Bivariate Option Pricing Using

Dynamic Copula Models

Rob W. J. van den Goorbergh∗ Christian Genest†
Bas J. M. Werker‡

This version: February 1, 2005

Abstract
This paper examines the behavior of bivariate option prices in the
presence of association between the underlying assets. Parametric fam-
ilies of copulas offering various alternatives to the normal dependence
structure are used to model this association, which is explicitly as-
sumed to vary over time as a function of the volatilities of the assets.
These dynamic copula models are applied to better-of-two-markets and
worse-of-two-markets options on the S&P500 and Nasdaq indexes. Re-
sults show that option prices implied by dynamic copula models differ
substantially from prices implied by models that fix the dependence
between the underlyings, particularly in times of high volatilities. Fur-
thermore, the normal copula produces option prices that differ signif-
icantly from non-normal copula prices, irrespective of initial volatility
levels. Within the class of non-normal copula families considered, op-
tion prices are robust with respect to the copula choice.

JEL classification: G13.

Keywords: Time-varying dependence, best-of-two-markets options, non-
normality.
∗
Research Department, ABP Investments, P.O. Box 75753, 1118 ZX Schiphol, The
Netherlands. Corresponding author. E-mail: [Link]@[Link], Phone:
+31 (0)20 405 5892, Fax: +31 (0)20 405 9809.
†
Département de mathématiques et de statistique, Université Laval, Québec, Canada
G1K 7P4.
‡
Department of Finance, Department of Econometrics and Operations Research, and
CentER for Economic Research, Tilburg University, Tilburg, The Netherlands.

1
1 Introduction
In today’s economy, multivariate (or rainbow) options are viewed as excellent
tools for hedging the risk of multiple assets. These options, which are written
on two or more underlying securities or indexes, usually take the form of calls
(or puts) that give the right to buy (or sell) the best or worst performer of
a number of underlying assets. Other examples include forward contracts
whose payoff is equal to that of the best or worst performer of its underlyings,
and spread options on the difference between the prices of two assets.
One of the key determinants in the valuation of multivariate options
is the dependence between the underlying assets. Consider for instance a
bivariate call-on-max option, namely a contract that gives the holder the
right to purchase the more valuable of two underlying assets for a pre-
specified strike price. Intuitively, the value of such an option should be
smaller if the underlyings tend to move together than when they move in
opposite directions. More generally, the dependence between the underlyings
could change over time. Accounting for time variation in the dependence
structure between assets should prove helpful in providing a more realistic
valuation of multivariate options.
Over the years, various generalizations of the Black–Scholes (1973) Brow-
nian motion framework have been used to model multivariate option prices.
Examples include Margrabe (1978), Stulz (1982), Johnson (1987), Reiner
(1992), and Shimko (1994). In these papers, the dependence between as-
sets is modelled by their correlation. However, unless asset returns are well
represented by a multivariate normal distribution, correlation is often an
unsatisfactory measure of dependence; see, for instance, Embrechts, McNeil
and Straumann (2002). Furthermore, it is a stylized fact of financial mar-
kets that correlations observed under ordinary market conditions differ sub-
stantially from correlations observed in hectic periods. In particular, asset
prices have a greater tendency to move together in bad states of the economy
than in quiet periods; see, for instance, Boyer, Gibson and Loretan (1999)
and Patton (2003, 2004) and references therein. These “correlation break-
downs,” associated with economic downturns, suggest a dynamic model of
the dependence structure of asset returns.
In this paper, the relation between bivariate option prices and the depen-
dence structure of the underlying financial assets is modelled dynamically
through copulas. A copula is a multivariate distribution function each of
whose marginals is uniform on the unit interval. It has been known since
the work of Sklar (1959) that any multivariate continuous distribution func-
tion can be uniquely factored into its marginals and a copula. Thus while

