T H E Q U A N T C L A S S R O O M B Y AT T I L I O M E U C C I

A Short,
Comprehensive, Practical
Guide to Copulas
Visually introducing a powerful risk management tool
to generalize and stress-test correlations.

he stochastic behavior of one single financial vari- retical results. In Section 3, we address copulas implementaable say, prices, or implied volatilities (etc.) tion issues.
is fully described by its probability distribution,
which is called the marginal distribution. In a Section 1: Univariate Results
market of multiple financial variables, all the in- In this section, we
cover well-known
that
prepare
Consider
an arbitraryresults
random
variable
,the
with a fully arbitrar
formation on the stochastic behavior of the market is fully ground for the
of, and
thedistribution
intuition behind,
Alldefinition
the features
of the
of copulas.
are described by its prob
inX,
such
function
(pdf)
, dened
described by the joint probability distribution.
Consider an
arbitrary
random
variable
witha away
fullythat,
arbi-for any set of pote
for
the
variable
,
the
following
identity
for
the pdf holds
The multivariate distribution of a set of financial
variables
trary
distribution.
All
the
features
of
the
distribution
of
X
Consider an arbitrary random variable , with a fully arbitrary distribution. are
Consider
an
arbitrary
random
variable
,
with
a
fully
arbitrary
distribution.
Z
is fully specified by the separate marginalAll
distributions
of
the
described
by
its
probability
density
function
(pdf)
f
,
defined
the features of the distribution of are described by its probabilityXdensity in
All
the
features
of
distribution
are
described
itsofprobability
density
variables and by their copula, or, loosely speaking,
the correlasuchina such
wayofthat,
forthat,
any for
set of
potential
() .
P values
{ Xvalues
}for
theXvari, dened
a
way
anyby
set
potential
function
(pdf)
the
X
,
dened
in
such
a
way
that,
for
any
set
of
potential
values
X
function
(pdf)

tions among the variables.

X, theidentity
following
the pdf holds:
for the variable , the able
following
foridentity
the pdffor
holds
for separately
the variable
, the following identity
Zfor the pdf
Modeling the marginals and the copula
provides
Accordingly,
withholds
mild abuse of notation we write to den
Z
greater flexibility for the practitioner to model randomness.
way to represent t
()whose
. pdf is . An alternative
(1) (1)
P { Xa}distribution

.
(1) (cdf), dened as follo
P { X }of isXits
()
Therefore, copulas have been used extensively in finance,
cumulative
distribution function
X
both on the sell-side to price derivativesAccordingly,
(see, e.g., Li (2000)),
Accordingly,
with
write
ZX~f
to denotewethat
has X to
with mild abuse
of notation
wemild
writeabuse
of
notation,
to
denote
that

has
with
mild
abuse
of
notation
we
write

and on the buy-side to model portfolioAccordingly,

risk
(see
e.g.
Meucci,
denote
that
X
has
a
distribution
whose
pdf
is
f
.
An
alternative
X
distribution
a distribution whose pdf is . An alternative way to representthe

() .
()
An alternative
way to
represent
the
distribution
a of
distribution
whose pdfdistribution
is to. represent
distriGan, Lazanas and Phelps (2007)).
way
the distribution
of as
X isfollows
its cumulative
is its cumulative
function
(cdf),
dened
of of iscopulas
its cumulative
function
(cdf),
dened
as follows
Here, we provide a review of the theory
proving distribution
bution function
(cdf),
defined
as follows:
ZIf
we feed the random variable into a generic function we
1
Z
the most useful results. We also provide a guide to the practi

() . (). For instance,

(2)if we set () sin (
()
random
variable
() random
variable
() . sin (), which is bound
(2) (2)in the interval [1

cal implementation of copulas. For a detailed discussion of

implementation issues and for the code, please

to the
arises function
when wetransform
with its own cdf, i.e.
If we refer
feed the
random variable situation
into a generic
we obtain
another
companion paper Meucci (2011).
If weFor
feed
the random
generic
function
If we feed
the random
variable
instance,
into
a generic
function
into
we
random
variable
().
if wevariable
set
()X
sinaobtain
()
we another
obtain
a g,
This article is organized as follows. Inrandom
Section
1,
we
review
we
obtain
another
random
variable
Y

g(X).
For
instance,
if
variable

().
For
instance,
if
we
set

()

sin
()
we
obtain
a
random variable sin (), which is bound
in the interval
special variable
Key concept.
If we [1
feed 1].
theAarbitrary
through
strictly univariate results that nonetheless
naturally
lead
to
the
we
set
g(x)

sin(X),
we
obtain
a
random
variable
Y

sin(X).
random
variable

sin
(),
which
is
bound
in
the
interval
[1
1].
A
special
situation arises when we transform we
with
its own
cdf,special
i.e.
.
obtain
a very
transformed
random variable, which
situation
arises
when wewhich
transform
with
itsinterval
own cdf,
i.e. Aspecial
. situation arises
multivariate concept of copulas. In Section
2, we
introduce
is bound
in the
[-1,1].
grade of
copulas, highlighting and proving the most important theo- when we transform X with its own cdf i.e., g FX.
cdf,
() .
Key concept. If we feed the arbitrary variable through itsown
Key
concept.
If
we
feed
the
arbitrary
variable

through
its
own
cdf,
we obtain a very special transformed
variable,
which
is called
the on the unit interval
Therandom
distribution
of the
grade
is uniform
we
obtain
a very special transformed random variable, which is called the
grade
of
the original distribution
grade of
(3)
O ()
C T .O B E R 2 0 1 1 RISK PROFESSIONAL
22
www.garp.org
(3)
() .

