Basics of Levy processes

Ole E. Barndorff-Nielsen
The T.N. Thiele Centre for Mathematics in Natural Science,
Department of Mathematical Sciences,
University of Aarhus, Ny Munkegade, DK-8000 Aarhus C, Denmark
& CREATES, University of Aarhus
[email protected]
Neil Shephard
Nuffield College, New Road, Oxford OX1 1NF, UK,
Department of Economics, University of Oxford
& Oxford-Man Institute, University of Oxford,
[email protected]
June 9, 2012

Abstract
This is a draft Chapter from a book by the authors on Levy Driven Volatility Models.

Contents
1 What is this Chapter about?
2 What is a L
evy process?
2.1 The random walk . . . . . . . . . . .
2.2 Brownian motion . . . . . . . . . . .
2.3 Infinite divisibility . . . . . . . . . .
2.4 Levy processes and semimartingales

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3 Non-negative L
evy processes subordinators
3.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Poisson process . . . . . . . . . . . . . . . . . . . .
3.1.2 Compound Poisson process . . . . . . . . . . . . .
3.1.3 Negative binomial process . . . . . . . . . . . . . .
3.1.4 Infinite activity subordinators . . . . . . . . . . . .
3.2 Levy measures for non-negative processes . . . . . . . . .
3.3 Levy-Khintchine representation for non-negative processes
3.3.1 Representation . . . . . . . . . . . . . . . . . . . .
3.3.2 Positive tempered stable process . . . . . . . . . .
3.4 Simulation for non-negative Levy processes . . . . . . . .
3.4.1 Simulating paths via the Levy density . . . . . . .
3.4.2 Simulation via a small jump approximation . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

4
4
4
5
6

.
.
.
.
.
.
.
.
.
.
.
.

7
8
8
9
10
11
14
17
17
17
18
18
20

We are grateful to the many colleagues and students who have over many years, given us comments on this
chapter. Of course we welcome further comments as we push it to completion.

4 Processes with real increments

4.1 Examples of Levy processes . . . . . . . . . . . . . . . . . . .
4.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Brownian motion . . . . . . . . . . . . . . . . . . . . .
4.1.3 Compound Poisson process . . . . . . . . . . . . . . .
4.1.4 Skellam process . . . . . . . . . . . . . . . . . . . . . .
4.2 Normal variance-mean mixture processes . . . . . . . . . . . .
4.2.1 Normal inverse Gaussian . . . . . . . . . . . . . . . . .
4.2.2 Normal gamma process . . . . . . . . . . . . . . . . .
4.2.3 Hyperbolic, Laplace and skewed Students t processes
4.2.4 Generalized hyperbolic process . . . . . . . . . . . . .
4.2.5 Normal positive stable and symmetric stable processes
4.2.6 Normal tempered stable process . . . . . . . . . . . .
4.3 Levy-Khintchine representation . . . . . . . . . . . . . . . .
4.4 Blumenthal-Getoor index . . . . . . . . . . . . . . . . . . . .
4.5 Power variation . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6 Levy-Ito representation . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

20
20
20
20
22
22
23
24
25
27
28
29
30
30
31
31
33

5 Time deformation and time-change

5.1 Basic framework . . . . . . . . . . . . .
5.2 Examples . . . . . . . . . . . . . . . . .
5.2.1 Compound Poisson process . . .
5.2.2 Normal inverse Gaussian process
5.2.3 Normal tempered stable process
5.2.4 Type G and P Levy processes . .
5.2.5 Negative binomial process . . . .
5.2.6 Discrete Laplace process . . . . .
5.2.7 Poisson-IG process . . . . . . . .
5.2.8 Time-changed Skellam processes

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

34
34
35
35
35
36
36
36
37
37
38

6 Empirical estimation and testing of GH L

evy processes
6.1 A likelihood approach . . . . . . . . . . . . . . . . . . . .
6.1.1 Estimation of GH Levy processes . . . . . . . . . .
6.1.2 Confidence intervals via profile likelihoods . . . . .
6.2 Model misspecification: robust standard errors . . . . . .
6.3 Further empirical results . . . . . . . . . . . . . . . . . . .
6.3.1 Six daily exchange rate movements . . . . . . . . .
6.3.2 Daily equity indexes . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

39
39
39
41
42
44
44
45

7 Quadratic variation
7.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Brownian motion and quadratic variation . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Realised QV process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

46
46
48
48

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

8 L
evy processes and stochastic analysis
50
8.1 Itos formula for Levy processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8.2 Stochastic exponential of a Levy process . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.3 Stochastic logarithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

9 Multivariate L
evy processes
9.1 Elemental constructions . . . . . . . . . . . . . . .
9.2 Multivariate generalised hyperbolic Levy process .
9.2.1 Background . . . . . . . . . . . . . . . . . .
9.2.2 Empirical fit of multivariate GH processes .
9.3 Stochastic discount factors . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

51
51
52
52
53
56

10 From L
evy processes to semimartingales

57

11 Conclusion

58

12 Exercises

58

13 Bibliographic notes
59
13.1 Levy processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
13.2 Flexible distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
13.3 Levy processes in finance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
13.4 Empirical fit of Levy processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Keywords: compound process; cumulant function; inverse Gaussian; jumps; Levy-Khintchine
representation; Levy measure; Levy process; Poisson process; quadratic variation; Skellam process;
stable process; time deformation.
JEL Classifications: C22, C63, G10

What is this Chapter about?

In this Chapter we provide a first course on Levy processes in the context of financial economics.
The focus will be on probabilistic and econometric issues; understanding the models and assessing
their fit to returns on speculative assets. The Chapter will delay the discussion of some of the
technical aspects of this material until Barndorff-Nielsen and Shephard (2012a). Our stochastic
analysis primer Barndorff-Nielsen and Shephard (2012b) may be of help to readers without a strong
background in probability theory. Throughout we hope our treatment will be as self-contained as
possible.
This long Chapter has 12 other Sections, whose goals are to:
Build continuous time random walk models using the concept of infinite divisibility.
Introduce Levy processes with non-negative increments.
Extend the analysis to Levy processes with real increments.
Introduce time change, where we replace calendar time by a random clock.
Use statistical methods to fit some common Levy processes to financial data.
Introduce quadratic variation, a central concept in econometrics and stochastic analysis.

 Briefly discuss stochastic analysis in the context of Levy processes.

Introduce various methods for building multivariate Levy processes.
Discuss the relationship between stochastic volatility and general semimartingales.
Draw conclusions to the Chapter.
Provide some exercises.
Discuss the literature associated with Levy processes.
Levy processes can only provide a rather partial description of asset prices, for they have
independent increments and so ignore volatility clustering. However, later developments in this
book will extend the setting to deal with that issue.

What is a L
evy process?

2.1

The random walk

The most basic model of the logarithm of the price of a risky asset is a discrete time random walk.
It is built by summing independent and identically distributed (i.i.d.) random variables C1 , C2 , ...
to deliver
Yt =

t
X

Cj ,

with

Y0 = 0,

t = 0, 1, 2, ... .

j=1

The process is moved by the i.i.d. increments

Yt+1 Yt = Ct+1 .

(1)

Hence future changes in a random walk are unpredictable.

What is the natural continuous time version of this process? There are at least two strong
answers to this question.

2.2

Brownian motion

The first approach is based on a central limit type result. Again suppose that C is an i.i.d. sequence
whose first two moments exist. Then, for a given n > 0, define the partial sum
tn

(n)
Yt

1 X
=
{Cj E(C1 )} ,
n
j=1

t R0 ,

(2)

where t represents time and a denotes the integer part of a. It means that over any fixed interval
for t the process is made up of centred and normalised sums of i.i.d. events. We then allow n,
4

the number of these events in any fixed interval of time of unit length, to go off to infinity. This
(n)

is often labelled in-fill asymptotics. As a result Yt

obeys a central limit theory and becomes

Gaussian. Further, this idea can be extended to show that the whole partial sum, as a random
function, converges to a scaled version of Brownian motion, as n goes to infinity. At first sight this
suggests the only reasonable continuous time version of a random walk, which will sum up many
small events, is Brownian motion. This insight is, however, incorrect.

2.3

Infinite divisibility

Our book follows a second approach. Suppose that the goal is to design a continuous time process
Yt such that Y1 , its value at time 1, has a given distribution D. We can divide the time from 0
until 1 into n pieces of equal length. The corresponding increments are assumed to be independent
from a common distribution D (n) such that the sum
tn
(n)
Yt

(n)

Cj ,

where

(n) i.i.d.

Cj

j=1

D (n) ,

has the distribution D when t = 1. Then as n increases the division of time between zero and one
becomes ever finer. In response, the increments and their distribution D (n) also change, but by
construction D, the distribution of the sum, is left unchanged. A simple example of this is where
Y1 P o(1), a Poisson random variable with mean 1, then if
tn
(n)
Yt

(n)

Cj ,

where

(n) i.i.d.

Cj

j=1

P o(1/n),

this produces a random walk with independent Poisson increments which sum to a Poisson random
variable. Hence this process makes sense even as n goes to infinity and so this type of construction
can be used to introduce a continuous time model the Poisson process. A second example is
where
tn
(n)
Yt

(n)

Cj ,

where

j=1

(n) i.i.d.

Cj

N (0, 1/n),

then as n this process converges to Brownian motion.

The class of distributions for which this general construction is possible is that for which D
is infinitely divisible. The resulting processes are called Levy processes. Examples of infinitely
divisible distributions include, focusing for the moment only on non-negative random variables, the
Poisson, gamma, reciprocal gamma, inverse Gaussian, reciprocal inverse Gaussian, F and positive
stable distributions. A detailed technical discussion of infinite divisibility will be given in BarndorffNielsen and Shephard (2012a).

2.4

L
evy processes and semimartingales

The natural continuous time version of the discrete time increment given in (1) is, for any value of
s > 0,
Yt+s Yt ,

t [0, ) .

Increments play a crucial role in the formal definition of a Levy process.

Definition 1 Levy process. A c`
adl`
ag stochastic process Y = {Yt }t0 with Y0 = 0 is a Levy process
if and only if it has independent and (strictly) stationary increments.
The c`
adl`
ag assumption is an important one which is discussed in detail below. For the moment
let us focus on the independence and stationarity assumption. This means that the shocks to the
process are independent over time and that they are summed, while the stationarity assumption
specifies that the distribution of Yt+s Yt may change with s but does not depend upon t. The
independence and stationarity of the increments of the Levy process implies that the cumulant
function of Yt
C { Yt } = log [E exp {iYt }]
= t log [E exp {iY1 }]
= tC { Y1 } ,
so the distribution of Yt is completely governed by the cumulant function of Y1 , the value of the
process at time one.
If a Levy process is used as a model for the log-price of an underlying asset then the increments
can be thought of as returns. Consequently Levy based models provide a potentially flexible
framework with which to model the marginal distribution of returns. However, the returns will be
independent and identically distributed when measured over time intervals of fixed length. Thus
important serial dependencies such as volatility clustering are ignored.
A feature of some stochastic processes, including Brownian motion, is that they have continuous
sample paths with probability one. We need to relax this assumption. We will allow jumps but
require the process Y to be, with probability one, right continuous
lim Ys = Yt
st

and have limits from the left

Yt = lim Ys .
st

For such processes the jump at time t is written as

Yt = Yt Yt ,
and the processes are said to be c`
adl`
ag (continu `
a droite, limites `
a gauche). We might, in particular, expect these types of jumps to appear in financial economics due to dividend payments,
microcrashes due to short-term liquidity challenges or news, such as macroeconomic announcements. The similarly named property c`
agl`
ad (continu `
a gauche, limites `
a droite), which has
lim Ys = Yt+
st

and

Yt = lim Ys ,
st

plays an important role in our stochastic analysis primer.

Semimartingales have a central role in modern stochastic analysis. They are c`
adl`
ag (by definition) and provide, in particular, a basis for the definition of a stochastic integral. Consequently it
is important to note that all Levy processes Y are semimartingales (written Y SM). They also
have the property that
Pr(Yt > 0) = Pr(Yt < 0) = 0,
i.e. there are no fixed discontinuities.
For Y SM, if H is c`
adl`
ag then H is locally bounded and the stochastic integral
Z t
Xt =
Hu dYu ,
0

is well defined. This is often written in the more abstract notation as the process
X = H Y.
More details of stochastic integrals are given in Barndorff-Nielsen and Shephard (2012b) and also
in Section 8.

Non-negative L
evy processes subordinators

We start our more detailed discussion of Levy processes by considering Levy processes with nonnegative increments. Such processes are often called subordinators the reason for this nomenclature will become clear in Section 5. This is our focus for two reasons: (i) they are mathematically
considerably simpler, (ii) many of the models we build in this book will have components which
are Levy processes with non-negative increments. The discussion of processes with arbitrary real
increments will be given in the next Section. As the processes are positive, it is natural to work
with the kumulant function in the form
K { Y1 } = log {E exp (Y1 )} ,

where 0.
7

Occasionally the more standard cumulant function (or log Laplace transform)
K { Y1 } = log {E exp (Y1 )} ,
where { : E exp(Y1 ) < }, will be used.

3.1
3.1.1

Examples
Poisson process

Suppose we count the number of events which have occurred from time 0 until t 0. The very
simplest continuous time model for this type of observation is a Levy process with independent
Poisson increments, so that
Y1 P o(),

> 0,

with density
fY1 (y) =

e y
,
y!

y = 0, 1, 2, ....

The process Y is called a (homogeneous) Poisson process and the mean of Y1 , , is often called
the intensity of this counting process. A simulated sample path of this process, when = 1, is
given in Figure 1(a). It shows a jumping process, where the jumps are irregularly spaced in time
and are of equal size. The times at which events occur are called arrival times, and are written as
1 , 2 , ..., Yt .
For the Poisson process
K { Y1 } = log [E exp {Y1 }]
= (e 1)
= (1 e ).
Now we have that t(1 e ), the kumulant function of Yt , corresponds to the kumulant function
for a P o(t).
Let Y be a Poisson process with Y1 P o().

The Poisson process as a semimartingale

Then Y is of finite variation and hence Y is a semimartingale. We can write it in the form of a
semimartingale, as
Y = Y0 + A + M,

(3)

where Y0 = 0,
At = E (Yt |F0 ) = t

and

Mt = Yt t.
8

(a) Simulated Poisson process

14

(b) Simulated CPP with IG(1,1) innovations

17.5
12
15.0
10
12.5
8
10.0
6
7.5
4

5.0

2.5

10

10

Figure 1: Horizontal axis is time t, vertical is Yt . (a) Sample path of a homogeneous Poisson
process with intensity = 1. (b) Corresponding compound Poisson process with Cj IG(1, 1).
code: levy graphs.ox.
The M process, which is often called a compensated Poisson process, is a mean-0 martingale, for
E(Mt+s |Ft ) = Mt and E(Mt ) = 0 for t, s 0. The A process is of finite variation.
As Y is a Poisson process, any stochastic integral X = H Y simplifies to the random sum
Xt =

Yt
X

H j

j=1

where 1 , 2 , ..., Yt are the arrival times of Y .

3.1.2

Compound Poisson process

Suppose N is a Poisson process and C is an i.i.d. sequence of strictly positive random variables.
Then define a compound Poisson process as (for t 0)
Yt =

Nt
X

Cj ,

where

Y0 = 0.

(4)

j=1

That is, Yt is made up of the addition of a random number Nt of i.i.d. random variables. This is a
Levy process for the increments of this process
Nt+s

Yt+s Yt =

Cj

j=Nt +1

are independent and are stationary as the increments of the Poisson process are independent and
stationary.
9

At this point it is important to note that there is no added flexibility if the distribution of the
C is allowed to have a positive probability of being exactly zero for this would, in effect, just knock
out or thin some of the Poisson process arrivals. This point will recur in our later exposition. It is
informative to note that for non-negative C
K { Y1 } = log [E exp {Y1 }]
= log [EN1 E exp {Y1 } |N1 ]



= log E exp N1 K { C1 }


= K K ( C1 ) N1


= 1 exp K ( C1 ) .

(5)
i.i.d.

Example 1 Figure 1(b) gives a simulation using Cj IG(1, 1), taking exactly the same Poisson
process draws as used in Figure 1(a). IG(1,1) is a special case of the inverse Gaussian law IG(, ),
with density
1
2 1
2

e y 3/2 e 2 ( y + y) ,
2

y R>0 .

The resulting K { Y1 } is

3.1.3


h n

1/2 oi
1 exp 2 + 2
.
Negative binomial process

A negative binomial Levy process Y requires the process at time one to follow a negative binomial
distribution Y1 N B(r, p), where r > 0, p (0, 1), with probability function
Pr(Y1 = y) =

(r + y) r
p (1 p)y ,
y! (r)

y = 0, 1, 2, ....

This implies
n
o
K { Y1 } = r log p r log 1 (1 p) e .

(6)

So this is infinitely divisible, so supports a Levy process. The form of the cumulant function implies
Yt N B(tr, p) so this has a finite number of jumps over a finite time interval. Such processes
are called finite activity and in this case it looks somwhat like a Poisson process. However, the
spaces in times between arrivals are not uniformly distributed and the variance of the number of
arrivals in a fixed time interval will be bigger than the corresponding expected value. We will
discuss the relationship between these two processes in the Section 5 on time-change.

