RJ 2021 060

C ONTRIBUTED R ESEARCH A RTICLES 467
gofCopula: Goodness-of-Fit Tests for

Copulae
by Ostap Okhrin, Simon Trimborn and Martin Waltz
Abstract The last decades show an increased interest in modeling various types of data through
copulae. Different copula models have been developed, which lead to the challenge of finding the
best fitting model for a particular dataset. From the other side, a strand of literature developed a list
of different Goodness-of-Fit (GoF) tests with different powers under different conditions. The usual
practice is the selection of the best copula via the p-value of the GoF test. Although this method is not
purely correct due to the fact that non-rejection does not imply acception, this strategy is favored by
practitioners. Unfortunately, different GoF tests often provide contradicting outputs. The proposed
R-package brings under one umbrella 13 most used copulae - plus their rotated variants - together
with 16 GoF tests and a hybrid one. The package offers flexible margin modeling, automatized
parallelization, parameter estimation, as well as a user-friendly interface, and pleasant visualizations
of the results. To illustrate the functionality of the package, two exemplary applications are provided.
Introduction
Being firstly introduced in 1959 by Abe Sklar (see Sklar (1959)), copulae gained enormous popularity in
applications in the last two decades. Researchers from different fields recognize the power of copulae
while working with multivariate datasets from insurance (Fang and Madsen, 2013; Shi et al., 2016),
finance (Salvatierra and Patton, 2015; Oh and Patton, 2018), biology (Konigorski et al., 2014; Dokuzoğlu
and Purutçuoğlu, 2017), hydrology (Liu et al., 2018; Valle and Kaplan, 2019), medicine (Kuss et al., 2014;
Gomes et al., 2019), traffic engineering, (Huang et al., 2017; Ma et al., 2017), etc. For a recent review,
we refer to Größer and Okhrin (2021). Unfortunately, the correct specification of the multivariate
distribution is not easy to find, and often interest in the understanding of the functional form of the
copula is dominated by the expected performance of the whole model. This is natural, taking into
account the huge list of different copula models proposed in the literature for different needs; see, e.g.,
Durante and Sempi (2010), Joe and Kurowicka (2010), or Genest and Nešlehová (2014). Although an
expanding list of R-packages devoted to copulae is existent, the issue of GoF testing is less frequently
addressed. Primarily, GoF tests for copulae are implemented in copula comparison packages as copula
(Hofert et al., 2020), TwoCop (Remillard and Plante, 2012), and VineCopula (Nagler et al., 2019), but
since Remillard and Scaillet (2009) and Genest et al. (2009), many other powerful tests were developed
that are not integrated into these packages. Most of the tests focus on the bivariate case, leaving a
further gap in the existing R-package landscape.
Given a variety of tests, the selection of the most appropriate copula seems simple at first glance.
However, the power of these tests varies significantly depending on the use case and the copula
tested for. The absence of the overall best GoF test leads researchers and practitioners often to the
selection of the test (and copula), which supports some subjective expectation, but not the copula that
fits the data at its best. Although GoF tests are not intended for model selection but rather to decide
whether the selected copula is not suitable for the data, the model selection strategy based on the rank
of the p-values is still commonly used. Following proper scoring rules (Gneiting and Raftery, 2007),
some tests still allow for selection, and even if not purely statistically sound, it is heavily advocated
among practitioners; see De Valpine (2014). An eloquent illustrative example of different powers
and contradictory decisions was provided in Zhang et al. (2016), where three different tests (Rn by
Zhang et al. (2016), Sn by Genest et al. (2009), and Jn by Scaillet (2007)) were applied for testing the
dependency between the standardized returns of the Bank of America and Citigroup. The model
selection was done from three copula models: normal, Gumbel, and t-copula, based on their respective
p-values. Interestingly,
1) for the year 2004, the Rn gave a favor for the Gumbel copula, while the Sn and Jn for the normal
one;
2) for the year 2006, the Sn gave the favor for the normal copula, while Rn and Jn for the t-copula;
3) for the year 2009, the Jn indicated that the dependency is close to the normal one, while Rn and
Sn were in favor of the Gumbel copula.
This implies that for each year, a different pair of tests returns consistent results. In an empirical study,
it is difficult to decide which copula is suitable and which test provides the most plausible results. An
extensive comparative study of different GoF tests was performed a decade ago by Genest et al. (2009),
intensively discussing all, up to that moment existing, tests for copulae. The main findings are that
The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