U
,
[01]
The distribution of the grade is uniform on the unit interval regardless of
The
of the grade
the distribution
original distribution
is uniform on the unit interval regardless of

P { X }

P {
()
.X }

X
X

() .
X

(1)

(1)

U FX X U 0,1

grade
denote
that
has
with Accordingly,
mild abuse ofwith
notation
we
write
to
to denote that has
mildwe
abuse
of
notation
wedenote
write

has
to
that
with
mild
abuse
of
notation
write

F
X U 0,1

.
An
alternative
way
to
represent
the
distribution
whose
pdf
is

. An
way
to represent the distribution
a distribution
whose
pdf is
alternative
way
to alternative
represent
the
distribution
whose
pdfdistribution
is . An
mulative
function
(cdf),
dened
as
follows
of distribution
is its cumulative
distribution function
(cdf), dened as follows
mulative
function
as follows
Z (cdf), dened
Z
() .
(2)
() Z
()

() .
(2)

(2)
() () .

T H E Q U A N T C L A S S R O O M B Y AT T I L I O M E U C C I

X random variable with arbitrary distribution

the random
variable
random
into a generic
function
we obtain
another
x
If wevariable
feed the
variablefunction
into a we
generic
function
we obtain another
he random
into a generic
obtain
another
ble
random
().
For
instance,
if
we
set

()

sin
()
we
obtain
a
variable
().
For
instance,
if ()
we set
obtain
() asin () we obtain a
X random variable with arbitrary distribution
ble

().
For
instance,
if wein
set
() sin
we A
ble random
sin
(),
which
is
bound
the interval
[1
1].
special [1 1]. A special
X random variable with arbitrary distribution
variable

sin
(),
which
is
bound
in
the
interval
ble
sin
whichis with
bound
thecdf,
interval
1]. A special
Figure 1: The cdf maps an arbitrary random variable into a uniform variable
es when
we (),
transform
itsin
own
i.e.
[1
own
. cdf, i.e.
.
situation
arises
when
we
transform

with
its

s when we transform
concept.
with its own
cdf,feed
i.e.the
arbitrary
.
The above result (3)-(4) also works backwards: if we feed
Key
If we
variable X through
-1 variable
Figure
1: Theacdf
maps an
arbitrary
random
variable
into a uniform
uniform
random
variable
U into
thefeed
inverse
cdf Frandom
its
own
cdf,
we
obtain
a
very
special
transformed
random
X , we
The above
result
(3)-(4)
backwards:
if we
ainto
uniform
ncept. If Key
we feed
the arbitrary
variable
through
its owncdf,
Figure
1: The
cdf also
mapsworks
an arbitrary
random
variable
a uniform
variable
concept.
If
we
feed
the
arbitrary
variable
through
its
own
cdf,
-1
1 variable X FX
1 fX,
ncept.
If we feed
the arbitrary
variable
thethrough
itscalled
a random
(U)variable
with distribution
variable,
which
is called
grade
of
Xown cdf,
we obtain a random

( )
variable into theobtain
inverse
cdf
very special
transformed
random
variable,
which
is
the
we obtain
a very special
transformed
random
variable,
which is called the
very special
transformed
random
variable, which
is called
the
as
follows:
with distribution
result
The
above
(3)-(4)
also
works
backwards:
if
we
feed
a
uniform
random

grade of
The above result (3)-(4) also
1 works backwards: if we feed a uniform random
1
(3)into
the
inverse
cdf cdf
() .
(3) variablevariable
1 obtain a random variable 1( )
we
1
obtain
variable (6)

into
the
inverse
()
.
(3)

( ) a random
(3) with distribution U[01] 7 we
() .
(6) ( )
.
with distribution
ution of
thedistribution
grade
is uniform
on
the
unit
interval
regardless
of unit interval,
The
of theon
grade
is
uniform
the unit
regardless of
The
distribution
of the
grade
is on
uniform
on interval
the
, say Student
ution
of
the
grade
is
uniform
the
unit
interval
regardless
of
1 1
Even
better,
we can choose
desired
target distribution
distribution
distribution
better,
any
U
7 choose
( ) .
[01]
we
U[01]
7 any

. distribution
(6)
the original
Even
can
desired
fX (6)

of theoriginal distribution fX :
1
( ) target
distribution
regardless
the inverse cdf
; and
or chi-square, etc.; then we compute the cdf and

say,
Student
t
or
chi-square,
etc.;
subsequently,
we
compute
U[01] ,
(4)
nally
transform
a uniform
todesired
generate
a random
variable
Student
with
say
Student
Even better,
canvariable
choose
any
target
distribution
, ,say
Even
better,
we
canwechoose
any desired
target
distribution

U[01] ,
(4)
-1
1
the distribution.
cdf
FX-we
andcompute
the
inverse
lastly,
we
transform
a ; and
U[01] ,
(4)
X ;
(4)
1
and,

and
the
inverse
cdf

orarbitrary
chi-square,
etc.;
then
we compute
the
the or
desired
and
the
inverse
cdf

chi-square,
etc.;
then
thecdf
cdfFcdf

; and
uniform
variable
U
to
generate
a
random
variable
X
with
the
transform
a uniform
variable
to generate
a randomvariable
variable
with
with
nally nally
transform
a uniform
variable
to generate
a random
desired
arbitrary
distribution.
the desired
arbitrary
distribution.
the
desired
arbitrary
distribution.
Key concept. Starting from an arbitrary target distribution and a
e proof of
this result
proceeds
as follows
The
simple
proof
of this
result isas follows
The
proof
of this
proceeds
e proof of
thissimple
result
proceeds
as result
follows
uniform random variable,
we canStarting
transform the
variable distribuinto a
concept.
anuniform
arbitrary
Key Key
concept.
Starting from from
an arbitrary
target target
distribution and a
}
= P{
()}
}= P { () } (5)
() P {
variable
with
the
desired
target
distribution
Key
concept.
Starting
from
an
arbitrary
target
distribution
and
a
()
P
{
(5)

}
} 1
()

and avariable,
uniformwe
random
variable,the
weuniform
can transform
(5) tionrandom
can transform
variable the
into a
= P {
() P {
1

uniformuniform
1 (5)
random
variable,
we
can
transform
the
uniform
variable
into
a

= P 1
=
1() = .