10

3.1.4

Infinite activity subordinators

Gamma process

The Poisson process and the compound Poisson process are by far the most

well-known non-negative Levy processes in economics. The jumps happen, typically, rather rarely.
Consequently increments to these processes are often exactly zero, even when measured over quite
large time intervals. This feature of the process is fundamentally different from the gamma Levy
process. A gamma Levy process Y makes Y1 obey a gamma law
Y1 (, ),

, > 0,

with density
fY1 (y) =

1
exp (y) ,
y
()

y R>0 .

(7)

Here 2 > 0 is thought of as a degrees of freedom parameter, controlling the skewness of the
distribution. The other parameter, , is a scale parameter.
The kumulant function of the gamma distribution is



K { Y1 } = log 1 +
,

implying


K { Yt } = t log 1 +

and so Yt (t, ). The gamma process has the useful property that it has increments which
are strictly positive whatever small time interval has elapsed. Such Levy processes are said to have
infinite activity (IA). This feature puts them apart from a compound Poisson process.
A sample path of a gamma process is drawn in Figure 2(a). The path is dominated by large
jumps and shows informally that the path of the process is not continuous (anywhere). It was
drawn by splitting time into intervals of length 1/2000 and sampling from the implied random
walk with increments taken from the (/2000, ) distribution. Very similar paths are produced
by using smaller time intervals. The process is a rough upward trend with occasional large shifts.
In the special case of = 1, then
fY1 (y) = exp (y) ,

y R>0 ,

which is the exponential distribution (1, ). Thus the exponential Levy process has Yt (t, ).
Inverse Gaussian process

An inverse Gaussian (IG) Levy process Y requires the process at

time one to follow an inverse Gaussian distribution Y1 IG(, ), where > 0, 0, with density



1
fY1 (y) = e y 3/2 exp 2 y 1 + 2 y , y R>0 .
2
2
11

(a) Simulated (4,200) Lvy process

(b) Simulated IG(0.2,100) Lvy process

0.200
0.175
0.175
0.150
0.150
0.125
0.125
0.100

0.100

0.075

0.075

0.050

0.050

0.025

0.025

10

10

Figure 2: Simulated and IG Levy processes, using intervals of length 1/2000.

levy graphs.ox.

Code

This implies
n

1/2 o
K { Y1 } = 2 + 2
.

Like the gamma process, the IG process has an infinite number of jumps in any small interval of
time. The form of the cumulant function implies Yt IG(t, ).
A sample path of an IG Levy process is drawn in Figure 2(b). The parameters were selected to
have the same mean and variance of Y1 as that used to draw the path of the gamma process given
in Figure 2(a). Again the process is a rough upward trend with occasional large shifts.
Some other non-negative processes

A reciprocal (or inverse) gamma (R) Levy process Y

requires the process at time one to be a reciprocal gamma variable Y1 R(, ), , > 0, so that
Y11 (, ), with density
fY1 (y) =

1
y
exp y 1 ,
()

y R>0 .

Only the moments of order less than exist for this distribution.
Sums of independent reciprocal gamma variables are not distributed as reciprocal gamma.
However, the reciprocal gamma is infinitely divisible (and so yields a Levy process), although we
do not know the distribution of Yt in closed form. This makes simulation of this process non-trivial.
A lognormal (LN) Levy process Y requires the process at time one to be a lognormal variable
12

Y1 LN (, 2 ), 2 0, with infinitely divisible density



1 1
1
2
fY1 (y) = y exp 2 (log y ) , y R>0 .
2
2
The proof that the lognormal is infinitely divisible is probabilistically challenging (cf. Thorin
(1977)) and not discussed here.
A reciprocal (inverse) Gaussian (RIG) Levy process Y requires the process at time one to be a
reciprocal inverse Gaussian variable Y1 RIG(, ), > 0, > 0, with density




1/2
1 2 1
2

fY1 (y) =
e y
exp y + y , y R>0 .
2
2
The corresponding kumulant function is

n



1/2 o
1
K { Y1 } = log 1 + 2/ 2 + 1 1 + 2/ 2
.
2

By construction, Y11 IG(, ) provided > 0. Again sums of independent RIG variables are
not distributed as RIG, but the RIG distribution is infinitely divisible.
Generalised inverse Gaussian L
evy process

Many of the above infinite activity processes

are special cases of the generalized inverse Gaussian (GIG) Levy process. This puts
Y1 GIG(, , ),
with GIG density


(/) 1
1 2 1
2
fY1 (y) =
y
exp ( y + y) ,
2K ()
2

y R>0 ,

(8)

where K () is a modified Bessel function of the third kind (cf. Barndorff-Nielsen and Shephard
(2012c)). This density has been shown (Barndorff-Nielsen and Halgreen (1977)) to be infinitely
divisible. Prominent special cases are achieved in the following ways:
IG(, ) = GIG( 12 , , ),
R(, 2 /2) = GIG(, , 0),
RIG(, ) = GIG( 12 , , ),
RP H(, ) = GIG(1, , ).

P H(, ) = GIG(1, , ),
(, 2 /2) = GIG( > 0, 0, ),
P HA(, ) = GIG(0, , )

Here all these distributions have been introduced above except for
the positive hyperbolic distribution (P H);
the positive hyperbola distribution (P HA);
reciprocal positive hyperbolic (RP H);
13

In order to obtain these results we have to allow or to be zero 0. In these cases the GIGs
density has to be interpreted in the limiting sense, using the well-known results that for y 0 we
have
K (y)

log y

if = 0
.

(||)2||1 y ||

if 6= 0.

In general, the density of the increments to this process is unknown in explicit form and we cannot
directly simulate from it without using computationally intensive methods.

3.2

L
evy measures for non-negative processes

It should be clear by now that the kumulant function of Y1 plays an important role in Levy
processes. In this subsection this observation will be further developed in order to build towards
the vital Levy-Khintchine representation which shows us the form characteristic functions of Levy
processes must obey. As this representation is so important, and is also mathematically involved,
its development will be carried out in stages. At first sight this looks unnecessary from a modelling
viewpoint, however we will see that practical modelling will sometimes be carried out directly via
some of the terms which make up the Levy-Khintchine representation. Hence a good understanding
of this Section is essential for later developments.
To start off with, think of a Poisson process Yt P o(t). Then, letting 1 (y) be the Dirac
delta centred at y = 1, we write
K{ Y1 } = (1 e )
Z
=
(1 ey ) 1 (dy)
0
Z
=
(1 ey )P (dy),
0

where P is the Dirac delta probability measure centred at one. The introduction of the probability
measure is entirely expository in this context; however, expressing kumulant functions in this
type of way will become essential later. Before proceeding another level of abstraction has to be
introduced. Instead of working with probability measures we will have to use more general measures
v concentrated on R>0 . An important point is that some of the measures that will be used later
R
will not be integrable (that is 0 v(dy) = ) and so probability measures are insufficient for a
discussion of Levy processes. In the simple Poisson case the measure is introduced by expressing
Z
K{ Y1 } =
(1 ey )v(dy)
(9)
0

where v = 1 is called the Levy measure. Of course this measure is integrable, indeed it integrates
to .
14

Let us now generalise the above setup to the compound Poisson process (4), but still requiring
C to be strictly positive ruling out the possibility that C can be exactly zero with non-zero
probability. Then, writing the distribution of C1 as P (dy C1 ),


K{ Y1 } = 1 exp K ( C1 )
Z
=
(1 ey )P (dy C1 )
0
Z
=
(1 ey )v(dy),

(10)

again, but now with v(dy) = P (dy C1 ). Again this measure is integrable as it is proportional to
the probability measure of C1 . In the simple case where C1 has a density we write
v(dy) = u(y)dy
and call u(y) (which is times the density of C1 ) the Levy density. In such cases the kumulant
function becomes
K{ Y1 } =

(1 ey )u(y)dy.
i.i.d.

A simple example of this is where Cj (, ). Then the Levy density is

u(y) =

1
y
exp (y) .
()

(11)

Of course this Levy density integrates to not one.

Remark 1 If they exist, the cumulants of Y1 satisfy

Z
j K{ Y1 } 
=
y j v(dy),
j =

j

0
=0

i.e. they equal the moments of v.

Although Poisson and compound Poisson processes have integrable Levy measures, for v is
proportional to a probability measure which integrates to one, theoretically more general Levy
processes can be constructed without abandoning the form (9). The non-integrable measures v
will not correspond to compound Poisson processes. To ensure that they yield a valid kumulant
R
function we require that 0 (1 ey )v(dy) exists for all > 0, while continuing to rule out the
possibility that v has an atom at zero. It is simple to prove that a necessary and sufficient condition
for existence is that
Z
min (1, y) v(dy) < .
0

If the Levy measure is absolutely continuous then we can define u(y) as the Levy density where
v(dy) = u(y)dy. However, as some Levy measures are not finite, it follows that Levy densities are
15

not necessarily integrable. This is at first sight confusing. This point comes up in the following two
examples.
Example 2 It turns out that the Levy density of Y1 IG(, ) is
u(y) = (2)1/2 y 3/2 exp( 2 y/2),

y R>0 .

(12)

This Levy density is not integrable as it goes off to infinity too rapidly as y goes to zero. This
implies an IG process is not a compound Poisson process it has an infinite number of jumps
in any finite time period. Although the Levy density is not integrable it does satisfy the finiteness
R
condition on 0 min (1, y) v(dy) for the addition of the y factor regularises the density near zero.

Example 3 It can also be shown that the Levy density of Y1 (, ) is

u(y) = y 1 exp(y),

y R>0 .

(13)

Again this is not an integrable Levy density although it is slower to go off to infinity than the inverse
Gaussian case. This means in practice that it will have fewer very small jumps than the IG process.
These two results are special cases of the result for the GIG(, , ) distribution (8). The
corresponding Levy density is then
 Z



1 1
21 2 y
u(y) = y
e
g ()d + max{0, } exp 2 y/2
2 0

(14)

where

g (y) =

o1
2 n 2
2
J
(
y)
+
N
(
y)
||
||
y2

and J and N are Bessel functions. This result is derived in Barndorff-Nielsen and Shephard
(2012a). Although this looks forbidding, when is a half integer these functions are analytically
simple.
Example 4 A case outside the GIG class is the positive stable P S(, ) process. Although the
probability density of this variable is in general unknown in simple form (see Barndorff-Nielsen and
Shephard (2012a) for details), the kumulant function is
K { Y1 } = (2) ,

0 < < 1,

> 0,

which implies it does not possess moments of order and above. The Levy density for the positive
stable Levy process is given by
u(y) = cy 1 ,

where

c = 2

,
(1 )
16

(15)

while Yt P S(, t). Feller (1971a) discusses many basic facts and examples of stable laws, while
Samorodnitsky and Taqqu (1994) give a very comprehensive account of these laws and the associated
Levy and other processes, in particular fractional Brownian motion. See also Sato (1999).

3.3

L
evy-Khintchine representation for non-negative processes

3.3.1

Representation

Having allowed the Levy measure not to be integrable, a single extra step is required in order to
produce a general setup. We allow a drift a > 0 to be added to the cumulant function. This is
carried out in the following fundamental theorem.
Theorem 2 Levy-Khintchine representation for non-negative Levy processes. Suppose Y is a Levy
process with non-negative increments. Then the kumulant function can be written as
Z 

K{ Y1 } = a
1 ey v(dy),

(16)

where a 0 and v is a measure on R>0 such that

Z
min{1, y}v(dy) < .

(17)

Conversely, any pair (a, v) with these properties determines a non-negative Levy process Y such
that Y1 has kumulant function determined by (16).
The importance of this representation is that the kumulant function of all non-negative Levy
processes can be written in this form. In other words, non-negative Levy processes are completely
determined by a and the Levy measure v (which has to satisfy (17)).
R
In the special case when 0 v(dy) < we say that Y is of finite activity (FA) indeed all

such processes can be written as a compound Poisson process. In cases where this does not hold,
Y is said to be an infinite activity (IA) process as it has an infinite number of very small jumps in
any finite time interval.
Importantly, when we move on to Levy processes on the real line the condition (17) has to be
R
strengthen to |y|>0 min{1, y 2 }v(dy) < . This will be discussed in some detail in Section 4.3.
3.3.2

Positive tempered stable process

An important implication of the Levy-Khintchine representation is that Levy processes can be built
by specifying a and v directly, implying the probability distribution of Y1 . An important example of
this is the positive tempered stable, P T S(, , ), class which exponentially tilts the Levy density
pS (y; , ) of the P S(, ), to deliver


1
p(y; , , ) = e exp 2 y pS (y; , ),
2
17

y R>0 .

The resulting Levy density is

u(y) =




1 2
1
y
exp y ,
() (1 )
2

y, > 0, 0 < < 1, 0,

(18)

which means the process has infinite activity. The density of Y1 is not generally known in simple
form, however the Levy density is simple and plays such a crucial role that this difficulty is not
overly worrying. Special cases of this structure include the IG Levy density (12) and the Levy
density (13), which is the limiting case of 0. Notice that the constraint < 1 is essential
in order to satisfy the condition (17) in the Levy-Khintchine representation. The corresponding
kumulant function can be calculated by solving (16) implying


K{ Y1 } =

1/

+ 2

while for > 0 all cumulants of Y1 exist, the first two cumulants being
2 (1)/

and 4 (1 ) (2)/ .

Finally the kumulant function implies the convenient property that Yt P T S(, t, ).

3.4

Simulation for non-negative L

evy processes

3.4.1

Simulating paths via the L

evy density

For subordinators the paths of Levy processes can be simulated directly off the Levy density u(y),
y R>0 . Define the tail mass function
Z
+
v (y) =
u(x)dx,
y

which is a decreasing function for all y R>0 . Denote the inverse function of v + by v , i.e.


v (y) = inf x > 0 : v + (x) y .

Then the desired result, called the series representation, is that

(m)

Yt

m
X
i=1

v (bi /T )I(ui t),

for

0 t T,

(19)
i.i.d.

converges uniformly to Y on [0, T ] as m . Here ui U (0, T ), where U (0, T ) is the uniform

distribution on (0, T ), is independent of b1 < ... < bi < ... which are the arrival times of a Poisson
process with intensity 1. Clearly the computational speed of these techniques will depend upon
how expensive it is to compute v and how quickly v (y) falls as y increases.

18

Example 5 Compound Poisson process. Let the CPP have intensity and probability density
f (y) for the positive jumps, then the Levy density and tail mass function are u(y) = f (y) and
v + (y) = {1 F (y)}, implying
n
 o
(
1 1 y
,
F

v (y) =
0,

y<
y .

Hence the inverse only involves computing the quantiles of the jumps. Overall this implies
(m)
Yt

m
X

bi T

bi
T



I(ui t).

Clearly if bi > T , then there is no approximation by using this series representation. This method
has a simple interpretation. If we sample from



bi
1
F
1
until bi > T ,
T
then an ordered sequence from f (y) of size P o(T ) is produced. The effect of the I(ui t) term is
to sample randomly from this ordered sequence a random share of the draws. So the infinite series
representation samples compound Poisson processes rather effectively.
As a special case, suppose u(y) = exp(y) so that v + (y) = ey , which has the convenient
property that it can be analytically inverted:

 y 
1

v (y) = max 0, log

.

Hence as soon as y > then v (y) = 0, implying

(m)
Yt

 
m
1 X
bi
=
log
I(ui t).

t
bi t

For some types of subordinators special methods have been devised to simulate their paths
without inverting v .
Example 6 Rosinskis method for dealing with P T S(, , ) processes. This method approximates
the path over [0, T ] by the sum of non-negative terms
(
)

m
X
AT 1/ 1 1/
(m)
Yt
=
min
, B ei vi
I(ui t),
bi
i=1

where
A = 2

,
(1 )

B=

1 1/
,
2

19

for

0 t T,

(20)

I() is an indicator function, {ei }, {vi }, {bi }, {ui } are independent of one another and over i except
i.i.d.

i.i.d.

for the {bi } process. Here ui U (0, T ), vi U (0, 1), the {ei } are exponential with mean 1.
Further the b1 < ... < bi < ... are the arrival times of a Poisson process with intensity 1. Then
(m)

as m the process Yt

converges uniformly on [0, T ] to a sample path of the tempered stable

Levy process.
3.4.2

Simulation via a small jump approximation

We will break up the subordinator into

Y = Y 1 + Y 2,
where Y 2 corresponds to a compound Poisson process with jumps of absolute size greater than
> 0, while Y 1 will contain the small jumps. A first order approximation is to take as very small
and to neglect Y 1 entirely and simply simulate the straightforward Y 2 .
A second order approximation to the path of Y 1 is by using a Brownian motion, based on the
central limit result
 1 1


Yt t t0 {Wt }t0 ,

where =

yv(dy)

and 2 =

2 v(dy),

which holds under certain conditions. A more detailed

discussion of this is given in Barndorff-Nielsen and Shephard (2012a).