2
correlation measures dependence through a single number, the dependence
between multiple assets is fully captured by the copula. From a practical
point of view, the advantage of the copula-based approach to modelling is
that appropriate marginal distributions for the components of a multivari-
ate system can be selected by any desired method, and then linked through
a copula or family of copulas suitably chosen to represent the dependence
prevailing between the components.
The use of copulas to price multivariate options is not new. For example,
in Rosenberg (1999), univariate options data are used to estimate marginal
risk-neutral densities, which are linked with a Plackett copula to obtain a
bivariate risk-neutral density from which bivariate claims are valuated. This
semiparametric procedure uses a particular identifying assumption on the
risk-neutral correlation to fix the copula parameter. Cherubini and Luciano
(2002) extend Rosenberg’s work by considering other families of copulas.
In Rosenberg (2003), a risk-neutral bivariate distribution is estimated from
nonparametric estimates of the marginal distributions and a nonparametric
estimate of the copula.
An innovating feature of the present paper, however, is that, contrary
to earlier works on multivariate option pricing, the dependence structure of
the underlying assets is not treated as fixed, but rather as possibly varying
over time. Taking into account this time variation is important because it
may influence option prices. This paper proposes a model for the time vari-
ation of the dependence structure, in which a parametric copula is specified
whose dependence parameter is allowed to change with the volatilities of
the underlying assets. A distinct advantage of the parametric approach is
that while the model may be misspecified, the robustness of the conclusions
can easily be verified by repeating the analysis for as many different copula
families as desired.
A similar dynamic-copula approach has already been used in the for-
eign exchange market literature by Patton (2003), who found time variation
to be significant in a copula model for asymmetric dependence between two
exchange rates where the dependence parameter followed a ARMA-type pro-
cess. While Patton’s goal was to study the effect of asymmetric dependence
on portfolio returns, the objective of the present paper is very different.
The main focus here is on the effect of time variation in the underlying
dependence structure on the price of multivariate options.
In the empirical study presented herein, multivariate options on two im-
portant American equity index returns are considered: the S&P500 and the
Nasdaq. An analysis of the results suggests that allowing for time variation
in the dependence structure of the underlyings produces substantially dif-

3
ferent option prices than under constant dependence, particularly in times
of increased volatility. Moreover, option prices implied by a normal dy-
namic dependence structure differ significantly from option prices implied
by non-normal dynamic dependence structures. These findings suggest that
unless the dependence between the S&P500 and Nasdaq stock indexes is well
described by a normal copula, alternative copula families should be consid-
ered. Option prices turned out to be robust among the alternative—i.e.,
non-normal—copula models considered in this study.
The remainder of this paper is organized as follows. Section 2 describes
the payoff structure of better-of-two-markets and worse-of-two-markets
claims, and explains in detail the proposed dynamic-dependence option val-
uation scheme. The empirical results are presented in Section 3, and con-
clusions are given in Section 4.

2 Option valuation with time-varying dependence

Multivariate options come in a wide variety of payoff schemes. The most
commonly traded options of this kind are basket options on a portfolio of
assets, such as index options. Other examples include spread options, some
of which are traded on commodity exchanges (see, for example, Rosenberg
(1998)), or dual-strike and multivariate-digital options.
This paper concentrates on European-type options on the best (worst)
performer of several assets, sometimes referred to as outperformance (un-
derperformance) options. As these are typically traded over the counter,
price data are not available. Therefore, valuation models cannot be tested
empirically. However, a robustness study comparing models with differ-
ent assumptions remains feasible, and this is the objective pursued herein.
While the study described in the sequel is restricted to options on better-
and worse-of-two-markets claims, the technique is sufficiently general to an-
alyze the aforementioned alternative multivariate options as well, and may
thus be of wider interest.
One can distinguish four types of better-of-two-markets or worse-of-two-
markets claims: call options on the better performer, put options on the
worse performer, call options on the worse performer, and put options on
the better performer. These may be referred to as call-on-max, put-on-min,
call-on-min, and put-on-max options, respectively. The expiration payoffs
of these claims may be defined in terms of the underlying asset prices. For
instance, the payoff of a call-on-max option would be max{max(S1 , S2 ) −
K, 0}, with Si being the price at maturity of asset i ∈ {1, 2} and K the strike