there is no superior blanket test, but several tests have very good power under different, often disjunct
conditions. A test proposed by Zhang et al. (2016) fills some gaps in the set of models under which
this test performs better than others under certain conditions. However, a common phenomenon in
empirical studies is the interpretation of the non-rejection of a copula as the correct model. Especially,
in situations where the used GoF test has low power, this is not necessarily the case. Tackling this issue,
Zhang et al. (2016) also developed the hybrid test, which is simple in construction and implementation.
It combines the power of different tests and is very helpful for practitioners; see Section 2.6. However,
even in this case, the interpretation of finding the correct copula should be treated with care.
We propose the R-package gofCopula to automatize the whole empirical procedure of selecting the
most suitable copula. Table 1 displays the broad range of available tests, copula models, and the maxi-
mum dimension. The latest version of this table is also accessible via the function CopulaTestTable()
in the package. Further details on the functionality of each test are provided in Section 2.3, while Table
A.1 of Appendix A contains some characteristics of the copulae implemented in the package.
huslerReiss
galambos
plackett
clayton
normal
gumbel
frank
tawn
joe
amh
tev
fgm
Test
t
gofCvM ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 2 2 2 2 2
gofKS ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 2 2 2 2 2
gofKendallCvM ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 2 2 2 2 2
gofKendallKS ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 2 2 2 2 2
gofRosenblattSnB ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 - - - 2 2
gofRosenblattSnC ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 - - - 2 2
gofRosenblattGamma ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 - - - 2 2
gofRosenblattChisq ≥2 ≥2 ≥2 ≥2 ≥2 ≥2 2 2 - - - 2 2
gofArchmSnB - - ≥2 ≥2 ≥2 ≥2 2 - - - - - -
gofArchmSnC - - ≥2 ≥2 ≥2 ≥2 2 - - - - - -
gofArchmGamma - - ≥2 ≥2 ≥2 ≥2 2 - - - - - -
gofArchmChisq - - ≥2 ≥2 ≥2 ≥2 2 - - - - - -
gofKernel 2 2 2 2 2 2 2 2 2 2 2 2 2
gofWhite 2 2 2 2 2 2 - - - - - - -
gofPIOSTn 3 2 3 3 3 3 2 2 - - - 2 2
gofPIOSRn 3 2 3 3 3 3 2 2 - - - 2 2
Table 1: Implemented tests, copula models (columns), and the maximum available dimension of each
test-copula combination. "-" means this combination is not available, "2" is available in dimension two,
"3" in dimensions two and three, and "≥ 2" in any dimension. amh corresponds to the Ali-Mikhail-Haq
copula, tev to the t-extreme value copula, and fgm to the Farlie-Gumbel-Morgenstern copula.
In summary, the package gofCopula offers the following attractive features which distinguish it
from other R-packages:
• Each of the 13 copulae in Table 1 is available in a rotated form for the bivariate case. Furthermore,
the flexible hybrid test is implemented to aggregate the results of the 16 tests.
• We provide an interface to integrate new GoF tests. The users can provide their own test
statistics and perform the tests with the integrated parametric bootstrap and also make use of
the automatized parallelization of gofCopula. The new tests can be further combined with
other tests via the hybrid test.
• The whole copula community relies justifiably on the R-package copula for conducting different
studies on copulae. Thus we provide an interface to use objects from the R-package copula and
perform the GoF tests with gofCopula.
• For the estimation of the margins, ten different parametric distributions are available, in addition
to the nonparametric estimation per default.
• GoF tests rely on bootstrapping methods which can result in substantially high computational
costs. In contrast to other R-packages, the package gofCopula comes with an integrated option
for automatized parallelization of the bootstrapping samples. For the convenience of the user,
the parallelization can be activated by specifying the number of parallel jobs as an argument of
the functions.

• As we believe in reproducible science, the user has the opportunity to specify the seeds for the
bootstrapping procedure in order to guarantee full reproducibility of all results gained from the
package gofCopula.
• An estimation function for the computation time is implemented, which fits a regression model
to give an estimate for the time adapted to the users’ machine.
• An informative console output is implemented, which keeps the user informed about the current
test and copula under estimation, as well as the remaining time until the derivation of the test
is performed. The latter functionality is supported by the R-package progress (Csárdi and
FitzJohn, 2019).
• The results of the tests are provided with the package’s own class "gofCOP", which allows
for a comprehensive overview of the test results. For a better comparison of the results, we
extend the generic plot function for objects of class "gofCOP", which illustrates the results in a
convenient manner. The plot function is supported by the R-package yarrr (Phillips, 2017), and
was customized to provide the user an insightful figure for the interpretation of the results.
The test statistics of six GoF tests were already implemented in R-packages. Thus, for the com-
putation of the test statistics of some tests, we use the functions gofTstat and BiCopGofTest from
the packages copula and VineCopula, respectively. For obvious reasons, we did not implement all
existing tests on copulae, but we will embed new tests in the proposed package as soon as they become
more relevant and actively used among academics and practitioners.
The paper, introducing the R-package gofCopula, is structured as follows: The tests and method-
ology implemented in the package are introduced in Sections 2.2 and 2.3 before presenting the
functionalities of the package in Section 2.4. We explain major functions, how to apply them, and elab-
orate the main arguments of each function. The explanations are supported by R-code and output. To
provide an impression of the runtime of various tests, we discuss the speed of the tests depending on
the copula to test for, the number of observations, and the number of bootstrap samples. A simulated
example (Section 2.5) contains a typical step-by-step procedure of how the package can be used in
practice, which is also applied to two real-world examples (Section 2.6), in which all corresponding
codes are given and explained. The cases are illustrated with interpretations of the console output and
plots, both generated directly from gofCopula, without any additional code. The results of the two
applications can be fully reproduced by the gofCopula package, which also contains the used datasets.
All illustrations, simulations, and applications in this paper are fully reproducible and designed to
guide the user into conducting their own research with the gofCopula package.
Estimation methods
Consider a d-dimensional random vector X = { X1 , . . . , Xd } with corresponding marginal distributions

Fj , j = 1, . . . , d. The multivariate distribution F ( x1 , . . . , xd ) can be decomposed via the copula function
Cθ (u1 , . . . , ud ) as
( X1 , . . . , Xd ) ∼ F ( x1 , . . . , xd ) = Cθ { F1 ( x1 ), . . . , Fd ( xd )}.
Having a sample X = { xij }, i = 1, . . . , n, j = 1, . . . , d of size n with the columns defined as x j , the main
task for empirical studies is to estimate the copula parameter θ and the margins Fj , j = 1, . . . , d for a
given copula specification. Since the properties and goodness of the estimator of θ heavily depend on
the estimators of the latter, their choice is also of importance.
In the bivariate case, one of the standard methods of estimation of the univariate parameter θ is
′ ′
based on Kendall’s τ by Genest and Rivest (1993). Following Joe (1997) for ( X1 , X2 ), and ( X1 , X2 )
being independent random pairs with continuous distribution F, Kendall’s τ is defined via:
′ ′ ′ ′
τ = P{( X1 − X1 )( X2 − X2 ) > 0} − P{( X1 − X1 )( X2 − X2 ) < 0}.
R
This can be written in terms of the underlying copula C in form of τ = 4 CdC − 1, linking Kendall’s
τ and the copula parameter of interest under a correct copula specification, e.g., for the normal copula,
the equality τ = π2 arcsin θ holds, with θ being the copula correlation, c.f. Demarta and McNeil (2005).
Thus, the parameter θ can be estimated via inversion of this relation and replacement of τ by its
empirical counterpart. However, as shown in Genest et al. (1995), the ML method leads to substantially
more efficient estimators. Therefore, we employ it as the first option in our study. The second reason
for ML estimation is the fact that several implemented tests are based on the likelihood ratios. Thus,
results based on Kendall’s τ will not be supported by the theory behind these tests. The ML procedure

can be performed for the parameters of the margins and of the copula function simultaneously:
(θ̂, α̂1 , . . . , α̂d )⊤ = argmax L(X , θ, α1 , . . . , αd ), (1)

