Professional Documents
Culture Documents
A Short,
Comprehensive, Practical
Guide to Copulas
Visually introducing a powerful risk management tool
to generalize and stress-test correlations.
T
he stochastic behavior of one single financial vari- retical results. In Section 3, we address copulas implementa-
able 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
Consider an arbitraryresults random that prepare
variable ,thewith a fully arbitrar
formation on the stochastic behavior of the market is fully ground for the Alldefinition
the features of, and thedistribution
of the intuition behind, of copulas.
are described by its prob
described by the joint probability distribution. Consider an arbitrary
function (pdf)random variable
, dened inX, witha away
such fullythat,
arbi-for any set of pote
The multivariate distribution of a set of financial trary for
distribution. the variable
All the features
, the
Consider an arbitrary random variable , with a fully arbitrary distribution. are
variables of following
the identity
distribution of for
X the pdf holds
is fully specified by the separate marginalAll Consider
distributions an arbitrary
of the random
described variable
by its with
probability
, a fully
the features of the distribution of are described by its probabilityXdensity in
density arbitrary
function distribution.
(pdf) f , defined
Z
All the
variables and by their copula, or, loosely speaking,
function features
(pdf) of
the
the correla- distribution wayofthat,
suchina such
, dened a wayare described
forthat,
any for anyby
set of itsofprobability
potential
set P values
potential
{ Xvalues density
}for theXvari-
() .
tions among the variables. function (pdf) , dened
for the variable , the able
in such a way
X, theidentity
following followingthat, for
foridentityany
the pdffor set of potential
the pdf holds:
holds values X
X
Then we can compute the marginal cdfs as in (2). Finally, we can feed
23 RISK PROFESSIONAL O C T O B E R 2 0 1 1 each cdf , which is a function, with the respectivewww.garp.org
entry of the vector ,
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
workNowtheweunivariate
are fully equipped to introduce
results (4) and (7). the copula, by extending to the multi- 1
ade U2
pdf fU1
a variate framework the univariate results (4) and (7).
-dimensional vector of random variables (1 ) x2
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 X fX cdf FX 2
inWe
therecall
multivariate
that in thecase
multivariate isthe
the pdf case dened
pdf inissuchdened a way that,
in such a way that,
potential
for any setjointT 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
of potential Rvalues
values Xjoint
forX(1 R for ) 1the
( following
) the following
identity holds
Z Z pdf f X 2
{(1 P
) {( )(
X1} X1}
)(
1 1
) .
. 1 (9) (9) x1
X X
joint X=(X1,X2) marginal X2
From the joint
t distribution distribution
we can in principle we can extract in principle all extract
the marginalall the marginal
distributions , where = 1 , by computing the marginal pdfs
, where = 1 , by computing the marginal pdfs
as follows
Z Figure 2: Copulas are non-linear standardizations of multivariate distributions
Z ( ) = and what is left is the pure joint information amongst the Xn's
( 1 ) 1 1 +1 . (10)(10)
) = (1 ) 1 1 +1 .
R 1
(10) i.e., the copula fU. We summarize this statement in an al-
R1 Then we can compute the marginal cdf s FX , as in (2). Fi- ternative, intuitive formulation of the copula.
Then we can compute the marginal cdfs as in (2). Finally, n we canKey feedconcept. the copula is the information missing from the individual
each cdfthe nally, we can feed each cdf F Xn, which is a function, with the re-
compute , which iscdfs
marginal a function,
aswith in the
(2).respective
Finally, entry we can of the
feedvector ,
spective entry of the vector X namely, the random marginals
variable to
Key complete
concept. theThejointcopula
distribution
is the information missing from
ulas:namely theory
the random variable . The
which is a function, with the respective entry of the vector ,
outcome of this operation are the grades,
which
ndom weXknow
variable n. The
from. outcome
The haveofa this
(4)outcome of operation
uniform thisdistribution
operation are the on
arethegrades,
theunit grades,which
intervalwe the individual marginals to complete the joint distribution:
" joint = copula + marginals " (13)
ully equipped know to from
introduce
w from (4) have a uniform distribution (4) havethe a uniform
copula, distribution
by
on extending
the unit onto
intervalthe
the unit
multi- interval
( ) U[01] . (11)
work the univariate results (4) and (7). " joint = copula + marginals " (13)
a -dimensional
However,the vector
mainpoint (of random U[01]variables
) remember
to . is that the (1of
entries 1) (11)
(11)
( )
general
are not multivariate
independent. distribution
Therefore, the represented
joint distributionby its pdf ofthe . The
grades intuitive
is not The denition (13) can (13)
intuitive definition be made can be rigorous. From theFrom
made rigorous. denition of
the main
inuniform point
the multivariate todomain,
on However,
its remember
case the the
which is pdf
main that
is the
point
theunit entries
isto dened
cube
remember ofin
[0 such
1] is ( a1 way
that
the
[0 that,
1], copula
see
)Figure
entries (12)
of 2. the definition of copula (12),
fendent.