4
price. Of course, initial asset prices need to be close for the option to make
sense. For expository reasons, it is assumed here that they are exactly equal
to an amount S, say, and the option premium is expressed as a percentage
(in basis points, or 0.01%) of this common S. This, by linearity, can be done
by valuing an option on S1 /S and S2 /S, with strike K/S. Effectively, all
initial prices are normalized to unity. Using this convention, the expiration
payoffs of the four types of contracts are given by:
call on max : max{max(R1 , R2 ) − E, 0},
put on min : max{E − min(R1 , R2 ), 0},
call on min : max{min(R1 , R2 ) − E, 0},
put on max : max{E − max(R1 , R2 ), 0},
where Ri = Si /S is the gross return at maturity on underlying i ∈ {1, 2},
and E = K/S denotes the normalized exercise price of the option.
The proposed scheme for valuating these options is as follows. Let
ri,t+1 = log Ri,t+1 be the log return on index i ∈ {1, 2} from time t to
time t + 1, and let It = σ((r1,s , r2,s ) : s ≤ t) denote all return informa-
tion available at time t. First, the objective bivariate distribution of the
log returns (r1,t+1 , r2,t+1 ) is specified conditional on past information It . It
is assumed that this conditional distribution has Gaussian margins and a
certain conditional copula Ct . The model allows volatilities and dependence
to be time varying in a non-deterministic way; the volatilities of the con-
ditional joint distribution are modelled as functions of past squared return
innovations, while the conditional copula is assumed to depend on the past
via Kendall’s tau as a function of past volatilities. The model is stationary;
see Comte and Lieberman (2003) for a discussion.
The next step in the valuation scheme is the derivation of the joint
risk-neutral return process from the objective bivariate distribution. The
specification of the objective marginals, in conjunction with the assumption
that the objective and the risk-neutral conditional copulas are the same,
allows for a particularly convenient transformation to risk neutrality; instead
of deriving the bivariate risk-neutral distribution directly, it is found by
transforming each of the marginal processes (and the copula) separately.
The fair value of the option is then determined by taking the discounted
expected value of the option’s payoff under the risk-neutral distribution.
The specification chosen for the objective marginal distributions is from
Duan (1995). It is general enough to capture volatility clustering, a stylized
fact of equity returns for which there is overwhelming empirical evidence
at the daily frequency, while still providing a relatively easy transformation
to risk-neutral distributions. Each of the objective marginal distributions

5
of the index returns is modelled by a GARCH(1,1) process with Gaussian
innovations. It is repeated here for the sake of completeness; see Bollerslev
(1986). For i ∈ {1, 2},
ri,t+1 = µi + ηi,t+1 ,
2
hi,t+1 = ωi + βi hi,t + αi ηi,t+1 ,
LP (ηi,t+1 |It ) = N (0, hi,t ),
where ωi > 0, βi > 0, and αi > 0, and LP (·|It ) denotes the objective
probability law conditional on the information set It , which includes all
realized returns on both indexes. The marginal distributions are specified
conditional on this common information set, so that copula theory can be
used to construct a joint conditional distribution. Failure to use the same
conditioning information for the margins and the copula will, in general, lead
to invalid joint density models. This point is emphasized by Patton (2003).
The GARCH parameters are estimated by maximum likelihood, using the
unconditional variance level ωi /(1 − βi − αi ) as starting value hi,0 .
It must be stressed that, in the light of Sklar’s theorem, in principle
any choice for the marginal distributions is consistent with the copula ap-
proach. The vast collection of alternatives that have been used by other au-
thors to model univariate index return distributions includes (variants of)
continuous-time geometric Brownian motion of Black and Scholes (1973),
and the discrete-time binomial model of Cox, Ross and Rubinstein (1979).
Again, the GARCH specification that is employed here is appealing as it
allows for an easy change of measure in addition to being able to capture
volatility clustering. In particular, Duan (1995) shows that, under certain
conditions, the change of measure comes down to a change in the drift. The
law of the returns under the risk-neutral probability measure (Q) is given
by:
ri,t+1 = rf − 12 hi,t + ηi,t+1
∗ ,
hi,t+1 = ωi + βi hi,t + αi (ri,t+1 − µi )2 ,
∗
LQ (ηi,t+1 |It ) = N (0, hi,t ),
where rf is the risk-free rate, which is assumed to be constant. Recall that
under the risk-neutral measure, actualized prices are martingales.
It is important to note here that the specified marginal distributions are
conditional on the common information set It , but that it is assumed that
both conditional margins only depend on their own past, i.e.,
LP (ri,t+1 |It ) = LP (ri,t+1 |Ii,t ),

