Professional Documents
Culture Documents
Research Paper N° 57
November 2002
Fame is about research, doctoral training, and executive education with “interfacing”
activities such as the FAME le ctures, the Research Day/Annual Meeting, and the Research
Paper Series.
FAME will strive to promote the research work in finance carried out in the three partner
Universities. These papers are distributed with a ‘double’ identification: the FAME logo and the
logo of the corresponding partner institution. With this policy, we want to underline the vital
lifeline existing between FAME and the Universities, while simultaneously fostering a wider
recognition of the strength of the academic community supporting FAME and enriching the
Lemanic region.
Each contribution is preceded by an Executive Summary of two to three pages explaining in non-
technical terms the question asked, discussing its relevance and outlining the answer provided.
We hope the series will be followed attentively by all academics and practitioners interested in
the fields covered by our name.
I am delighted to serve as coordinator of the FAME Research Paper Series. Please contact me
if you are interested in submitting a paper or for all suggestions concerning this matter.
Sincerely,
Jean-David FERMANIAN
Olivier SCAILLET
November 2002
NONPARAMETRIC ESTIMATION
OF COPULAS FOR TIME SERIES
1
Jean-David FERMANIANa and Olivier SCAILLETb
a
CDC Ixis Capital Markets, 56 rue de Lille, 75007 Paris, and CREST, 15 bd Gabriel Péri, 92245 Malako¤
scaillet@hec.unige.ch
November 2002
Abstract
Key words: Nonparametric, Kernel, Time Series, Copulas, Dependence Measures, Risk Man-
1
We would like to thank Marc Henry for many stimulating discussions and proof checking. We also
thank Paul Embrechts and Alexander McNeil for kindly providing the Danish data on …re insurance losses,
which were collected by Mette Rytgaard at Cop enhagen Re. We have received fruitful comments from P.
Doukhan, H. Hong and O. Renault. The second author receives support by the Swiss National Science
Foundation through the National Center of Competence: Financial Valuation and Risk Management. Part
of this research was done when he was visiting THEMA and IRES.
Downloadable at
http://www.hec.unige.ch/professeurs/SCAILLET_Olivier/pages_web/Home_Page.htm
Executive Summary
Knowledge of the dependence structure between financial assets or claims is crucial to achieve
performant risk management in finance and insurance. To measure dependence through
computation of correlations reveals adequate in the context of multivariate normally distributed
risks or in assessing linear dependence. Contemporary financial risk management however calls
for other tools due to the presence of an increasing proportion of nonlinear risks (derivative
assets) in trading books and the nonnormal behaviour of most financial time series (skewness and
leptokurticity).
Furthermore it is now well admitted that the choice of the dependence structure, or similarly of
the copula, is also often a key issue for numerous pricing models in finance and insurance. This is
especially true concerning the pricing and hedging of credit sensitive instruments such as
collateralised debt obligations (CDO) or basket credit derivatives.
In this paper we propose a nonparametric estimation method for copulas. The copula of a
multivariate distribution can be considered as the part describing its dependence structure as
opposed to the behaviour of each of its margins. One attractive property of the copula is its
invariance under strictly increasing transformation of the margins, and its direct link with scale
invariant measures of association such as Kendall's tau and Spearman's rho.
Estimation of copulas have mainly been developed in the context of bivariate i.i.d. samples either
through maximum likelihood methods or nonparametric methods based on empirical distributions.
In the following we use a kernel based approach. Such an approach has the advantage to provide a
smooth (differentiable) reconstitution of the copula function without putting any particular
parametric a priori on the dependence structure between margins and without losing the usual
parametric rate of convergence. The approach is developed in the context of multivariate
stationary processes satisfying strong mixing conditions. This class of processes includes standard
dynamic models such as ARMA, ARCH and stochastic volatility models.
Once estimates of copulas and their derivatives are available (need of differentiability explains
our choice of a kernel approach), concepts such as positive quadrant dependence and left tail
decreasing behaviour may be empirically analysed. These estimates are also useful to draw
simulated data satisfying the dependence structure inferred from observations. Besides,
nonparametric estimators of copulas may lead to statistics aimed to assess independence between
margins. These statistics are in the same spirit as kernel based tools used to test for serial
dependence for a univariate stationary time series.
The paper starts with a section outlining the framework. Definition and properties of copulas are
recalled. Then we present the kernel estimators of copulas and their derivatives, and characterize
their asymptotic behaviour. Their use in estimation of measures of dependence between margins
is also briefly described. A section containing Monte Carlo results for a stationary vector
autoregressive process of order one with Gaussian innovations shows that the proposed estimators
have good properties in terms of bias and MSE. Two empirical examples illustrate the empirical
relevance of the estimators to the analysis of dependencies. The first one concerns European and
US stock index returns. The second one deals with Danish data on fire insurance losses. The
empirical study reveals the very different type of dependence structures among the stock index
returns and the fire insurance losses.
1 Introduction
2
dependence and left tail decreasing behaviour may be empirically analysed. These esti-
mates are also useful to draw simulated data satisfying the dependence structure infered
from observations (see Embrechts, McNeil and Straumann (2002) for the description
of an algorithm). Besides, nonparametric estimators of copulas may lead to statistics
aimed to assess independence between margins. These statistics are in the same spirit as
kernel based tools used to test for serial dependence for a univariate stationary time series
(see the survey of Tjostheim (1996)).