()

variable
the
desiredinto
target
distribution
1
= =
P
uniform
variable
a variable
with the desired target distri () = () = .
= P ()
with

desired
() = .
variable with the
target distribution

1
bution,
as
follows:

(7)
() Therefore
= , which
is()the= cdf
of a uniform
distribution.
We distribution.
can
7 ( ) .
, Fof
which
isu, the
cdf
of
acdf
uniform
We can
Therefore,
()
=
,
which
is
the
cdf
a
uniform
distribution.
We
can
(u)
=
which
is
the
of
a
uniform
distribu1
U
( ) .

U
e univariate
miracle
as
follows.

(7)
[01]

summarize
the
univariate
miracle as follows.

e1univariate
miracle
as follows.
tion.
1 ( ) .
(7)
we sketch
the
intuition
behind
(4).
First
of
all
the
variable

(7)
[01]

In Figure
1 we sketch
the intuition
behind
(4).
First of
the variable

1 we sketch
the intuition
behind
First
of
all
variable
all
In
1 (see(4).
below),
we
sketch
the
intuition
behind (4).
()
1.the
Furthermore,
terval
[0 1]
because
theFigure
cdf
satises
0

U[01]
()

1.
Furthermore,
lives
in
the
interval
[0
1]
because
the
cdf
satises
0

in1. the
Furthermore,
[0 1] because
the
cdfall,
satises
0
First
of
thethe
variable
U ()
interval
[0,1], because
islives
steeper
where
there
erval
is the
uniform
on [0
1]
because
cdf
is steeper
where
there
variable
1]
is because
uniform the
on cdf
[0
1] because
the cdf
there
The proof
of (7)
reads
is
steeper
where
ntial
is uniform
on
[0

The proof
(7) reads
satisfies
0 outcomes
FX (x)1. are
Furthermore,
variable U The
is proof
outcomes
forthe
cdf
and
thus these
spread
out the
overare
of(7) of
reads
are
more
potential
outcomes
for

and
thus
these
outcomes
spread
out
over

ntialon
outcomes
foruniform
andthe
thus
these
outcomes
are Fspread
out
over
1
on
[0,1]
because
the
cdf
is
steeper
where
there
are
is
atter
where
are
few
val;
the
other
hand,
cdf

X there

() = () .

P
{

}
=
P

(
)

=
P
(8)
is
atter
where
there
are
few
a
wider
interval;
on
the
other
hand,
the
cdf

The
proof
of
(7)
reads
is atter
where
there
are
few
on the
the cdf
other
hand,
the all
cdfthe
outcomes
P { } = P ( ) = P () = () . (8) (8)
outcomes
dal;
thus
concentrates
occurring
over
a
large
more
potential
for
X,
and
thus
these
outcomes
are
outcomes,
and thus theallcdf
all the outcomes
occurring over a large
1generate Monte

from arbitrary
the
concentrates
theconcentrates
outcomes
occurring
over
a large
This
iscdf
useful
a thus
small
set.cdf
The
balance
between
two
eects,on
i.e.
dilution
of result
spread
out
over
athese
wider
interval;
the
other
hand,
theThis
Pextremely
{
of
}
= P to
( ) to
generate
= P Carlo
scenarios
= ()
. arbitrary
(8)
()scenarios
result
is
extremely
useful
Monte
Carlo
from
interval
into
a
small
set.
The
balance
between
these
two
eects,
i.e.
dilution

smalland
set.concentration
The balance
between
these
twoare
eects,
i.e.todilution
ofthus distributions
desired
using
as
input
only
a
uniform
number
generator.
narios
of
scarce
scenarios,
gives
rise
a
uniform
This
result
is
extremely
useful
to
generate
Monte
Carlo sceF
is
flatter
where
there
few
outcomes,
and
the
cdf
X
desired
distributions using as input only a uniform number generator.
abundant
scenarios
and
concentration
of
scarce
scenarios,
gives
rise
to
a
uniform
arios and concentration of scarce scenarios, gives rise to a uniformThis result is extremely useful to generate Monte Carlo scenarios from arbitrary
distribution. concentrates all the outcomes occurring over a large interval narios from arbitrary desired distributions, using as input only
desired
using
as input
only a uniform number generator.
number
generator.
into a small set. The balance between these two effects
distributions
i.e., a uniform
3
3
dilution 2of abundant scenarios and concentration of scarce
2
Section 2: Copulas (Theory)
scenarios2 gives rise to a uniform
distribution.
2 Copulas:
theory 3
Now we are fully equipped to introduce the copula, by extending
to the multivariate
framework
the by
univariate
results
Figure 1: Mapping an Arbitrary Random Now we are
fully equipped
to introduce
the copula,
extending
to the multi(3)-(4) and
(7).univariate results (4) and (7).
Variable into a Uniform Variable via cdf variate framework
the
a N-dimensional
vector
of random
variables
X
Consider Consider
a -dimensional
vector of
random
variables
(
1 )
(X1general
,...,XN) with
a fully general
multivariate
distribution
with a fully
multivariate
distribution
represented
by itsreprepdf .
is dened
in such
a way that,
We recall that
in by
theitsmultivariate
case
the pdf
sented
pdf X~ fX. We
recall
that in the
multivariate
case,
(1 set
of potential
)
the
following
for any setthe
ofpdf
potential
joint invalues
that,
R for
fX is defined
such aXway
for any

N
identity holds
joint values X
for (X1,...,XN), the following identity holds:
Z
P {(1 ) X }
(1 ) 1 . (9)
(9)
X

From
the joint distribution
fX, principle
we can inextract
principle
extract
all the
marginal
From the joint
distribution
we can in
all
the
N
marginal
distributions
X
~
f
,
where
n=1,...,N,
by
n
X
pdfs
distributions , where = 1 , by n computing the marginal
computing
the
marginal
pdfs,
as
follows:
as follows
Z
(1 ) 1 1 +1 .
(10)
( ) =
R1

23

RISK PROFESSIONAL O C T O B E R 2 0 1 1

Then we can compute the marginal cdfs as in (2). Finally, we can feed
entry of the vector ,
each cdf , which is a function, with the respectivewww.garp.org
namely the random variable . The outcome of this operation are the grades,
which we know from (4) have a uniform distribution on the unit interval

pdf fU1

x2
X fX

ade U2

Now
are fully equipped
to introduce
work
theweunivariate
results (4)
and (7). the copula, by extending to the multiframework the univariate results (4) and (7).
a variate
-dimensional
vector of random variables (1 )
Consider a -dimensional vector of random variables (1 ) x2
eneral
multivariate distribution represented by its pdf . .
with a fully general multivariate distribution represented by its pdf

dened
a way
that,
inWe
therecall
multivariate
the pdf case
isthe
dened
in such
a way that,
that in thecase
multivariate
pdf inissuch

forX(
(
) 1the
potential
joint
values Xjoint
Rvalues
following
) the following
for any set
of potential
1 R for

cdf FX 2

T H E Q U A N T C L A S S R O O M B Y AT T I L I O M E U C C I

identity holds

pdf f X 2

)(
X1}
)
.