Processes with real increments

4.1
4.1.1

Examples of L
evy processes
Motivation

In this Section the focus will be on Levy processes with innovations which are on the real line.
Many of them play important roles in financial economics as direct models of financial assets.
4.1.2

Brownian motion

In Brownian motion we write

Y1 N (0, 1),
with density


1
1 2
fY1 (y) = exp y ,
2
2

y R,

while
K { Y1 } = log [E exp {Y1 }] =

1 2
.
2
20

The implication of this is that marginally Yt N (0, t), while the increments are independent as
usual with
Yt+s Yt N (0, s).
A standard Brownian motion, written W , can be generalised to allow for the increments to have
a non-zero mean and a different scale than one. A drift and a volatility term can be introduced
to deliver the Levy process
(a) Simulated scaled Brownian motion

0.4

(b) Simulated NIG Lvy process

0.5

0.3
0.4

0.2
0.1

0.3

0.0

0.2

0.1
0.1
0.2
0.0
0.3
0.1

0.4
0.5

0.2
0

10

10

Figure 3: (a) Sample path of

0.02 times standard Brownian motion. (b) Sample path of a
NIG(0 .2 , 0 , 0 , 10 ) Levy process. Thus the increments of the processes have a common variance.
Code: levy graphs.ox.

Yt = t + Wt ,
with increments
Yt+s Yt N (s, 2 s).
The associated kumulant function for Y1 is + 12 2 2 .
A graph of a sample path from standard Brownian motion is displayed in Figure 3(a). It
illustrates that the path is continuous. In a moment we will see that, except for the pure linear
drift case Yt = t, Brownian motion is the only Levy process with this property all other Levy
processes have jumps.

21

4.1.3

Compound Poisson process

Compound Poisson processes were introduced in (4), but there we required the shocks {Cj } to be
strictly positive. Here this condition is relaxed, just ruling out that they have an atom at zero. In
this case, again,
K { Y1 } = log [E exp {Y1 }]
= {exp K { C1 } 1}
(so long as K { C1 } exists).
i.i.d.

Example 7 Suppose that Cj N (, 2 ), then





1 2 2
K { Y1 } = exp + 1 .
2
So here the Levy process is constant until there is a new arrival from the Poisson process. The
arrival then moves the Levy process by a Gaussian variable. This variable can have a non-zero
mean and a non-unit variance. Quite a lot of effort has been expended on working on the derivative
pricing theory associated with this simple structure.
4.1.4

Skellam process

Let
(1)

Yt = Nt
(i)

where Nt

(2)

Nt ,
(i)

are independent Poisson processes with E(N1 ) = i > 0. This is a Skellam process

and has the attractive feature that it can be scaled to have a fixed tick size which may be helpful
for certain types of high frequency financial data.
The essential nature of this discreteness at the microscopic financial level is shown in Figure 4 which reports the first 80 best bid (squares) and best ask (crosses) rescaled prices from a
Euro/Dollar futures contract for 10th November 2008. For this contract the tick size is 0.0001 of a
unit, i.e. prices move from, for example, 1.2768 to 1.2767 U.S. Dollar to the Euro.
Now for the Skellam process
n
o
n
o 

(1)
(2)
K { Y1 } = K N1
+ K N1
= 1 e + 2 e 1 2 ,

so is infinitely divisible. The Yt is distributed as a Skellam Sk(t 1 , t 2 ) variable the distribution

of the difference between two independent Poisson variables with means 1 t and 2 t. A Sk( 1 , 2 )
probability function is
1 2

pk = e

 k/2

X
p
k1 k+n
1
1 2
2
=e
Ik (2 1 2 ),
k!(k + n)!
2

n=0

22

k = 0, 1, 2, ...,

Euro 10 Nov
2

-1

-2

Ask

10

20

30

40
Price number

50

60

70

80

Figure 4: From Barndorff-Nielsen, Pollard, and Shephard (2012)

where Ik (x) is a modified Bessel function of the first kind. Importantly E(Yt ) = ( 1 2 ) t and
Var(Yt ) = ( 1 + 2 ) t. Hence if 1 = 2 the process is a martingale.

4.2

Normal variance-mean mixture processes

A rather general way of building densities on the real line is by using a normal variance-mean
mixture
Y = + 2 + U,

(21)

where U N (0, 1) and U 2 ( denoting stochastic independence). In this mixture the

constraint = 0 would imply Y would be symmetrically distributed, while < 0 typically delivers
a negatively skewed density. Further, E(Y ) = + E( 2 ) so long as E( 2 ) exists and


1 2
2
C { Y1 } = i + K i ,
2
where for complex z,





K z 2 = log E exp z 2 .

Provided 2 is infinitely divisible this normal variance-mean mixture has an elegant time deformation intepretation which will be an important common theme to the threads of this book. Many of
the commonly used parametric densities used in financial economics fall within this class by specific
choices for the density for 2 . Here we discuss some of these important special cases.
23

4.2.1

Normal inverse Gaussian

If we assume 2 IG(, ) then Y N IG(, , , ), where =

Gaussian (N IG) distribution. The N IG Levy process puts
Y1 N IG(, , , ),

R, R>0 ,

2 + 2 , has a normal inverse

|| <

which has the density

fY1 (y) = a(, , , )q
p

1




y
K1 q
exp { (y )} ,

1 + y 2 , K1 () is a modified Bessel function of the third kind and

 q

1
2
2
a(, , , ) = exp .

where q(y) =

Sometimes it is convenient to reparameterise this model, noting that

= ,

= ,

are invariant under changes of the location and scale parameters and . A popular location-scale
invariant choice is achieved by defining

1/2
q

2
2
= 1+
and = ,

(22)

where , the steepness parameter, and , the asymmetry parameter, obey a triangular constraint
{(, ) : 1 < < 1,

|| < < 1} .

(23)

The flexibility of the model is shown in Figure 5, which displays the log-density for a variety of
values of , . Such a plot is called a shape triangle. As 0 so the log-density becomes more
quadratic, while for values of around 0.5 the tails are approximately linear. For larger values of
the tails start decaying at a rate which looks appreciably slower than linear. In the limit as 1
while = 0 the density becomes that of a Cauchy variable.
Note that it is sometimes more convenient to think of = / = /, rather than , as an
asymmetry parameter. A fixed value of corresponds to a straight proportionality line in the shape
triangle.
One of the NIGs attractions is that the cumulant function has the simple form
q

p
2
2
2
2
K { Y1 } =
( + ) + ,

which means that

Yt N IG(, , t, t).
24

Figure 5: Shape triangle for the NIG model. That is we graph the shape of the log-density for
the NIG model for a variety of values of the steepness parameter and the asymmetry parameter
. The representative distributions shown all have variance of one. This graph was kindly made
available to us by Preben Blsild.
In particular this implies that the increments of the NIG Levy process are non-zero NIG with
probability one. A sample path of a N IG Levy process is drawn in Figure 3(b). For this case we
have chosen there is symmetry and no drift (since = = 0) and the variance per unit of time is
the same as for the Brownian motion given in Figure 3(a). The process moves only by jumps, with
infinitely many jumps in any time interval, however small. The irregular size of the jumps implies
the very jagged shape of the picture.
4.2.2

Normal gamma process

If we assume 2 (, 2 /2) then Y N (, , , ), which we call the normal gamma (written

N ) distribution. From the cumulant function


+ 2 /2
K { Y1 } = log 1
,
2 /2

(24)

it follows that Yt N (t, , , t). This means this process is analytically simple to handle. The
density of the process at time one is
2 12
fY1 (y) =
K1/2
2()21


2
|y | exp { (y )} ,
2

(x) = x K (x).
where K

Note that 2 is a scale parameter and that and / are invariant under location-scale changes.

25

(a) Simulated N Lvy process

0.3

(b) Simulated Laplace Lvy process

0.35
0.30

0.2

0.25
0.1
0.20
0.0
0.15
0.10

0.1

0.05
0.2
0.00
0.3
0.05
0

10

10

Figure 6: (a) Sample path of a N (4,200,0,0) Levy process. Such processes are often called variance gamma processes in the literature. (b) Sample path of a La(0.2,0,0) Levy process. Code:
levy graphs.ox.
The special case of = 0 (symmetry) has been used extensively in the finance literature where it
is often called the variance gamma (V G) process. It was introduced by Madan and Seneta (1990),
while the 6= 0 case was developed by Madan, Carr, and Chang (1998). Later we will mention
a generalisation of the normal gamma process, the extended Koponen (or KoBol) class, which is
often referred to in the finance literature as the CMGY Levy process.
Figure 6(a) graphs a simulated path from a N (4, 200, 0, 0) process. As we would expect, the
sample path has some of the features of the N IG process we drew in Figure 3(b). In particular
both of these infinite activity processes are very jagged. More detailed mathematical analysis of
the corresponding Levy measures shows that the N IG process has more very small jumps than the
N process. In particular near zero the Levy density u(y) of the N IG process behaves like y 3/2 ,
while the corresponding result for the N process is y 1 .
An interesting feature of the normal gamma process is that it can be written in the form Y+ Y
where Y+ and Y are independent subordinators (see Exercise 8).
For comparison with the NIG log-density (see Figure 5), note that N log-density either has a
cusp at zero, when (0, 1], or is concave.

26

4.2.3

Hyperbolic, Laplace and skewed Students t processes

If we assume 2 P H(, ) then Y H(, , , ), where =

2 + 2 , has the hyperbolic

distribution. This distribution can be shown to be infinitely divisible, although the proof of this
is difficult see Barndorff-Nielsen and Shephard (2012a). The hyperbolic process puts Y1
H(, , , ), where the density is

fY1 (y) = p
2 2 + 2 K1 ()

 q

2
2
exp + (y ) + (y ) ,

y R.

All the moments of this process exist so long as > 0, while the cumulant function is
 q


2 ( + )2



K

1
2
+
log
K { Y1 } = log
+ .

2
K1 ()
2 ( + )2

(25)

(26)

We can compare the hyperbolic model to the N IG density using the shape triangle. In particular

reparameterise into the location-scale invariant parameters given in (22), then Figure 7 shows the
log-densities for this model. We see that again as 0 we get the normal quadratic log-density.
For higher values the log-density gets increasingly linear decay in the tails as 1. Indeed in the
limit we get the Laplace densities, see below. This contrasts with the N IG density which has the
ability to have thicker tails. Hyperbolic laws have the interesting and important feature that the

Figure 7: Shape triangle for the hyperbolic model. That is we graph the shape of the log-density
for the hyperbolic model for a variety of values of the steepness parameter and the asymmetry
parameter . This graph was kindly made available to us by Preben Blsild.

27

log-density is a hyperbola (hence their name) and so behaves approximately linearly in the tails of
the distribution.
Hyperbolic Levy processes have the disadvantage that we do not have an exact expression for
the density of Yt for t 6= 1, nor can we simulate from the process in a non-intensive manner. Both
of these properties are inherited from the fact that this is also the case for the positive hyperbolic
process we discussed in the previous Section.
The Laplace distributions (symmetric and asymmetric) occur as limiting cases of (25) for ,
and fixed and 0. We write this as La(, , ). The corresponding density is
2 2
exp { |y | + (y )} ,
2

where =

2 + 2,

(27)

which is achieved by 2 E( 2 /2) = (1, 2 /2). One of the main features of this model is that
Yt N (t, , , t).

Finally, if we assume 2 R(, 2 /2) then Y T (, , , ), a skewed Students t distribution,

which is infinitely divisible and so can be used as the basis of a Levy process. The skewed Students
t Levy process puts Y1 T (, , , ), where the density is

 q

2 ||12
2
2

K1/2 || + (y ) exp { (y )} ,
2()21

where, for any R, K (x) = x K (x). The more familiar Students t distribution is found when
we let 0, then the density becomes
( + 1/2)

2 ()

1+

2 )1/2

In the symmetric case only moments of order less than will exist at any time horizon. However,
we do not know the distribution of Yt for this process in any explicit form, while simulation has
to be carried out in quite an involved manner. Hence this process is not as easy to handle as the
N IG or normal gamma Levy processes.
4.2.4

Generalized hyperbolic process

If we assume 2 GIG(, , ) then Y GH(, , , , ), where =

hyperbolic (GH) distribution. Its density is
 q

2 12
2
2
K 1 + (y ) exp { (y )} ,
K () 2 2

p
2 + 2 , the generalised
(28)

where K is the modified Bessel function of the third kind and K (x) = x K (x). It is helpful
to recall that K (y) = K (y). This distribution includes as special cases many of the above

28

distributions in the following way:

N (, 2 ) = lim GH(, , 0, , 2
),
N RIG(, , , ) = GH 12 , , , , ,
T (, , , ) = GH(, , , , ),
N (, , , ) = GH(, , , , 0),

N IG(, , , ) = GH 12 , , , , ,
H(, , , ) = GH(1, , , , ),
La(, , ) = GH(1, , , , 0),
RH(, , , ) = GH(1, , , , ),

for > 0. The GH distribution is infinitely divisible (a proof is given in Barndorff-Nielsen and
Shephard (2012a)) and so can be used as the basis of a rather general Levy process. The special
cases not introduced above are
normal reciprocal inverse Gaussian distribution (NRIG), which happens when 2 RIG(, ).
reciprocal hyperbolic (RH ), which happens when 2 RP H(, ).
The cumulant function of the GH is

 q




K ( + )

n
o
K { Y1 } = log
+
log
+ ,
p

2
2 2
2 ( + )2

| + | < ,

while the first two moments (when they exist) are

E(Y1 ) = +
Var(Y1 ) = 2

K+1 ()
K ()

and
"

 #!
K+1 () 2 K+2 ()
K+1 () 2
+

.
K () 2
K ()
K ()

Not surprisingly, in general we do not know the GH density of Yt for t 6= 1, nor can we simulate
from the process in a non-intensive manner. This model is so general that it is typically difficult to
manipulate mathematically and so is not often used empirically. Instead special cases are usually
employed.
4.2.5

Normal positive stable and symmetric stable processes

If we assume 2 P S(/2, ) then Y has a normal stable distribution, S(, , , ), which is

infinitely divisible and so supports a Levy process. The important special case where = = 0
is the well known symmetric stable Levy process with index . Except for the boundary case of
= 2, the symmetric stable distribution has the empirically unappealing feature that the variance
of Y1 is infinity. The density of this variable is unknown in general, with exceptions being the
Gaussian variable ( = 2) and the Cauchy variable ( = 1).
Despite the complexity of the density of a stable random variable the cumulant function is
simply
C { Y1 } = || ,
29

which implies Yt S(, 0, 0, t). Note that for this process the mean exists only if > 1. The
Levy density for a symmetric stable process is given by
u(y) = |y|1 ,

0 < < 2.

(29)

Stable Levy processes have the remarkable property that for > 0
{Yt }t0

n
o
1/
= Yt
L

t0

(30)

Thus, in particular, increments of the Levy process over time t are, in distribution, just scaled
versions of increments over time t. This fractal like property is called self-similarity, and stable
Levy processes (symmetric or not) are the only Levy processes which possess this feature.
Although stable processes have received considerable attention in financial economics since their
introduction into that subject in the early 1960s, it has been known since the late 1960s that they
provide a poor fit to the empirical data we usually see in practice. This is because returns over long
time intervals, which are sums of returns over finer time intervals, tend to be more Gaussian than
the ones over short horizons. Hence our interest in this type of process will usually be to provide
theoretical illustrations, rather than as practical models.
4.2.6

Normal tempered stable process

If we assume 2 P T S(, , ) then Y N T S(, , , , ), a normal tempered stable distribution.

The N T S distribution is infinitely divisible and so can be used to generate a Levy process. Special
cases of the normal tempered stable process is the N IG Levy process and the normal gamma Levy
process (when 0). This process will be discussed in more detail in Example 5.2.3.

4.3

L
evy-Khintchine representation

The Levy-Khintchine representation for positive variables given in (16) can be generalised to cover
Levy processes with increments on the real line. Four basic developments are needed. First, the
Levy measure must be allowed to have support on the real line, not just the positive half-line, but
still excluding the possibility that the measure has an atom at zero. Second, the parameter a needs
to be allowed to be a real, not just positive. Third, we imagine that an independent Brownian
motion component is added to the process. Fourth, the below technical condition on the Levy
measure (32) must be satisfied. The result is the celebrated Levy-Khintchine representation for
Levy processes.
Theorem 3 Levy-Khintchine representation. Suppose Y is a Levy process. Then
Z n
o
1 2 2
C{ Y1 } = ai +
eiy 1 iy1[1,1] (y) v(dy),
2
R
30

(31)

where a R, 0, and the Levy measure v must satisfy

Z
min{1, y 2 }v(dy) <

(32)

and v has no atom at 0.

Levy processes are completely determined by the characteristic triplet (a, 2 , v): the drift a, the
variance 2 of the Brownian motion and the Levy measure v (which has to satisfy (32)). Importantly
only processes with v = 0 do not have jumps but in that case Y is a scaled Brownian motion
with drift a. The representation implies that all Levy processes can be decomposed into
Yt = at + Wt + Ldt ,
a drift, a scaled Brownian motion W and an independent pure jump process Ldt .
There are many interesting additional features of the Levy-Khintchine representation which we
bring out in Barndorff-Nielsen and Shephard (2012a). Important points are that the centring function 1[1,1] (y) is one choice amongst many and that many interesting properties of the distribution
of Y1 can be deduced from direct inspection of v (e.g. existence of moments and unimodality).

4.4

Blumenthal-Getoor index

A key measure of the degree of variation of Levy processes is the Blumenthal-Getoor index defined
by
 Z 



= inf :
1 |y| v(dy) < .
R

This index has the property that

n 

X


p lim
Y i Y i1  <
n

i=1

for any > .

Clearly for any CPP we have = 0; but what about infinite activity processes? In general we
know 2 as the limit exists for quadratic variation, but more generally? The Blumenthal-Getoor
index is determined by the behaviour of the Levy measure near zero, with larger indexes arising
when the Levy density goes to infinity quickly as |y| 0. Informally the higher the index the
larger the frequency of small jumps in the Levy process. Some examples of the index are given in
Table 1.