The paper is organized as follows. In Section 2 we outline our framework and re-
call the de…nition of copulas and some of their properties. In Section 3 we present the
kernel estimators of copulas and their derivatives, and characterize their asymptotic be-
haviour. Their use in estimation of measures of dependence between margins is also brie‡y
described. Section 4 contains some Monte Carlo results for a stationary vector autoregres-
sive process of order one with Gaussian innovations. An empirical illustration on European
and US stock index returns is provided in Section 5. Another empirical illustration deals
with Danish data on …re insurance losses. Section 6 concludes. All proofs are gathered in
an appendix.
2 Framework
3
of the elements constituting Yt. The marginal p.d.f. and c.d.f. of each element Yj t at point
y j, j = 1; :::; n, will be written fj (yj), and Fj (yj ), respectively. A copula describes how
the joint distribution F is "coupled" to its univariate margins Fj , hence its name. Before
de…ning formally a copula and reviewing various useful dependence concepts, we would
like to refer the reader to Nelsen (1999) and Joe (1997) for more extensive treatments.
De…nition. (Copula)
A n-dimensional copula is a function C with the following properties:
2. C is grounded, i.e. for every u in [0; 1]n , C(u) = 0 if at least one coordinate uj = 0,
j = 1; :::; n.
3. C is n-increasing, i.e. for every a and b in [0; 1]n such that a · b, the C-volume
VC ([a; b]) of the box [a; b] is positive.
The reason why a copula is useful in revealing the link between the joint distribution
and its margins transpires from the following theorem.
If F1; :::; Fn are all continuous, then C is uniquely de…ned. Otherwise, C is uniquely
determined on range F1 £::: £range Fn . Conversely, if C is an n-copula and F1; :::; Fn are
distribution functions, then the function F de…ned by (1) is an n-dimensional distribution
function with margins F1; :::; Fn.
4
where F1¡1; :::; Fn¡1 are quasi inverses of F1; :::; Fn, namely
Note that, if Fj is strictly increasing, the quasi inverse is the ordinary inverse. Copulas
are thus multivariate uniform distributions which describe the dependence structure of
random variables. Besides as already mentioned, strictly increasing transformations of the
underlying random variables result in the transformed variables having the same copula.
From expression (2), we may observe that the dependence structure embodied by the
copula can be recovered from the knowledge of the joint distribution F and its margins
Fj . These are the distributional objects that we propose to estimate nonparametrically
by a kernel approach in the next section. Before turning our attention to this problem,
let us review some relevant uses of copulas.
First copulas characterize independence and comonotonicity between random variables.
Qn
Indeed, n random variables are independent if and only if C(u) = j =1 uj , for all u, and
each random variable is almost surely a strictly increasing function of any of the others
(comonotonicity) if and only if C(u) = min(u1; :::; un ), for all u. In fact copulas are
intimately related to standard measures of dependence between two real valued random
variables Y1t and Y2t , whose copula is C. Indeed, the population versions of Kendall’s tau,
Spearman’s rho, Gini’s gamma and Blomqvist’s beta can be expressed as:
Z 1Z 1
@C(u1 ; u2 ) @C(u1 ; u2)
¿Y1;Y2 = 1 ¡ 4 du1du2 ; (3)
0 0 @u1 @u2
Z 1Z 1
½Y1;Y2 = 12 C(u1 ; u2 )du1 du2 ¡ 3; (4)
0 0
Z 1
°Y1;Y2 = 4 [C(u1; 1 ¡ u1) + C(u1; u1)]du1 ¡ 2; (5)
0
¯Y1;Y2 = 4C(1=2; 1=2) ¡ 1: (6)
Second copulas can be used to analyse how two random variables behave together
when they are simultaneously small (or large). This will be useful in examining the joint
behaviour of small returns, especially the large negative ones (big losses), which are of
particular interest in risk management. This type of behaviour is best described by the
5
concept known as positive quadrant dependence after Lehmann (1966). Two random
variables Y1t and Y2t are said to be positively quadrant dependent (PQD) if, for all (y 1; y2 )
in IR2 ,
This states that two random variables are PQD if the probability that they are simultane-
ously small is at least as great as it would be if they were independent. Inequality (7) can
be rewritten in terms of the copula C of the two random variables, since (7) is equivalent
to the condition C(u1; u2) ¸ u1 u2, for all (u1; u2) in [0; 1]2.
Finally inequality (7) can be rewritten P[Y1t · y 1jY2t · y 2] ¸ P[Y1t · y1] by applica-
tion of Bayes’ rule. The PQD condition may be strengthened by requiring the conditional
probability being a non increasing function of y 2. This implies that the probability that
the return Y1t takes a small value does not increase as the value taken by the other return
increases. It corresponds to particular monotonicities in the tails. We say that a random
variable Y1t is left tail decreasing in Y2t, denoted LT D(Y1jY2), if P[Y1t · y1jY2t · y2 ] is a
non increasing function of y2 for all y 1. This in turn is equivalent to the condition that,
for all u1 in [0; 1], C(u1; u2 )=u2 is non increasing in u2 , or @C(u1; u2 )=@u2 · C(u1 ; u2)=u2
for almost all u2 .