6
where Ii,t = σ(ri,s : s ≤ t) denotes the information on index i available at
time t. What this means in particular is that return spillovers or volatility
spillovers from one index to the other are excluded. This restriction is
necessary for the application of Duan’s change of measure.
An alternative, nonparametric approach is to use univariate option price
data to obtain arbitrage-free estimates of the marginal risk-neutral densities,
as in Ait-Sahalia and Lo (1998). This route is taken by Rosenberg (2003).
Clearly, an advantage of this approach is that it does not impose restrictions
on the asset return processes or on the functional form of the risk-neutral
densities. However, this flexibility comes at the cost of imprecise estimates,
especially if the distributions are time-varying.
The description of the joint distribution of the index returns under the
objective probability measure is completed by fixing the conditional cop-
ula. A set of well-known one-parameter copula families is considered for
this purpose. They are the Frank, Gumbel–Hougaard, Plackett, Galambos,
and normal families. Their cumulative distribution functions are given in
Appendix A. For all of these copulas, there is a one-to-one relation between
the dependence parameter—denoted θ—and Kendall’s nonparametric mea-
sure of association. For any copula Cθ , Kendall’s tau is related to θ in the
following way:
τ (θ) = 4ECθ (U, V ) − 1, (1)
where (U, V ) is distributed as Cθ , and E denotes the expectation operator
with respect to U and V . Appendix B displays closed-form formulas for
the population value of Kendall’s tau for some of the copula models under
consideration.
This relation suggests a natural way to estimate the copula. An esti-
mate of θ is readily obtained by computing the sample version of tau on a
(sub)sample of paired index-return observations, inverting Relation (1), and
plugging in the sample tau. See Appendix C for a definition of the sample
version of Kendall’s tau. This method-of-moment type procedure yields a
rank-based estimate of the association parameter which is consistent, under
the assumption that the selected family of copulas describes accurately the
dependence structure of the equity indexes. Other methods could be used
without fundamentally altering this approach, e.g., inversion of Spearman’s
rho, or the maximum pseudo-likelihood method.
The proposed technique assumes that the objective and risk-neutral cop-
ulas are identical, so that the objective joint returns process is easily trans-
formed into its risk-neutral counterpart, using the risk-neutral marginals and
the copula. Rosenberg (2003) makes this assumption as well. If bivariate

7
option price data were available, equality of the objective and risk-neutral
copulas could be tested or the appropriate risk-neutral copula could be es-
timated. Only data on prices of bivariate claims would reveal information
about the risk-neutral dependence structure. Information about the risk-
neutral dependence structure can never be extracted from univariate option
prices—which are available—as these only bear relevance to the marginal
risk-neutral processes, and not to the joint risk-neutral process. Identi-
fication of the bivariate density requires knowledge of both the marginal
densities and the dependence function that links them together.
Variation of the dependence structure through time can be modelled by
means of the conditional copula, which was introduced by Patton (2003,
2004) and recently extended by Fermanian and Wegkamp (2004). In the
present paper, time variation in the copula is modelled by allowing the de-
pendence parameter to evolve through time according to a particular equa-
tion. The forcing variables in this equation are the conditional volatilities
of the underlying assets. These are also the forcing variables that are typi-
cally chosen to model time-varying correlations; see, e.g., the BEKK model
introduced by Engle and Kroner (1995). Additional motivation is provided
by the evidence on correlation breakdowns, which suggests that financial
markets exhibit high dependence in periods of high volatility. Patton (2003)
proposes an ARMA-type process linking the dependence parameter to ab-
solute differences in return innovations, which is another way to capture the
same idea.
To be more specific, let τt be Kendall’s measure of association at time
t, and let hi,t be the objective conditional variance estimate at time t of
underlying index return i ∈ {1, 2} implied by Duan’s GARCH option pricing
model. It is assumed that
τt = γ(h1,t , h2,t ) (2)
for some function γ(·, ·) to be specified later. This conditional measure of
association governs the degree of dependence for the risk-neutral copula
under consideration.
The proposed valuation scheme is implemented using Monte Carlo simu-
lations. Pairs of random variates are drawn from the copula implied by the
estimated conditional risk-neutral measure of association, which are then
transformed to return innovations using Duan’s GARCH model. Subse-
quently, the payoffs implied by these innovations are averaged and dis-
counted at the risk-free rate. The result then constitutes the fair value
of the option. Algorithms for random variable generation from the non-
normal copulas are given in Genest and MacKay (1986), Genest (1987),

8
Ghoudi, Khoudraji and Rivest (1998), and Nelsen (1999). For the normal
copula, a straightforward Cholesky decomposition may be used.