θ,α1 ,...,αd
 
n d
with L(X , θ, α1 , . . . , αd ) = ∑ log c{ F1 (xi1 , α1 ), . . . , Fd (xid , αd ), θ } ∏ f j (xij , α j ) ,
i =1 j =1
(2)
where c(·) is the copula density, α1 , . . . , αd are parameters of the margins, f j (·) are marginal densities,
and L(·) is the full log-likelihood function. Nevertheless, simultaneous maximization of the function
in (1) is very computationally intensive. Therefore, we consider only two-stage procedures, where at
the first stage, we estimate margins parametrically (c.f. Joe (1997) and Joe (2005)) as
( )
n
Fj ( x, α̂ j ) = Fj x, argmax ∑ log f j ( xij , α) , for j = 1, . . . , d, (3)
α i =1
or nonparametrically (c.f. Chen and Fan (2006) and Chen et al. (2006)) as
n
Fbj ( x ) = (n + 1)−1 ∑ 1{xij ≤ x}, j = 1, . . . , d, (4)
i =1
with 1 being the indicator function. Afterward, the copula parameter is estimated in the second step
as
n
θ̂ = argmax ∑ log c Fe1 ( xi1 ), . . . , Fed ( xid ), θ ,

(5)
θ i =1
where Fe( x ) ∈ { Fb( x ), F ( x, α̂)} are parametrically or nonparametrically estimated margins. In the
case of parametric margins, one shall be aware that the two-step approach does not lead to efficient
estimators, though the loss in the efficiency is moderate and mainly depends on the strength of
dependencies (Joe, 1997). The method of nonparametric estimation of the marginal distributions for
copula estimation was first used in Oakes (1994) and further investigated in Genest et al. (1995) and
Shih and Louis (1995).
Furthermore, Fermanian and Scaillet (2003) and Chen and Huang (2007) consider a fully non-
parametric estimation of the copula, which is heavily used in the GoF testing. It is called an empirical
copula and is shown to be a consistent estimator of the true underlying copula, c.f. Gaensler and Stute
(1987) and Radulovic and Wegkamp (2004). This estimator is defined as
n d
Cn (u1 , . . . , ud ) = n−1 ∑ ∏ 1{ Fbj (xij ) ≤ u j }.
i =1 j =1
Goodness-of-fit tests for copulae
Having a list of different copulae, the most suitable one for real applications still needs to be found
and motivated. For this purpose, a series of different GoF tests has been developed in the last decades.
Several authors, e.g., Genest et al. (2009), tested the power of those tests against each other and showed
that no superior test for all possible situations exist. We cover 16 tests and implement them into
the gofCopula package. Most of these tests work with the parametric family of copulae denoted by
C0 = {Cθ ; θ ∈ A ⊂ R p } for some integer p ≥ 1 and the copula C, under the general H0 -hypothesis:
H0 : C ∈ C0 .
We differentiate seven groups of GoF tests for copulae based on: (1) empirical copula process; (2)
Kendall’s process; (3) Rosenblatt integral transform; (4) transformation for Archimedean copulae; (5)
Kernel density; (6) White’s information matrix equality; and (7) pseudo in-and-out-of-sample (PIOS)
estimator.
Empirical copula process
The first group is based on the most natural approach: the deviation of the empirical √ copula Cn from
the parametric copula C (u1 , . . . , ud ; θ ), captured by the empirical copula process n{Cn (u1 , . . . , ud ) −
C (u1 , . . . , ud ; θ )}. Based on an estimation of the parametric copula C (u1 , . . . , ud ; θ̂ ), the following

process can be defined:

√
Cn ( u1 , . . . , u d ) = n{Cn (u1 , . . . , ud ) − C (u1 , . . . , ud ; θ̂ )}.
Different measures of divergence can be constructed to evaluate Cn ; see Fermanian (2005) and Genest
and Rèmillard (2008). We implemented two commonly applied approaches using the Cramér-von
Mises and Kolmogorov-Smirnov statistics:
Z
SnE = Cn (u1 , . . . , ud )2 dCn (u1 , . . . , ud ), TnE = sup |Cn (u1 , . . . , ud )|.
[0,1]d u1 ,...,ud ∈[0,1]d
Notice that the Cramér-von Mises statistic yields better performances in most cases (Genest et al.,
2009). The evaluation of the d-dimensional integral in practice uses numerical approximations, and
the test statistic SnE has been already implemented in the copula package as function gofTstat, so we
included it into our package. The tests are later denoted by gofCvM and gofKS, respectively.
Kendall’s process
The tests from the second group were developed and investigated by Genest and Rivest (1993), Wang
and Wells (2000), Genest et al. (2006). The main idea behind them is to use the copula-based random
variable:
C { F1 ( X1 ), . . . , Fd ( Xd ); θ } ∼ K (·, θ ), (6)
where K (·, θ ) is the univariate Kendall’s distribution (not uniform in general); see Barbe et al. (1996),
Jouini and Clemen (1996). The empirical version of K (·) is given through:
n h i
K n ( v ) = n −1 ∑1 Cn { Fb1 ( xi1 ), . . . , Fbd ( xid )} ≤ v , v ∈ [0, 1].
i =1
√
Based on the definition of Kendall’s process n{Kn (v) − K (·, θ )} and a parametric K (·, θ̂ ) estimated
with the parameter θ̂, we can define an empirical process as
√
Kn (v) = n{Kn (v) − K (v, θ̂ )}. (7)
On this basis, two applicable test statistics are Cramér-von Mises and Kolmogorov-Smirnov; see
Genest et al. (2006).
Z 1
(K ) (K )
Sn = Kn (v)2 dK (v, θ̂ ), Tn = sup |Kn (v)|.
0 v∈[0,1]
′′
Worth mentioning are the different null hypotheses H0 : K ∈ K0 = {K (·, θ ) : θ ∈ Θ} of these tests.
′′
Since H0 ⊂ H0 , the non-rejection of H0′′ does not imply non-rejection of H0 . However, for bivariate
′′
Archimedean copulae, H0 and H0 are equivalent (Genest et al., 2009). Both tests are later denoted as
gofKendallCvM and gofKendallKS, respectively.
Rosenblatt transform
Under the assumption of copula dependency, the conditional distribution of Ui given U1 , . . . , Ui−1 is
specified through:
Cd (ui |u1 , . . . , ui−1 ) = P{Ui ≤ ui |U1 = u1 . . . Ui−1 = ui−1 }