In short, concepts such as independence, PQD or LTD, may be characterized in terms
of copulas, and thus may be checked (see for example the testing procedures developped in
Denuit and Scaillet (2002) and Cebrian, Denuit and Scaillet (2002)), once copulas
are empirically known. In the next section we develop estimation tools for that purpose.
3 Kernel estimators
We start with the de…nition of kernel estimators before moving to their asymptotic distri-
butions.
6
3.1 De…nitions
For given uj 2 (0; 1), j = 1; : : : ; n, we assume that the c.d.f. Fj of Yjt , is such that the
equation Fj (y) = uj admits a unique solution denoted ³j(uj ), or more compactly ³j (if
there is no ambiguity).
To build our estimators we need to introduce kernels, i.e. real bounded functions kj on
R such that
Z
kj(x) dx = 1; j = 1; : : : ; n:
For the sake of simplicity, we choose to work here with products of univariate kernels. We
could however extend easily our results to more general k and K. Let us denote further
n
à ! n
à !
Y xj Y xj
k(x; h) = kj ; K(x; h) = Kj ;
j=1
hj j=1
hj
where the bandwidth h is a diagonal matrix with elements (hj )nj=1 and determinant jhj
(for a scalar x, jxj will denote its absolute value), while the individual bandwidths hj are
positive functions of T such that hj ! 0 when T ! 1. Moreover, we denote by h¤ the
largest bandwidth among h1; : : : ; hn. The p.d.f. of Yjt at yj , i.e. fj (yj ), will be estimated
as usually by
T
à !
X yj ¡ Yjt
f^j (yj ) = (T hj ) ¡1
kj ;
t=1
hj
while the p.d.f. of Yt at y = (y1; :::; yn )0, i.e. f(y), will be estimated by
T
X
f^(y) = (Tjhj)¡1 k(y ¡ Yt; h):
t=1
7
Hence, an estimator of the cumulative distribution of Yjt at some point yj is obtained
as
Z yj
F^j (yj ) = f^j (x)dx; (8)
¡1
and
T Y
X n
^
F(y) = T ¡1 © ((yj ¡ Yjt )=hj ) ;
t=1 j=1
where ' and © denote the p.d.f. and c.d.f. of a standard Gaussian variable, respectively.
In order to estimate the copula at some point u, we use a simple plug-in method, and
exploits directly expression (2):
^
C(u) ^ ^³);
= F( (10)
where ^³ = (³^1; :::; ³^n)0 and ³^j = infy2Rfy : F^j (y) ¸ uj g. In fact ³^j corresponds to a kernel
estimate of the quantile of Yj t with probability level uj (see e.g. Gourieroux, Laurent
and Scaillet (2000) and Scaillet (2000) for use of this type of smooth quantile estimates
in risk management and portfolio selection as well as Azzalini (1981) for the asymptotic
properties in the i.i.d. case).
The asymptotic normality of kernel estimators for copulas can be established under suit-
able conditions on the kernel, the asymptotic behaviour of the bandwidth, the regularity
of the densities, and some mixing properties of the process.
8
Assumption 1. (kernel and bandwidth)
(a) Bandwidths satisfy Th2¤ ! 0, or
(a’) Bandwidths satisfy T h4¤ ! 0 and the kernel k is even,
(b) The kernel k has a compact support.
Assumption 1 (b) could in fact be weakened, by controlling the tails of k, for instance
by assuming supj jkj (x)j · (1+jxj)¡® for every x and some ® > 0 (as in Robinson (1983)).
This type of assumption is satis…ed by most kernels, in particular by the Gaussian kernel.
Assumption 2. (process)
(a) The process (Yt) is strong mixing with coe¢cients ®t such that ®T = o(T ¡a) for some
a > 1, as T ! 1.
(b) The marginal c.d.f. Fj , j = 1; : : : ; n are continuously di¤erentiable on the intervals
[Fj¡1(a) ¡ "; Fj¡1(b) + "] for every 0 < a < b < 1 and some " > 0, with positive derivatives
f j. Moreover, the …rst partial derivatives of F exist and are Lipschitz continuous on the
product of these intervals.
9
process Á0(G) in l 1([0; 1]n) endowed with the sup-norm, where the limiting process is given
by
and à ! à !
T
^ ^³) = 1
X ³^p ¡ Ypt Y ³^l ¡ Ylt
@pF( k Kl :
Thp t=1 p hp l6=p
hl
10
^
The estimators C(u) ^
and @pC(u) will help to detect positive quadrant dependence and
left tail decreasing behaviour through the empirical counterparts of the aforementioned
inequalities.
Assumption 4. (process)
(a) The process (Yt ) is strong mixing with coe¢cients ®t such that ®T = O(T ¡2 ), as
T ! 1.
(b) The …rst derivative @pF and the marginal density fp are Lipschitz continuous.