4.5

Power variation

Related to the Blumenthal-Getoor index is power variation, which is

n 
r
X


[r]
r/21
{Y } = p lim n
Y i Y i1  .
n

i=1

31

name
IG

PTS
NIG
N
PS

w(y)
cy 3/2 exp(y)
cy 1 exp(y)
cy 1 exp (y)
c |y|2 for small y
c |y|1 (for small y)
c |y|1

B-G index
1/2
0

1
0

Table 1: Blumenthal-Getoor index for various infinitely divisible distributions.

In a moment we will see its most famous case where r = 2, which is quadratic variation
{Y }[r] = 2 +

(Ys )2 ,

s1

but the other cases are interesting too. In particular when r < 2 we have that
{Y }[r] = r r ,

r = E |U |r ,

U N (0, 1),

which reveals the volatility of the Brownian motion component of the Levy process, while when
r > 2 then power variation goes off to infinity. This suggests that sums of power of absolute
returns maybe useful ways of assessing the importance of jumps in financial economics. This was
first formalised in finance by Barndorff-Nielsen and Shephard (2004), although the mathematics
behind this goes back further (see, for example, Jacod and Protter (2012)).
2.25
x

(a) Density of IG

2.25

(b) Levy-Ito form

2.00

2.00

1.75

1.75

1.50

1.50

1.25

1.25

1.00

1.00

0.75

0.75

0.50
0.50
0.25
0.25
0.0

0.5

1.0

1.5

density

2.0

10

15

20

time

25

Figure 8: (a) rotates the IG(1 , 2 ) density for 1 . (b) displaying the Poisson basis N (dy, dt).
Filename is levy graphs.ox.

32

4.6

L
evy-Ito representation

An important application of Levy bases is in the Levy-Ito representation of Levy processes.

Theorem 4 Levy-Ito representation1 .

Let L be a Levy process with Levy measure v. Then L has

the representation
Lt = at +

bWt +

Z tZ
0

y{N (dy, ds) 1[1,1] (y)v(dy)ds}.

(33)

where W is a Brownian motion and N is a Poisson basis on R>0 R, independent of W and with
mean measure v(dy)dt.
That N is a Poisson basis means that it is a random measure which attaches a random number
N (A) to any (bounded Borel) subset of the underlying space, in this case R>0 R, and that the
numbers are independent for disjoint subsets. Furthermore, since we are talking about a Poisson
basis the law of N (A) is Poisson with mean equal to the integral over A of the measure v(dy)dt.
The representation may be given the alternative form

Lt = at + bWt
Z
+
y{Nt (dy) tv(dy)}
{|y|<}
Z
yNt (dy),
+
{|y|}

where
Nt (dy) =

Ndy,ds
0

and is an arbitrarily chosen positive number (here a will generally depend on .).
This is insightful in a number of ways. For example, it demonstrates that Levy processes are
always semimartingales for we can decompose
At = at +

1{|Ls |} Ls

and

Mt =

bWt +

0<st

{|y|<}

y{Nt (dy) tv(dy)}.

The latter is obviously a martingale, while A is of finite variation (since the number of jumps Ls
R
of absolute size is locally finite, as follows from the fact that {|y|} v(dy) < .)
The mean value of L1 exists if and only if (cf. Sato (1999, p. 39))
Z
|y|v(dy) < .
|y|>1

In this case the representation (33) can be recast in the form

Z tZ

Lt = E (Lt ) + bWt +
y{N (dy, ds) v(dy)ds}.
0

For a proof see, for instance, Sato (1999, Ch. 4).

33

(34)

If, moreover, the Levy measure is a member of the infinite activity finite variation (denoted IAFV)
class then
Lt = a0 t +

Z tZ
0

yN (dy, ds),

(35)

where the drift a0 is given by

Z
yv(dy).
a0 = E (L1 )
R

Recall that finite variation means here that the sums of the absolute values of the infinitesimal
increments of the process are bounded with probability one.
Any non-negative Levy process L without drift is representable in the Levy-Ito form
Z tZ
Lt =
yN (dy, ds)
0

(36)

with v satisfying
Z
min{1, y}v(dy) < .
0

Time deformation and time-change

5.1

Basic framework

Financial markets sometimes seem to move more rapidly than in other periods. One way of starting
to model this is to allow the relationship between calendar time and the pace of the market to be
random. We call a stochastic process which models the random clock a time-change, while the
resulting process is said to be time deformed.
Definition 5 A time-change is any non-decreasing random process T with T0 = 0. The special
case where the time-change has independent and stationary increments is called a subordinator.
The requirement that the time-change is non-decreasing rules out the chance that time can go
backwards. A special case of a time-change is a subordinator, while subordinators are special cases
of Levy processes (e.g. Poisson or IG Levy processes are subordinators). All subordinators are pure
upward jumping processes plus non-negative drift. We should note here that the finance literature
often labels time-changes subordinators, while in the probability literature the term subordinator
is reserved for positive Levy processes and we stick to the latter convention in this book.
In this Section we will study what happens when a subordinator is used to change the clock,
that is deform, a stochastically independent Levy process L. The result is
Yt = LTt ,
34

which is often written as

Yt = L Tt ,
or
Y = L T.
The increments of this process are
Yt+s Yt = L Tt+s L Tt ,
which are independent and stationary and so Y is a Levy process.
Brownian motion is the only Levy process with continuous sample paths, however this property
does not survive being deformed by a pure subordinator for such a process has to be a pure jump
process.

5.2
5.2.1

Examples
Compound Poisson process

Let T be a Poisson process, independent of a scaled Brownian motion W and so that

Y = T + W T,
then Y1 |T1 N (T1 , 2 T1 ). This is a compound Poisson process
Yt =

Tt
X

Cj ,

j=1

i.i.d.

Cj N (, 2 ),

(37)

with shocks which are Gaussian. The implication is that when the process jumps, the jumps are
independent of the time we have waited until the jump.
5.2.2

Normal inverse Gaussian process

Suppose T is an IG(, ) Levy process and Yt = t + Tt + WTt , where W is Brownian motion,

then
Y1 |T1 N ( + T1 , T1 )
and so unconditionally the increments are independent with
Y1 N IG(, , , ),

2 = 2 + 2 .

Hence this deformed Brownian motion is the NIG Levy process, which we simulated in Figure 2(b).
Likewise the N Levy process can be obtained by subordinating Brownian motion plus drift by a
gamma process.
35

5.2.3

Normal tempered stable process

Suppose T is a P T S(, , ) Levy process. If Yt = t + Tt + WTt , then Y is called a normal

tempered stable (N T S) Levy process. We write
q
Y1 N T S(, , , , ), where = 2 + 2 ,
but the corresponding probability density is generally unknown (except for an infinite series representation, see Feller (1971b, p. 583)). The cumulant function, on the other hand, is rather
simple
n
o
K { Y1 } = + 2 ( + )2 .
The form of this function implies Yt N T S(, , , t, t). It can be shown (after some considerable
work), using the cumulant function of the N T S process, that the Levy density is

2+1 + 1 (+ 1 )
2 K
2 |y|
u(y) =
+ 21 ( |y|) exp {y} .
2 (1 )
The direct use of this Levy density is obviously going to be difficult due to its complexity. The
deformation interpretation will mean that we can usually sidestep this, instead employing the simple
Levy density of the P T S(, , ) Levy process given in (18).
5.2.4

Type G and P L
evy processes

Definition 6 We call Levy processes which can be written as Yt = t + Tt + WTt , for some
subordinator T , type G Levy processes.
This is the subset of Levy processes for which there is a deformation of Brownian motion with
drift interpretation.
Definition 7 We call Levy processes which can be written as Yt = NTt , for some subordinator T
and where N is a Poisson process, type P Levy processes.
This is the subset of Levy processes for which there is a deformation of Poisson process representation. We will now give some examples of this.
5.2.5

Negative binomial process

Suppose T is a (, ) Levy process and let Yt = NTt , where N is a standard Poisson process with
unit intensity and N T . Then Y is a negative binomial Levy process with Yt having a negative
binomial distribution


1
N B t,
,
1+
36

which follows from the well known Poisson-gamma mixture. Of course if = and then
Tt t and so Yt becomes like a Poisson process with intensity . For this model
h
n
oi
K { Y1 } = log E exp (e 1) , where (, ),
n

o
= K e 1


1 e
= log 1 +
,

which is a reparameterisation of (6). Then E(Y1 ) = /.

5.2.6

Discrete Laplace process

A special case of this is where T is an exponential (1, ) Levy process. Then Y1 has a geometric
probability function
Pr(Y1 = y) = p (1 p)y ,

y = 0, 1, 2, ...,

p=

1
,
1+



(1)
(2)
1
while generally Yt N B t, 1+
. Further let Yt = N (1) N (2) , where N (1) , N (2) , T (1) and T (2)
Tt

Tt

are independent processes, the first two of which are Poisson processes and the last two exponential
processes. Then Yt is a discrete Laplace Levy process while
Pr(Y1 = y) =

1 p |y|
p , y = 0, 1, 2, ....
1+p

(38)

This has a zero mean and variance of 2p/(1 p)2 . It is a heavier tailed alternative to the Skellam
distribution discussed in Section 4.1.4, but again it only has one parameter.
5.2.7

Poisson-IG process

More generally if T is a GIG Levy process and we let Yt = NTt , then we call the result a PoissonGIG Levy process. The most important special case is the gamma one just discussed and the
IG(, ) case. Writing 2 = 2 + 2, the P-IG(, ) variable has


Z y


e 3/2
1 2 1
2
e
Pr(Y1 = y) =
exp + d,
y!
2
2
0


Z

1 2 1
y 12 )1

2
(
e
=

exp + dy,
2
y! 2
0
p

K
1 ( ) (/ )
2 / y 2

=
e
, y = 0, 1, 2, ...,
y!
2
where K () is a modified Bessel function of the third kind. This follows from using the properties
of the GIG distribution. The Yt is marginally P-IG(t, ).
Of course this is a finite activity process with
n

o
K { Y1 } = K 1 e , where IG(, ),
37

PoissonIG and Poisson process

40

Yt
30
PoissonIG
Poisson

20
10

10

12

Logprobability function for PoissonIG and Poisson

10

5
PoissonIG
Poisson

10
15
20
0

14

16

18 k

20

Figure 9: Top graph: Simulation from a Poisson and Poisson-IG(1.5, 0.5) process with mean intensity of 3. It shows the larger gaps in the P-IG process and larger jumps. Bottom graph: log of
the probability function for Y1 for the Poisson and P-IG distributions. It shows the approximate
quadratic tails of the Poisson distribution and the approximate linear tails of the P-IG distribution.
Code: poisson.ox.


= +2 1e

o1/2 

Now E(Y1 ) = /. For 1/2 this can be thought of as a compound Poisson process with intensity
and a discrete mixing distribution (5) for C1 whose cumulant function is
n

o1/2


exp K ( C1 ) = 1 + 2 + 2 1 e
, 1/2.
From this E(C1 ) = 1/.

The top part of Figure 9 shows a simulated sample path from a Poisson-IG process, it illustrates
the longer gaps between arrivals and the occasional very large new arrival, compared to a Poisson
process. The two processes are simulated to have the same mean, but the Poisson-IG process has
= 1.5 and = 0.5. The bottom part of the Figure show the log of the probability function for
the Poisson-IG and Poisson distributions for Y1 , which has equal means of 3. It shows that the
Poisson distribution decays approximately quadratically in the tails, while the Poisson-IGs decay
is more like linear.
5.2.8

Time-changed Skellam processes

Suppose that


Yt = N (1) N (2)

Tt

(1)

(2)

= NTt NTt ,
38

where N (i) are independent Poisson processes with unit intensity and Tt is a gamma process. Now


N (1) N (2) t is marginally a Skellam variable with
(1)

Pr(Nt

(2)

Nt

= k) = e2t I|k| (2t),

k = 0, 1, 2, ...,

whose cumulant function is



K { Yt } = t e + e 2 .
This means that for a random time change T independent of Y we have that
n
o
K { YT1 } = K e + e 2 T1 ,
which has an analytic solution for all cases where the cumulant function of T1 is known. Leading
examples are the gamma and inverse Gaussian cases.
Although the cumulant function is simple the probability function is less easy to work with.
For example in the Skellam-gamma case
Z

(1)
(2)
e2 I|k| (2) 1 e d
Pr(NT1 NT1 = k) =
() 0
Z

=
I (2) 1 e(+2) d.
() 0 |k|
An analytic expression for this integral does not seem to be available.

Empirical estimation and testing of GH L

evy processes

6.1
6.1.1

A likelihood approach
Estimation of GH L
evy processes

Here we will assess how well Levy processes fit the marginal distribution of financial returns. Their
flexibility allows important improvements over conventional Brownian motion, however they clearly
neglect the dynamics of returns.
Throughout we assume that Y is observed at unit time intervals which we will think of as
representing a day. Then inference will be based on the increments
yi = Yi Yi1 ,

i = 1, 2, ..., n.

Under the Levy assumption

f (y1 , ..., yn ) =

n
Y

fY1 (yi ),

i=1

where fY1 is the density of Y1 .

39

Typically fY1 is indexed by some parameters which are written as . Inference on is carried
out using the likelihood
log f (y1 , ..., yn ; ) =

n
X

log fY1 (yi ; ).

i=1

For concreteness we will now focus on the case where Y is a GH Levy process. Then fY1 (y; ) is
 q

2 12
K 1 2 + (y )2 exp { (y )} .
2
K () 2
p
The parameters of the GH distribution are = (, , , , ) , where = 2 + 2 . The maximum
likelihood (ML) estimator of is given by
b
=arg max log f (y1 , ..., yn ; ),

which has to be determined by numerical optimisation, either directly or via the EM algorithm.
The latter is particularly effective in the multivariate case and is used in Section 9.2.2 while the
associated theory is detailed in Barndorff-Nielsen and Shephard (2012c). In the case where Y is
univariate the optimisation is carried out using the Broyden, Fletcher, Goldfard and Shanno (BFGS)
quasi-Newton algorithm made available in the matrix programming language Ox by Doornik (2001)
and takes a handful of seconds for sample sizes of around 3, 000.
Confidence intervals for the parameters can be constructed via the asymptotic distribution of
b
based on the Levy assumption of i.i.d. increments. As already mentioned, the independence
assumption is clearly unrealistic. In a later subsection we will discuss the impact of this misspecification on confidence intervals. For now we stand by the Levy assumption.
The asymptotic theory for ML estimators means that


L
T b
N (0, I 1 ),

as

n ,

(39)

where I is the expected information per observation which is



Si
log f (yi ; )
I = E
= Cov (Si ) , where Si =
.

(40)

For inference expected information is usually replaced by averaged observed quantities

n

IS =

1 X Si
n

or

i=1

IO =

1X
Si Si .
n

(41)

i=1

The terms Si and Si / are found by numerical differentiation. This allow us to construct
asymptotically valid t statistics for elements of b
. In particular a 95% asymptotic confidence
interval for can be found as
r
1 1
1.96
b
(I ) ,
n

(42)
40

ML estimates
Outer product (IO )
Second derivative (IS )
Robust: m=250
Robust: m=500

-0.0133
(.0088)
(.0087)
(.013)
(.016)

MLE of GH parameters

0.179
0.419
1.95
(.113) (.053) (.43)
(.115) (.049) (.39)
(.14)
(.071) (.39)
(.17)
(.076) (.39)

-1.62
(.71)
(.65)
(.74)
(.76)

GH
-626.74

Likelihoods
=0
N (, 2 )
-627.97 -819.32

Table 2: ML estimates of GH for the Canadian daily exchange rate. Brackets are the asymptotic
standard errors computed using different estimates of the expected information: IO and IS . GH
column denotes the likelihood for the unrestricted model. = 0 imposes symmetry.
Robust,
denotes robust standard errors computed using m lags, which will be explained in a moment.
where I 1

denotes the diagonal element of I 1 corresponding to .

