] is the Laplace exponent of X.

84
 • Case of arbitrary payoffs. Let the option payoff be H = f (ST ) with f
of the form Z
f (s) = sz Π(dz)

for some finite complex measure Π on a strip {z ∈ C : R0 ≤ <z ≤ R},

0
where R0 , R ∈ R are chosen such that E[e2RX1 ] < ∞ and E[e2R X1 ] <
∞. Then H admits a Föllmer-Schweizer decomposition with coeffi-
cients
Z
H
φt = φ(z)t Π(dz)
Z
H
Lt = L(z)t Π(dz).

Example 7. Let N 1 and N 2 be two standard Poisson processes with intensity

1 under P and suppose that the stock price is given by

St = γt + 2Nt1 + Nt2 − 3t,

and that the contingent claim to be hedged is

H = 5NT1 .

Define
Lt = Nt1 − 2Nt2 + t
Then L is a P-martingale and

[L, S]t = 2Nt1 − 2Nt2

which means that L is orthogonal to the martingale part of S under P . It is

now easy to check that the Föllmer-Schweizer decomposition for H is given
by
H = (5 − 2γ)T + LT + 2ST .
The locally risk-minimizing strategy therefore consists in
• Buying 2 units of the risky asset at date t = 0 (at zero price) and
holding them until maturity.

• Placing (5−2γ)T at the bank account and dynamically adding/withdrawing

money according to the value of L.

85
The initial cost of this strategy is thus equal to H0 = (5 − 2γ)T , which can
be both positive and negative (if γ > 25 ), and therefore cannot be interpreted
as the price of the claim H. An intuitive explanation is that when the stock
returns are very high, one can obtain a terminal pay-off which is (on average)
positive even with a negative initial capital.
The minimal martingale measure in this setting is defined by

dQM |Ft dZt γ

= Zt , = − (2dNt1 + dNt2 − 3dt).
dP|Ft Zt− 5

From Proposition 9, we deduce that QM is a probability measure if and only

if γ < 52 , in which case N 1 and N 2 are independent Poisson processes under
QM , with intensities
2γ γ
λ1 = 1 − and λ2 = 1 − .
5 5
Easy calculations show that

• The martingale property of L is preserved under QM , and in particular,

we can compute
M
HtM = E Q [H|Ft ] = 5λ1 T + 5(Nt1 − λ1 t)
dhH M , SiPt
and φH
t = = 2.
dhS, SiPt

• On the other hand, the orthogonality of S and L is not preserved under

QM : this would require [L, S]t = 2Nt1 − 2Nt2 to be a QM -martingale,
which holds if and only if λ1 = λ2 .

Variance-optimal hedging An alternative approach is to choose a self-

financing strategy φ and the initial capital V0 such as to minimize
 
2
E (V0 + GT (φ) − H) .
P

under the statistical measure P. This approach, known as mean-variance

hedging or variance optimal hedging, is described in many papers including
[12, 52, 65, 11, 39, 18]. The general results concerning existence of optimal

86
strategies are given in [18]. Schweizer [63] studies the case where the mean-
variance tradeoff process K is deterministic and shows that in this case, the
variance-optimal hedging strategy is also linked to the Föllmer-Schweizer
decomposition. Hubalek et al. [39] exploit these results to derive explicit
formulas for the hedging strategy in the case of Lévy processes. The following
proposition uses the notation of Proposition 21.
Proposition 22 (Mean variance hedging in exponential Lévy models [39]).
Let the contingent claim H be as in the second part of Proposition 21. Then
the variance optimal initial capital and the variance optimal hedging strategy
are given by
V0 = H0
λ
φt = φH
t + (Ht− − V0 − Gt− (φ)), (99)
St−
κ(1)
where λ = κ(2)−2κ(1)
and
Z
Ht = Stz eη(z)(T −t) Π(dz).

In the case of exponential Lévy models, and in all models with determin-
istic mean-variance tradeoff, the variance optimal initial wealth is therefore
equal to the initial value of the locally risk minimizing strategy. This allows to
interpret the above result as a “stochastic target” approach to hedging, where
the locally risk minimizing portfolio Ht plays the role of a “stochastic tar-
get” which we would like to follow because it allows to approach the option’s
pay-off with the least fluctuations. Since the locally risk-minimizing strategy
is not self-financing, if we try to follow it with a self-financing strategy, our
portfolio may deviate from the locally risk minimizing portfolio upwards or
downwards. The strategy (99) measures this deviation at each date and tries
to compensate it by investing more or less in the stock, depending on the
sign of the expected return (λ is the expected excess return divided by the
square of the volatility).