potential Therefore,
Ujoint
(Uvalues the joint not
1,...,UN)Xare
distribution
Rindependent. of
the
for (1 Therefore, ) grades
the the is
following
joint notdistri-
domain, which isfUthe unitgradescube [0 1] [0distribution
1],itsseedomain,
Figure 2.the joint () P {1 1 } (14)
Key bution
concept. of The the copula is not
of an uniform
arbitrary on iswhich is
the unit
distribution cube grades... [0,1] (see Figure 2, below).
of itsZ[0,1] = P {1 (1 ) 1 ( ) }
{(
ncept.
1 ) X }
Thecopula of an arbitrary ( 1 )
distribution
1
is . the joint (9) = P 1 1
(1 ) 1
( ) (14)
1 1 (1 ) 1 1
sort of non-linear
s of the distribution can bez-score, which as
interpreted forces allofthe
a sort entries Xn inverse
non-linear to We now
cdfs , derive
1
another
we obtain useful
new transformed
1 1 ( (1 ))
result.random If we feed the
variables withgrades
a given
joint distribution, which we denote by1 1
have a uniform distribution on the unit interval
h forces all the entries to have a uniform distribution on the [0,1]. By feed- (12) into arbitrary inverse cdf
.. s F Y n
-1
, we obtain new trans-
( . ) (17)
ing eacheach
[0 1]. By feeding random variable
random Xn into
variable itsinto
ownits
cdf,own
all the
cdf,informa-
all formed random variables 1
1 (
1 with( a
)) given joint distribution,
1 (1 ( ))
1
on containedtion
in contained in each
each marginal marginal distribution
distribution is sweptfXnaway,
is swept
andaway, which we denote ( by fY .. ) (17)
.
s the pure joint information amongst the s, i.e. the copula The joint distribution has marginals
1 whose cdfs are , which follows from
( ( ))
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
The joint distribution has marginals whose cdfs are , which follows from
the copula
applying (11) (12) we obtain
and (6) in sequence. Furthermore, the copula of is the same as
www.garp.org the copula of , because from
OC Tthe
O denition
B E R 2of0 1 in
1 (1 ) = 1 (1 )
RISK
1 (17) PROFESSIONAL
and the denition of 24
the copula (12) we obtain ..
4 ( . ) . (18)
( ) = ( )
(
la , and |the 1 joint
marginals (1) , as
) , or equivalently follows
( )
{z } | 1 1
{z 1 } .
1 pure
( 1 joint
) 1
( pure
) m arginal
( )
1 (1 ) | {z( ) } | 1 1
1 1
{z } = {z (16) } .
lars theorem
joint justies the intuitive copula denition
pure joint (13).
pure m arginal
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 justies
) 1 (1 ) copula 1
denition
(13). ( ) 1
klars
| theorem{z } |the1 intuitive {z } .
klars theorem
pure joint provides the pdf of the copula
pure m arginal from the joint pdf and the
ginal pdfs. This allows us to use maximum likelihood to t copulas to
Sklars
heorem
irical theoremthe
justies
data. provides
intuitivethe pdf of
copula the copula
denition (13). from the joint pdf and the
We now derive another useful result. If we feed thelikelihood
rginal pdfs. This allows us to use maximum grades (12) to into
t copulas
arbitraryto
pirical
rse data.