To illustrate the above methods we go back to the daily exchange rate series for the Canadian
Dollar rate against the US Dollar. The results are given in Table 2. The ML estimate of is quite
negative, while and are close to zero. The asymptotic standard errors for and are quite
large and suggest both and are not significantly different from zero. Interestingly the standard
errors based on IO and IS are very similar indeed.
The Table also gives the likelihood when is constrained to be zero. The likelihood drops by
around one, which again suggests can be set to zero. Finally the Table shows the GH model
improves upon the Gaussian likelihood fit by around 193, which is a very large improvement.
6.1.2

Confidence intervals via profile likelihoods

An alternative way of quantifying uncertainty is based on the likelihood ratio statistic. Again
suppose our focus is on . Define = (, , , ) , so that = (, ) , and

c =arg max log f (y1 , ..., yn ; , ).

c is a constrained ML estimator of , imposing on an a priori fixed value of . Likelihood theory

tells us that if we constrain correctly then the likelihood ratio statistic
o
n
L
) log f (y1 , ..., yn ; ,
c ) 21 , as n ,
2 log f (y1 , ..., yn ; b

then the ratio should typically take on unusually large values. We will be plotting the profile
likelihood
log f (y1 , ..., yn ; ,
c ) log f (y1 , ..., yn ; b
) against

to indicate plausible values of . Of course, while the 21 distribution is only valid if the model is
accurate, which certain features are demonstrably not, it is still interesting to plot out the profile.
The top left graph in Figure 10 draws the profile likelihood function for for the Canadian
Dollar example. This gives a similar result to the t statistics given in Table 2 with ranges of
41

Canadian Dollar

German Mark

French Franc

Generalised hyperbolic
Student or normal gamma

5
10

10
10
15

20
Generalised hyperbolic
Student or normal gamma

2.5
0.0
Swiss Franc

20
Generalised hyperbolic
Student or normal gamma

20
2.5
0

2.5
0.0
Japanese Yen

2.5
0

2.5
0.0
UK Sterling

2.5

Generalised hyperbolic
Student or normal gamma

Generalised hyperbolic
Student or normal gamma

10

10
10
15
15

20

20

Generalised hyperbolic
Student or normal gamma

20
2.5

0.0

2.5

2.5

0.0

2.5

2.5

0.0

2.5

Figure 10: Daily exchange rate data. Profile likelihood (truncated at -25) for the parameter of the
GH. Also profile likelihood for the skewed Student ( < 0) and the normal gamma ( > 0). NIG
case corresponds to the generalised hyperbolic curve at the point = 0.5.
approximately 2.5 to 0.2 supported by the data. The Figure also shows the profile likelihoods
for the normal gamma and skewed Students t, special cases of the GH model. Recall in the normal
gamma model is set to zero, while in the skewed Students t case = 0. Of course the likelihoods
for these models cannot exceed that of the GH model, but this plot shows how far these models
fall behind the GH model. We can see that for very negative the likelihood for the GH model is
the same type of that as the skewed Students t model, for the ML of turns out to be zero. The
same effect can be seen for large values of for then the ML of is zero. The Figure shows that
the skewed Students t model performs quite well, but the normal gamma process has some very
significant difficulties.

6.2

Model misspecification: robust standard errors

Levy processes allow us to flexibly model the distribution of i.i.d. increments, however in financial
economics returns exhibit volatility clustering. Later models will be developed which deal with
this feature, but for now our Levy models are misspecified. Even though our models are incorrect,
estimation by ML methods makes sense. We will now be modelling the marginal distribution of
the increments.
The theory of estimating equations implies




L
n b
N 0, I 1 J I 1 ,

as

n ,
42

(43)

where
n
X
1
Si
J = lim Cov
n n
i=1

|y i |

0.050

and

n
X
1
Si
I= E
n

i=1

0.15

0.050
0.025

0.10

0.025
0.000

0.05

0.025

0.00

0.15

0.000

200

400

0.025

200

400

0.15

0.15

0.10

0.10

0.05

0.05

0.00

0.00

200

400

200

400

0.10
0.05
0.00

200

400

200

400

Figure 11: Correlograms for the Canadian data drawn against lag length. Top left is of |yi |, other
graphs are for the five elements of Si . Code: em gh.ox.
In order to construct robust standard errors for b
we estimate the elements of the sandwich

I 1 J I 1 . The empirical average IS will consistently estimate I so long as the process is ergodic.
However, unless the returns, and so the scores Si , are i.i.d. IO will not correctly estimate J . Figure

11 shows the correlograms in the Canadian Dollar case (evaluated at b

) of the elements of the score

and |yi |. It shows the scores for , and have correlograms which are close to that of |yi |. The

scores for and are much less dependent.

There is a large literature on estimating J in the presence of autocorrelation. J is just the zero
frequency of the spectral matrix of the vector Si process, i.e.

X
J = Cov (Si ) +
{Cov (Si , Sis ) + Cov(Sis , Si )} ,
s=1

which can be estimated by

n
m
X
1 X

JO =
Si Si +
K(j; m)
n
s=1
i=s+1

n
n
1 X
1 X

Si Sis +
Sis Si
n
n
i=s+1

i=s+1

where K(j; m) denotes a non-negative Bartlett smoothing window






 j 
1  j  ,
 m+1  1,
m+1


K(j; m) =
 j 
0,
 m+1  > 1,
43

while m is called the lag truncation parameter.

Canadian Dollar

German Mark

French Franc

0
-2.5

-2

-2.5

-5
-4

-5

-7.5

Fitted GH model
Non-parametric estimator

-7.5

-1
0
Swiss Franc

1
0

Fitted GH model
Non-parametric estimator

Fitted GH model
Non-parametric estimator

-6

-2.5
0
Japanese Yen

2.5
0

-2.5
UK Sterling

2.5

-2.5

-2

-2
-5

-4

-4

-7.5
-6

Fitted GH model
Non-parametric estimator

-2.5

2.5

Fitted GH model
Non-parametric estimator

-10

Fitted GH model
Non-parametric estimator

-6
-5

-2.5

2.5

-2.5

2.5

Figure 12: Log of the estimates of the unconditional density of the returns for six exchange rates
against the U.S. Dollar. Also plotted is the log-density for fitted GH model.
Table 2, considered earlier, reports the results from using the above methods to compute the
robust standard errors for the Canadian Dollar dataset. The results do not vary much with m.
The corresponding fitted generalised hyperbolic log-density for the Canadian Dollar is given in
the upper left hand graph in Figure 12. This shows sub-log-linear tails in the marginal distribution.
The fit of the model is very close to the drawn non-parametric estimate of the log-density. The
non-parametric estimator is the log of the Gaussian kernel estimator coded in Applied Statistics
Algorithm AS 176 by Bernard Silverman, which is available at StatLib and NAG and in many
P 2
statistical software environments. The bandwidth is chosen to be 1.06b
n1/5 , where
b2 =
yi /n

(this is an optimal choice against a mean square error loss for Gaussian data).

6.3
6.3.1

Further empirical results

Six daily exchange rate movements

Table 3 gives the estimates of the parameters, together with their standard and robust standard
errors, for our daily exchange rate return data sets. The corresponding fitted log-density for all six
series was given in Figure 12. This shows sub-log-linear tails in the marginal distributions for all the
fitted distributions except for and German Mark and Sterling, which has approximately log-linear
tails. And, as already notes, the fit of the model is very close to the drawn non-parametric estimate
of the log-density.
44

Partly repeating some of the discussion above, there are a number of common features across
these results. First all the non-Gaussian models provide dramatic improvements over the fit of
the normal likelihood. The parameters seems to take values between 2 and 0.5, while neither
nor are close to zero. To reinforce this, Figure 10 showed the profile likelihood function for
each of the datasets. Also drawn are the corresponding profile likelihoods for the skewed Students
t and normal gamma models. The results indicates that the normal gamma model is not really
supported by the data. The skewed Students t model fits better primarily as it has fatter tails.
Typically when = 0, is around 2, which corresponds to 4 degrees of freedom for the Students
t distribution. The fit of the distribution is very sensitive to this value. The skewed Students t
is dominated by GH models with > 0. The likelihood function is typically flat for GH models
with between 2 and 2. Overall, however, the values between 2.0 and 0 seem best. Finally,
the special cases of imposing = 0 seems not to harm the fit a great deal for exchange rate data,
although there is slight statistical significance in the negative skewness in the UK Sterling, Swiss
Franc and Japanese Yen series.
Rate
Canada
DM
FF
SF
JY
Pound

-0.013
(.016)
0.024
(.033)
0.030
(.029)
0.073
(.037)
0.044
(.026)
0.034
(.018)

MLE of

0.179
(.172)
-0.064
(.063)
-0.074
(.060)
-0.143
(.065)
-0.120
(.041)
-0.089
(.039)

GH parameters

0.419
1.95
(.076) (.399)
0.873
1.40
(.129) (.265)
0.930
0.923
(.083) (.208)
1.08
1.25
(.094) (.214)
0.800
0.841
(.069) (.144)
0.430
1.89
(.090) (.097)

-1.62
(.769)
-0.979
(1.10)
-1.55
(.614)
-1.27
(.834)
-1.12
(.372)
0.145
(.439)

Likelihoods
=0
=0
-638.19
-627.61

GH
-626.74

=0
-627.97

-3,903.1

-3,903.8

-3,909.8

-3,905.0

-4,052.2

-3,797.1

-3,798.1

-3,809.3

-3,798.1

-3,988.2

-4,296.9

-4,300.6

-4,302.1

-4,298.3

-4,428.2

-4,022.6

-4,027.1

-4,042.8

-4,025.5

-4,310.3

-3485.7

-3487.5

-3492.5

-3495.1

-3704.8

N
-819.32

Table 3: Fit of GH for daily exchange rates. GH denotes unrestricted model. = 0 imposes
symmetry, = 0 skewed N model, = 0 skewed Student. Robust S.E.s ( m = 500) are in brackets.

6.3.2

Daily equity indexes

Table 4 gives the estimates of our parameters for the daily equity return data. The corresponding
profile likelihoods are given in Figure 13. These results are more mixed, with values of between
1 and 1 being roughly necessary. Overall again the normal inverse Gaussian usually does pretty
well, never fitting really poorly. One conclusion from these fitted models is that there seems very
little asymmetry in this data. This is perhaps surprising as this is always an important possibility
for equity data. The improvement over the Gaussian fit is picked up very well in the discrepancy
between the Gaussian and the GH likelihood fits. This holds across all the assets, but is less severe

45

for FTSE which is not surprising given its normal gamma like behaviour.
Index

0.234
(.091)
0.043
(.041)
0.124
(.027)
0.009
(.038)

DAX 30
FTSE 100
S&P 500
Nikkei 500

MLE of

-0.113
(.037)
-0.004
(.037)
-0.059
(.019)
-0.007
(.029)

GH parameters

0.983
0.821
(.126) (.133)
0.734
1.99
(.328) (.190)
1.05
0.655
(.191) (.236)
1.08
0.780
(.293) (.322)

-0.206
(.836)
1.72
(1.95)
-1.03
(.872)
-0.654
(1.75)

Likelihoods
=0
=0
-2,739.1 -2,740.3

GH
-2,735.5

=0
-2,740.6

-2,488.7

-2,488.7

-2,488.8

-2,490.4

-2,515.4

-2,314.6

-2,315.6

-2,322.3

-2,315.8

-2,444.9

-2,531.0

-2,531.0

-2,536.1

-2,532.4

-2,638.1

N
-2,849.4

Table 4: Fit of GH for daily equities. GH denotes unrestricted model. = 0 imposes symmetry,
= 0 skewed N model, = 0 skewed Student. Robust S.E.s ( m = 500) are in brackets.

DAX 30

FTSE 100
Generalised hyperbolic
Student or normal gamma

5
10

10
15

20

20

Generalised hyperbolic
Student or normal gamma

4
2
S&P 500 Composite

4
0

4
2
Nikkei 500

5
10

10
15

20

20

Generalised hyperbolic
Student or normal gamma

Generalised hyperbolic
Student or normal gamma

Figure 13: Daily index return data. Profile likelihood for in GH model. Also profile for the skewed
Student ( < 0) and the skewed N ( > 0).

7
7.1

Quadratic variation
Basics

A central concept in financial econometrics, derivative pricing and stochastic analysis is the Quadratic
Variation (QV) process. This has two steps. First, time is split into small intervals
tr0 = 0 < tr1 < ... < trmr = t.
Then the QV process is defined as
2
X
[Y ]t =p lim
Ytri+1 Ytri ,

(44)

46



where supi tri+1 tri 0 for r .

It is sometimes helpful to work with an alternative, and equivalent2 (to (44)), definition of QV

which is written in terms of a stochastic integral. It is that

Z t
2
[Y ]t = Yt 2
Yu dYu .

(45)

This is discussed in some detail in our primer on stochastic analysis.

Example 8 Let N be a Poisson process and let us check the consistency of the formulae (45) and
(44). Suppose Nt = n. It is immediate from (44) that
[N ]t = n
while, on the other hand,
Nt2 2

t
0

Nu dNu = n2 2

n
X


(i 1) = n2 2 n2 = n.
i=1

In general the QV process of a Levy process is a subordinator, for the increments are nonnegative, independent and stationary. If Y has a Levy density u, recalling this means its Levy
measure v can be written as v(dy) = u(y)dy, then for y > 0 the Levy density of the QV process is

u( y) u( y)
+

.
2 y
2 y
This follows from the Levy-Ito representation of Y .
Example 9 Suppose Y = W T where W is Brownian motion and T is a Poisson process
P t
i.i.d.
Cj , where Cj N (0, 2 ), is a compound
independent of W with intensity , then Yt = Tj=1
process (37) while
[Y ]t =

Tt
X

Cj2 .

j=1

Hence [Y ] is a compound Poisson process subordinator with Levy density

1 f ( 1 y 21 ),
where f (y 21 ) denotes the probability density function of the 21 distribution with one degree of
freedom.
2

That they are exactly equivalent follows from Itos lemma. Many probabilists use this second form as the original
definition of QV, rather than via the limit of sums of squares. However, their formal equivalence implies we can use
either definition provided we are in the semimartingale framework.

47

7.2

Brownian motion and quadratic variation

Suppose Y is a scaled Brownian motion with drift, such that Y1 N (, 2 ). Then

Ytri+1 Ytri N ( tri+1 tri , tri+1 tri 2 ).

For small values of tri+1 tri the variation in the series dominates the standard deviation and




pr
drift are O
ti+1 tri and O tri+1 tri , respectively. As a result
[Y ]t = t 2 ,

whatever the value of . The QV is non-stochastic and is the only non-trivial example of a Levy
process where the QV degenerates to a deterministic function of time.
From a statistical viewpoint it means we can theoretically estimate 2 without error using a
tiny path of Brownian motion even in the presence of drift. Of course in practice this is a highly
misleading argument for the continuous time model is unlikely to be perfectly specified at very
short time horizons.

7.3

Realised QV process

The realised QV process is defined, for > 0 and any stochastic process Y , by
t/

[Y ]t =

X
j=1

Yj Y(j1)

2

where t denotes the largest integer less than or equal to t. We can see that if Y SM then
p lim [Y ]t = [Y ]t ,
0

that is the realised QV is a consistent estimator of QV. However, if Y is a Levy process [Y ] is

not a Levy process rather it jumps upwards at specified points in time and so is a discrete time
random walk.
Example 10 A numerical example of the realised QV process is given in Figure 14, which computes
it for a N IG Levy process. In this picture we have taken = 1 and = 1/10, so taking 1 and 10
squared observations per unit of time, respectively. Also given is the corresponding limit, the QV.
We see that as gets small so [Y ] becomes a good approximation of [Y ].
In applied economics it is often inappropriate to study returns over tiny time intervals for our
models tend to be highly misspecified at that scale due to market frictions. In particular the idea
of a unique price is a fiction, for the transaction price tends to depend upon, for example, the
volume of the deal, the reputation of the buyer and seller, prevailing liquidity (and so time of day)
48

(b) Realised QV process: = 1

(a) Simulated NIG Lvy process

0.20
0.4
0.15
0.2

0.10

0.05

0.0
0

2
4
6
(c) Realised QV process: = 0.1

10

2
(d) QV process

10

10

0.3

0.3
0.2
0.2

0.1

0.1

10

Figure 14: Figure (a) Sample path of NIG(0.2,0,0,10) Levy process. (b) Corresponding realised QV
process taking = 1. (c) Same but with = 1/10. (d) QV of the process. Code: levy code.ox.
and the initiator (i.e. was it the buyer or the seller). These issues will be discussed at more length
in later Chapters. To avoid the worst effects of misspecification, realised quadratic variation [Y ]
with not too small is used.
In order to understand the connection between the Levy, realised QV and QV processes it is
helpful to think about the following calculation

Yt
1
Yt
2
3
3
E [Y ]t = t 2 , Cov [Y ]t = t 3 4 + 222 4 ,
[Y ]t
2
[Y ]t
3
4
4

where r denotes the r-th cumulant of Y1 . The only one of these results which is not straightforward
is
t

t
=

= t

Var([Y ]t ) =

4 (Y ) 2 (Y )2



4 (Y ) + 222 (Y )


4 + 222 ,

where, generically, r (X) is the r-th cumulant of X. Finally we notice the implication that [Y ][Y ]
has a zero mean, while
Var ([Y ]t [Y ]t ) = 222 t.
Hence we could use [Y ] as an estimator of [Y ].
49

Example 11 Suppose Y is standard Brownian motion, then 3 = 4 = 0, which means that

Yt
1
Yt
2
0
0
E [Y ]t = t 2 , Cov [Y ]t = t 0 222 0 ,
0
0
0
[Y ]t
2
[Y ]t
which makes sense, for [Y ]t = t.

Example 12 Suppose Y is the homogeneous Poisson process, then all the cumulants are equal to
1 . Thus, writing = (1 1 ... 1) ,

1
1
1
Yt
Yt
E [Y ]t = t1 , Cov {Y }t = t1 1 1 + 21 1 .
[Y ]t
[Y ]t
1
1
1

Notice the covariance is singular, for Y = [Y ] as was noted in Example 8.

Example 13 Suppose Y = W T . Then


1 2 2
T1 .
C { Y1 } = K
2
This implies
1 = 0,

2 = 2 1 (T ),

3 = 0,

4 = 3 2 2 (T ).

Hence in this case

Var ([Y ]t [Y ]t ) = 2 4 21 (T )t,
which only depends upon the mean of the subordinator, not its variance.