) {(
X1}
)(
1 1
. 1 (9)
{(1 P
X

x1

(9)

joint X=(X1,X2)

marginal X2

in principle
all the marginal
From the joint
we can extract
we can in principle
all extract
the marginal
t distribution
distribution
distributions , where = 1 , by computing the marginal pdfs
, where = 1 , by computing the marginal pdfs
as follows
Figure 2: Copulas are non-linear standardizations of multivariate distributions
Z
and what is left is the pure joint information amongst the Xn's
Z ( ) =

1
1
+1
. (10)(10)
1
i.e., the copula fU. We summarize this statement in an alR
) =
(1 ) 1 1 +1 .
(10)
R1 Then we can compute the marginal cdf s FX , as in (2). Fiternative, intuitive formulation of the copula.
n
we canKey
feedconcept. the copula is the information missing from the individual
Then we can compute the marginal cdfs as in (2). Finally,
nally,
we
can
feed
each
cdf
F
Xn, which is a function, with the rea function,
of the
,
each cdfthe
marginal
in the
(2).respective
Finally, entry
we can
feedvector
compute
aswith
, which iscdfs
marginals
to
complete
theThe
jointcopula
distribution
spective
entry
of
the
vector
X

namely,
the
random
variable
Key
concept.
is the information missing from
.
The
outcome
of
this
operation
are
the
grades,
namely
the
random
variable

ulas:
theory
which is a function, with the respective entry of the vector ,
are the
grades,
which
which
weXknow
(4)outcome
haveofa this
uniform
on
intervalwe the individual marginals to complete the joint distribution:
n. The
. outcome
The
of operation
thisdistribution
operation
arethe
theunit
grades,
ndom
variable
from
" joint = copula + marginals "
(13)
know
from
(4)
have
a
uniform
distribution
on
the
unit
interval
ully
equipped
to
introduce
the
copula,
by
extending
to
the
multiw from (4) have a uniform distribution
on
the
unit
interval
(11)
( ) U[01] .
" joint = copula + marginals "
work the univariate results (4)
and (7).
(13)

a -dimensional
vector
of random

(1of
point
(
) remember
U[01]variables
. is that the
(11)
1) (11)
)
However,the
main
to
entries
(

intuitive
(13) can (13)
be made
rigorous.
From theFrom
denition of
general
multivariate
distribution
represented
by its pdf
grades
. The
The denition
intuitive definition
can be
made rigorous.
is not
are not
independent.
Therefore, the
joint distribution
ofthe
copula
(12)
the
1],
that,
see
main
point
todomain,
remember
is pdf
that
the
entries
ofin
such
on However,
its
which
is the
unit
cube
[0
1]
is
(

[0
isto
dened
a1 way
the
inuniform
the multivariate
case
the

)Figure
the
main
point
remember
that
entries
of 2. the definition of copula (12),
of
is
notdistriTherefore,
the joint not
distribution
for (1 Therefore,
the
the the
following
fendent.
potential
Rindependent.
) grades
Ujoint
(Uvalues
joint
1,...,UN)Xare
() P {1 1 }
(14)
domain,
which
isfUthe
unitgrades
cube [0
1]
[0distribution
1],its
seedomain,
Figure
2.the joint
bution
of The
the
is not
uniform
on
is
Key
concept.
copula
of
an
arbitrary
iswhich
= P {1 (1 ) 1 ( ) }
of itsZ[0,1]
grades... [0,1] (see Figure 2, below).
distribution
cube
the unit

1
1
=
P 1
(1 )
( ) (14)

X
}

.
(9)
{(
1

is
the
joint
ncept.
Thecopula of an arbitrary
distribution

1 1 (1 )

1
X
1
Figure 2: Copulas

Standardiza of its grades

.. Non-linear
(1 )
( ) .
=
1

(12)
(

. Distributions
tions
we Multivariate
can in principle
extract all the marginal
t distribution
of

(
)

(
)

Dierentiating
this expression
obtain Sklars
theorem
(for two
variables, the
Differentiating
thisweexpression,
we obtain
Sklars
theorem
11 1 , by computing the marginal pdfs
, where1 =
..
general
case
follows
immediately)
(for
two
variables,
the
general
case
follows
immediately):
)

(12)
(

.
1

Z
1
( )
(1 )
(2 )
( 2 ) = 21 2 (1 2 ) = 21 2

2
distribution

can be
interpreted
as .a sort(10)
of non-linear
generalcase1 follows
immediately)

( ) =The grades
of
(the

1
1
+1

1
1
1
1
2

= 1 2 1 (1 ) 2 (
1 ) 22 (2 )
z-score,
which forces all the entries to have a uniform distribution on the
2 )11 1 (
R1
1
2
2 1
(1 )
(2 )
(1 2 ) = 1 2 2(1 2 ) =1
1 2)
unit interval [0 1]. By feeding each random variable into its own cdf, all
2
(
)
12 1 (
1
2
2
1
1
1
1
2

(15)
=
as in (2).
Finally, we
can
feed
compute
the marginal
cdfs
marginal
is
swept
away,
and
the information
contained
in each
distribution
(1 ) 2 (
= 12
1
1
1 (

(1 ) 2 2 (2 )
2 )
s of the distribution can be interpreted as a sort ofnon-linear
1 1 1
2 2
1 (1 )
(15)
2 12 )
1
1
, which
is left
a function,
with
theinformation
respectiveamongst
entry ofthe
thevector
s, i.e. ,
the copula
what is
is the pure
joint