3 Pricing options on two equity indexes

The dynamic-dependence valuation scheme outlined in Section 2 is applied
to better-of-two-markets and worse-of-two-markets options on the S&P500
and Nasdaq indexes. A sample consisting of pairs of daily returns on the
S&P500 and Nasdaq from January 1, 1993 to August 30, 2002 was obtained
from Datastream. The sample size is T = 2422. The maximum likelihood
estimates of the GARCH parameters for the marginal index return processes
may be found in Table I. These estimates are based upon an initial variance
level hi,0 = ωi /(1 − βi − αi ). The values for α and β nearly add up to one.
These estimates are in line with previously reported values.
Figure 1 depictsp the time series
p
of the estimated standardized GARCH
innovations (η1,t+1 / h1,t , η2,t+1 / h2,t ) for the last 250 trading days in the
sample. (For clarity, the picture is restricted to a subsample; other episodes
show a similar pattern.) Note that outliers typically occur simultaneously
and in the same direction. This positive dependence between the two series
is even more apparent from Figure 2, which displays the support set of the
empirical copula of the standardized return innovations. This scatter plot
consists of the observed pairs of ranks (divided by T + 1) for the estimated
standardized GARCH innovations of the two markets. In case of static
dependence, the empirical copula estimates the copula function; under reg-
ularity conditions, the empirical copula function converges to the true (here,
objective) copula function, see Van der Vaart and Wellner (1996). In case of
time-varying dependence, it is an estimate of the copula of the unconditional
distribution. Notice the pronounced positive dependence, particularly in the
tails. The sample version of Kendall’s tau for the entire sample amounts to
0.60, confirming positive dependence. Figure 3 gives an impression of how
this dependence measure of the standardized return innovations evolves over
time. It shows rolling-window estimates of Kendall’s tau using window sizes
of two months, i.e., Kendall’s tau at day t is computed using the 20 trading
days prior to day t, day t itself, and the 20 trading days after day t. While
the estimates show considerable variation, a slight upward trend over the
sample period is discernable.
The time variation in the copula is governed by Equation (2). It models
the dependence measure as a function of the conditional volatilities of the

9
index returns. The following specification of this function is proposed:

γ(h1 , h2 ) = γ0 + γ1 log max(h1 , h2 ). (3)

To motivate this specification, recall that the evidence on correlation break-

downs suggests that increased dependence occurs in hectic periods. Hence,
theory predicts a positive value of γ1 . The maximum operator reflects that
hectic periods in either market may cause dependence to go up. Since volatil-
ity in both markets is highly dependent, the actual specification is likely not
to affect the results in the present section too much. The parameters γ 0
and γ1 were estimated by regressing the non-parametric (rolling-window)
estimates of Kendall’s tau from Figure 3 on the estimated log maximum
conditional volatility. The regression is illustrated in Figure 4. The slope
coefficient, γ1 , was estimated at 0.063; positive, as expected. The estimated
dependence measure implied by these parameter estimates,

γ(h1,t , h2,t ) = γ0 + γ1 log max(h1,t , h2,t ),

was then used to fix the conditional risk-neutral copula at time t.

Return innovations were sampled from this conditional copula to com-
pute the price of the option. In total, the Monte Carlo study was based on
100, 000 replications, leading to simulation errors in the order of magnitude
of 1 basis point for one-month maturity claims. The option prices turned
out to be robust with respect to the choice of the window used to calculate
Kendall’s tau. Halving or doubling the window size only led to slight price
changes in the order of magnitude of the simulation errors.
Clearly, the option price depends on the initial levels of volatility of
the underlyings. Prices for three levels of initial volatility were computed:
low, medium, and high volatility, where medium volatility is defined as the
estimated unconditional variance ω/(1 − β − α), and low and high volatility
are one-fourth of and four times this amount, respectively. Furthermore,
different maturities were considered, ranging from one day to one month
(i.e., 20 trading days). The strike price was set at levels between 0.98 and
1.02. Finally, the risk-free rate was assumed to be 4 percent per annum.
The results show that allowing for time varying dependence leads to
different option prices than under static dependence, in particular in times
of high volatility. This is illustrated in Figure 5 which displays, for various
copula parametrizations, the price (measured in basis points) of a one-month
put-on-max option as a function of the exercise price implied by dynamic
dependence, and compares it to the option price under three levels of static
dependence: low, medium, and high static dependence. The medium level