8
8.1

L
evy processes and stochastic analysis
Itos formula for L
evy processes

Any Levy process Y is a semimartingale and so Itos formula for semimartingales immediately
applies such that, for any real function f which is twice continuously differentiable, then f (Yt )
SM and
1
df (Yt ) = f (Yt )dYt + f (Yt )d[Y c ]t + f (Yt ) f (Yt ) f (Yt )Yt ,
2

(46)

where Y c is the continuous part of the Levy process. For Levy processes [Y c ] is exactly zero unless
Y has a Brownian component, in which case [Y c ] is proportional to t.
Example 14 Suppose Y is a pure jump Levy process and V = exp(Y ), then
dVt = Vt dYt + Vt Vt Yt .
50

8.2

Stochastic exponential of a L
evy process

The stochastic exponential E(Y ) of Y is defined as

1

E(Y )t = eYt 2 [Y ]t

(1 + Yu )eYu + 2 Yu

(47)

0<ut

and V = E(Y ) is the unique solution to the SDE

dVt = Vt dYt

(48)

with initial condition V0 = 1. The process E(Y )t is a semimartingale. Some of the above is familiar


from the result on exponentiating a diffusion, where the result is E(Y )t = exp Yt 12 [Y ]t .

8.3

Stochastic logarithm

The stochastic logarithm L(Y ) of Y SM, where Y is assumed to live on R>0 , is defined as
Z t
1
dYu .
(49)
L(Y )t =
0 Yu
Clearly, with V = E(Y ),
Z t
Z t
1
1
L(E(Y ))t =
dVu =
Vu dYu = Yt ,
V
V
u
u
0
0
where we have used formula (48).

Multivariate L
evy processes

9.1

Elemental constructions

An important question is how to generate multivariate Levy processes, that is processes with
independent and stationary multivariate increments. Here we discuss just two approaches: linear
transformation and time deformation, of independent Levy process U and V .
If is some deterministic matrix, then the linear combination of the original Levy processes


U
Y =
,
V
is a bivariate Levy process. The elements of Y are marginally Levy processes. This type of argument
generalises to any dimension.
We saw in Section 5 that subordination can be used to generate compelling Levy processes.
Here we use this idea to put


U
Y =
T,
V
51

where T is an independent, common subordinator. This means, for Yt1 = UTt and Yt2 = VTt , and Y
is a multivariate Levy process. A concrete example of this is where (U, V ) are independent standard
Brownian motions, then
Yt |Tt N (0, Tt I),
which implies the elements of Yt are uncorrelated but dependent. In particular
 1 


Yt
1 0
Cov
= E (Tt )
,
Yt2
0 1
while
E

Yt1

2

Yt2

2 o

n



2 o n 2
2 o
= E Tt2 =
6 E (Tt )2 = E Yt1
E Yt
.

More generally we could consider a bivariate Levy process T = T (1) , T (2) , with both T (1) and


T (2) being subordinators, setting Y = U T (1) , V T (2) .

9.2

Multivariate generalised hyperbolic L

evy process

9.2.1

Background

Suppose we take Vt as a d 1 vector of correlated Brownian motions generated by

Vt = t + 1/2 Wt ,
where W is a d 1 vector of independent, standard Brownian motions and is a positive definite
d d matrix. Further we take T to be an independent subordinator and define the deformed series
Yt = t + VTt . Then Y is a Levy process with
Yt |Tt N (t + Tt , Tt ).
Suppose we choose to make T a GIG(, , ) Levy process, then we say that Y is a multivariate
generalised hyperbolic Levy process, following our earlier work on the univariate process discussed
in Section 4.2.4. In particular the increments of such a process are independent and stationary
while the density of Y1 is known to follow a multivariate GH(, , , , , ) density
fY1 (y) =
where

 q



(2)d/2 2 d2
Kd/2 2 + q exp (y ) ,
|| K ()

q = (y ) 1 (y )

and

(50)

2 + 1 .

Here allows us to model the correlation between the processes, while , , and controls the tails
of the density. The whole vector freely parameterises the skewness of the returns. In order to
52

(b) NIG(1,0,0,1) density

0.1

0.2

0.3

0.05 0.10 0.15

(a) Standard normal density

2.5

0.0

2.5

0.0

2.5

0.0

0.0

2.5

2.5
(d) NIG(1,0,0,1) logdensity

0.0

2.5

10.07.5 5.0 2.5

7.5 5.0 2.5 0.0

2.5
(c) Standard logdensity

2.5

2.5

2.5

0.0

2.5

2.5

2.5

0.0

2.5

2.5

0.0

2.5

Figure 15: Densities and log-densities of N(0,I) and NIG(1,0,0,1,I) variables. Densities: (a) N(0,I).
(b) NIG. Log-densities: (c) N(0,I). (d) NIG. Code: levy graphs.ox.
enforce identification on this model it is typical to assume that det () = 1, although Mencia and
Sentana (2004) have normalised the model by setting = 1. The multivariate GH density has many
interesting special cases such as the multivariate skewed Students t, normal gamma, normal inverse
Gaussian, hyperbolic and Laplace. Of course, as the multivariate GH is a normal variance-mean
mixture, then linear combinations of Y are also GH, while if we write Y = (X , Z ) then Z|X,
X|Z, X and Z are all GH. Hence many of the important attractive features of the multivariate
Gaussian distribution carry over to the multivariate generalised hyperbolic distribution.
A simple example of the above multivariate distributions of Y1 is given in Figure 15 which draws
the density and log-density for the bivariate standard normal and the corresponding N IG(1, 0, 0, 1, I)
variables (chosen so that the marginal variances of the variables are 1). Again the log-densities
show that the tails of the N IG variables are much thicker looking roughly linear in all tails. This
has a very big impact on the chance of observing two observations in the tails of the distribution.
9.2.2

Empirical fit of multivariate GH processes

Experience suggests that it is computationally convenient to compute the ML estimator of multivariate GH models using the EM algorithm. This approach, which is detailed in Barndorff-Nielsen
and Shephard (2012c), becomes particularly attractive when d is large for the EM algorithm quickly
converges to the ML estimator as the degree of missing data, which in this context is the unknown
scale , lessens as d increases.

53

We start by fitting some bivariate models. The first example of this is a fit of the German
DM and French Franc against the US Dollar which is reported in the first two lines of Table 5
and Figure 16. The Table shows the expected dramatic improvement in fit associated with these
DM
+
FF
Can
SF
JY
BP

.0221
.0241
.0130
-.0116
.0429
.0548
.0185
.0578
.0315
.0449

MLE of GH parameters

-.0310 .365 1.03 -1.06

-.0143
-.0434 .655 1.78 -1.46
.167
.0283
.660 1.36 -1.62
-.129
.0424
.737 1.37 -.848
-.175
-.303
.543 1.95 -.391
-.361

3.83
3.59
2.31
.049
2.24
2.27
1.15
0.60
1.57
-0.98

3.59
3.62
.049
.432
2.27
2.76
0.60
1.18
-0.98
1.24

GH
-2,865
4,835
-4,518
11.8
-4,592
3,607
-7,258
667
-5,975
1,413

=0
-2,867
4,834
-4,524
7.4
-4,605
3,599
-7,264
666
-6,026
1,365

Likelihoods
=0
=0
-2,933 -2,869
4,785
4,834
-4,535 -4,520
12.6
12.2
-4,615 -4,594
3,596
3,609
-7,281 -7,279
671
651
-5,993 -5,986
1,408
1,414

N
-3,913
4,127
-4,868
3.2
-5,023
3,179
-7,730
632
-6,457
1,298

Table 5: Bivariate GH models. Fits the pair of the DM plus another currency against US Dollar.
In italics are improvement in the log-likelihood compared to fit of the univariate models. = 0
imposes symmetry, = 0 gives skewed N model, = 0 gives skewed Student distribution. Code:
em gh mult.ox.
multivariate models, for the DM and FF are highly related currencies. This is shown up by the
estimated matrix. Again is estimated to be negative, while the fit of the GH model is very close
to the bivariate skewed Students t in this case. The normal gamma model is quite a lot poorer in
this multivariate setting. The result in the italics gives the likelihood for the multivariate model
minus the sum of the likelihoods for the DM and FF univariate models. So the number for the
normal case shows an improvement in the likelihood of 4, 127. Although this is very substantial,
the improvements for the other models are much higher. Hence the gains in using GH models is
even higher in the multivariate case than one might have expected from the univariate analysis.
Figure 16 shows the fitted bivariate normal and GH densities for the DM and FF returns. The
graphs have been drawn to show the densities in places where the log-density does not drop 12
from the mode. This gives an impression of the plausible scatter of points from this variable. The
bivariate normal density is tightly packed, while the GH model gives a wider range of possible
points while the tails of the log-density appear linear or slower in each direction.
Table 5 gives the results for all the bivariate exchange rate relationships which involve the Dollar.
Broadly similar conclusions follow from the above, except the degree of dependence between the
currencies is smaller in these other cases. Interestingly the UK Sterling is negatively related to the
DM returns. Throughout the table the estimated values of ranges between about 0.5 and 1.5.
This is an important common theme, again suggesting evidence against the use of normal gamma
models.
Table 6 gives the GH fit to all six exchange rate return series. This high dimensional model has
54

GH Logdensity
2.5
5.0
7.5

(b): DM/FF GH density

N Logdensity
0.5
1.0

(a): DM/FF Normal density

2
0

0
2
2
(d): DM/FF GH logdensity

0
FF

GH Logdensity
5
0

N Logdensity
10
5
0

2
2
(c): DM/FF Normal logdensity

0
FF

2
0
2

0
FF

0
2

0
FF

Figure 16: Fit of bivariate Gaussian and GH models for the DM and FF against the Dollar. (a) ML
fit of bivariate Gaussian density, (c) log-density. (b) ML fit of bivariate GH density, (d) log-density.
Code: em gh mult.ox.
a low value of , which suggests the fit is very close to a skewed multivariate Students t distribution.
The skewness parameters are important. The N IG fits worse than the skewed Students t but is
much better than the normal gamma model. All these models are enormous improvements over
the multivariate normal fit to the data.
Table 7 gives the corresponding result for the four dimensional equity return data. Here the
elements of are all estimated to be negative, indicating common negative skewness. That is, the
large negative movements have a tendency to occur in all the markets at the same time. In this
case the non-symmetries are important, while the normal gamma is again considerably worse than
the Students t or the N IG distributions.
= .638,

GH
-9,671

= .586,
.628
.0692
.0692
3.09
.0754
2.86
.0760
3.13
.0199
1.66
.1157 1.95
=0
-9,726

= 2.22,
.0754 .0760
2.86
3.13
2.91
2.99
2.99
3.82
1.57
1.85
1.86 2.08

=0
-9,791

.0199 .115

1.66
1.95

1.57
1.86
, =

1.85
2.08

3.45
1.19
1.19
2.54
Likelihoods
=0
N IG
N
-9,672 -9,697 -11,635

.0115
.0307
.0346
.0494
.0605
.0239

, =

.120
.0875
.0647
.182
.148
.254

Table 6: Multivariate GH model for CD, DM, FF, SF, JY and Sterling. = 0 imposes symmetry.
= 0 gives skewed N model, = 0 gives skewed Student t distribution. Code: em gh mult.ox.

55

 = 1.47,

1.54
.778
=
.654
.145

= .660, = 1.72,

.778 .654 .145

.201

.132
1.18 .510 .181
, =

.161
.510 .977 .128
.181 .128 1.26
.0872
Likelihoods
=0
=0
=0
N IG
-9,285
-9,293
-9,270 -9,271

GH
-9,268

.0668

, = .0105 .

.0438
.0580
N
-9,694

Table 7: Multivariate GH model for DAX 30, FTSE 100, S&P500, Nikkei 100. = 0 imposes symmetry, = 0 implies the skewed N model, = 0 is the skewed Student t. Code: em gh mult.ox.

9.3

Stochastic discount factors

In financial economics we typically price contingent payoffs g(YT ) as

!
 
f
M
T
f = exp M
ft ,
Ct = E
g(YT )|Ft , M
t
f
M
t

f is called the stochastic discount factor

the expected discounted value of the claim where T > t, M
f is the log-SDF. For this setup to rule out trivial arbitrages we require that
(SDF) and M


ft Yt
exp M

and

ft ) exp (tr)
exp(M

are local martingales, in order to avoid arbitrage, and where r is a riskless interest rate. A book
length exposition of this approach to asset pricing is given by Cochrane (2001).
f and Y together constitute a bivariate Levy process. What constraints are
Suppose that M

ft ) exp (tr) to be a martingale we need that

imposed on the Levy process in this setup? For exp(M
n
o
f1 + Y1 = 0 and
K 1M

n
o
f1 = r.
K 1M

(51)

Example 15 Suppose






f
M
M
M
f
fB
=
t+
,
Y
Y W t
Y
t

where B, W are independent Brownian motions. Then (51) imply

1
Y = r 2Y
2

and

1 2
M
.
f = r M
2 f

f Y and M
f, Y are Levy
Hence Ct does not depend upon the value of 2f. More generally, if M
M

processes then we require

n
o
f =r
K {1 Y1 } + K 1 M

and

n
o
f = r.
K 1M

56

10

From L
evy processes to semimartingales

Levy processes are determined by their characteristic triplets. For any Levy process Y with characteristic triplet (a, b, v) the law of Yt has characteristic triplet (ta, tb, tv). Suppose we think of
extending the concept of Levy processes by having a triplet of predictable processes (at , bt , vt ) instead of (a, b, v). It turns out that the class of semimartingales can be seen as the natural answer
to this quest which in turn places them at the centre of modern continuous time asset pricing
for arbitrage freeness and semimartingales are synonymous.
We will indicate the character of this, leaving the rather formidable technicalities aside.
Subject to minor regularity assumptions, if Y is a univariate semimartingale then there is a
unique triplet Y = (a, b, v) of predictable processes such that Y has representation
Z t
Z t
Z tZ
Yt = Y0 +
as ds +
s dWs +
c (x) (dsdx)
Rd
0
0
0
Z tZ
+
c (x) ( (dsdx) s (dx) ds)
0

(52)

Rd

where W is Brownian motion, b = 2 , c is a truncation function, c = 1 c, is the random measure

given by
((0, t] A) =

1A (Ys )

0<st

(where A is an arbitrary Borel set in R) and vt (dx) dt is the compensator of (dtdx), i.e.
vt (dx) dt = E { (dtdx)} .
Example 16 If there are no jumps then
Z t
Z t
Yt = Y0 +
as ds +
s dWs ,
0

(53)

which is often called a Brownian semimartingale. The characteristic triplet of Y is in this case


Rt
at , 2t , 0 . The term 0 s dBs is a stochastic volatility process and such processes are the subject
of the next Chapter.

Example 17 Suppose N is a Poisson process jumping with intensity and let

Yt =

Cs 1{1} (Ns )

(54)

0st

where {Ct }t0 is a family of independent but not necessarily identically distributed random variables.
Then Y is semimartingale of the form (52) with triplet (0, 0, t ) where t is the law of Ct and is
assumed to depend continuously on t. In this case,
Yt = ((0, t] {1}) .
57

If, more generally, the process N is a Poisson process with deterministic intensity t we are still in
the semimartingale setting but the structure is slightly more complicated than in (52) and a more
general triplet concept is needed.
Example 18 Adding (53) and (54), assuming they are independent, gives a semimartingale as in


(52) and with triplet at , 2t , t .

The process (54) is an example of an additive process, that is a process with independent

increments. Note that all such processes have deterministic triplet (at , bt , t ).

11

Conclusion

This Chapter has provided an informal introduction to Levy processes. Starting with subordinators, where the Levy-Khintchine representation is more accessible, we have typically built models
using the idea of time deformation. Such models are then imbedded within the familiar normal
variance-mean mixture class of processes. We have emphasised the role of QV and have discussed
the development of various multivariate Levy processes. Barndorff-Nielsen and Shephard (2012a)
provides a more detailed mathematical treatment of the material we have developed in this Chapter, as well as other topics involving Levy processes, that we develop later. Background information
of semimartingale theory is provided in Barndorff-Nielsen and Shephard (2012b).

12

Exercises

Exercise 8 Suppose that Y and Y are independent gamma processes indexed by , and , ;
show that Y Y is a skewed normal gamma process. Hint, recall that



K { Y1 } = log 1 +
.

Exercise 9 Suppose that Y and Y are independent IG processes indexed by , and , ; show
that Y Y is not a N IG process.
Exercise 10 Suppose that Y is a compound Poisson process
Yt =

Nt
X

Cj .

j=1

Prove directly that

[Y ]t = Yt2 2

Yu dYu =
0

Nt
X

Cj2 .

j=1

58

Exercise 11 Let X be a random variable such that

Y = sign(X) |X|1/2
follows the standard normal distribution. Compare the law of X to that of the symmetric NRIG
distribution, as regards the behaviour of the probability densities at 0 and .
Exercise 12 Let X and Y be semimartingales and suppose that they are related by
E(Y )t = eXt .
Show that Y is a Levy process if and only if X is a Levy process.
Exercise 13 Show that the Cauchy motion (that is the Levy process which at time one follows
the Cauchy law) can be represented as the sum of an N IG motion and an independent compound
Poisson process. Give an extension of this result.
Exercise 14 Show that if Y is a Levy process of type G and if its subordinator T has BlumenthalGetoor index then the Blumenthal-Getoor index of Y is 2.