13 Calibration of exp-Lévy models

In the Black-Scholes setting, the only model parameter to choose is the
volatility σ, originally defined as the annualized standard deviation of loga-
rithmic stock returns. The notion of model calibration does not exist, since

87
 0.32
Mean jump = ï10%
Mean jump = 10%
Mean jump=0
0.5
0.3
0.45

0.4
0.28
0.35

0.3

0.25 0.26

0.2

0.15
0.24
0.1
0
0.5
1 0.22

1.5 600
800
2 1000
2.5 1200
1400 0.2
3 1600 80 85 90 95 100 105 110 115 120

Figure 8: Left: implied volatilities of options on S&P 500 index as a function

of their strikes and maturities. Right: implied volatilities as a function of
strike for different values of the mean jump size in Merton jump diffusion
model. Other parameters: volatility σ = 0.2, jump intensity λ = 1, jump
standard deviation δ = 0.05, option maturity T = 1 month.

after observing a trajectory of the stock price, the pricing model is com-
pletely defined. On the other hand, since the pricing model is defined by a
single volatility parameter, this parameter can be reconstructed from a single
option price (by inverting the Black-Scholes formula). This value is known
as the implied volatility of this option.
If the real markets obeyed the Black-Scholes model, the implied volatility
of all options written on the same underlying would be the same and equal
to the standard deviation of returns of this underlying. However, empirical
studies show that this is not the case: implied volatilities of options on the
same underlying depend on their strikes and maturities (figure 8, left graph).
Jump-diffusion models provide an explanation of the implied volatility
smile phenomenon since in these models the implied volatility is both dif-
ferent from the historical volatility and changes as a function of strike and
maturity. Figure 8, right graph shows possible implied volatility patterns (as
a function of strike) in the Merton jump-diffusion model.
The results of calibration of the Merton model to S&P index options
are presented in figure 9. The calibration was carried out separately for
each maturity using the routine [8] from Premia software. In this program,
the vector of unknown parameters θ is found by minimizing numerically the

88
 0.9 0.30
Market 0.28 Market
0.8
Model 0.26 Model
0.7
0.24
0.6 0.22
0.5 0.20
0.4 0.18
0.16
0.3
0.14
0.2 0.12
0.1 0.10
60 70 80 90 100 110 120 50 60 70 80 90 100 110 120 130

0.7 0.30
Market Market
0.6 Model Model
0.25
0.5

0.4 0.20

0.3
0.15
0.2

0.1 0.10
30 40 50 60 70 80 90 100 110 120 130 40 50 60 70 80 90 100 110 120 130 140

Figure 9: Calibration of Merton jump-diffusion model to market data sepa-

rately for each maturity. Top left: maturity 1 month. Bottom left: maturity
5 months. Top right: maturity 1.5 years. Bottom right: maturity 3 years.

squared norm of the difference between market and model prices:

N
X
θ∗ = arg inf kP obs − P θ k2 ≡ arg inf wi (Piobs − P θ (Ti , Ki ))2 , (100)
i=1

where P obs denotes the prices observed in the market and P θ (Ti , Ki ) is the
Merton model price computed for parameter vector θ, maturity Ti and strike
1
Ki . Here, the weights wi := (P obs )2
were chosen to ensure that all terms in
i
the minimization functional are of the same order of magnitude. The model
prices were computed simultaneously for all strikes present in the data using
the FFT-based algorithm described in section 8. The functional in (100) was
then minimized using a quasi-newton method (LBFGS-B described in [13]).
In the case of Merton model, the calibration functional is sufficiently well
behaved, and can be minimized using this convex optimization algorithm.
In more complex jump-diffusion models, in particular, when no parametric
shape of the Lévy measure is assumed, a penalty term must be added to
the distance functional in (100) to ensure convergence and stability. This
procedure is described in detail in [21, 22, 67].
The calibration for each individual maturity is quite good, however, al-
though the options of different maturities correspond to the same trading
day and the same underlying, the parameter values for each maturity are
different, as seen from table 3. In particular, the behavior for short (1 to
5 months) and long (1 to 3 years) maturities is qualitatively different, and