10
of dependence is equal to the average measure of dependence found in the
sample, 0.60; the low and high levels are 0.40 and 0.80, respectively. Note
that a static model for the dependence structure, which uses the sample
measure of dependence of 0.60, underestimates the option price generated by
the dynamic model considerably for all copula parametrizations and over the
entire range of strike prices considered. The difference is significant since the
95% confidence intervals of the price estimates do not overlap. In the interest
of clarity, confidence intervals are not displayed here, but available from the
authors upon request. Note that the prices implied by dynamic copulas are
between the high and the medium static-dependence prices, suggesting that
the dynamic model implies a dependence that is on average stronger than in
the medium static-dependence case. Interestingly, price differences between
the dynamic and static model vanish as initial volatilities are at a medium
level; see Figure 6. The same holds for low initial volatilities (not shown),
again, across a broad range of copula families and strike prices.
It is also interesting to compare option prices produced by different dy-
namic copula families. It turns out that prices implied by the normal cop-
ula deviate substantially from prices implied by the other copula families.
Outside the normal class, the copula choice appears to be irrelevant. This
suggests that unless the dependence between index returns can be described
by a normal model, alternative specifications should be considered. These
findings are illustrated in Figures 7 and 8 which depict dynamic-dependence
one-month call-on-max and put-on-min option prices respectively, as a func-
tion of their strike under medium initial volatilities. The prices implied by
the normal copula are significantly lower than the prices implied by the other
copulas across the whole range of strike prices. The effect is there at other
maturities as well. The difference between normal and non-normal prices is
also found for high and low initial volatility levels. The differences are less
significant for call-on-min and put-on-max options.

4 Conclusions
This paper studies the relation between multivariate options prices and the
dependence structure of the underlying assets. A copula-based model was
proposed for the valuation of claims on multiple assets. A novel feature of
the proposed model is that, contrary to earlier works on multivariate option
pricing, the dependence structure is not taken as fixed, but rather as poten-
tially varying with time. The time variation in the dependence structure was
modelled using various parametric copulas by letting the copula parameter

11
depend on the conditional volatilities of the underlyings.
This dynamic copula model was applied to better- and worse-of-two-
markets options on the S&P500 and Nasdaq indexes for a variety of copula
parametrizations. Option prices implied by the dynamic model turned out
to differ substantially from prices implied by a model that fixes the depen-
dence between the underlying indexes, especially in high-volatility market
conditions. Hence, the application suggests that time variation in the de-
pendence between the S&P500 and the Nasdaq is important for the price
of options on these indexes. A comparison of option prices computed from
different copula families shows that the normal family produces prices that
differ significantly from the ones implied by the non-normal alternatives.
These findings suggests that if the dependence between the index returns
is not well represented by a normal copula, alternative copulas need to be
considered. The empirical relevance of such alternatives is apparent given
the evidence of non-normality in financial markets.

12
A One-Parameter Copula Families
The table below displays several one-parameter copula families.

(eθu −1)(eθv −1)

n o
1
Frank Cθ (u, v) = θ log 1 + eθ −1

  1 
θ
Gumbel–Hougaard Cθ (u, v) = exp − | log u|θ + | log v|θ

√
1+(θ−1)(u+v)− [1+(θ−1)(u+v)]2 −4uvθ(θ−1)
Plackett Cθ (u, v) = 2(θ−1)

 1 
θ
Galambos Cθ (u, v) = uv exp | log u|θ + | log v|θ

Normal Cθ (u, v) = Nθ (Φ−1 (u), Φ−1 (v))

Note: Φ is the standard (univariate) normal distribution function, and Nθ denotes the
standard bivariate normal distribution function with correlation coefficient θ.

B Kendall’s tau
The table below provides expressions—closed-form if available—of the re-
lation between Kendall’s tau and the dependence parameter for the copula
families considered in Appendix A.