13

Bibliographic notes

13.1

L
evy processes

Levy processes were introduced by Levy (1937) who pioneered the theory of infinite divisibility.
Modern accounts of the probability theory of Levy processes are given in Bertoin (1996), Sato (1999)
and Applebaum (2004). See also Ito (2004), Rogers and Williams (1994, pp. 7384) and Bertoin
(1999). In his notes after each chapter Sato (1999) gives a detailed discussion of the historical
development of the subject. A reasonably accessible overview of the theory and uses of Levy
processes is given in Barndorff-Nielsen, Mikosch, and Resnick (2001). A compact account in the
context of finance is presented by Shiryaev (1999, pp. 200-206), while a more extensive discussion
of the uses of Levy processes in finance is given in Cont and Tankov (2004) and Boyarchenko and
Levendorskii (2002b).
The simulation of Levy processes has to be carried out with some care. There are extensive
results available. Some of the most useful are the infinite series representation developed by Rosinski
(2001), Rosinski (2002) and Rosinski (2007). The special case of gamma process simulation is
discussed by Wolpert and Ickstadt (1999), while some more general discussion is given in Walker
and Damien (2000). We should also note the important contribution of Asmussen and Rosinski
(2001). Cont and Tankov (2004, Ch. 6) is an excellent source on these numerical issues.

59

The connections between Levy processes and semimartingales are well discussed in some technical depth by for example Kallsen (2006). In turn this relies strongly on results in Jacod and
Shiryaev (2003), see also Jacod (1979).

13.2

Flexible distributions

Most of modern financial economics is built out of Brownian motion and the corresponding It
o
calculus. In this Chapter we have discussed many familiar alternative Levy processes like the
Poisson, normal gamma, Students t, Laplace and normal inverse Gaussian laws. All these processes
have been used as empirical models for log-prices. Throughout we have emphasised the normal
variance-mean mixture distributions
Y = + 2 + U,
where U is a standard normal variable which is independent from a random 2 . When = 0 these
variables are called normal variance mixtures and are very familiar in econometrics. The extension
to 6= 0 is important in financial economics as this allows us to model skewness. This approach
to building non-Gaussian densities is attractive for it leads naturally into Levy processes which
have a subordination interpretation. A general discussion of these types of mixtures in statistics
is given in Barndorff-Nielsen, Kent, and Srensen (1982). When the distribution of 2 is infinitely
divisible a natural name for this distribution is type G. On the other hand, Steutel and Van Harn
(2004) call this type of distribution B(V), while Chaumont and Yor (2003) use the label Gaussian
transforms.
In early work Praetz (1972) and Blattberg and Gonedes (1974) suggested modelling the increments to log-prices using a Students t distribution. This model was not set in continuous time, but
we have seen above that it is possible to construct a Levy process to justify this type of modelling.
Further the model can be extended to allow for asymmetry. More recently Granger and Ding (1995)
have advocated the use of Laplace distributions to model discrete time returns, while the non-linear
Brownian motion based Cox, Ingersoll, and Ross (1985) processes have gamma marginals and so
normal gamma distributions are often implicitly used in econometrics. It turns out that fitted
values of the normal gamma distribution are typically thinner tailed, in fact sub-log-linear, than
the corresponding Student, normal inverse Gaussian or the Laplace.
In this Chapter we have placed a great deal of emphasis on generalised hyperbolic and generalised inverse Gaussian distributions. We have carried this out for they support Levy processes,
are empirically flexible, encompass many of the familiar models econometricians are accustomed to
and are mathematically tractable. However, their generality and some of the special cases are not
so familiar.
60

The hyperbolic distribution and its extension to the generalised hyperbolic distribution was
introduced in Barndorff-Nielsen (1977) in order to describe the size distribution of sand grains in
samples of windblown sands. This was motivated by empirical observations due to R. A. Bagnold
who noted that in double logarithmic plotting (that is both the horizontal and vertical axes are
plotted on the logarithmic scale) the histograms looked strikingly as following hyperbolae, the slopes
of the asymptotes being related to the physical conditions under which the sand was deposited; see
Bagnold (1941) (note the similarity to the Granger and Ding (1995) empirics). Subsequently, it
was discovered that the hyperbolic shape, or shapes very close to that, occur in a very wide range
of empirical studies, for instance in other areas of geology, in turbulence, in paleomagnetism, in
relativity theory and in biology. For a survey of developments up till the mid-1980ies, see BarndorffNielsen, Blsild, Jensen, and Srensen (1985). The generalised inverse Gaussian distribution is due

to Etienne
Halphen in 1946 (see the review article by Seshadri (1997)), while it was briefly discussed
by Good (1953). A detailed discussion of this distribution was given by Jrgensen (1982).
Following a suggestion by Barndorff-Nielsen, Ernst Eberlein and coworkers began an investigation of the applicability of the generalised hyperbolic laws in finance and this has developed into a
major project. For their work on this, see Eberlein and Keller (1995), Eberlein, Keller, and Prause

(1998), Eberlein (2001), Eberlein and Ozkan

(2003b) and Prause (1999). Bauer (2000) discusses
the use of these models in the context of value at risk.
When deviations from the hyperbolic shape occurred they typically showed somewhat heavier tails than the hyperbolic. This led Barndorff-Nielsen to consider more closely another of the
generalised hyperbolic laws, the normal inverse Gaussian, which had until then received no attention, but turned out not only to fit a much wider range of data but also to possess various nice
mathematical properties not shared by the hyperbolic (Barndorff-Nielsen (1997), Barndorff-Nielsen
(1998b), Barndorff-Nielsen (1998a)). Srensen (2006) presents a model for the development of the
size distribution of sand under transport by wind, leading to the log N IG distribution.
The class of tempered stable distributions was introduced by Tweedie (1984). Hougaard (1986)
discussed their applicability in survival analysis. See also Jrgensen (1987) and Brix (1999). The
normal variance-mean mixtures with T S mixing was introduced by Barndorff-Nielsen and Shephard
(2001b), who also extended this concept to the normal modified stable distributions. BarndorffNielsen and Shephard (2001a) used some of these distributions in their work on stochastic volatility.
Multivariate Students t distributions have been used since their introduction by Mardia (1970,
p. 92) and Zellner (1971, pp. 383-389). The multivariate GH distribution was first defined in
Barndorff-Nielsen (1977) and was discussed in considerable detail and with a biological application
by Blsild (1981). A rather extensive discussion of the use of multivariate GH distributions in

61

financial econometrics is given by Mencia and Sentana (2004). Alternative skewed multivariate
Students t distributions are provided by Fern
andez and Steel (1998), Bauwens and Laurent (2005)
and Bera and Premaratne (2002). Jones and Faddy (2003) provides a discussion of much of the
literature. Applications of the multivariate NIG distribution have been discussed by Aas, Haff, and
Dimakos (2006) and igard, Hanssen, Hansen, and Godtliebsen (2005).
In a series of papers Kou (2002), Kou and Wang (2003), Kou and Wang (2004), Kou, Petrella,
and Wang (2005), Heyde and Kou (2004) and Glasserman and Kou (2003).
The Skellam distribution was introduced by Irwin (1937). Related mathematical finance work
is carried out by Kirch and Runggaldier (2004). Sichel (1973) and Sichel (1975) studied Poisson-IG
distributions. The Skellam process was introduced by Barndorff-Nielsen, Pollard, and Shephard
(2012) together with various integer valued extensions.

They used it in the context of high

frequency financial data.

Bondesson (1992) is also a key reference for infinite divisibility.

It also includes a detailed

account of Thorins pioneering work (e.g. Thorin (1977) and Thorin (1978)) on the log-normal
distribution and some of the important work that followed.
A simple generic method to sample from the GIG distribution has been derived by Dagpunar
(1988, pp. 133-5) (see also Atkinson (1982)). Tempered stable variables are simulated by Devroye
(2009) (see also Zhang (2011)).

13.3

L
evy processes in finance

The use of normal gamma based Levy processes in finance was pioneered by Madan and Seneta
(1990) and Madan, Carr, and Chang (1998) who paid particular attention to their use in option
pricing. Recent extensions of this work include Carr, Geman, Madan, and Yor (2002).
The thicker tailed hyperbolic distribution and Levy process was studied extensively in the
context of finance by Eberlein and Keller (1995), who also discussed the corresponding option
pricing theory and practice in Eberlein, Keller, and Prause (1998) and Eberlein (2001). This
work is possible because the generalised inverse Gaussian distribution were shown to be infinitely
divisible by Barndorff-Nielsen and Halgreen (1977). In fact it has the stronger property of being
selfdecomposable and the same holds for the GH distribution, see Halgreen (1979).
The even thicker tailed normal inverse Gaussian process is studied by Barndorff-Nielsen (1995),
Barndorff-Nielsen (1997), while Rydberg (1997b) and Rydberg (1997a) discusses both fitting the
process to financial data and simulating from such processes. Prause (1999) and Raible (2000)
have recently written first rate Ph.D. theses on generalised hyperbolic Levy processes under the
supervision of Ernst Eberlein. Both of these theses have a wealth of information on this topic.
Bingham and Kiesel (2000) looks at the use of hyperbolic processes in finance, while Bibby and
62

Srensen (2003) reviews the area of generalised hyperbolic processes in finance.

The idea of time deformation or subordination is due to Bochner (1949) and Bochner (1955),
while it was introduced into economics by Clark (1973) who suggested the use of volume statistics
as a subordinator, placing particular weight on studying the implications of using a lognormal
subordinator. At that stage it was not known that this was a valid mathematical construction for
it was not until Thorin (1977) that the lognormal was shown to be infinitely divisible. See also
Bondesson (2002) for up to date treatment of lognormal Levy processes. Epps and Epps (1976)
and Tauchen and Pitts (1983) further studied the relationship between volume and the variance of
the increments to prices. Recent discussions of this includes Ane and Geman (2000). Stock (1988)
used the concept of subordination in a wider economic context outside finance. The StudentOU process, and other stationary processes with Student marginals, are discussed in Heyde and
Leonenko (2005).
Mandelbrot (1963) and Mandelbrot and Taylor (1967) introduced the concepts of self-similarity
and stable Levy processes into financial economics. Almost immediately the main stream academic
profession rejected these models, after some initial support from Fama (1965), due to their lack of
empirical fit, as most research papers suggested the existence of at least two moments for returns.
An elegant discussion of the move away from these models and its importance is given in Campbell,
Lo, and MacKinlay (1997, pp. 17-19). However, there still remains a small group of researchers
who push in this area. Recent work is discussed by Rachev and Mittnik (2000). Taleb (2007) has
popularised some direct uses of stable models, although without any detailed empirical work to
back them up in the financial context nor discussion of heavy tailed alternatives which are both
tractable mathematically and empirically more appealing. Put simply stable processes are a poor
way to produce financial Black Swans.
Truncated Levy flights were introduced by Mantegna and Stanley (1994), while it has been
pioneered in finance in Mantegna and Stanley (1996) and Mantegna and Stanley (2000). The
extended Koponen class has been considered by Novikov (1994), Koponen (1995), Mantegna and
Stanley (2000), Boyarchenko and Levendorskii (2002b), Boyarchenko and Levendorskii (2000), Boyarchenko and Levendorskii (2002c), Boyarchenko and Levendorskii (2002a), Barndorff-Nielsen and
Levendorskii (2001), Carr, Geman, Madan, and Yor (2002), and Rosinski (2001). Carr, Geman,
Madan, and Yor (2002) called these models CGMY processes after their own initials. We have
not followed that nomenclature. Meixner distributions were introduced by Schoutens and Teugels
(1998) and have been studied in the context of Levy based models for finance by Schoutens and
Teugels (2001) and Grigelionis (1999). Ben-Hamou (2000) has studied estimating the parameters
of the Levy process from option prices.

63

The use of Levy processes for term structure and credit risk has been an active recent area.

Work includes, for example, Eberlein and Raible (1999), Ozkan

(2002), Eberlein, Jacod, and Raible

(2005), Barndorff-Nielsen, Christiansen, and Nicolato (2001) and Eberlein and Ozkan
(2003a),
Schoutens and Cariboni (2009) and Eberlein and Kluge (2007). See also Bocker and Kl
uppelberg
(2007) for an application to operational risk. This highly stimulating literature is beyond the scope
of this book and so will not be discussed further.
Kijima (2002), Cont and Tankov (2004) and Schoutens (2003) are general books which discuss
the use of Levy processes and finance.

13.4

Empirical fit of L
evy processes

There is a large literature on studying the fit of various parametric models to the marginal distribution of returns of speculative assets. Most of these papers are not based on a background of a
Levy process and so risked fitting an incoherent (from a continuous time viewpoint) model. An
example of this is Praetz (1972) in his work on the Student t distribution. Notable exceptions are
Mandelbrot (1963), where he used stable distributions and related this to stable processes. See also

Eberlein and Keller (1995), Eberlein (2001) and Eberlein and Ozkan
(2003b).
The likelihood methods we used to fit the models are entirely standard. We have used profile
likelihoods. A discussion of this literature is given in Barndorff-Nielsen and Cox (1994, Section
3.4). The use of profile likelihoods for in the generalised hyperbolic is new as was the use of the
EM algorithm in this context. As well as our own work on the EM algorithm for GH distributions,
independent and concurrent work on the use of the EM algorithm for this problem was carried out
by Protassov (2004). An elegant discussion of the EM algorithm is given in Tanner (1996). The
theory of robust standard errors for maximum likelihood estimation is standard in econometrics
and statistics. Leading references are White (1982) and White (1994).
Barndorff-Nielsen and Prause (2001) showed that the Olsen scaling law is explainable by the
NIG Levy process.

References
Aas, K., I. H. Haff, and X. K. Dimakos (2006). Risk estimation using the multivariate inverse Gaussian
distribution. Journal of Risk 8.
Ane, T. and H. Geman (2000). Order flow, transaction clock and normality of asset returns. Journal of
Finance 55, 22592284.
Applebaum, D. (2004). Levy Processes and Stochastic Calculus. Cambridge: Cambridge University Press.
Asmussen, S. and J. Rosinski (2001). Approximation of small jumps of Levy processes with a view towards
simulation. Journal of Applied Probability 38, 482493.
Atkinson, A. C. (1982). The simulation of generalised inverse Gaussian and hyperbolic random variables.
SIAM Journal of Scientific and Statistical Computing 3, 50217.
Bagnold, R. A. (1941). The Physics of Blown Sand and Desert Dunes. London: Chapman and Hall.

64

Barndorff-Nielsen, O. E. (1977). Exponentially decreasing distributions for the logarithm of particle size.
Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences 353, 401
419.
Barndorff-Nielsen, O. E. (1995). Normal inverse Gaussian processes and the modelling of stock returns.
Research Report 300, Dept. Theor. Statistics, Aarhus University.
Barndorff-Nielsen, O. E. (1997). Normal inverse Gaussian distributions and stochastic volatility modelling.
Scandinavian Journal of Statistics 24, 114.
Barndorff-Nielsen, O. E. (1998a). Probability and statistics: selfdecomposability, finance and turbulence.
In L. Accardi and C. C. Heyde (Eds.), Probability Towards 2000. Proceedings of a Symposium held 2-5
October 1995 at Columbia University, pp. 4757. New York: Springer-Verlag.
Barndorff-Nielsen, O. E. (1998b). Processes of normal inverse Gaussian type. Finance and Stochastics 2,
4168.
Barndorff-Nielsen, O. E., P. Blsild, J. L. Jensen, and M. Srensen (1985). The fascination of sand. In
A. C. Atkinson and S. E. Fienberg (Eds.), A Celebration of Statistics, pp. 5787. New York: SpringerVerlag.
Barndorff-Nielsen, O. E., B. J. Christiansen, and E. Nicolato (2001). General stochastic volatility and explicit option prices in Heath-Jarrow-Morton term structure analysis. Unpublished paper, Department
of Mathematical Sciences, Aarhus University.
Barndorff-Nielsen, O. E. and D. R. Cox (1994). Inference and Asymptotics. London: Chapman & Hall.
Barndorff-Nielsen, O. E. and C. Halgreen (1977). Infinite divisibility of the hyperbolic and generalised
inverse Gaussian distributions. Zeitschrift fur Wahrscheinlichkeitstheorie und verwandte Gebiete 38,
309311.
Barndorff-Nielsen, O. E., J. T. Kent, and M. Srensen (1982). Normal variance-mean mixtures and z
distributions. International Statistical Review 50, 145159.
Barndorff-Nielsen, O. E. and S. Levendorskii (2001). Feller processes of normal inverse Gaussian type.
Quantitative Finance 1, 318331.
Barndorff-Nielsen, O. E., T. Mikosch, and S. Resnick (2001). Levy Processes Theory and Applications.
Boston: Birkhauser.
Barndorff-Nielsen, O. E., D. G. Pollard, and N. Shephard (2012). Integer-valued Levy processes and low
latency financial econometrics. Quantitative Finance 12, 587605.
Barndorff-Nielsen, O. E. and K. Prause (2001). Apparent scaling. Finance and Stochastics 5, 103113.
Barndorff-Nielsen, O. E. and N. Shephard (2001a). Non-Gaussian OrnsteinUhlenbeck-based models and
some of their uses in financial economics (with discussion). Journal of the Royal Statistical Society,
Series B 63, 167241.
Barndorff-Nielsen, O. E. and N. Shephard (2001b). Normal modified stable processes. Theory of Probability
and Mathematical Statistics 65, 119.
Barndorff-Nielsen, O. E. and N. Shephard (2004). Power and bipower variation with stochastic volatility
and jumps (with discussion). Journal of Financial Econometrics 2, 148.
Barndorff-Nielsen, O. E. and N. Shephard (2012a). Mathematics of Levy based processes. Unpublished
chapter of the authors book on Levy Driven Volatility Models.
Barndorff-Nielsen, O. E. and N. Shephard (2012b). Primer on stochastic analysis. Unpublished chapter of
the authors book on Levy Driven Volatility Models.
Barndorff-Nielsen, O. E. and N. Shephard (2012c). Toolboxes for financial processes: distributions. Unpublished chapter of the authors book on Levy Driven Volatility Models.
Bauer, C. (2000). Value at risk using hyperbolic distributions. Journal of Economics and Business 52,
455467.
Bauwens, L. and S. Laurent (2005). A new class of multivariate skew densities, with application to generalised autoregressive conditional heteroscedasticity models. Journal of Business and Economic Statistics 23, 346354.
Ben-Hamou, E. (2000). Option pricing with Levy process. Financial Markets Group Working Paper.
Bera, A. and G. Premaratne (2002). Modelling asymmetry and excess kurtosis in stock return data.
Unpublished paper, Economics Department, University of Illinois.