2
(1 )1
(
(2 )
1 2 1(
1
2
distributionformulation
on the
forces
allsummarize
the entries
1)
2)
to haveinaanuniform

==
(15)
thisstatement
alternative,
. We
outcome
of this
operationintuitive
are the grades, of the
ndom
variable
. The
11
) 2 1 1(
1
. 2 )

1 1 1( )1 2 2 (2)
0 1].
By feeding each random variable into its own cdf, all
1 1
1
2
copula
2 2
w from
(4) have a uniform distribution on the unit interval

on contained in each marginal distribution is swept away, and

s the pure joint
( ) amongst
U[01] . the s, i.e. the copula
(11)
information
arize this statement in an alternative,4intuitive formulation of the
the main point to remember is that the entries of (1 )
endent. Therefore, the joint distribution of the grades is not
s domain, which is the unit cube [0 1] [0 1], see Figure 2.
4

joint fX
ncept. The Key
copula
of an arbitrary
distribution
is the
concept.
The copula
of an arbitrary
distribution
of its grades
the joint distribution fU of its grades, as follows:
(

1 1 (1 )
..
.
( )

(12)

1
1
(1 )
(2 )

1 1
52 1
.
=

(
)

2 )original joint distribution ,

1
2
1 theorem links
2 (
Key
Sklars
the
Keyconcept.
concept.
Sklars theorem
links
the original joint disthe copula , and the marginals , or equivalently , as follows
tribution fX, the copula fU and the marginals fXn or, equiva 1 Sklars theorem
the original joint distribution ,
Key concept.
links
1
lently,
as(follows:
FX
1 ) ( )
n,and
1
,
the
marginals

the copula

= equivalently , as follows (16)

|
{z
} , or
joint
1

1
1

( ) 1 ( )
1 (1 )
({z

1
(
)
1
)
(16)
|
|
{z
} } | =1 1
{z
} .
joint
pure joint
pure m arginal 1

1
is
1 1 (1 ) ( )
(1 )
{z
}theintuitive
|
{z (13).
} .
|
Sklars theorem
justies
copula denition
pure joint

pure m arginal

Sklars
theorem
justies
the intuitive
copula
denition
(13).
Sklars
theorem
provides
thethe
pdfintuitive
of the copula
from
the joint pdf
Sklars
theorem
justifies
copula
definition
(13).and the
(12)marginal pdfs. This allows us to use maximum likelihood to t copulas to
empirical
data. provides the pdf of the copula from the joint pdf and the
Sklars
theorem
Sklars
theorem
provides
theIf pdf
of the
thegrades
copula
fromarbitrary
the
We
now
derive
another
we feedlikelihood
marginal pdfs.
This
allows useful
us to result.
use maximum
to(12)
t into
copulas to
1
joint
pdf
and
the
marginal
pdfs.
This
allows
us
to
use
maxi,
we
obtain
new
transformed
random
variables
with
a given
inverse
cdfs

empirical data.
distribution,
which
we
denote
by

as ajoint
mum
likelihood
to
fit
copulas
to
empirical
data.

We now derive another useful result. If we feed the grades (12) into arbitrary

The grades of the distribution fX can be interpreted

1
we obtain
new transformed
random
variables
withgrades
a given
cdfs
sort
of non-linear
z-score,
which as
forces
allofthe
entries Xn inverse
to
We now
another
useful
result.
If we
feed the
, derive
s of the distribution
can be
interpreted
a sort
non-linear
1 1
( (1 ))
1 1
-1
joint distribution,
which we denote
by

have
a
uniform
distribution
on
the
unit
interval
[0,1].
By
feed(12)
into
arbitrary
inverse
cdf
s
F
,
we
obtain
new
transY
..
n
h forces all the entries to have a uniform distribution on the
)
(
(17)
.
ing eacheach
random
variable
Xn into
ownits
cdf,own
all the
formed random variables
with
a
given joint distribution,
1
1
(
(
))
cdf,informaall
[0 1]. By feeding
random
variable
itsinto
1

1
1 (1 ( ))

..
tion
in each
marginal distribution
is swept
andaway, which we denote
on contained
in contained
each marginal
distribution
is sweptfXnaway,
)
( by fY
(17)
.
has
marginals
whose cdfs are , which follows from
The
joint
distribution

1
s the pure joint information amongst the s, i.e. the copula
( ( ))
applying (11) and (6) insequence.
Furthermore, the copula of is the same as
marize this statement in an alternative, intuitive formulation of the
the copula of , because from the denition of in (17) and the denition of

www.garp.org
4

The joint distribution has marginals whose cdfs are , which follows from
the copula
we obtain
applying
(11) (12)
and (6)
in sequence. Furthermore, the copula of is the same as
the copula of , because from
and
the denition of
O
C Tthe
O denition
B E R 2of0 1 in
1 (17)
RISK
PROFESSIONAL
1 (1 ) = 1 (1 )
the copula (12) we obtain
..
) .
(18)
(
.
( ) = ( )

24

 (1) , as
follows
( )
) , or equivalently

(
1 joint
marginals
la , and |the

{z
} | 1 1
{z
} .
1

1
1
pure
joint
pure
m
arginal
1

(
)

(
)

1 (1 ) | {z( ) } | 1 1
{z
} .
(16)
{z
} =

pure joint
pure m arginal
lars theorem
(13).
joint justies the intuitive copula denition