∂i−1 C (u1 , . . . , ui , 1, . . . , 1)/∂u1 . . . ∂ui−1
= .
∂ i −1 C ( u
1 , . . . , ui −1 , 1, . . . , 1) /∂u1 . . . ∂ui −1
The Rosenblatt transform (c.f. Rosenblatt, 1952) is then defined as follows.
Definition 1 Rosenblatt’s probability integral transform of a copula C is the mapping R : (0, 1)d → (0, 1)d ,
R(u1 , . . . , ud ) = (e1 , . . . , ed ) with e1 = u1 and ei = Cd (ui |u1 , . . . , ui−1 ), ∀i = 2, . . . , d.
If the copula is correctly specified, the variables (e1 , . . . , ed )⊤ resulting from the Rosenblatt transform
should be independent from each other and uniformly distributed. Therefore, the null hypothesis
H0 : C ∈ C0 is equivalent to
H0R : (e1 , . . . , ed )⊤ ∼ Π, (8)
where Π(u1 , . . . , ud ) = u1 · . . . · ud is the product (independence) copula.

Two different types of tests may be constructed using this property. In the first type, similar to
the previous two groups, we measure the deviation of the product copula of (e1 , . . . , ed )⊤ from the
corresponding empirical copula:
n d
Dn ( u 1 , . . . , u d ) = n − 1 ∑ ∏ 1{eij ≤ u j }.
i =1 j =1
Thus, following Genest et al. (2009), two Cramér-von Mises statistics result:
Z
( B)
Sn =n { Dn (u1 , . . . , ud ) − Π(u1 , . . . , ud )}2 du1 · · · dud ,
[0,1]d
Z
(C )
Sn =n { Dn (u1 , . . . , ud ) − Π(u1 , . . . , ud )}2 dDn (u1 , . . . , ud ).
[0,1]d
Since the H0 changed to H0R , the tests evaluate the difference of Dn (u) to the product copula. In the
package, these tests are defined as gofRosenblattSnB and gofRosenblattSnC, respectively.
The second type of test uses the fact that a specific combination of independent uniformly dis-
tributed random variables follows some known distribution. Based on this, two further Anderson-
Darling type tests were introduced by Breymann et al. (2003). By defining
 
 d 
Gi,Γ = Γd ∑ (− log eij ) ,
 j =1 
where Γd (·) is the Gamma distribution with shape d and scale 1 and
 
d
Gi,χ2 = χ2d  ∑ {Φ−1 (eij )}2  ,
j =1
where χ2d (·) is the Chi-squared distribution with d degrees of freedom and Φ being the standard
normal distribution. It results:
n
2i − 1
Tn = −n − ∑ n
[log G(i) + log{1 − G(n+1−i) }],
i =1
where G(i) is the i-th ordered observation of the Gi,Γ or Gi,χ2 . One should note that Anderson-
Darling type tests have almost no power and even do not capture the type 1 error (Dobrić and
Schmid, 2007), while the Cramér-von Mises tests behave much more satisfactory (Genest et al., 2009).
Furthermore, the basic assumption of uniformly distributed and independent observations after
applying the Rosenblatt transform is violated since those variables are not mutually independent and
only approximately uniform. The latter two tests are denoted in the package as gofRosenblattGamma
and gofRosenblattChisq, respectively, and are obtained via the function gofTstat from the package
copula.
Transformation for Archimedean copulae
Recently, Hering and Hofert (2015) proposed a procedure of GoF testing based on a transformation
similar to the one of Rosenblatt (1952) specifically designed for Archimedean copulae.
Definition 2 Hering and Hofert’s transformation of an Archimedean copula C of dimension d ≥ 2 with

d-monotone generator ψ and continuous Kendall distribution K is the mapping T : (0, 1)d → (0, 1)d ,
i
i
ψ −1 ( u )
T(u1 , . . . , ud ) = (v1 , . . . , vd ) with vi = ∑ik+=11 −1 k , ∀i = 1, . . . , d − 1 and vd = K {C (u1 , . . . , ud )}.
∑ k =1 ψ (uk )
Distribution function K is estimated empirically, and the variables (v1 , . . . , vd )⊤ are independent and
uniformly distributed if the copula is correctly specified. Note that this transformation was originally
considered in Wu et al. (2007) as a method for generating random numbers from Archimedean
copulae, such as the inverse of the Rosenblatt transform can be used for sampling copulae. Following
Hering and Hofert (2015), the main advantage of this approach in comparison to tests based on
the Rosenblatt transform is the more convenient computation in higher dimensions, in which the
Rosenblatt procedure is numerically challenging and unstable.
The null hypothesis equals (8) from the tests based on Rosenblatt’s probability integral transform:
H0T : (v1 , . . . , vd )⊤ ∼ Π with Π being the product copula. Consequently, the approaches to test it are

identical. In analogy to the naming introduced in Section 2.3.3, we denoted the tests as gofArchmSnB,
gofArchmSnC, gofArchmGamma, and gofArchmChisq in the package.
Kernel density estimator
A test from this group has been introduced by Scaillet (2007). Following his approach, a d-variate
quadratic kernel K with bandwidth H = 2.6073n−1/6 Σ b 1/2 is used, with Σ b being a sample covariance
matrix with Σ
b 1/2 its Cholesky decomposition. Using K H (y1 , . . . , yd ) = K( H −1 {y1 , . . . , yd }⊤ )/ det( H ),
the copula density is nonparametrically estimated by
n
ĉ(u1 , . . . , ud ) = n−1 ∑ K H [(u1 , . . . , ud )⊤ − { Fe1 (xi1 ), . . . , Fed (xid )}⊤ ],
i =1
where under Fei (·), we consider nonparametric as well as parametric estimators of the margins. The
test statistic is then:
Z
Jn = {ĉ(u1 , . . . , ud ) − K H ∗ c(u1 . . . , ud ; θ̂ )}2 w(u1 , . . . , ud )du1 · · · dud , (9)
[0,1]d
with “∗” being a convolution operator, w(u1 , . . . , ud ) a weight function, and c(u1 , . . . , ud ; θ̂ ) the cop-
ula density under the H0 , with estimated copula parameter θ̂. Note that the integral is computed
numerically using the Gauss-Legendre quadrature method; see Scaillet (2007). The number of knots
can be specified via the argument nodes.Integration. A scaling parameter for H is implemented
via delta.J, and the internal size of the bootstrapping samples can be controlled via MJ. This test is
denoted by gofKernel in the package.
White test
This test was introduced by Huang and Prokhorov (2014) and had its foundation in the information
matrix equality stated by White (1982). Given the presence of certain regularity conditions, the White
equality establishes a connection between the negative sensitivity matrix S(θ ), and the variability
matrix V(θ ) defined as
∂2