cdfs 1
, wereturns
obtain new transformed random variables with a given
theorem Compounded
provides the C=(C
pdf of1,C2) copula from the
the Linear
joint returns
pdf R=(R
and1,R 2)
the
We now derive
distribution, another
which we useful
denote result.
by If we feed the grades (12) into arbitrary
dfs. This allows c2 us to use maximum likelihood to r2t copulas to
erse cdfs
data.
f C2 1 , we obtain newreturns
Compounded transformed
C=(C1,C2) random variables withR=(R
Linear returns a given
1,R2)
C f
t distribution, which we denote1 by 1
1 c Ifwe
C
(
R
e C
ecopula
joint distribution which has from
marginals applying
whose (11) FR, FR (6)
and in sequence. Fur-
of inare
cdfs , which follows from
1 C1 C2
of , because
from ( the( ))
denition (17) and the denition
1
of 2
istribution
copula of , has marginals
because
because from whose
fromthethe u2 cdfs
denition are
definition , which
ofof YininU=(U
Copula
(17)follows
(17) ,U
1 2
and
and
) from
the
the definition
denition of
of display Same Copula
11) and (6)
(12)in we
sequence. Furthermore,
the copula the copula u of U is fthe same as
copula obtain 1 ((12),
1 ) = we obtain
1 (1 )
2 U
of , because from the denition fU 2
.. of in (17) )and the denitionU of f
. f 1 (1 ) . (18) U
(12) we obtain ( 1 (1 ) = U2
( ) .. ( )
( ( )= ( .) ) u1 . (18)
(18)
1 1 1 1 u1
.. (copula fU
refore, we derive ( thatthe . ) of )an( arbitrary
.) random variable
(18)f
1
U1
) does not(change when we transform each into a new variable
erefore, ) the copula
( ) of an arbitrary
(we derive
that random
) by means Therefore,
of functions we derive
() that the 1 copula of anvariable
( ()), where are
arbitraryran-
we ) does
derive that not
dom change
variable when we
Xarbitrarytransform
(X1,...,Xreturns
) does each
not into
change a new
when
we variable
trans-
3: the copulaandofcompounded
an Nrandom variable
Figure Linear have samecopula These two types of returns, though calculated on the same securities prices
(not
) does ) by means
changeform
when of
Figure
each functions
we 3:transform
XLinear
n into and
a each
()
new 1
compounded
variable
into
(aY returns
new ()),
n variable
where
gn (Xhave
n) bysamemeans are
copula ofare dierent. Therefore, their distributions are dierent, refer to
,
functions () 61
+1
) by means of functions ( ()), where where Fare arearbitrary cdfs.(2010) for more details on the pitfalls of disregarding such dier-
Yn Meucci
itrary cdfs.arbitrary
It It
is iseasy
cdfs. to verify
easy toItverifyis easy that
that to such
such functions
functions
verify that g
such
are are
a a
very
functions very
broad broad
class
6 n ences.a For
are veryexample,
broad if the prices distribution is multivariate log-normal,
ss, namely class, all the increasing
namely,
namely alltransformations,
all6 the the increasing
increasing also known also
transformations,
transformations, as also
co-monotonic
known known as as
co-monotonic
the linear returns distribution is multivariate shifted-lognormal and the
nsformations. Thus we obtain
transformations.
co-monotonic Thus the wefollowing
obtain the following
transformations. Thus, we obtain the following: compounded returns distribution is multivariate normal, as illustrated in
Figure 3 for two negatively correlated stocks.
Key concept. KeyKeyCo-monotonic
concept. transformations
concept.Co-monotonic
Co-monotonic transformations
= (
transformations )Y= the
nof gHowever,
(Xn)) of the
n( the copula of the linear returns and the copula of the com-
ntries of doentries not
of alter
the
of the
entries copula
do not ofalter
X do ofthenotcopula
alter the of copula of X pounded returns However, the copula
are identical, seeofagain
the linear returns
Figure 3. Thisand the copula
result of sur-
is not
prising, because the compounded
= returns
1 is anare identical transformation
increasing (see, again, Figure 3).
of and
1 1 1 1
is an increasing transformation of Cn .