65

Bertoin, J. (1996). Levy Processes. Cambridge: Cambridge University Press.

Bertoin, J. (1999). Subordinators: Examples and Applications. Berlin: Springer. Presented at Ecole dEte
de St-Flour, 1997. Lecture Notes in Mathematics.
Bibby, B. M. and M. Srensen (2003). Hyperbolic processes in finance. In S. Rachev (Ed.), Handbook of
Heavy Tailed Distributions in Finance, pp. 211248. Elsevier Science.
Bingham, N. and R. Kiesel (2000). Modelling asset returns with hyperbolic distributions. In J. Knight
and S. E. Satchell (Eds.), Asset Return Distributions, pp. 120. Butterworth-Heinemann.
Blsild, P. (1981). The two-dimensional hyperbolic distribution and related distributions, with an application to the Johannsen bean data. Biometrika 68, 251263.
Blattberg, R. C. and N. J. Gonedes (1974). A comparison of the stable and student distributions as models
for stock prices. Journal of Business 47, 244280.
Bochner, S. (1949). Diffusion equation and stochastic processes. Proceedings of the National Academy of
Science of the United States of America 85, 369370.
Bochner, S. (1955). Harmonic Analysis and the Theory of Probability. Berkeley: University of California
Press.
Bocker, K. and C. Kl
uppelberg (2007). Modelling and measuring multivariate operational risk with Levy
copulas. Journal of Operational Risk 3, 327.
Bondesson, L. (1992). Generalized Gamma Convolutions and Related Classes of Distributions and Densities. Heidelberg: Springer-Verlag.
Bondesson, L. (2002). On the Levy measure of the lognormal and the logcauchy distributions. Methodology
and Computing in Applied Probability 4, 243256.
Boyarchenko, S. I. and S. Z. Levendorskii (2000). Option pricing for truncated Levy processes. International Journal of Theoretical and Applied Finance 3, 549552.
Boyarchenko, S. I. and S. Z. Levendorskii (2002a). Barrier options and touch-and-out options under regular
Levy processes of exponential type. Annals of Applied Probability 12, 12611298.
Boyarchenko, S. I. and S. Z. Levendorskii (2002b). Non-Gaussian Merton-Black-Scholes Theory. Singapore:
World Scientific.
Boyarchenko, S. I. and S. Z. Levendorskii (2002c). Perpetual American options under Levy processes.
SIAM Journal on Control and Optimization 40, 16631696.
Brix, A. (1999). Generalized gamma measures and shot-noise Cox processes. Advances in Applied Probability 31, 929953.
Campbell, J. Y., A. W. Lo, and A. C. MacKinlay (1997). The Econometrics of Financial Markets. Princeton, New Jersey: Princeton University Press.
Carr, P., H. Geman, D. B. Madan, and M. Yor (2002). The fine structure of asset returns: an empirical
investigation. Journal of Business 75, 305332.
Chaumont, L. and M. Yor (2003). Exercises in Probability. Cambridge: Cambridge University Press.
Clark, P. K. (1973). A subordinated stochastic process model with fixed variance for speculative prices.
Econometrica 41, 135156.
Cochrane, J. H. (2001). Asset Pricing. Princeton: Princeton University Press.
Cont, R. and P. Tankov (2004). Financial Modelling with Jump Processes. London: Chapman and Hall.
Cox, J. C., J. E. Ingersoll, and S. A. Ross (1985). A theory of the term structure of interest rates.
Econometrica 53, 385407.
Dagpunar, J. (1988). Principles of Random Number Generation. Oxford: Oxford University Press.
Devroye, L. (2009). Random variable generation for exponentially and polynomially tilted stable distributions. ACM Transactions on Modeling and Computer Simulation 19, 18:118:20.
Doornik, J. A. (2001). Ox: Object Oriented Matrix Programming, 5.0. London: Timberlake Consultants
Press.
Eberlein, E. (2001). Application of generalized hyperbolic Levy motion to finance. In O. E. BarndorffNielsen, T. Mikosch, and S. Resnick (Eds.), Levy Processes Theory and Applications, pp. 319337.
Boston: Birkhauser.

66

Eberlein, E., J. Jacod, and S. Raible (2005). Levy term structure models: no-arbitrage and completeness.
Finance and Stochastics 9, 6788.
Eberlein, E. and U. Keller (1995). Hyperbolic distributions in finance. Bernoulli 1, 281299.
Eberlein, E., U. Keller, and K. Prause (1998). New insights into smile, mispricing and value at risk: the
hyperbolic model. Journal of Business 71, 371405.
Eberlein, E. and W. Kluge (2007). Calibration of Levy term structure models. In M. Fu, R. A. Jarrow,
J.-Y. Yen, and R. J. Elliott (Eds.), In Advances in Mathematical Finance In Honor of D. B. Madan,
pp. 147172. Birkhauser.

Eberlein, E. and F. Ozkan

(2003a). The defaultable Levy term structure: ratings and restructing. Mathematical Finance 13, 277300.

Eberlein, E. and F. Ozkan

(2003b). Time consistency of Levy models. Quantitative Finance 3, 4050.
Eberlein, E. and S. Raible (1999). Term structure models driven by general Levy processes. Mathematical
Finance 9, 3153.
Epps, T. W. and M. L. Epps (1976). The stochastic dependence of security price changes and transaction
volumes: implications for the mixture-of-distributions hypothesis. Econometrica 44, 305321.
Fama, E. F. (1965). The behaviour of stock market prices. Journal of Business 38, 34105.
Feller, W. (1971a). An Introduction to Probability Theory and Its Applications (2 ed.), Volume 1. New
York: John Wiley.
Feller, W. (1971b). An Introduction to Probability Theory and Its Applications (2 ed.), Volume 2. New
York: John Wiley.
Fernandez, C. and M. F. J. Steel (1998). On Bayesian modelling of fat tails and skewness. Journal of the
American Statistical Association 93, 359371.
Glasserman, P. and S. G. Kou (2003). The term structure of simple foward rates with jump risk. Mathematical Finance 13, 383410.
Good, I. J. (1953). The population frequencies of species and the estimation of population parameters.
Biometrika 40, 237264.
Granger, C. W. J. and Z. Ding (1995). Some properties of absolute returns, an alternative measure of risk.
Annals dEconomie et de Statistique 40, 6791.
Grigelionis, B. (1999). Processes of Meixner type. Lithuanian Mathematics Journal 39, 3341.
Halgreen, C. (1979). Self-decomposability of the generalized inverse Gaussian and hyperbolic distributions.
Zeitschrift fur Wahrscheinlichkeitstheorie und verwandte Gebiete 47, 1317.
Heyde, C. C. and S. G. Kou (2004). On the controversy over tailweight of distributions. Operations
Research Letters 32, 399408.
Heyde, C. C. and N. N. Leonenko (2005). Student processes. Advances in Applied Probability 37, 342365.
Hougaard, P. (1986). Survival models for heterogeneous populations derived from stable distributions.
Biometrika 73, 387396.
Irwin, J. O. (1937). The frequency distribution of the difference between two independent variates following
the same Poisson distribution. Journal of the Royal Statistical Society, Series A 100, 415416.
Ito, K. (2004). Stochastic Processes: Lectures given at Aarhus University. Springer.
Jacod, J. (1979). Calul Stochastique et Problems de Martingales. Number 714 in Lecture Notes in Mathematics. Berlin: Springer.
Jacod, J. and P. Protter (2012). Discretization of Processes. Heidelberg: Springer.
Jacod, J. and A. N. Shiryaev (2003). Limit Theorems for Stochastic Processes (2 ed.). Springer: Berlin.
Jrgensen, B. (1982). Statistical Properties of the Generalised Inverse Gaussian Distribution. New York:
Springer-Verlag.
Jrgensen, B. (1987). Exponential dispersion models (with discussion). Journal of the Royal Statistical
Society, Series B 49, 127162.
Jones, M. C. and M. J. Faddy (2003). A skew extension of the t-distribution, with applications. Journal
of the Royal Statistical Society, Series B 65, 159174.

67

Kallsen, J. (2006). A didactic note on affine stochastic volatility models. In Y. Kabanov, R. Lipster, and
J. Stoyanov (Eds.), From Stochastic Calculus to Mathematical Finance. The Shiryaev Festschrift, pp.
343368. Heidelberg: Springer.
Kijima, M. (2002). Stochastic Processes with Applications to Finance. London: Chapman and Hall.
Kirch, M. and W. Runggaldier (2004). Efficient hedging when asset prices follow a geometric Poisson
process with unknown intensities. SIAM Journal on Control and Optimization 43, 11741195.
Koponen, I. (1995). Analytic approach to the problem of convergence of truncated Levy flights towards
the Gaussian stochastic process. Physics Review E 52, 11971199.
Kou, S. G. (2002). A jump-diffusion model for option pricing. Management Science 48, 10861101.
Kou, S. G., G. Petrella, and H. Wang (2005). Pricing path-dependent options with jump risk via laplace
transforms. Kyoto Economic Review 74, 123.
Kou, S. G. and H. Wang (2003). First passage time of a jump diffusion process. Advances in Applied
Probability 35, 504531.
Kou, S. G. and H. Wang (2004). Option pricing under a double exponential jump-diffusion model. Management Science 50, 11781192.
Levy, P. (1937). Theories de LAddition Aleatories. Paris: Gauthier-Villars.
Madan, D. B., P. Carr, and E. Chang (1998). The variance gamma process and option pricing. European
Finance Review 2, 79105.
Madan, D. B. and E. Seneta (1990). The VG model for share market returns. Journal of Business 63,
511524.
Mandelbrot, B. (1963). The variation of certain speculative prices. Journal of Business 36, 394419.
Mandelbrot, B. and H. Taylor (1967). On the distribution of stock price differences. Operations Research 15, 10571062.
Mantegna, R. and H. E. Stanley (2000). Introduction to Econophysics: Correlation and Complexity in
Finance. Cambridge: Cambridge University Press.
Mantegna, R. N. and H. E. Stanley (1994). Stochastic process with ultraslow convergence to a Gaussian:
the truncated Levy flight. Physics Reviews Letters 73, 29462949.
Mantegna, R. N. and H. E. Stanley (1996). Turbulence and financial markets. Nature 383, 587588.
Mardia, K. V. (1970). Families of Bivariate Distributions. Darien, Connecticut: Hafner Publishing Co.
Mencia, F. J. and E. Sentana (2004). Estimating and testing of dynamic models with generalised hyperbolic
innovations. Unpublished paper: CEMFI, Madrid.
Novikov, E. A. (1994). Infinitely divisible distributions in turbulence. Physics Reviews E 50, R3303R3305.
igard, T. A., A. Hanssen, R. E. Hansen, and F. Godtliebsen (2005). EM algorithm and modeling of heavytailed processes with the multivariate normal inverse Gaussian distribution. Signal Processing 85,
16551673.

Ozkan, F. (2002). Levy processes in credit risk and market models. Unpublished Ph.D. thesis, Freiburg
University.
Praetz, P. D. (1972). The distribution of share prices. Journal of Business 45, 4555.
Prause, K. (1999). The generalised hyperbolic models: estimation, financial derivatives and risk measurement. PhD Thesis. Mathematics Faculty, Freiburg University.
Protassov, R. S. (2004). EM-based maximum likelihood parameter estimation for multivariate generalized
hyperbolic distributions with fixed . Statistics and Computing 14, 6777.
Rachev, S. and S. Mittnik (2000). Stable Paretian Models in Finance. New York: Wiley.
Raible, S. (2000). Levy processes in finance: Theory, numerics and empirical facts. PhD Thesis. Mathematics Faculty, Freiburg University.
Rogers, L. C. G. and D. Williams (1994). Diffusions, Markov Processes and Martingales. Volume 1,
Foundations (2 ed.). Chichester: Wiley.
Rosinski, J. (2001). Series representations of Levy processes from the perspective of point processes. In
O. E. Barndorff-Nielsen, T. Mikosch, and S. Resnick (Eds.), Levy Processes Theory and Applications,
pp. 401415. Boston: Birkhauser.

68

Rosinski, J. (2002). Tempered stable processes. In O. E. Barndorff-Nielsen (Ed.), Miniproceedings from

2nd MaPhySto Conference on Levy Processes: Theory and Applications, pp. 215220. Aarhus:
http://www.maphysto.dk/publications/MPS-misc/2002/22.pdf.
Rosinski, J. (2007). Tempering stable processes. Stochastic Processes and Their Applications 117, 677707.
Rydberg, T. H. (1997a). Existence of unique equivalent martingale measures in a Markovian setting.
Finance and Stochastics 1, 251257.
Rydberg, T. H. (1997b). The normal inverse Gaussian Levy process: simulation and approximation.
Communications in Statistics: Stochastic Models 13, 887910.
Samorodnitsky, G. and M. S. Taqqu (1994). Stable Non-Gaussian Random Processes. New York: Chapman
and Hall.
Sato, K. (1999). Levy Processes and Infinitely Divisible Distributions. Cambridge: Cambridge University
Press.
Schoutens, W. (2003). Levy Processes in Finance: Pricing Financial Derivatives. New York: Wiley.
Schoutens, W. and J. Cariboni (2009). Levy Processes in Credit Risk. New York: Wiley.
Schoutens, W. and J. L. Teugels (1998). Levy processes, polynomials and martingales. Statistics and
Stochastic Models 14, 335349.
Schoutens, W. and J. L. Teugels (2001). Meixner processes in finance. Unpublished paper: Eurandom,
K.U. Leuven.
Seshadri, S. (1997). Halphens laws. In S. Kotz (Ed.), Encyclopedia of Statistical Sciences, pp. 302306.
New York: Wiley.
Shiryaev, A. N. (1999). Essentials of Stochastic Finance: Facts, Models and Theory. Singapore: World
Scientific.
Sichel, H. S. (1973). Statistical valuation of diamondiferous deposits. Journal of South African Institute
of Mining and Metallurgy 73, 235243.
Sichel, H. S. (1975). On a distribution representing sentence-length in written prose. Journal of the Royal
Statistical Society, Series A 137, 2534.
Srensen, M. (2006). On the size distribution of sand. Unpublished paper, Department of Department of
Applied Mathematics and Statistics, University of Copenhagen.
Steutel, F. W. and K. Van Harn (2004). Infinitely Divisibility of Probability Distributions on the Real Line.
Marcel Dekker.
Stock, J. H. (1988). Estimating continuous-time processes subject to time deformation: an application to
postwar U.S. GNP. Journal of the American Statistical Association 83, 7785.
Taleb, N. N. (2007). The Black Swan. Penguin.
Tanner, M. A. (1996). Tools for Statistical Inference: Methods for Exploration of Posterior Distributions
and Likelihood Functions (3 ed.). New York: Springer-Verlag.
Tauchen, G. and M. Pitts (1983). The price variability-volume relationship on speculative markets. Econometrica 51, 485505.
Thorin, O. (1977). On the infinite divisibility of the lognormal distribution. Scandinavian Actuarial Journal 47, 121148.
Thorin, O. (1978). An extension of the notion of a generalized -convolution. Scandinavian Actuarial
Journal 48, 141149.
Tweedie, M. (1984). An index which distinguishes between some important exponential families. In
J. Ghosh and J. Roy (Eds.), Statistics: Applications and New Directions: Proc. Indian Statistical
Institute Golden Jubilee International Conference, pp. 579604.
Walker, S. and P. Damien (2000). Representations of Levy processes without Gaussian components.
Biometrika 87, 477483.
White, H. (1982). Maximum likelihood estimation of misspecified models. Econometrica 50, 125.
White, H. (1994). Estimation, Inference and Specification Analysis. Cambridge: Cambridge University
Press.

69

Wolpert, R. L. and K. Ickstadt (1999). Simulation of Levy random fields. In D. D. Dey, M. Muller, and
D. Sinha (Eds.), Practical Nonparametric and Semiparametric Bayesian Statistics, pp. 227242. New
York: Springer-Verlag.
Zellner, A. (1971). An Introduction to Bayesian Inference in Econometrics. New York: John Wiley and
Sons.
Zhang, S. (2011). Exact simulation of tempered stable ornstein-uhlenbeck processes. Journal of Statistical
Computation and Simulation 81, 15331544.

70