S(θ ) = − E0 log c { F1 ( x 1 ) , ..., Fd ( x d ) ; θ } ,
∂θ∂θ ⊤
⊤ !
∂ ∂
V(θ ) = E0 log c{ F1 ( x1 ), ..., Fd ( xd ); θ } log c{ F1 ( x1 ), ..., Fd ( xd ); θ } ,
∂θ ∂θ
where E0 is the expectation under correct model specification, which is represented by the null
hypothesis to be specified. The equality states:
S( θ ) = V( θ ).
Using this approach, the testing problem can be formulated as follows:
H0W : S(θ ) = V(θ ) vs. H1W : S(θ ) ̸= V(θ ).
Following Schepsmeier (2015), a test statistic is based on empirical versions of the two information
S(θ̂ ) and V
matrices, denoted by b S(θ̂ ) + V
b (θ̂ ). These are aggregated via d(θ̂ ) = vech{b b (θ̂ )} with vech
denoting vectorization of the lower triangular of a matrix. As a result, d(θ̂ ) is a vector of dimension
p ( p +1)
2 given the copula parameter vector is of dimension p. It can be shown that the constructed test
,
statistics:
TW = n{d(θ̂ )}⊤ Â−1 d(θ̂ ),

θ̂
√
with Â−1 being the sample estimator of the asymptotic covariance matrix of nd(θ̂ ), follows asymp-
θ̂
p ( p +1)
totically a χ2 distribution with degrees of freedom. For the derivation of the test statistic, this
2
test relies on the function BiCopGofTest from the VineCopula package, again, in order to avoid code
redundancy. Note that the implementation of the test can be unstable for the t-copula; see Nagler et al.
(2019). This is the reason why it could not be computed in some cases of the second empirical example
in Section 2.6.2. This test is called gofWhite in the package.

Cross-validated tests
A recent test using a leave-one-block strategy, and its approximation were introduced by Zhang et al.
(2016). Authors derive θ̂ as in (5) and compare it with θ̂−b , 1 ≤ b ≤ B, which are delete-one-block
pseudo ML estimates:
B m
θ̂−b = argmax ∑ ∑ log c{ Fe1 (xi1 ), . . . , Fed (xid ); θ }, b = 1, . . . , B,
θ ∈Θ ′
b ̸ = b i =1
where B is the number of non-overlapping blocks and m the length of each block. Note that in the
general setting, these blocks need not be of the same size. However, we follow here the approach of
Zhang et al. (2016), who restrict themselves to the same length case of each block. This assumption
also simplifies the usage in terms of many parameters. The resulting test statistics,
" #
B m
c{ Fe1 ( xi1 ), . . . , Fed ( xid ); θ̂ }
Tn (m) = ∑ ∑ log , (10)
i =1
b =1 c{ Fe1 ( xi1 ), . . . , Fed ( xid ); θ̂−b }
compares the full likelihood, “in-sample”, against the resulting likelihoods from the leave-one-block
out estimation, “out-of-sample”. If the data in each block significantly influence the estimation of the
copula parameter under the null hypothesis, then the chosen copula model is inadequate to represent
the data.
Depending on the number of blocks, B, a possibly huge amount of dependence parameter esti-
n
mations have to be performed to get (10). In the case of equal length of each block, [ m ] parameters
should be computed. To overcome this drawback, under suitable regularity conditions, Zhang et al.
(2016) proposed the test statistic asymptotically equivalent to (10):
S(θ̂ )−1 V
Rn = tr{b b (θ̂ )}. (11)
As we see, this result is very similar to the White (1982) test, but the power of the test is much higher.
Both exact and asymptotic test statistics are denoted in the package as gofPIOSTn and gofPIOSRn,
respectively.
Hybrid test
Many power studies including Genest et al. (2009) showed that no overall single optimal test exists for
testing for copula models. Zhang et al. (2016) introduced a Hybrid test to combine the testing power
of several tests. Having q different tests and the corresponding p-values, p(1) , . . . , p(q) , the combined
p-value is defined to be:
phybrid = min{q · min ( p(1) , . . . , p(q) ), 1}. (12)
In Zhang et al. (2016), it is shown that the consistency of (12) is ensured as long as at least one of the q
tests is consistent.
Bootstrapping test statistics
As the distribution of the test statistics is in most cases unknown, we perform a parametric bootstrap
to receive the p-values. The necessary steps are described as follows:
n o
(m)
Step 1. Generate bootstrap sample ϵi , i = 1, . . . , n from copula C (u1 , . . . , ud ; θ̂ ) under H0 with θ̂
and estimated marginal distributions Fe obtained from original data;
n o
(m)
Step 2. Based on ϵi , i = 1, . . . , n from Step 1, estimate θ of the copula under H0 and compute test
statistics under consideration, say ;
Step 3. Repeat M-times Steps (1. – 2.) and obtain M statistics Tnm , m = 1, . . . , M;
Step 4. Compute an empirical p-value as pe = M−1 ∑m M m
=1 1 {| Tn | ≥ | Tn |} with Tn being the test
statistics estimated from original dataset.
Depending on the different tests, variants of the described steps have to be performed. For example,
the Kernel density estimation test of Scaillet (2007) described in Section 2.3.5 relies on a double
bootstrapping procedure, in which for the computation of each test statistic, Tn and Tnm in the steps
above, an additional bootstrapping is utilized. Thus, the double bootstrapping approach consists of
one bootstrap to calculate the p-value from a given test statistic and a second bootstrap to calculate
the test statistic from an estimated copula. For further details, we refer to Scaillet (2007). Both
bootstrapping procedures can be controlled via the arguments M and MJ, respectively.