89
 0.7 0.30
Market 0.28 Market
0.6 Model Model
0.26
0.24
0.5
0.22
0.4 0.20
0.18
0.3
0.16
0.14
0.2
0.12
0.1 0.10
60 70 80 90 100 110 120 50 60 70 80 90 100 110 120 130

0.7 0.30
Market Market
0.6 Model Model
0.25
0.5

0.4 0.20

0.3
0.15
0.2

0.1 0.10
30 40 50 60 70 80 90 100 110 120 130 40 50 60 70 80 90 100 110 120 130 140

Figure 10: Calibration of Merton jump-diffusion model simultaneously to

4 maturities. Calibrated parameter values: σ = 9.0%, λ = 0.39, jump
mean −0.12 and jump standard deviation 0.15. Top left: maturity 1 month.
Bottom left: maturity 5 months. Top right: maturity 1.5 years. Bottom
right: maturity 3 years.

for longer maturities the mean jump size tends to increase while the jump
intensity decreases with the length of the holding period.
Figure 10 shows the result of simultaneous calibration of Merton model to
options of 4 different maturities, ranging from 1 month to 3 years. As we see,
the calibration error is much bigger than in figure 9. This happens because
for processes with independent and stationary increments (and the log-price
in Merton model is an example of such process), the law of the entire process
is completely determined by its law at any given time t (this follows from the
Lévy-Khintchine formula — equation 10). If we have calibrated the model
parameters for a single maturity T , this fixes completely the risk-neutral
stock price distribution for all other maturities. A special kind of maturity
dependence is therefore hard-wired into every Lévy jump diffusion model,
and table 3 shows that it does not always correspond to the term structures
of market option prices.
To calibrate a jump-diffusion model to options of several maturities at the
same time, the model must have a sufficient number of degrees of freedom to
reproduce different term structures. This is possible for example in the Bates
model (101), where the smile for short maturities is explained by the presence
of jumps whereas the smile for longer maturities and the term structure of
implied volatility is taken into account using the stochastic volatility process.
Figure 11 shows the calibration of the Bates model to the same data set as

90
 Maturity σ λ jump mean jump std. dev.
1 month 9.5% 0.097 −1.00 0.71
2 months 9.3% 0.086 −0.99 0.63
5 months 10.8% 0.050 −0.59 0.41
11 months 7.1% 0.70 −0.13 0.11
17 months 8.2% 0.29 −0.25 0.12
23 months 8.2% 0.26 −0.27 0.15
35 months 8.8% 0.16 −0.38 0.19
Table 3: Calibrated Merton model parameters for different times to maturity.

above. As we see, the calibration quality has improved and is now almost as
good as when each maturity was calibrated separately. The calibration was
once again carried out using the tool [8] from Premia.

14 Limits and extensions of Lévy processes

Despite the fact that Lévy processes reproduce the implied volatility smile for
a single maturity quite well, when it comes to calibrating several maturities at
the same time, the calibration by Lévy processes becomes much less precise.
This is clearly seen from the three graphs of Figure 12. The top graph shows
the market implied volatilities for four maturities and different strikes. The
bottom left graphs depicts implied volatilities, computed in an exponential
Lévy model calibrated using a nonparametric algorithm to the first maturity
present in the market data. One can see that while the calibration quality
is acceptable for the first maturity, it quickly deteriorates as the time to
maturity increases: the smile in an exponential Lévy model flattens too fast.
The same effect can be observed in the bottom right graph: here, the model
was calibrated to the last maturity, present in the data. As a result, the
calibration quality is poor for the first maturity: the smile in an exponential
Lévy model is more pronounced and its shape does not resemble that of the
market.
It is difficult to calibrate an exponential Lévy model to options of several
maturities because due to independence and stationarity of their increments,
Lévy processes have a very rigid term structure of cumulants. In partic-
ular, the skewness of a Lévy process is proportional to the inverse square
root of time and the excess kurtosis is inversely proportional to time [49]. A