(c) Let ft;t0 be the density function of (Yt ; Yt0 ) with respect to the Lebesgue measure. There
exists an integrable function à : R2n¡2 ¡! R such that, for every vectors u; v in Rn , and
for some open subset O of R2 containing Ap £ Ap , we have
Then we get:
Again the asymptotic covariance matrix §¤ of this random vector admits a rather
complex explicit form (see Equation (13) in the appendix).
Note that the extension of the previous propositions to higher derivatives
^
@ k C(u)=(@u 1@u2:::@uk ), k · n, is straightforward. Such estimates are for example re-
11
^ ¡ C) at some
As clear from Theorem 4, a random vector of derivatives of T 1=2(C
points u1; : : : ; ud does not in general tend weakly to a vector of independent Gaussian
variables. This is however the case when the components corresponding to the indices
of the derivatives are all di¤erent. A similar result is also true for successive derivatives.
Indeed, under some technical assumptions, the random vector
³ ´
(Thm mn 1=2
1 : : : : :hn )
1
@1m1 : : : @mn ^ m1 mn ^
n (C ¡ C)(u1); : : : ; @1 : : : @n (C ¡ C)(u d)
1. ml ¸ 1 for every l = 1; : : : ; n, or
2. for the indices l such that ml 6= 0, uli 6= ulj for every pair (i; j).
We end up this section with the description of how to build the sample analogues of
^ for the
the dependence measures (3)-(6). First we may substitute a kernel estimator C
unknown C in (4)-(6) to estimate ½Y1 ;Y2 , °Y1;Y2 , and ¯Y1;Y2 . According to Theorem 3, these
estimators are consistent and asymptotically normally distributed. Second we may replace
^ 1 ; u2 )=@u1 and @C(u
the unkown derivatives by @ C(u ^ 1; u2)=@u2 in order to estimate ¿Y1;Y2 .
In this section we wish to investigate the …nite sample properties of our kernel estimators.
The experiments are based on a stationary vector autoregressive process of order one:
Yt = A + BYt¡1 + ºt ;
where ºt » N(0; §) and all the eigenvalues of B are less than one in absolute value. The
marginal distribution of the process Y is Gaussian with mean ¹ = (Id ¡ B)¡1A and
covariance matrix - satisfying vec - = (Id ¡ B - B)¡1vec § (vec M is the stack of the
columns of the matrix M).
We start with a bivariate example where the components Y1t and Y2t are independent.
The parameters are A = (1; 1)0 , vec B = (0:25; 0; 0; 0:75)0 and
12
vec § = (:75; 0; 0; 1:25)0. The parameters of the Gaussian stationary distribution are then
¹ = (1:33; 4)0 and vec - = (0:8; 0; 0; 2:86)0 . The number of Monte Carlo replications is
5,000, while the data length is T = 45 = 1024 (roughly four trading years of quotes).
The nonparametric estimators make use of the product of two Gaussian kernels. Band-
^ iT ¡1=5, which uses the empirical stan-
width values are based on the rule of thumb hi = ¾
dard deviation of each series. Table 1 gives bias and mean squared error (MSE) of the
^ 1 ; u2) for several pairs (u1; u2) in the tails and center of the distribu-
kernel estimates C(u
tion. All bias and MSE …gures are expressed as percents of percents (10 ¡4 ).
We may observe that the results are satisfactory both in terms of bias and MSE.
The pairs located in the tails of the distribution are well estimated. The pairs C(:01; :01);
C(:05; :05); in the left tail are of particular interest in risk management since they measure
the dependence of joint extreme losses.
We continue with a bivariate example where the components Y1t and Y2t are dependent.
The parameters are A = (1; 1)0 , vec B = (0:25; 0:2; 0:2; 0:75)0 and vec § = (:75; :5; :5; 1:25)0.
The parameters of the Gaussian stationary distribution are then ¹ = (3:05; 6:44)0 and
vec - = (1:13; 1:49; 1:49; 3:98)0 (correlation of 0.70). Results are reported in Table 2. Bias
and MSE are higher when compared with the previous independent case. They are how-
ever still satisfactory.
13
Table 2: Bias and MSE of kernel estimates of copulas
£10¡4 C(.01,.01) C(.05,.05) C(.25,.25) C(.5,.5) C(.75,.75) C(.95,.95) C(.99,.99)
True 27.08 197.95 1511.74 3747.68 6511.74 9197.95 9827.08
Bias -7.47 -34.88 -130.32 -172.28 -130.53 -35.25 -7.65
MSE 0.01 0.18 1.98 3.36 1.99 0.18 0.01
5 Empirical illustrations
This section illustrates the implementation of the estimation procedure described in Sec-
tion 3. We provide two empirical illustrations, the …rst one in …nance and the second one
in insurance. They concern data on stock index returns and Danish data on …re insurance
losses.
We analyse return data on two pairs of major stock indices: CAC40-DAX35, and SNP500-
DJI. The data are one day returns recorded daily from 03/01/1994 to 07/07/2000, i.e. 1700
observations. We also report results for return with a holding period of ten days instead of
one day. This holding period is favoured by the Basle Committee on Banking Supervision
when quantifying trading risks. We have selected bandwidth values according to the usual
rule of thumb (empirical standard deviation over the sample length at the power one …fth),
and used a Gaussian kernel.