Frank τ (θ) = 1 − 4 {D1 (−θ) − 1} /θ

Gumbel-Hougaard τ (θ) = 1 − 1/θ

R 1R 1
Plackett τ (θ) = 4 0 0 Cθ (u, v)dCθ (u, v) − 1
 −1
θ+1 R 1 1 1
Galambos τ (θ) = θ 0 t1/θ
+ (1−t)1/θ
−1 dt

2
Normal τ (θ) = π arcsin θ
Rθ t
Note: D1 denote the first-order Debye function, D1 (−θ) = 1
θ 0 et −1
dt + θ2 .

13
C Sample version of Kendall’s tau
Let {(X1 , Y1 ), . . . , (Xn , Yn )} be a random sample of n observations from a
vector (X, Y ) of continuous random variables. Two distinct pairs (Xi , Yi )
and (Xj , Yj ) are said to be concordant if (Xi − Xj )(Yi − Yj ) > 0, and
discordant if (Xi − Xj )(Yi − Yj ) < 0. Kendall’s tau for the sample is then
defined as t = (c − d)/(c + d), where c denotes the number of concordant
pairs, and d is the number of discordant pairs.

14
References
Ait-Sahalia, Y. and Lo, A. W.: 1998, Nonparametric estimation of state-
price densities implicit in financial asset prices, Journal of Finance
53(2), 499–547.
Black, F. and Scholes, M. S.: 1973, The pricing of options and corporate
liabilities, Journal of Political Economy 81, 637–654.
Bollerslev, T.: 1986, Generalized autoregressive conditional heteroskedas-
ticity, Journal of Econometrics 31, 307–327.
Boyer, B. H., Gibson, M. S. and Loretan, M.: 1999, Pitfalls in tests for
changes in correlations. International Finance Discussion Papers, No.
597, Board of Governors of the Federal Reserve System.
Cherubini, U. and Luciano, E.: 2002, Bivariate option pricing with copulas,
Applied Mathematical Finance 9(2), 69–86.
Comte, F. and Lieberman, O.: 2003, Asymtotic theory for multivariate
Garch processes, Journal of Multivariate Analysis 84(1), 61–84.
Cox, J. C., Ross, S. A. and Rubinstein, M.: 1979, Option pricing: A simpli-
fied approach, Journal of Financial Economics 7, 229–263.
Duan, J.-C.: 1995, The Garch option pricing model, Mathematical Finance
5(1), 13–32.
Embrechts, P., McNeil, A. J. and Straumann, D.: 2002, Correlation and
dependence in risk management: properties and pitfalls, in M. A. H.
Dempster (ed.), Risk Management: Value at Risk and Beyond, Cam-
bridge University Press, Cambridge, England, pp. 176–223.
Engle, R. F. and Kroner, K. F.: 1995, Multivariate simultaneous generalized
ARCH, Econometric Theory 11, 122–150.
Fermanian, J.-D. and Wegkamp, M.: 2004, Time-dependent copulas.
CREST Working Paper.
Genest, C.: 1987, Frank’s family of bivariate distributions, Biometrika
74, 549–555.
Genest, C. and MacKay, R. J.: 1986, Copules archimédiennes et familles
de lois bidimensionnelles dont les marges sont données, The Canadian
Journal of Statistics 14(2), 145–159.

15
Ghoudi, K., Khoudraji, A. and Rivest, L.-P.: 1998, Propriétés statistiques
des copules de valeurs extrêmes bidimensionnelles, The Canadian Jour-
nal of Statistics 26(1), 187–197.
Johnson, H.: 1987, Options on the maximum or the minimum of several
assets, Journal of Financial and Quantitative Analysis 22(3), 277–283.
Margrabe, W.: 1978, The value of an option to exchange one asset for
another, Journal of Finance 33(1), 177–186.
Nelsen, R. B.: 1999, An Introduction to Copulas, Lecture Notes in Statistics
139, Springer-Verlag, New York.
Patton, A. J.: 2003, Modelling asymmetric exchange rate dependence. Uni-
versity of California, San Diego, Discussion Paper 01-09.
Patton, A. J.: 2004, On the out-of-sample importance of skewness and asym-
metric dependence for asset allocation, Journal of Financial Economet-
rics 2(1), 130–168.
Reiner, E.: 1992, Quanto mechanics, From Black-Scholes to Black Holes:
New Frontiers in Options, RISK Books, London, pp. 147–154.
Rosenberg, J. V.: 1998, Pricing multivariate contingent claims using esti-
mated risk-neutral density functions, Journal of International Money
and Finance 17, 229–247.
Rosenberg, J. V.: 1999, Semiparametric pricing of multivariate contingent
claims. NYU, Stern School of Business, Working Paper S-99-35.
Rosenberg, J. V.: 2003, Nonparametric pricing of multivariate contingent
claims, Journal of Derivatives 10(3).
Shimko, D. C.: 1994, Options on futures spreads: Hedging, speculation, and
valuation, Journal of Futures Markets 14, 183–213.
Sklar, A.: 1959, Fonctions de répartition à n dimensions et leurs marges,
Publ. Inst. Statist. Univ. Paris 8, 229–231.
Stulz, R. M.: 1982, Options on the minimum or the maximum of two
risky assets: Analysis and applications, Journal of Financial Economics
10, 161–185.
Van der Vaart, A. W. and Wellner, J. A.: 1996, Weak Convergence and
Empirical Processes, Springer-Verlag, New York.