T H E Q U A N T C L A S S R O O M B Y A T T I L I O M E U C C I
1

denition
(13).
( )
1 (1 ) copula
(1 justies
)
klars
| theorem
{z
} |the1 intuitive
{z
} .
klars theorem
copula
from
the
joint pdf and the
pure joint provides the pdf of the
pure m arginal
ginal pdfs. This allows us to use maximum likelihood to t copulas to
Sklars
theoremthe
provides
pdf of
the copula
heorem
justies
intuitivethe
copula
denition
(13). from the joint pdf and the
irical
data.
rginal
pdfs.
This
allows
us
to
use
maximum
to into
t copulas
We now derive another useful result. If we feed thelikelihood
grades (12)
arbitraryto
1
pirical
data.
obtain
new
transformed random
variables
with
a given
rse
cdfs
, wereturns
Compounded
C=(C
Linear
returns
theorem
provides
the
pdf
of
the
joint
pdf R=(R
and1,R
the
1,C
2) copula from the
2)
We
now derive
another
useful
result.
If we feed the grades (12) into arbitrary
distribution,
which
we
denote
by

likelihood to r2t copulas to

dfs. This allows
us to use maximum
c2
Compounded
C=(C1,C2) random variables
f
Linear returns
, we obtain
newreturns
transformed
withR=(R
a given
erse cdfs
1
1,R2)

data.
C f
R

e
1 by 1
r
t
distribution,
which
we
denote
R fR
1 c Ifwe
(
(1grades
)) f R (12) into arbitrary
1 the
w derive another usefulf result.
feed
1
C f
.
R
e 1 with a given

C
ln
1

R
,
we
obtain
new
transformed
random
variables
s 1

.
( 1. ( ( )) )
R (17)
f (17)
f

1
1
1
1
bution, which we denote by
c11
r

C
ln
1

( (... ( )) )
(17)
C2

C2

R2

fC 1 (

c1

( ))

r1

f R1

1
1
11
(whose
)) are , which follows from
1 . marginals
cdfs
joint distribution has
(

.. distribution
joint
marginals whose cdfs f are FYn
f
) fY has
( (6)The
ying (11) and
in sequence.
Furthermore,
the copula of (17)
is the same as
F follows
,F
FR, FR (6)
1
which
from
applying
(11)
and
in sequence.
Furhas
marginals
whose
cdfs
follows from
ecopula
joint distribution

and

(
(
))
, which

the

of , because
denition
of inare
(17)
the
denition
of

from

F , F of Y isthe
F copula
, F same
thermore,
the
copula
the
same
as
the
of
X,
lying
(11)
and
(6)
in
sequence.
Furthermore,
copula
of

is
the
as
Copula
U=(U
,U
)
1 2
copula (12) we obtain
marginals
whose
are
follows
from
istribution
has
u2 cdfs
because
from
definition
YininU=(U
(17)
and
the
definition
of
(17)
and
the
denition
of
copula of ,
because
from
thethe
denition
ofof
, which

Copula
,U
)
1 2
11)
and (6)
sequence.
the copula
u of U is
copula
(12)in we
obtain
fthe same as
the copula
obtain
Furthermore,
1 ((12),
1 ) = we
1 (1 )
fU
the denitionU of
of , because from the denition
f
.. of in (17) )and
(18)
(
(12) we obtain
. f 1 (1 ) .
1 (1 ) =
( ) .. ( )
(18)
( )=
) u1 .
(18)
( .)
(
1

R1

C1

C1

C2

C1

C2

R1

R2

the pitfalls of disregarding such differences). For example, if

the prices distribution fP is multivariate log-normal, the linear
returns distribution fR is multivariate shifted-lognormal and
the compounded returns distribution fC is multivariate normal, as illustrated in Figure 3 (see below) for two negatively
correlated stocks.

Figure 3: Linear and Compounded Returns

display Same Copula

U2

u1

fU
.. (copula
(
refore, we derive
thatthe
arbitrary
random
variable

) of

)an

.)
(18)
(
f
.
) does not(change
when
we
transform
each

into
a
new
variable

copula
( ) of an arbitrary
that
) the
1 copula
erefore,
derive
random

Therefore,
we derive
of anvariable
arbitraryran
(we
) by
means
of functions
() that
the
( ()), where are
we
derive
) does
not
change
when
we
transform
each

into
a
new
variable

dom
variable
Xarbitrary
(X1,...,Xreturns
) does
not
change
when
transthat
copula
an
variable
we
Nrandom
Figure
3: the
Linear
andofcompounded
have
samecopula
These
two types of returns, though calculated on the same securities prices
(not
means
of
functions
a each
()

1
(
()),
where
copula
are
Figure
3:transform
compounded
returns
) by

change
when
we
variable
a
new
form
each
XLinear
new
Y
gn (Xhave
means
ofare dierent. Therefore, their distributions are dierent, refer to
n into and
n variable
n) bysame

) does
into
,

+1
) by means of functions
functions () 61
( ()), where
where Fare
are
arbitrary
cdfs.(2010) for more details on the pitfalls of disregarding such dier
Yn
Meucci
itrary cdfs.arbitrary
It It
is iseasy
to
verify
that
such
functions

are
a
very
broad
functions
easy toItverify
that
such
functions
g
are
a
very
broad
class
n
cdfs.
is easy
to
verify
that
such

are
veryexample,
broad if the prices distribution is multivariate log-normal,
6

ences.a For
ss, namely class,
all the
increasing
also known also
as also
co-monotonic
namely
all6 the
increasing
transformations,
known
as
co-monotonic

namely,
alltransformations,
the
increasing
transformations,
known
as
the linear returns distribution is multivariate shifted-lognormal and the
nsformations.
Thus
we obtain
the
transformations.
Thus
wefollowing
obtain the following
co-monotonic
transformations.
Thus, we obtain the following:
compounded returns distribution is multivariate normal, as illustrated in
Figure 3 for two negatively correlated stocks.
Key concept.
Co-monotonic
transformations
= (
Key
concept.Co-monotonic
Co-monotonic
transformations
the
gHowever,
(Xn)) of the
Key
concept.
transformations
)Y=
nof
n(
the copula of the linear returns and the copula of the com
ntries of doentries
not
the
copula
ofthe

of
do not
of copula of X
of alter
the
entries
ofalter
X do
notcopula
alter the
However,
the copula
the linear
returns
and the
copula
of surpounded returns
are identical,
seeofagain
Figure
3. This
result
is not
1 is anare
increasing
of
prising, because
= returns
the compounded
identical transformation
(see, again, Figure
3).
and
1
1
1
1
Cn .
=
ln
(1
+

)
is
an
increasing
transformation
of

This
result
is
not
surprising,
because
R
=
e
1
is
an
increasn

1
1
.
.