Functionality of the R-package
The core of the gofCopula package is the function gof, which computes different tests for different
copulae for a given dataset, based on the user’s choice.
R> library("gofCopula")
R> data("IndexReturns2D", package = "gofCopula")
R> system.time(result <- gof(IndexReturns2D, M = 100, seed.active = 1))
The margins will be estimated as: ranks

normal copula
Test gofCvM is running
Test gofKendallCvM is running
Test gofKendallKS is running
Test gofKernel is running
Test gofKS is running
t copula
clayton copula
gumbel copula
frank copula
joe copula
amh copula
The copula amh is excluded from the analysis since the parameters do not fit its
parameter space. See warnings and manual for more details.
galambos copula
Progress: [===>--------------------------------------] 15% | time left: 3s
...
user system elapsed

629.26 1.08 628.94

Warnings:
...
gof considers all 13 available copula models if no copulae or tests are specified. If a copula is unsuitable
in the sense that the estimated parameter is at the boundary of the parameter space, the copula is
automatically excluded, and the user is informed via a console statement (see above) and additional
warnings. In the given example, this is the case for the AMH, tawn, and FGM copulae because the
used IndexReturns2D dataset exhibits an estimated Kendall’s τ̂ = 0.611, which none of these three
copulae can model adequately; see Table A.1. The object result is of class "gofCOP" and has length 10,
which is the number of copulae used in testing (here: 13) minus the ones excluded during calculation
(here: 3). Following Table 1, five tests are available for all of these copula models in d = 2, and these
are used in the given function call.
If the user specifies copulae, tests, or both, the intersection of possible tests and copulae following
Table 1 is considered. For example, if copula = c("normal","tawn") is specified, the function
calculates the five tests which are implemented for both copulae (assuming d = 2). If, on the
other hand, tests = c("gofKernel","gofArchmSnB") is selected, the five Archimedean copulae
implemented for both tests are computed. In the case when both copulae and tests are defined, the
function provides results for the possible combinations. During the calculation, the user is informed
about the computation progress by statements about the running test and copula. Furthermore, a
progress bar indicates the percentage of progress for this specific test as well as a dynamically updated
estimated remaining time.
R> result$normal
$method
[1] "Parametric bootstrap goodness-of-fit test with hybrid test and normal copula"
$copula
[1] "normal"
$margins
[1] "ranks"
$param.margins
list()
$theta
[,1]
[1,] 0.8347428
$df
NULL
$res.tests
p.value test statistic
CvM 1.00 0.01520542
KendallCvM 0.41 0.06286712
KendallKS 0.11 0.80800000
Kernel 0.39 0.56012429
KS 1.00 0.31392428
hybrid(1, 2) 0.82 NA
hybrid(1, 3) 0.22 NA
...
The first element of result provides results for the normal copula. Note that in the field res.tests
the hybrid tests, starting after the individual ones, contain numbers in brackets indicating which tests
are considered for this hybrid. Thus, hybrid(1,2) means that this is the hybrid of CvM and KendallCvM
tests. The p-value 0.82 in testing for normality is obtained following formula (12) and therefore is
min{2 · min(1.00, 0.41), 1} = 0.82. To access the rotated versions of the copulae, one can set, for
example, copula = c("clayton","gumbel") together with flip = c(0,180), which would test for
the Clayton copula and the 180 degrees rotated Gumbel copula.

gofCOP class
Objects of class "gofCOP" are generated by the function gof or a single test function like, e.g., gofPIOSTn.
They consist of different sub-elements - one for each copula - as, e.g., result$normal in the given
example. These sub-elements are lists of length seven and contain the estimation and test results
for the specific copula. They present in the field method a description of the test scenario. The field
margins lists the defined marginal distribution that can also be a vector of distributions where each
element is applied to the respective data column, whereas param.margins returns the estimates of the
parameters of the marginal distributions if a parametric approach was specified. Field theta contains
the ML estimate of the copula parameter. In case of the t- and t-EV-copulae, the section df is the
estimated number of degrees of freedom for the copula. The values of these parameters are identical
for all the tests. In res.tests, the p-values and test statistics (only for individual tests) are given for
each of the executed tests. Each row corresponds to one test from the individual to the hybrid tests.
p-values of all the individual tests are computed via the bootstrap method described in Section 2.3.9.
The number of bootstrap samples M can be adjusted via the parameter M.
Plotting gofCOP class objects
"gofCOP" objects can be called by a generic plot function allowing the user to get the p-values of the
single, and the hybrid tests visualized in a pirateplot of the R-package yarrr. It enables the user to
select which copulae and hybrid testing sizes are desired for plotting. The remaining customization
options are equal to those of the function pirateplot from the package yarrr, except for the arguments
formula, data, sortx, xaxt, xlim, ylim, and ylab.
R> plot(result, copula = c("clayton", "joe", "plackett"), hybrid = c(1, 3, 5))

singleTests hyb3 hyb5
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
clayton
joe
plackett
clayton
joe
plackett
clayton
joe
plackett
Figure 1: Resulting p-values of different hybrid tests for the Clayton, Joe, and Plackett copula visualized
in a pirateplot.
Specifying hybrid = c(1,3,5) means that the p-values of the single tests (column singleTests
in Figure 1), the p-values of hybrid tests of size three (column hyb3), and size five (column hyb5)
should be plotted, separated by selected copulae. For example, we focus on the column hyb3 for the
Plackett copula. It contains information of all hybrid tests, which include three single tests for the
Plackett copula. In this case, we can see that the mean of these tests is approximately 0.76, as shown
by the thick horizontal line. All test p-values are shown by light-grey points in the column, indicating
the heterogeneity of the tests ranging from 0.6 to 1. Finally, the green bar around the mean line is a
Bayesian highest density interval, which provides the user, together with the shown density estimate
in the grey continuous lines, further information about the distribution of the p-values. For more
details on the pirateplot and its customization options, we refer to Phillips (2017).
Fire-and-forget
The R-package gofCopula includes all the discussed tests in Section 2.3. For each of the tests, a separate
function is implemented with a variety of arguments. We give shortly the most important arguments
all the tests share before we go into details about the structure of the package.
• copula: The copula to test for. Possible options depend on the test and dimension.