91
 0.8 0.40
Market Market
0.7 0.35
Model Model
0.6
0.30
0.5
0.25
0.4
0.20
0.3

0.2 0.15

0.1 0.10
60 70 80 90 100 110 120 50 60 70 80 90 100 110 120 130

0.7 0.35
Market Market
0.6 Model Model
0.30

0.5
0.25
0.4
0.20
0.3

0.15
0.2

0.1 0.10
30 40 50 60 70 80 90 100 110 120 130 40 50 60 70 80 90 100 110 120 130 140

Figure 11: Calibration of the Bates stochastic volatility jump-diffusion model

simultaneously to 4 maturities. Top left: maturity 1 month. Bottom left:
maturity 5 months. Top right: maturity 1.5 years. Bottom right: matu-
rity
√ 3 years. Calibrated parameters (see equation (101)): initial volatility
V = 12.4%, rate of volatility mean reversion ξ = 3.72, long-run volatility
√ 0
η = 11.8%, volatility of volatility θ = 0.501, correlation ρ = −48.8%, jump
intensity λ = 0.038, mean jump size −1.14, jump standard deviation 0.73.

number of empirical studies have compared the term structure of skewness

and kurtosis implied in market option prices to the skewness and kurtosis
of Lévy processes. Bates [6], after an empirical study of implicit kurtosis in
$/DM exchange rate options concludes that “while implicit excess kurtosis
does tend to increase as option maturity shrinks, . . . , the magnitude of ma-
turity effects is not as large as predicted [by a Lévy model]”. For stock index
options, Madan and Konikov [49] report even more surprising results: both
implied skewness and kurtosis actually decrease as the length of the hold-
ing period becomes smaller. It should be mentioned, however, that implied
skewness/kurtosis cannot be computed from a finite number of option prices
with high precision.
A second major difficulty arising while trying to calibrate an exponential
Lévy model is the time evolution of the smile. Exponential Lévy models
belong to the class of so called “sticky moneyness” models, meaning that in
an exponential Lévy model, the implied volatility of an option with given
moneyness (strike price to spot ratio) does not depend on time. This can
be seen from the following simple argument. In an exponential Lévy model
Q, the implied volatility σ of a call option with moneyness m, expiring in τ

92
 0.28
Implied volatility

0.26
0.24
0.22
0.2
0.18
0.16
0 5000
5500
0.5 6000
Maturity Strike
6500
1 7000

0.3
0.28 0.28
Implied volatility
Implied volatility

0.26 0.26
0.24 0.24
0.22 0.22

0.2 0.2
0.18
0.18
0.16
0.16 0 5000
0 5000 5500
5500 0.5 6000
0.5 6000
6500
Maturity 6500 Strike 1 7000
1 7000 Maturity Strike

Figure 12: Top: Market implied volatility surface. Bottom left: implied
volatility surface in an exponential Lévy model, calibrated to market prices of
the first maturity. Bottom right: implied volatility surface in an exponential
Lévy model, calibrated to market prices of the last maturity.

93
 30ïday ATM options
450ïday ATM options
50

45

40

35

30

25

20

15

2/01/1996 2/01/1997 2/01/1998 30/12/1998

Figure 13: Implied volatility of at the money European options on CAC40

index.

years, satisfies:
σ2
e−rτ E Q [(St erτ +Xτ − mSt )+ |Ft ] = e−rτ E[(St erτ +σWτ − 2
τ
− mSt )+ |Ft ]
Due to the independent increments property, St cancels out and we obtain an
equation for the implied volatility σ which does not contain t or St . Therefore,
in an exp-Lévy model this implied volatility does not depend on date t or
stock price St . This means that once the smile has been calibrated for a given
date t, its shape is fixed for all future dates. Whether or not this is true in real
markets can be tested in a model-free way by looking at the implied volatility
of at the money options with the same maturity for different dates. Figure 13
depicts the behavior of implied volatility of two at the money options on the
CAC40 index, expiring in 30 and 450 days. Since the maturities of available
options are different for different dates, to obtain the implied volatility of an
option with fixed maturity T for each date, we have taken two maturities,
present in the data, closest to T from above and below: T1 ≤ T and T2 > T .
The implied volatility Σ(T ) of the hypothetical option with maturity T was
then interpolated using the following formula:
T2 − T T − T1
Σ2 (T ) = Σ2 (T1 ) + Σ2 (T2 ) .
T1 − T T2 − T1
As we have seen, in an exponential Lévy model the implied volatility of an
option which is at the money and has fixed maturity must not depend on