( ... ) ( .. )
( ... ) ( .. )
ing transformation of Cn, and Cn =1n (1+ Rn) is an increasing

transformation of Rn.

U1

&

&

Copulas:
(19) practice

(19)

Section 3: Copulas (Practice)

The implementation of the copula-marginal decomposition in practice relies on
(19) The implementation of the copula-marginal decomposition
two distinct processes, which appear in multiple steps in the theoretical discusin practice relies on two distinct processes, which appear in
sion of Section 2.

multiple steps
in thewhich
theoretical
discussion
of Sectionof2.the copula

First, the separation

process,
led us
to the denition
First,
the
separation
process,
which
led
us
to the definition
(12).
To illustrate how the copula is not affected by increasing of the copula (12).
To illustrate how consider,
the copulaasis follows,
not aected
increasing
transformations,
transformations,
the by
linear
returns
and
To illustrate
how
the
copula
isreturns
not aected
bycompounded
increasing
transformations,
consider
the
linear
and
the
returns
between
time Key
andThe
Key
separation
S strips
an Sarbitrary
the compounded
betweenreturns
time t and
time ttime
+ 1 for
theconcept.
concept.
The process
separation
process
strips an distribution
arbitrary
consider the linear
thereturns
compounded
between
and
time returns
+ 1 for and
the same
securities
into
its
marginals

and
its
copula

same
securities:
distribution
f
into
its
marginals
f
and
its
copula
f
,
as
follows:
X
X
U
n
time + 1 for the same securities
+1
+1

ln(
).
(20)

+1
+1
1

(20)

1

).
(20)

ln(

.. )
1
S
:
(

7
(21)

.
(21)
These two types of returns, though
calculated on the same
7

.. ) ,

securities prices Pn,t+17, are different. Therefore, their distribu

tions are different (refer to Meucci (2010) for more details on

1
( ... )

1
( ... )

where are the grades

25

RISK PROFESSIONAL O C T O B E R 2 0 1 1

( ) .

(22)

www.garp.org
The separation process can be reverted, similarly to the univariate case (6).
By feeding each grade back into the respective inverse cdf 1 we obtain

irst, the separation process, which led us to the denition of the copula
ration process, which led us to the denition of the copula

T H E Q U A N T C L A S S R O O M B Y AT T I L I O M E U C C I

Key concept. The separation process S strips an arbitrary distribution

pt.into
The
S strips
an arbitrary
distribution
itsseparation
marginalsprocess
and
its copula

ginals and its copula

1
(21)
S:
( ... )
7
.
1

.. ) ,(21)
:
( .. ) 7

.
.

( .. )
,

where Un are the grades

here are the grades
e grades
(22)
( ) .
(22)
( ) .
(22)

The implementation of
the theory of copulas
relies on two fundamental
steps: "separation"
and "combination."

The separation process can be reverted, similarly to the

The separation process can be reverted, similarly to the univariate case (6).
univariate
case similarly
(6). By feeding
each gradecase
Un back
into the
processeach
can
be reverted,
the univariate
(6).
1
eeding
grade
back
into theto
inverse cdf

we obtain
-1 respective 1

respective
inverse
cdf FXinverse
, we obtain
a
random
variable
X
back
into
the
respective
cdf

we
obtain
rade

is exactly the
ndom variable (1 ) whose joint distribution
joint
distribution
is
exactly
the
original
dis 1,...,X

whose
joint
distribution
is
exactly
the
enal
distribution
(1(X
N) )whose
:
on :
tribution fX:
1
1
(1 )
1
1
1

(
)
1
.. ) 1 71 ( 1
..
)
(

(23)(23)
.. )
..
) .1
7. (
(23)

Implementing all the steps involved in such processes ana.

.

( )
1
lytically
is impossible, except in trivial cases, such as within the

(
)

However,
we
do
not
need
to
limit
ourselves
to
reverting
to
normal
family.
However, we do not need to limit ourselves to reverting to the original disotion
notneed
to limit
ourselves
to reverting
to the
dis- because
the
original
distribution
fX. Copulas
areoriginal
so powerful
Therefore, all practical applications of copulas rely on nu . Copulas are so powerful because they can be glued with arbitrary
ulas are so they
powerful
because
be glued
with arbitrary
can be
glued they
with can
arbitrary
marginal
distributions, with merical techniques, most notably the representation of distria technique that generalizes the univariate case (7), which we butions by means of Monte Carlo scenarios.
8
used to prove
co-monotonic
invariance of the copula (19).
Within the Monte Carlo framework, scenarios that repre8 the
rginal distributions,
with
a
technique
that
generalizes
the1.univariate
case (7),
Accordingly,
we
start
with
two
ingredients:
an arbitrary
sent copulas are obtained by feeding joint distributions sce_
d which we copula
used tofUprove
thegrades
co-monotonic
invariance
of the copula (19).
i.e.,
U (U1,...,U
N), each of which has a
narios into the respective marginal cdfs, as in (22), and joint
Accordingly,
we
start
with
two
ingredients:
an
arbitrary
copula
, i.e.
uniform distribution and joint distribution structure
specified
scenarios with a given copula are obtained by feeding grades
_
des (1 ) each of which has a uniform distribution, and joint
by fU; and 2. arbitrary
marginal
distributions
f
Xn. Then, we
scenarios into the respective inverse cdfs. as in (25).
_
_
.
tribution structure specied by ; and arbitrary marginal distributions
-1
compute the marginal cdfs'
F
F
Xn and their inverses
Xn , and we
1
en we compute the marginal cdfs , their inverses and we feed eachHowever, even the above numerical operations present diffeed each grade into the respective marginal cdf. The output ficulties. First, the computation of the cdfs in (22) requires
de into the respective marginal cdf. The output is a -variate random
is a N-varitate random variable X (X1,...,X
has the univariate integrations, as in (2). Second, the computation
_ N)that
and the desired
iable (1 ) that has the desired copula
desired
copula
f
and
the
desired
marginals
f
.
U
X
of the inverse cdf in (25) requires univariate integrations foln
rginals .
Weassummarize,
assecond
follows,process.
this second process.
lowed by search algorithms. Finally the extraction of the marWe summarize
follows this
ginal distributions from the joint distribution in (21) requires
Key concept.
The combination process C glues arbitrary multivariate integrations, as in (10).
_
Key concept.
combination
process
C f
glues
arbitrary marginals The first two problems, namely the univariate integration
marginalsThe
fXn and
an arbitrary
copula
U into a new joint dis into a new joint distribution
and an arbitrary copula

tribution fX, as follows:

and inversion, have been addressed for a broad variety of

parametric distributions, for which cdf and quantile are avail1

able either analytically or in terms of efficient quadratures.