• x: A matrix containing the data with rows being observations and columns being variables.
• M (default: 1000): Number of bootstrapping loops.
• param (default: 0.5): The copula parameter to use if it shall not be estimated. In case of the
Gumbel copula, the default value is set to 1.5.
• param.est (default: TRUE): Boolean. TRUE means that param will be estimated.
• margins (default: ranks): Specifies which estimation method shall be used. The default ranks
stands for formula (4), which is the standard approach to convert data. Alternatively, the
following distributions can be specified: beta, cauchy, Chi-squared (chisq), f, gamma, Log
normal (lnorm), Normal (norm), t, weibull, Exponential (exp).
• flip (default: 0): The parameter to rotate the copula by 90, 180, 270 degrees clockwise. Only
applicable for bivariate copula.
• seed.active (default: NULL): Sets the seeds for the bootstrapping procedure. It has to be either
an integer or a vector of M+1 integers. If an integer is provided, the seeds for the bootstrap
samples will be simulated based on it. If M+1 seeds are given, these are used in the bootstrapping
procedure. In the default case (seed.active = NULL), R generates the seeds from the computer
runtime.
• processes (default: 1): The number of parallel processes which are performed to speed up the
bootstrapping. Should not be larger than the number of logical processors.
The package is coded as a fire-and-forget package. Each of the single tests just requires the input of
a dataset x and a copula to test for. All the other function parameters have reasonable default values
such that quick first results can be achieved easily. The calculation steps of each GoF test function are
the following:
• Estimation of margins: At first, the function transforms the data nonparametrically or parametri-
cally to U [0, 1]; see Section 2.2. This transformation is performed automatically, and a console
statement informs the user about the transformation.
• Estimation of copula: Afterward, the parameters of the copula model are estimated, so a two-stage
estimation is applied; see Section 2.2. Since a full ML estimation is computationally demanding,
canonical ML estimation or inference for margins is applied. In case the ML estimation fails, the
package automatically changes to inversion of Kendall’s tau (see Section 2.2), which guarantees
a result. The user is informed about that switch by a warning message.
• Bootstrapping: Following the estimation of the copula parameters, the bootstrapping procedure
will be performed, and the empirical p-value will be derived according to the test statistics in
Section 2.3.9. Since the bootstrapping procedure can require a long computational time, it can
pay out to parallelize the bootstrapping via the argument processes.
Hybrid testing and further functionality
Besides gof and the single tests, the package gofCopula offers additional functionality for the user.
Next to descriptions, illustrative examples are provided, assuming the following was called before-
hand:
R> data("IndexReturns2D", package = "gofCopula")
R> (res <- gof(copula = "normal", x = IndexReturns2D, M = 10, seed.active = 1,
+ tests = c("gofPIOSRn", "gofCvM", "gofKernel")))

normal copula
Test gofPIOSRn is running
-------------------------------------------------------------------------------
Parametric bootstrap goodness-of-fit test with hybrid test and normal copula
Parameters:
theta.1 = 0.834742824340301
Tests results:
PIOSRn 0.5 -0.11032857

Sn 1.0 0.01520542
Kernel 0.3 0.56012429
hybrid(1, 2) 1.0 NA
hybrid(1, 3) 0.6 NA
hybrid(2, 3) 0.6 NA
hybrid(1, 2, 3) 0.9 NA
Please use the functions gofGetHybrid() and gofOutputHybrid() for

display of subsets of Hybrid tests. To access the results, please
obtain them from the structure of the gofCOP object.
The functions are:

• gofCustomTest: This function has - next to the standard arguments listed in Section 2.4.3 -
the argument customTest, a character string referencing one customized test loaded in the
workspace. The function containing this test should have two arguments: a matrix named
x for the dataset and a character string named copula for the copula to test for. The whole
calculation process, including the estimation of the margins and the copula, the calculation of
the test statistics, and the bootstrapping of the p-value, is performed for this customized test.
The procedure is shown using the test statistics of the gofCvM test.
R> Testfunc = function(x, copula) {
+ C.theo = pCopula(x, copula = copula)
+ C.n = F.n(x, X = x)
+ CnK = sum((C.n - C.theo)^2)
+ return(CnK)
+ }
R> gofCustomTest(copula = "normal", x = IndexReturns2D, M = 10,

+ customTest = "Testfunc", seed.active = 1)
-------------------------------------------------------------------------------
Parametric bootstrap goodness-of-fit test with Testfunc test and normal copula
Parameters:
theta.1 = 0.834742824340301
Tests results:
Testfunc 1 0.01520542
• gofGetHybrid: Allows calculating hybrid test p-values for given p-values from customized
tests with an object of class "gofCOP" generated in the package. Through the combination of
gofCustomTest and gofGetHybrid, the users are not limited to the implemented tests in the
package and have the opportunity to include their own tests in the analysis. Note that the
function gofOutputHybrid has slightly different but comparable functionality, which is the
reason it is not separately shown.
R> gofGetHybrid(result = res, nsets = 5, p_values = c("MyTest" = 0.7,
+ "AnotherTest" = 0.3))
-------------------------------------------------------------------------------
Hybrid test p-values for given single tests.
Parameters:
theta.1 = 0.834742824340301
Tests results:
p.value
PIOSRn 0.5
Sn 1.0
Kernel 0.3
MyTest 0.7
AnotherTest 0.3
hybrid(1, 2, 3, 4, 5) 1.0

• gofTest4Copula: Returns for a given copula and a given dimension the list of applicable imple-
mented tests.
R> gofTest4Copula("gumbel", d = 5)
[1] "gofArchmChisq" "gofArchmGamma" "gofArchmSnB"

[4] "gofArchmSnC" "gofCustomTest" "gofCvM"
[7] "gofKendallCvM" "gofKendallKS" "gofKS"
[10] "gofRosenblattChisq" "gofRosenblattGamma" "gofRosenblattSnB"
[13] "gofRosenblattSnC"
• gofCopula4Test: Returns for a given test the list of applicable implemented copulae.
R> gofCopula4Test("gofPIOSTn")
[1] "normal" "t" "clayton" "gumbel" "frank" "joe"