94
time or stock price. Figure 13 shows that in reality this is not so: both graphs
are rapidly varying random functions.
This simple test shows that real markets do not have the “sticky money-
ness” property: arrival of new information can alter the form of the smile.
The exponential Lévy models are therefore “not random enough” to account
for the time evolution of the smile. Moreover, models based on additive
processes, that is, time-inhomogeneous processes with independent incre-
ments, although they perform well in calibrating the term structure of im-
plied volatilities for a given date [20], are not likely to describe the time
evolution of the smile correctly since in these models the future form of the
smile is still a deterministic function of its present shape [20]. To describe
the time evolution of the smile in a consistent manner, one may need to
introduce additional stochastic factors (e.g. stochastic volatility).
Several models combining jumps and stochastic volatility appeared in the
literature. In the Bates [5] model, one of the most popular examples of the
class, an independent jump component is added to the Heston stochastic
volatility model:
p
dXt = µdt + Vt dWtX + dZt , St = S0 eXt , (101)
p
dVt = ξ(η − Vt )dt + θ Vt dWtV , dhW V , W X it = ρdt,

where Z is a compound Poisson process with Gaussian jumps. Although Xt

is no longer a Lévy process, its characteristic function is known in closed
form [20, chapter 15] and the pricing and calibration procedures are similar
to those used for Lévy processes.

References
[1] Abramowitz, M. and Stegun, I., eds., Handbook of Mathematical Func-
tions, Dover: New York, 1968.

[2] Ansel, J. P. and Stricker, C., Lois de martingale, densités et

décomposition de Föllmer-Schweizer, Annales de l’Institut Henri
Poincaré, 28, pp. 375–392.

[3] Appelbaum, D., Lévy Processes and Stochastic Calculus, Cambridge Un-
viersity Press, 2004.

95
 [4] Appelbaum, D., Martingale-valued measures, Ornstein-Uhlenbeck pro-
cesses with jumps and operator self-decomposability in Hilbert space, in
In Memoriam Paul-André Meyer, Seminaire de Probabilités XXXIX,
Springer, 2006, pp. 171–196.

[5] D. Bates, Jumps and stochastic volatility: the exchange rate processes
implicit in Deutschemark options, Rev. Fin. Studies, 9 (1996), pp. 69–
107.

[6] D. S. Bates, Testing option pricing models, in Statistical Methods in

Finance, vol. 14 of Handbook of Statistics, North-Holland, Amsterdam,
1996, pp. 567–611.

[7] Belomestny, D. and Reiss, M., Spectral calibration of exponential Lévy

models, Finance and Stochastics, 10 (2006), pp. 449–474.

[8] A. Ben Haj Yedder, Calibration of stochastic volatility model

with jumps. A computer program, part of Premia software. See
www.premia.fr, 2004.

[9] Benth, F. E., Di Nunno, G., Løkka, A., Øksendal, B., and Proske, F., Ex-
plicit representation of the minimal variance portfolio in markets driven
by Lévy processes, Mathematical finance, 13 (2003), pp. 55–72.

[10] Bertoin, J., Lévy Processes, Cambridge University Press: Cambridge,

1996.

[11] Biagini, F. and Guasoni, P., Mean-variance hedging with random volatil-
ity jumps, Stochastic Analysis and Applications, 20 (2002).

[12] Bouleau, N. and Lamberton, D., Residual risks and hedging strategies in
Markovian markets, Stochastic Process. Appl., 33 (1989), pp. 131–150.

[13] R. Byrd, P. Lu, and J. Nocedal, A limited memory algorithm for

bound constrained optimization, SIAM Journal on Scientific and Statis-
tical Computing, 16 (1995), pp. 1190–1208.

[14] Carr, P., Geman, H., Madan, D., and Yor, M., The fine structure of
asset returns: An empirical investigation, Journal of Business, 75 (2002),
pp. 305–332.

96
[15] Carr, P. and Madan, D., Option valuation using the fast Fourier trans-
form, J. Comput. Finance, 2 (1998), pp. 61–73.

[16] Carr, P. and Wu, L., The finite moment logstable process and option
pricing, Journal of Finance, 58 (2003), pp. 753–778.

[17] Carr, P. and Wu, L., What type of process underlies options? A simple
robust test, J. Finance, 58 (2003).