.
. = ln (1 + This
)result is not surprising, because R n = e
1 is an increas-
( ... ) ( .. ) ( ... ) ( .. )
1 1
ing transformation of Cn, and Cn =1n (1+ Rn) is an increasing
transformation of Rn.
1 1 1 1 3 (19)
Copulas:
(19) practice
& & . .
1 Section 3: Copulas (Practice)
1 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 discus-
in practice relies on two distinct processes, which appear in
sion of Section 2.
multiple steps
First, the separation process,in thewhich
theoretical
led us discussion of Sectionof2.the copula
to the denition
(12). First, the separation process, which led us to the definition
To illustrate how the copula is not affected by increasing of the copula (12).
To illustrate how consider,
transformations, the copulaasis follows,
not aected
the by increasing
linear returnstransformations,
and
To illustrate how
considerthe copula
the linear isreturns
not aected
and the bycompounded
increasing transformations,
returns between time Key
andThe
the compounded Key theconcept. separation The process S strips an Sarbitrary
strips an distribution
consider the linear
time returns
+ 1 for and thereturns
the same
betweenreturns
compounded
securities
time t and time ttime
between + 1 for and concept. separation process arbitrary
same securities: into its marginals
distribution f and
into its
its copula
marginals f and its copula f , as follows:
time + 1 for the same securities X X n U
+1 +1
+1 1
+1
ln( ). (20) 1
1 ln(
). (20)
(20)
1
S : ( .. )
7 1 (21)
These two types of returns, though 7 calculated on the same .
.. ) , (21)
securities prices Pn,t+17, are different. Therefore, their distribu-
( .
tions are different (refer to Meucci (2010) for more details on
The 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 itsseparation
marginalsprocess
andS strips an arbitrary
its copula distribution
ginals and its copula
1
1
1
1
. S: ( ... )
7 1 (21)
: ( .. ) 7 1
( .. ) ,(21)
.
.
( .. ) ,
e grades
where Un are the grades
here are the grades
( ) . (22)
The implementation of
( ) . (22) (22)
the theory of copulas
processeach
can
The separation process can be reverted, similarly to the
The separation process can be reverted, similarly to the univariate case (6).
univariate
be reverted, case similarly
(6). By feeding each gradecase Un back
relies on two fundamental
into the
eeding grade back into theto the univariate
-1 respective 1 inverse cdf (6). 1
we obtain
rade
enal
(1(X
respective
back into
1,...,X
distribution
theinverse
N) )whose
whose
cdf FXinverse
respective
joint
joint
n , we obtain
ndom variable (1 ) whose joint distribution
distribution
cdf
distribution
is
a
is
random
we
exactly
obtain
exactlythe
variable
original
the
X
steps: "separation"
is exactly the
dis-
:
on : tribution fX:
1 1 1
(1 )
and "combination."
1
1
.. ) 1 71 ( 1 ( ) ..
1
.. ) (
7. (
.. ) .1
)
(23) (23)(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
However, we do not need to limit ourselves to reverting to the original dis-ourselves to reverting to normal family.
otion
notneed the
to limitoriginal distribution
ourselves fX. Copulas
to reverting to the areoriginal
so powerful dis- because 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 can bebecause
glued theywith can be glued
arbitrary with arbitrary
marginal distributions, with merical techniques, most notably the representation of distri-
a technique that generalizes the univariate case (7), which we butions by means of Monte Carlo scenarios.
used to prove 8
rginal distributions, 8 the
with a
co-monotonic
technique that
invariance of the copula (19).
generalizes the1.univariate case (7),
Within the Monte Carlo framework, scenarios that repre-
Accordingly,
_ we start with two ingredients: an arbitrary sent copulas are obtained by feeding joint distributions sce-
d which we copula
used tofUprove thegrades
i.e., co-monotonic
U (U1,...,U invariance of the copula (19).
N), each of which has a narios into the respective marginal cdfs, as in (22), and joint
Accordingly, we start with two ingredients:
uniform distribution and joint distribution structure an arbitrary copula , i.e.