We start with the pair of European indices CAC40-DAX35. The linear correlation co-
e¢cient between the two indices is 67.03%. Figure 1 reports plots of observed returns, and
3D and contour plots of estimated copulas for this …rst pair. The left column corresponds
to one day returns while the right column corresponds to ten day returns. Estimated
shapes show that the dependence is more pronounced for the ten day returns. In Figure 2
we use estimates of the copulas and their derivatives to analyse positive quadrant depen-
dence (PQD) and left tail decreasing behaviour (LTD). The …rst line of graphs shows that
C(u1; u2) ¡ u1 u2 is greater than zero, which means that one day and ten day returns ex-
hibit PQD. The di¤erence is larger in the center of the distribution and decreases when we
14
move to the extremes. The two other lines carry graphs of C(u1; u2)=u2 ¡@C(u1; u2)=@u2,
and C(u1; u2)=u1 ¡ @C(u1; u2)=@u1, respectively. Again we get positiveness, and LTD is
thus present for both stock indices and both holding periods. Besides, PQD and LTD are
heavier for the ten day holding period.
For the pair of US indices SNP500-DJI, we get Figures 3-4 and Table 6. The linear
correlation coe¢cient is equal to 93.25%. Comments made for European indices carry
over. However the dependence is higher for US indices, especially the positive quadrant
dependence, as seen on Figures 3-4.
The Danish data consist of 1794 losses over one million Danish Krone (DKK) for the years
1980 to 1990. Loss …gures are amounts lost, and are classi…ed as damage to buildings,
and damage to furniture and personal property. The linear correlation coe¢cient between
the two types of damage is 26.92%. Again the usual rule of thumb is applied to select
the bandwidth, while we keep a Gaussian kernel. Figure 5 plots the data, estimated
copula, and an analysis of PQD and LTD. The dependence structure is very di¤erent
when compared with the previous return data. The dependence is rather weak in the
center of the distribution, but is larger for a small probability in one type of damage and a
high probability in the other one. This is explained by the clustering of very large damage
to furniture but with small damage to buildings, and vice-versa.
6 Concluding remarks
In this paper we have proposed simple nonparametric estimation methods of copulas and
their derivatives. The procedure relies on a kernel approach in the context of general
stationary strong mixing multivariate processes, which provides smooth di¤erentiable es-
timators. These estimators have proven to be empirically relevant to the analysis of
dependencies among stock index returns and …re insurance losses. In particular they have
allowed revealing the very di¤erent type of dependence structures present in these data.
Hence they complement ideally the existing battery of nonparametric tools by providing
15
speci…c instruments dedicated to dependence measurement and joint risk analysis. They
should also help to design goodness-of-…t tests for copulas. This is under current research.
16
APPENDIX
Proof of Theorem 2
First, the equicontinuity of the process T 1=2(FT ¡ F) is a consequence of its weak conver-
gence (see Rio (2000)) :
and such that the empirical process T 1=2(FT ¡ F ) tends weakly to G in l 1(Rn ). The
covariance structure of the limiting process is given by Equation (11).
Thus, under such assumptions, and since k has a compact support, sup y jA1j = oP (1).
Second, since F is Lipschitz continuous, the second term A2 is O(T 1=2 h¤), or further
O(T 1=2 h2¤) if k is even. Thus, we have proved that
17
hence the stated result. Note that we have proved Theorem 2 using the fact that F is
Lipschitz continuous which is weaker than Assumption 2. 2
Proof of Theorem 3
^ ¡ C) in
By the functional Delta-Method, we deduce the weak convergence of T 1=2(C
l 1([a; b]n ), for every a; b, 0 < a < b < 1, exactly as in Van der Vaart and Wellner
(1996, p. 389), and we obtain the convergence of the …nite dimensional distributions.
^ ¡ C)(u) is zero when one component of u is zero. Moreover, when one
Note that T 1=2(C
^ ¡ C)(u) tends to a Gaussian
component of u is 1, say the j-th component, then T 1=2(C
random variable that would be obtained if we had forgotten all the j-th components (in
other words, as if the observed random variables had been (Ykt )k6=j ). Let us …nally remark
that the weak convergence of T 1=2 (C^ ¡ C) can be proved in l 1([0; 1]n). This can be done
by exploiting the proximity between F^ and FT provided by Theorem 2, and by mimicking
the proof of Theorem 10 in Fermanian, Radulovic and Wegkamp (2002). 2
Proof of Theorem 4
Let us remark that the Delta-Method provides us the limiting behaviour of the smoothed
empirical quantiles.