[18] Černý, A. and Kallsen, J., On the structure of general mean-variance

hedging strategies, The Annals of Probability, 35 (2007), pp. 1479–1531.

[19] Cherny, A. S. and Shiryaev, A. N., Change of time and measure for Lévy
processes. MaPhySto lecture notes series 2002-13, 2002.

[20] Cont, R. and Tankov, P., Financial Modelling with Jump Processes,
Chapman & Hall / CRC Press, 2004.

[21] Cont, R. and Tankov, P., Non-parametric calibration of jump-diffusion

option pricing models, J. Comput. Finance, 7 (2004), pp. 1–49.

[22] R. Cont and P. Tankov, Retrieving Lévy processes from option

prices: Regularization of an ill-posed inverse problem, SIAM Journal
on Control and Optimization, 45 (2006), pp. 1–25.

[23] , Constant proportion portfolio insurance in the presence of jumps

in asset prices, Mathematical Finace, 19 (2009), pp. 379–401.

[24] Cont, R., Tankov, P., and Voltchkova, E., Hedging with options in models
with jumps, in Stochastic Analysis and Applications - the Abel Sympo-
sium 2005, Springer, 2007.

[25] R. Cont and E. Voltchkova, Integro-differential equations for op-

tion prices in exponential Lévy models, Finance and Stochastics, 9
(2005), pp. 299–325.

[26] J. W. Cooley and J. W. Tukey, An algorithm for the machine

calculation of complex Fourier series, Math. Comp., 19 (1965), pp. 297–
301.

97
[27] Eberlein, E., Glau, K., and Papapantoleon, A., Analysis of Fourier
transform valuation formulas and applications, Appl. Math. Finance,
13 (2010), pp. 211–240.

[28] Föllmer, H. and Elliott, R., Orthogonal martingale representation, in

Stochastic Analysis, Mayer-Wolf, E., Merzbach, E., and Schwartz, A.,
eds., Academic Press: Boston, 1991, pp. 139–152.

[29] Föllmer, H. and Schweizer, M., Hedging of contintgent claims under

incomplete information, in Applied Stochastic Analysis, Davis, M. H. A.
and Elliott, R. J., eds., Gordon and Breach, 1991, pp. 389–414.

[30] Föllmer, H. and Sondermann, D., Hedging of non-redundant contingent

claims, in Contributions to Mathematical Economics, Hildenbrand, W.
and Mas-Colell, A., eds., North Holland: Amsterdam, 1986, pp. 205–
223.

[31] Friz, P. and Benaim, S., Smile asymptotics II: Models with known MGF,
Journal of Applied Probability, 45 (2008), pp. 16–32.

[32] Friz, P. and Benaim, S., Regular variation and smile asymptotics, Math-
ematical Finance, 19 (2009), pp. 1–12.

[33] Galtchouk, L., Représentation des martingales engendrées par un pro-

cessus à accroissements indépendants (cas des martingales de carré
intégrable), Ann. Inst. H. Poincaré Sect. B (N.S.), 12 (1976), pp. 199–
211.

[34] H. Geman, D. Madan, and M. Yor, Asset prices are Brownian

motion: Only in business time, in Quantitative Analysis in Financial
Markets, M. Avellaneda, ed., World Scientific, River Edge, NJ, 2001,
pp. 103–146.

[35] Goll, T. and Kallsen, J., Optimal portfolios for logarithmic utility,
Stochastic Processes and their applications, 89 (2000), pp. 31–48.

[36] Jacod, J., Méléard, S., and Protter, P., Explicit form and robustness of
martingale representations, Annals of Probability, 28 (2000), pp. 1747–
1780.

98
[37] Jakubenas, P., On option pricing in certain incomplete markets, Pro-
ceedings of the Steklov Mathematical Institute, 237 (2002).

[38] J. Kallsen and P. Tankov, Characterization of dependence of multi-

dimensional Lévy processes using Lévy copulas, Journal of Multivariate
Analysis, 97 (2006), pp. 1551–1572.

[39] Kallsen, J., Hubalek, F., and Krawczyk, L., Variance-optimal hedging
for processes with stationary independent increments, The Annals of
Applied Probability, 16 (2006), pp. 853–885.

[40] Kardaras, C., No-free-lunch equivalences for exponential Lévy models

under convex constraints on investment, Math. Finance, 19 (2009), pp.
161–187.