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 .
compute the marginal cdfs' F Xn and their inverses F Xn , and we
-1
en we compute the marginal cdfs , their inverses and we feed eachHowever, even the above numerical operations present dif-
1
feed 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
iable (1 ) that has the desired copula _ N)that has the univariate integrations, as in (2). Second, the computation
and the desired
desired
rginals . copula f U and the desired marginals f X n . of the inverse cdf in (25) requires univariate integrations fol-
We summarize Weassummarize,
follows this assecond
follows,process.
this second process. lowed by search algorithms. Finally the extraction of the mar-
ginal distributions from the joint distribution in (21) requires
Key concept. _ The combination process C glues arbitrary multivariate integrations, as in (10).
Key concept.
marginalsThe combination
fXn and an arbitrary process
copula glues
C f arbitrary marginals The first two problems, namely the univariate integration
U into a new joint dis-
and an arbitrary copula into a new joint distribution
tribution fX, as follows: and inversion, have been addressed for a broad variety of
parametric distributions, for which cdf and quantile are avail-
1
1 able either analytically or in terms of efficient quadratures.
On the other hand, the multivariate integration to extract
C: 1 7 ( ... ) , (24)
.
( .. )
(24) the marginal from the joint distribution represents a signifi-
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 im- An alternative to avoid the multivariate integration is to
To implementplement copulas in practice,process
the separation we must (21),beinable to to
order implement
obtain suit-the sep-
draw 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 sce-
ios. www.garp.org O C T O B E R 2 0 1 1 RISK PROFESSIONAL 26
Within the Monte Carlo framework, scenarios that represent copulas are
ained by feeding joint distributions scenarios into the respective marginal
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
(1993)) and few other extensions. Furthermore, the param- Embrechts, P., F. Lindskog and A. J. McNeil, 2003. "Modelling
eters of the Archimedean family are not immediate to inter- Dependence with Copulas and Applications to Risk Management,
pret. Finally, simulating grades scenarios from the Archime- Handbook of Heavy-Tailed Distributions in Finance."
dean family when the dimension N is large is computationally
Genest, C., and R. Rivest, 1993. "Statistical Inference Procedures
challenging. for Bivariate Archimedean Copulas," Journal of the American Sta-
To summarize, traditional implementations of copulas tistical Association 88, 1034 1043.
mainly proceed as follows: first, Monte Carlo scenarios are
drawn from elliptical or related distributions; next, the sce- Jaworski, P., F. Durante, W. Haerdle and T. Rychlik (Editors), 2010.
narios are channelled through the respective (quasi- )analyti- Copula Theory and its Applications (Springer, Lecture Notes in
Statistics - Proceedings).
cal marginal cdfs, as in (22), thereby obtaining grade scenar-
ios; then, the grade scenarios are fed into flexible parametric Li, D. X., 2000. "On Default Correlation: A Copula Function Ap-
quantiles, as in (25), thereby obtaining the desired joint sce- proach," Journal of Fixed Income 9, 4354.
narios.
To avoid the restrictive assumptions of the traditional cop- Meucci, A., 2010. "Linear vs. Compounded Returns - Common
Pitfalls in Portfolio Management," Risk Professional, "The Quant
ula implementation and circumvent all the above problems,
Classroom by Attilio Meucci," April, 5254. Article and code
Meucci (2011) proposes the Copula-Marginal Algorithm available at http://symmys.com/node/141.
(CMA), which simulates Monte Carlo scenarios with flex-
ible probabilities from arbitrary distributions, computes the ibid, 2011. "New Breed of Copulas for Risk and Portfolio Manage-
marginal cdfs without integrations, and avoids the quantile ment," Risk 24, September, 122-126. Article and code available at
computation. CMA is numerically extremely efficient. For all http://symmys.com/node/335.
the details, the code, and an application to stress-testing with Meucci, A., Y Gan, A. Lazanas and B. Phelps, 2007. A Portfolio
panic markets, please refer to Meucci (2011). Manager's Guide to Lehman Brothers Tail Risk Model, Lehman
Brothers Publications.