Theorem 6. Under Assumptions 1 and 2, for every uj 2]0; 1[ and every j 2 f1; : : : ; ng,
³^j is a consistent estimator of ³j ´ Fj¡1 (uj ). Moreover,
law
T 1=2 (³^j ¡ ³j) ¡! N (0; ¾2(uj ));
where P
2 t2Z Cov(1fYj0 · uj g; 1fYjt · ujg)
¾ (uj ) = 2 ¢
fj (³j )
In particular, each ³^j tends to ³j at the parametric rate T ¡1=2 . Note that Assumptions
3 and 4 imply Assumptions 1 and 2. We prove in the next theorem that the quantities
@p F^ and f^p converge at the slower rate (Thp)¡1=2. Thus, it will be convenient to replace
18
each random quantity ³^j by its limit ³j in @pF^ (³)
^ and f^p(³^p ). The proof of the following
tends weakly to a centered Gaussian vector whose covariance matrix § = (¾i;j )1·i;j·2d, is
characterized by
Z
¾i;j = ¾j;i = @pF (yi ^ yj) k2p ; if ypi = y pj ;
Z
¾i+d;j+d = ¾j+d;i+d = fp (ypi ) k2p ; if ypi = ypj ;
Z
¾i;j+d = ¾j+d;i = @p F(yi ) kp2; if ypi = ypj ;
where (u1; : : : ; ud) is some point in ]0; 1[nd . Since every ³^j tends to ³j ´ Fj¡1(uj ) at the
rate T ¡1=2, and since f^p(³p ) tends to f p(³p) at the slower rate (Thp)¡1=2, the asymptotic
behaviour of f^p(³^p ) will be the same as the one of f^p (³p). We however still need a continuity
argument. If kp is twice continuously di¤erentiable, kp00 being bounded,
à !
(³^p ¡ ³p)2 X
T ¤
00 ³ p ¡ Ypt
f^p(³^p ) = f^p(³p) + f^p0 (³p ):(³^p ¡ ³p) + k
2Th3p t=1 p hp
(³^p ¡ ³p)2
= f^p(³p) + f^p0 (³p ):(³^p ¡ ³p) + O(T);
2Th3p
where j³p¤ ¡ ³pj · j³^p ¡ ³p j. It can be proved easily that f^p0 (³p) is OP (1) (for instance by
use of an exponential inequality, or even the Markov inequality). Thus,
+ OP ((Thp)1=2T ¡1h¡3
p ) = (T hp)
1=2 ^
(f p ¡ fp)(³p) + OP (h1=2
p ) + OP (T
¡1=2 ¡5=2
hp ):
19
This quantity tends to zero when T h5p tends to zero. Exactly in the same way, we get
under the same assumptions:
Proof of Theorem 7
= O(h¤);
20
since @p F is Lipschitz continuous. Then, this bias term is negligible under our assumptions.
Pd
Similarly, (Thp)1=2 ^
i=1 ¹i(E[fp] ¡ f )(ypi) is o(1) under the same assumptions.
Second, to obtain the asymptotic normality, the simplest way to proceed is to apply
Lemma 7.1 in Robinson (1983) (see e.g. Bierens (1985) or Bosq (1998) for alternative
sets of assumptions). In his notations, set p = 2d, ai = hp , as well as
and
¤
VitT = ¹i fkp (ypi ¡ Ypt ; hp) ¡ E [kp (ypi ¡ Ypt ; hp)]g:
21
If yip 6= y jp, then kp ((ypi ¡ypj )=hp +up) and k0p((ypi ¡ypj )=hp +up) is zero for every up 2 Ap,
for hp su¢ciently small. We get that E[VitT VjtT ] = 0 in this case, for T su¢ciently large.
Thus, let us assume ypi = ypj . For any index l 6= p, notice that Kl ((ylj ¡ y li )=hl + ul )
is zero if ylj < yli and is one if ylj > yli, for T su¢ciently large. In the …rst case,
set the change of variable yli ¡ hl :ul = ylj ¡ hl :vl . The l-th factor in brackets becomes
kl (vl )Kl ((yli ¡ ylj )=hl + vl ), which is kl (vl ) for T su¢ciently large. In the second case, the
latter factor is kl (ul ). Thus, E[VitT Vj tT ]) is nonzero only if ypi = ypj and, for T su¢ciently
large,
8 9
<Z Y =
E[VitT VjtT ] = ¸i¸j F(yi ^ yj ¡ h ¢ u)2kp0 (up)kp (up) ¢ kl (ul ) du + O(h2¤): (14)
: ;
l6=p
2
By an integration by parts with respect to up (similarly as for E[VitT ]), we get easily
2
E[VitT Vj tT ] = ¸i ¸j hp¾ij + o(h¤):
¤ ¤
if yip = yjp, and E[VitT VjtT ] = oP (hp) otherwise. Moreover,
Z Ã !
¤ y ip ¡ up
E[VitT VjtT ] = ¹i¸j kp (@pK) (yj ¡ u; h) F(du) + OP (h2p)
hp
Z ( Ã !
Y yip ¡ yj p
= ¹i ¸j F(yj ¡ h ¢ u) kl (ul ) kp (up)kp0 + up
l6=p
hp
à !)
yip ¡ yjp
+ kp0 (up)kp + up du + O P (h2p):
hp
If yip 6= yjp , the latter quantity is zero for T su¢ciently large. Otherwise, it is equivalent
R
to ¹i ¸j hp @p F(yj) kp2.