16
Table I: Maximum likelihood estimates of the GARCH parameters for the
marginal index return processes. Figures in brackets are robust quasi-
maximum likelihood standard errors.

Parameter S&P500 Nasdaq

µ × 102 0.0674 (0.0168) 0.0812 (0.0246)
ω × 105 0.0680 (0.0398) 0.1895 (0.0987)
β 0.9258 (0.0220) 0.8906 (0.0309)
α 0.0680 (0.0198) 0.1015 (0.0288)

S&P500
5

−5
Q4−01 Q1−02 Q2−02 Q3−02

Nasdaq
5

−5
Q4−01 Q1−02 Q2−02 Q3−02

Figure 1: Daily standardized GARCH innovations for S&P500 and Nasdaq

for the last 250 trading days in the sample.

17
 1

0.9

0.8

0.7

Nasdaq innovations
0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
S&P500 innovations

Figure 2: Support set of the empirical copula of the standardized GARCH

innovations.

0.9

0.8

0.7
Kendall’s tau

0.6

0.5

0.4

0.3

0.2
1992 1994 1996 1998 2000 2002 2004

Figure 3: Rolling-window estimates of Kendall’s tau for the standardized

return innovations using a window size of 41 trading days.

18
 1

0.9

0.8

0.7
Kendall’s tau

0.6

0.5

0.4

0.3

0.2
−10.5 −10 −9.5 −9 −8.5 −8 −7.5 −7 −6.5 −6 −5.5
Log maximum volatility

Figure 4: Regression of rolling-window estimates of Kendall’s tau for the

standardized return innovations on the logarithm of the maximum return
volatility.

19
 normal

400 static
low
high
300 dyna.

200

0.98 0.99 1 1.01 1.02

frank gumhou

400 static 400 static

low low
high high
300 dyna. 300 dyna.

200 200

0.98 0.99 1 1.01 1.02 0.98 0.99 1 1.01 1.02

plackett galambos

400 static 400 static

low low
high high
300 dyna. 300 dyna.

200 200

0.98 0.99 1 1.01 1.02 0.98 0.99 1 1.01 1.02

Figure 5: One-month maturity put-on-max prices in basis points as a func-

tion of the strike under high initial volatilities for dynamic dependence and
for low, medium, and high static dependence for various copulas.

20
 normal
250
static
200 low
high
150 dyna.

100

50
0.98 0.99 1 1.01 1.02
frank gumhou
250 250
static static
200 low 200 low
high high
150 dyna. 150 dyna.

100 100

50 50
0.98 0.99 1 1.01 1.02 0.98 0.99 1 1.01 1.02
plackett galambos
250 250
static static
200 low 200 low
high high
150 dyna. 150 dyna.

100 100

50 50
0.98 0.99 1 1.01 1.02 0.98 0.99 1 1.01 1.02

Figure 6: One-month maturity put-on-max prices in basis points as a func-

tion of the strike under medium initial volatilities for dynamic dependence
and for low, medium, and high static dependence for various copulas.

21
 500
normal
frank
gumhou
450 plackett
galambos

400

350

300

250

200

150
0.98 0.99 1 1.01 1.02 1.03 1.04

Figure 7: One-month call-on-max prices in basis points as a function of the

strike under dynamic dependence and medium initial volatilities for various
copula models.
500
normal
frank
gumhou
450 plackett
galambos

400

350

300

250

200

150
0.98 0.99 1 1.01 1.02 1.03 1.04

Figure 8: One-month put-on-min prices in basis points as a function of the

strike under dynamic dependence and medium initial volatilities for various
copula models.

22