On the other hand, the multivariate integration to extract

1
7 ( ... ) ,
C:
(24)
the
marginal from the joint distribution represents a signifi
(24)
.

( .. )

cant hurdle, because numerical integration is practical only

in markets of very low dimension N. For larger markets, one

where
must resort to analytical or quasi-analytical formulas. Such
where
1
formulas are available only for a handful of distributions,
( ) .
(25) (25)
most notably the elliptical family, which includes the normal
and the Student t distributions.
To implement copulas in practice, we must be able to imAn alternative to avoid the multivariate integration is to
To implement
copulas
in practice,process
we must
beinable
to to
implement
the sepplement
the separation
(21),
order
obtain suitdraw scenarios directly from parametric copulas. However,
tion process (21), in order to obtain suitable copulas, and the combination
able copulas, and the combination process (24), in order to the parametric specifications that allow for direct simulation
cess (24), in order to glue those copulas with suitable marginals.
glue those copulas with suitable marginals.
are limited to the Archimedean family (see Genest and Rivest
Implementing all the steps involved in such processes analytically is impos-

e, except in trivial cases, such as within the normal family.

Therefore, all practical applications of copulas rely on numerical techniques,
st notably the representation of distributions by means of Monte Carlo sceios.
www.garp.org
Within the Monte Carlo framework, scenarios that represent copulas are
ained by feeding joint distributions scenarios into the respective marginal

O C T O B E R 2 0 1 1 RISK PROFESSIONAL

26

T H E Q U A N T C L A S S R O O M B Y AT T I L I O M E U C C I

The Copula-Marginal
Algorithm is a flexible tool
to implement in practice the
crucial separation step and
combination step with fully
flexible distributions.
(1993)) and few other extensions. Furthermore, the parameters of the Archimedean family are not immediate to interpret. Finally, simulating grades scenarios from the Archimedean family when the dimension N is large is computationally
challenging.
To summarize, traditional implementations of copulas
mainly proceed as follows: first, Monte Carlo scenarios are
drawn from elliptical or related distributions; next, the scenarios are channelled through the respective (quasi- )analytical marginal cdfs, as in (22), thereby obtaining grade scenarios; then, the grade scenarios are fed into flexible parametric
quantiles, as in (25), thereby obtaining the desired joint scenarios.
To avoid the restrictive assumptions of the traditional copula implementation and circumvent all the above problems,
Meucci (2011) proposes the Copula-Marginal Algorithm
(CMA), which simulates Monte Carlo scenarios with flexible probabilities from arbitrary distributions, computes the
marginal cdfs without integrations, and avoids the quantile
computation. CMA is numerically extremely efficient. For all
the details, the code, and an application to stress-testing with
panic markets, please refer to Meucci (2011).
FOOTNOTE
1. For more background on the subject, the reader is referred to
articles such as Embrechts, A. and Straumann (2000); Durrleman,
Nikeghbali and Roncalli (2000); and Embrechts, Lindskog and McNeil (2003). The reader is also referred to monographs, such as
Nelsen (1999); Cherubini, Luciano and Vecchiato (2004); Brigo,
Pallavicini, and Torresetti (2010); and Jaworski, Durante, Haerdle
and Rychlik (2010).

27

RISK PROFESSIONAL O C T O B E R 2 0 1 1

REFERENCES
Brigo, D., A. Pallavicini and R. Torresetti, 2010. Credit Models
and the Crisis: A Journey Into CDOs, Copulas, Correlations and
Dynamic Models (Wiley).
Cherubini, U., E. Luciano and W. Vecchiato, 2004. Copula Methods in Finance (Wiley).
Durrleman, V., A. Nikeghbali and T. Roncalli, 2000. "Which Copula is the Right One?" Working Paper.
Embrechts, P., A. McNeil and D. Straumann, 2000. "Correlation:
Pitfalls and Alternatives," Working Paper.
Embrechts, P., F. Lindskog and A. J. McNeil, 2003. "Modelling
Dependence with Copulas and Applications to Risk Management,
Handbook of Heavy-Tailed Distributions in Finance."
Genest, C., and R. Rivest, 1993. "Statistical Inference Procedures
for Bivariate Archimedean Copulas," Journal of the American Statistical Association 88, 1034 1043.
Jaworski, P., F. Durante, W. Haerdle and T. Rychlik (Editors), 2010.
Copula Theory and its Applications (Springer, Lecture Notes in
Statistics - Proceedings).
Li, D. X., 2000. "On Default Correlation: A Copula Function Approach," Journal of Fixed Income 9, 4354.
Meucci, A., 2010. "Linear vs. Compounded Returns - Common
Pitfalls in Portfolio Management," Risk Professional, "The Quant
Classroom by Attilio Meucci," April, 5254. Article and code
available at http://symmys.com/node/141.
ibid, 2011. "New Breed of Copulas for Risk and Portfolio Management," Risk 24, September, 122-126. Article and code available at
http://symmys.com/node/335.
Meucci, A., Y Gan, A. Lazanas and B. Phelps, 2007. A Portfolio
Manager's Guide to Lehman Brothers Tail Risk Model, Lehman
Brothers Publications.
Nelsen, R. B., 1999. An Introduction to Copulas (Springer).
:
Attilio Meucci is the chief risk officer at Kepos Capital LP. He runs the 6-day "Advanced Risk and Portfolio Management Bootcamp" (see www.symmys.com). He is
grateful to Garli Beibi.

www.garp.org