22
Thus, condition A:7:3 of Robinson (1983) is satis…ed. It remains to verify condition
A:7:4. For every i 6= j and t 6= t0 ,
Z ¯ µZ ¶¯
¯ ¯
¯ (@pK) (yi ¡ u; h) ¡
E[jVitT Vjt0 T j] = ¸i ¸j (@ p K) (y i ¡ u; h) F(du) ¯
¯ ¯
¯ µZ ¶¯
¯ ¯
¢ ¯¯ (@pK) (yj ¡ v; h) ¡ (@pK) (yj ¡ v; h) F(dv) ¯¯ F(Yt;Yt0 ) (du; dv)
0¯ ¯ 1 0¯ ¯ 1
Z ¯ µ ¶¯ ¯ µ ¶¯
¯ Y yil ¡ ul ¯ ¯ Y yjl ¡ vl ¯
· Cst:h2p ¸i ¸j @ ¯¯kp(up) Kl ¯ + 1A @ ¯kp(vp)
¯ ¯ Kl ¯ + 1A
¯
¯ l6=p
h ¯ ¯ l6=p
h ¯
ft;t0 (u¡p; y ip ¡ hpup; v¡p; yjp ¡ hpvp) du dv:
Under our assumptions, the latter quantity if O(h2¤ ). Using similar boundings the same
property is satis…ed for E[jVitT Vjt¤0 T j] and E[jVitT
¤ V ¤ j].
jt 0T
The other conditions of Lemma 7.1 in Robinson (1983) are clearly satis…ed. Note
P+1
that our condition 4 (a) implies Robinson’s one on the mixing coe¢cients, i.e. t=T ®t =
O(T ¡1). Hence we have indeed proved the stated result. 2
23
Asymptotic covariance matrix in Corollary 1
References
24
[4] Bosq, D. (1998). Nonparametric Statistics for Stochastic Processes: Estimation and
Prediction. Lecture Notes in Statistics, Springer-Verlag, New-York.
[5] Caperaa, P., A. Fougeres, and C. Genest (1997). Estimation of Bivariate Ex-
treme Value Copulas. Biometrika, 84, 567-577.
[6] Cebrian, A., M. Denuit, and O. Scaillet (2002). Testing for Concordance Or-
dering. DP FAME 41.
[7] Denuit, M., and O. Scaillet (2002). Nonparametric Tests for Positive Quadrant
Dependence. DP FAME 44.
[9] Doukhan, P. (1994). Mixing: Properties and Examples. Lecture Notes in Statistics,
Springer-Verlag, New-York.
[10] Embrechts, P., A. McNeil, and D. Straumann (2002). Correlation and Depen-
dence in Risk Management: Properties and Pitfalls, in Risk Management: Value at
Risk and Beyond edited by Dempster M., Cambridge University Press, Cambridge.
[12] Frees, E., and E. Valdez (1998). Understanding Relationships Using Copulas.
North American Actuarial J., 2, 1 ¡ 25.
[13] Frey, R., and A. McNeil (2001). Modelling dependent defaults. Working paper.
[14] Genest, C., and L.P. Rivest (1993). Statistical Inference Procedures for Bivariate
Archimedean Copulas. Journal of the American Statistical Association, 88, 1034-
1043.
[15] Ghoudi, K., A. Khoudraji, and L.P. Genest (1998). Proprietes Statistiques des
Copules de Valeurs Extremes. Canadian Journal of Statistics , 26, 187-197.
25
[16] Gouriéroux, C., J.P. Laurent and O. Scaillet (2000). Sensitivity Analysis of
Values at Risk. Journal of Empirical Finance, 7, 225-245.
[17] Joe, H. (1997). Multivariate Models and Dependence Concepts. Chapman & Hall,
London.
[20] Rio, E. (2000). Théorie Asymptotique des Processus Aléatoires Faiblement Dépen-
dants. SMAI, Springer.
[21] Robinson, P. (1983). Nonparametric Estimators for Time Series. Journal of Time
Series Analysis, 4, 185-207.
[24] Van der Vaart, A., and J. Wellner (1996). Weak Convergence and Empirical
Processes. Springer-Verlag, New York.
26
RESEARCH PAPER SERIES
Extra copies of research papers are available to the public upon request. In order to obtain copies of past or future works, please
contact our office at the following address: International Center FAME, 40 bd. du Pont d’Arve, Case Postale 3, 1211 Geneva 4.
As well, please note that these works are available on our website www.fame.ch, under the heading "Research" in PDF format for
your consultation. We thank you for your continuing support and interest in FAME, and look forward to serving you in the future.
N° 56: Interactions Between Market and Credit Risk: Modeling the Joint Dynamics of Default-Free
and Defaultable Bond Term Structures
Roger WALDER, University of Lausanne, International Center FAME and Banque Cantonale Vaudoise;
November 2002
N° 53: Permanent and Transitory Factors Affecting the Dynamics of the Term Structure of Interest Rates
Christophe PÉRIGNON, Anderson School, UCLA and Christophe VILLA, ENSAI, CREST-LSM and
CREREG-Axe Finance; June 2002
N° 48: A Geometric Approach to Multiperiod Mean Variance Optimization of Assets and Liabilities
Markus LEIPPOLD, Swiss Banking Institute, University of Zurich; Fabio TROJANI, Institute of Finance,
University of Southern Switzerland; Paolo VANINI, Institute of Finance, University of Southern Switzerland;
April 2002
N° 38: Indirect Estimation of the Parameters of Agent Based Models of Financial Markets
Manfred GILLI, University of Geneva; Peter WINKER, International University in Germany; November 2001
N° 37: How To Diversify Internationally? A comparison of conditional and unconditional asset allocation methods.