[41] Koponen, I., Analytic approach to the problem of convergence of trun-

cated Lévy flights towards the Gaussian stochastic process., Physical Re-
view E, 52 (1995), pp. 1197–1199.

[42] Kou, S., A jump-diffusion model for option pricing, Management Sci-
ence, 48 (2002), pp. 1086–1101.

[43] Kunita, H. and Watanabe, S., On square integrable martingales, Nagoya

Math. J., 30 (1967), pp. 209–245.

[44] Kyprianou, A., Introductory Lectures on Fluctuations of Lévy Processes

with Applications, Springer, 2006.

[45] Lee, R., The moment formula for implied volatility at extreme strikes,
Mathematical Finance, 14 (2004), pp. 469–480.

[46] Lee, R. W., Option pricing by transform methods: extensions, unifica-

tion and error control, J. Comput. Finance, 7 (2004).

[47] Lewis, A., A simple option formula for general jump-

diffusion and other exponential Lévy processes. available from
http://www.optioncity.net, 2001.

[48] Madan, D., Carr, P., and Chang, E., The variance gamma process and
option pricing, European Finance Review, 2 (1998), pp. 79–105.

99
[49] D. Madan and M. Konikov, Option pricing using variance gamma
Markov chains, Rev. Derivatives Research, 5 (2002), pp. 81–115.

[50] D. Madan and F. Milne, Option pricing with variance gamma mar-
tingale components, Math. Finance, 1 (1991), pp. 39–55.

[51] Merton, R., Option pricing when underlying stock returns are discontin-
uous, J. Financial Economics, 3 (1976), pp. 125–144.

[52] Pham, H., Hedging and optimization problems in continuous financial

models, in Mathematical Finance: Theory and Practice, Cont, R. and
Yong, J., eds., Vol. 1 of Series in Contemporary Applied Mathematics,
Higher Education Press: Beijing, 2000, pp. 353–381.

[53] Protter, P., Stochastic integration and differential equations, Springer:

Berlin, second ed., 2004.

[54] Raible, S., Lévy Processes in Finance: Theory, Numerics and Empirical
Facts, PhD thesis, Freiburg University, 2000.

[55] Rockafellar, R. T., Convex Analysis, Princeton University Press, 1970.

[56] Rogers, L. C. G. and Tehranchi, M., Can the implied volatility surface
move by parallel shifts?, Finance Stoch., 14 (2010), 235–248.

[57] Rudin, W., Real and complex analysis, McGraw-Hill, 3 ed., 1987.

[58] Rüschendorf, L. and Woerner, J. H., Expansion of transition distribu-

tions of Lévy processes in small time, Bernoulli, 8 (2002), pp. 81–96.

[59] Sato, K., Lévy Processes and Infinitely Divisible Distributions, Cam-
bridge University Press: Cambridge, UK, 1999.

[60] Schoutens, W., Lévy Processes in Finance: Pricing Financial Deriva-

tives, Wiley: New York, 2003.

[61] Schweizer, M., Risk-minimality and orthogonality of martingales,

Stochastics Stochastics Rep., 30 (1990), pp. 123–131.

[62] Schweizer, M., Option hedging for semimartingales, Stochastic Process.

Appl., 37 (1991), pp. 339–363.

100
[63] Schweizer, M., Approximating random variables by stochastic integrals,
Annals of Probability, 22 (1994), pp. 1536–1575.

[64] Schweizer, M., On the minimal martingale measure and the Föllmer-
Schweizer decomposition, Stochastic Anal. Appl., 13 (1995), pp. 573–
599.

[65] Schweizer, M., A guided tour through quadratic hedging approaches,

Cambridge University Press: Cambridge, 2001, pp. 538–574.

[66] Selivanov, A. V., On the martingale measures in exponential Lévy mod-

els, Theory of Probability and its Applications, 49 (2005), pp. 261–274.

[67] P. Tankov, Lévy Processes in Finance: Inverse Problems and Depen-

dence Modelling, PhD thesis, Ecole Polytechnique, France, 2004.

[68] Tehranchi, M. R., Asymptotics of implied volatility far from maturity,

J. Appl. Proba., 46 (2009), 261–274.

[69] Tehranchi, M. R., Implied volatility: long maturity behavior, in Encyclo-

pedia of Quantitative Finance, Wiley, 2010.

101