Laurent BARRAS, HEC-University of Geneva, International Center FAME; Dušan ISAKOV, HEC-University of
Geneva, International Center FAME; November 2001
N° 35: Country, Sector or Style: What matters most when constructing Global Equity Portfolios? An empirical
investigation from 1990-2001.
Foort HAMELINK, Lombard Odier & Cie and Vrije Universiteit; Hélène HARASTY, Lombard Odier & Cie;
Pierre HILLION, Insead (Singapore), Academic Advisor to Lombard Odier & Cie; October 2001
N° 30: Serial and Parallel Krylov Methods for Implicit Finite Difference Schemes Arising in Multivariate
Option Pricing
Manfred GILLI, University of Geneva, Evis KËLLEZI, University of Geneva and FAME,
Giorgio PAULETTO, University of Geneva; March 2001
N° 24: Assessing Market Risk for Hedge Funds and Hedge Funds Portfolios
François-Serge LHABITANT, Union Bancaire Privée and Thunderbird, the American
Graduate School of International Management; March 2001
N° 23: On the Informational Content of Changing Risk for Dynamic Asset Allocation
Giovanni BARONE-ADESI, Patrick GAGLIARDINI and Fabio TROJANI, Università della Svizzera Italiana;
March 2000
N° 22: The Long-run Performance of Seasoned Equity Offerings with Rights: Evidence From the Swiss Market
Michel DUBOIS and Pierre JEANNERET, University of Neuchatel; January 2000
N° 21: Optimal International Diversification: Theory and Practice from a Swiss Investor's Perspective
Foort HAMELINK, Tilburg University and Lombard Odier & Cie; December 2000
N° 12: Do fixed income securities also show asymmetric effects in conditional second moments?
Lorenzo CAPIELLO, The Graduate Institute of International Studies; January 2000
N° 11: Dynamic Consumption and Portfolio Choice with Stochastic Volatility in Incomplete Markets
George CHACKO, Harvard University; Luis VICEIRA, Harvard University; September 1999
N° 6: Systematic patterns before and after large price changes: Evidence from high frequency data
from the Paris Bourse
Foort HAMELINK, Tilburg University; May 1999
N° 5: Who Should Buy Long-Term Bonds?
John CAMPBELL, Harvard University; Luis VICEIRA, Harvard Business School; October 1998
N° 3: Real Options as a Tool in the Decision to Relocate: An Application to the Banking Industry
Pascal BOTTERON, HEC-University of Lausanne; January 2000
N° 1: Enhancing portfolio performance using option strategies: why beating the market is easy.
François-Serge LHABITANT, HEC-University of Lausanne; December 1998
International Center FAME - Partner Institutions
More than 13’000 students, the majority being foreigners, are enrolled in the various programs from the
licence to high-level doctorates. A staff of more than 2’500 persons (professors, lecturers and assistants) is
dedicated to the transmission and advancement of scientific knowledge through teaching as well as
fundamental and applied research. The University of Geneva has been able to preserve the ancient European
tradition of an academic community located in the heart of the city. This favors not only interaction between
students, but also their integration in the population and in their participation of the particularly rich artistic and
cultural life. http://www.unige.ch
Lausanne is a comprehensive university composed of seven Schools and Faculties: religious studies; law; arts;
social and political sciences; business; science and medicine. With its 9’000 students, it is a medium-sized
institution able to foster contact between students and professors as well as to encourage interdisciplinary
work. The five humanities faculties and the science faculty are situated on the shores of Lake Leman in the
Dorigny plains, a magnificent area of forest and fields that may have inspired the landscape depicted in
Brueghel the Elder's masterpiece, the Harvesters. The institutes and various centers of the School of
Medicine are grouped around the hospitals in the center of Lausanne. The Institute of Biochemistry is located
in Epalinges, in the northern hills overlooking the city. http://www.unil.ch
The Institute attempts to be both international and pluridisciplinary. The subjects in its curriculum, the
composition of its teaching staff and the diversity of origin of its student body, confer upon it its international
character. Professors teaching at the Institute come from all regions of the world, and the approximately 650
students arrive from some 60 different countries. Its international character is further emphasized by the use
of both English and French as working languages. Its pluralistic approach - which draws upon the methods of
economics, history, law, and political science -reflects its aim to provide a broad approach and in-depth
understanding of international relations in general. http://heiwww.unige.ch
Prospect Theory
and Asset Prices
Nicholas BARBERIS
University of Chicago
Ming HUANG
40, Bd. du Pont d’Arve Stanford University
PO Box, 1211 Geneva 4
Switzerland Tano SANTOS
University of Chicago
Tel (++4122) 312 09 61
Fax (++4122) 312 10 26
http: //www.fame.ch
E-mail: admin@fame.ch