[7] "amh" "galambos" "fgm" "plackett"
• gofCheckTime: Estimates the time necessary to compute a selected single or group of GoF tests
for a given number of bootstrapping rounds. This function uses an underlying regression model,
so the results may vary from reality and also from the progress bar predictions. See Section
2.4.5.
R> gofCheckTime("normal", x = IndexReturns2D, tests = "gofRosenblattSnC",
+ M = 10000, seed.active = 1)

An estimate of the computational time is under derivation.
Depending on the tests chosen, dimensionality and complexity of the data, this
might take a while.
The computation will take approximately 0 d, 0 h, 10 min and 7 sec.
• gofco: In the case a copula is already estimated with the package copula, one can provide
an object of class "copula" to this function, and the parameter estimates are taken from the
respective object.
R> copObject = normalCopula(param = 0.8)
R> gofco(copObject, x = IndexReturns2D, M = 10, seed.active = 1,
+ tests = c("gofPIOSRn", "gofKernel"))

Test gofPIOSRn is running
-------------------------------------------------------------------------------
Parametric bootstrap goodness-of-fit test with hybrid test and normal copula
Parameters:
theta.1 = 0.8
Tests results:
PIOSRn 0.9 -0.03641543
Kernel 0.2 0.57115224
hybrid(1, 2) 0.4 NA
Please use the functions gofGetHybrid() and gofOutputHybrid() for

display of subsets of Hybrid tests. To access the results, please
obtain them from the structure of the gofCOP object.

Computational time and parallelization
One of the main drivers of long computation times is the high number of bootstrapping loops to achieve
an asymptotically reliable result. As mentioned in Section 2.4.4, the build-in function gofCheckTime
allows estimating the necessary computation time for a given test, copula, dataset, and number of
bootstrapping rounds. Since different machines may have highly varying computation times for
tests, the function relies on a regression using the number of bootstrapping loops as the independent
variable. To ensure that the linear model is a valid assumption, we investigated the case using the
functions gofKendallKS and gofKernel; see Section 2.5.2 for the results.
Enabling parallelization of the bootstrapping is necessary for computationally demanding tests
as gofPIOSTn where, e.g., the computation for a dataset of 500 observations and 1000 bootstrapping
loops for the t-copula can take, depending on the engine, up to several hours. However, even for
tests with faster computation time, parallelization is useful given the sample size and the number of
bootstrapping loops is sufficiently high. This is shown in Table 2 in the form of a comparison between
the computation times of five tests for five copulas without and with parallelization on four cores.
The dataset contained n = 500 observations randomly generated from a bivariate standard normal
distribution, and the number of bootstrapping loops was set to M = 1000.
Test normal t clayton gumbel frank

#processes = 1:
gofKendallCvM 200.33 530.22 166.85 249.86 167.81

gofKendallKS 115.43 450.50 100.69 172.26 88.83
gofPIOSRn 84.90 433.32 67.20 148.09 50.98
gofRosenblattChisq 75.79 431.02 73.09 143.94 55.97
gofRosenblattGamma 75.28 420.31 75.77 142.16 52.17
#processes = 4:
gofKendallCvM 106.19 288.86 104.29 140.25 94.69

gofKendallKS 68.73 238.55 56.13 89.02 53.00
gofPIOSRn 48.81 235.35 47.37 83.71 33.66
gofRosenblattChisq 49.43 230.33 45.83 85.70 35.47
gofRosenblattGamma 47.45 234.50 48.41 82.04 37.32
Table 2: Computation times in seconds without and with parallelization using an Intel Core i7-4712MQ
CPU with 2.3 GHz on a 64-Bit Windows 10 system.
Simulations
Empirical process of using the package
We would like to illustrate the power of the GoF tests with the use of the gofCopula package. In
practice, one is often confronted with realizations of random variables for which an adequate copula
model has to be found, as, e.g., in the two examples from the financial domain provided in Section 2.6.
To illustrate the procedure, we focus in this Section on an easy replicable example. For this purpose,
we start by simulating n = 1000 observations from a Clayton copula with Kendall’s τ = 0.5.
R> param = iTau(copula = claytonCopula(), tau = 0.5)
R> n = 1000; set.seed(1)
R> x = rCopula(n = n, copula = claytonCopula(param = param))
To gain a better understanding of the data, Figure 2 shows the simulated data with different margins,
reflecting the typical shape of the Clayton copula.
R> par(mfrow = c(1,2))

R> u = cbind(ecdf(x[,1])(x[,1]), ecdf(x[,2])(x[,2])) * n / (n + 1)
R> plot(u, col = "blue3", pch = 19, cex.lab = 1.25,
+ xlab = expression(u[1]), ylab = expression(u[2]))
R> plot(qnorm(u), col = "blue3", pch = 19, cex.lab = 1.25,

+ xlab = expression(x[1]), ylab = expression(x[2]))

1.0
3
0.8
2
1
0.6
u2
x2
0
0.4
−1
0.2
−2
−3
0.0
0.0 0.2 0.4 0.6 0.8 1.0 −3 −2 −1 0 1 2 3
u1 x1
Figure 2: n = 1000 observations sampled from a Clayton copula with τ = 0.5. Margins are transformed
using ranks on the left plot and are standard normal on the right plot.
To make an adequate decision on which copula should be used in the respective modeling task,
the GoF testing should involve more than looking at one or two test results and should consider a
reasonable amount of potential copula models. We structure our procedure by testing for three groups
of copulae separately: Elliptical, Archimedean, and extreme value (EV) copulae. In the function call,
we select the FGM and Plackett copulae together with the EV category, although they do not belong
to any of the three categories. Elliptical copulae include the normal and t-copula, while the Clayton,
Gumbel, Frank, Joe, and AMH copulae are the Archimedean ones. Galambos, Husler-Reiss, Tawn,
and t-EV belong to the EV category. Notice that this categorization could be modified, as, e.g., the
Gumbel copula is also an EV copula. However, the given approach offers not only a logical structuring
of the modeling task, but leads to using a close to maximal number of tests via only three function
calls. The bootstrap parameters were set to M = 100 and MJ = 1000. As this task is computationally
demanding, we set the argument processes = 7 to speed up the calculation using parallelization on 7
cores. We use the default margins = "ranks" and set seed.active = 10 for reproducibility.
R> cop_1 = gof(x = x, M = 100, MJ = 1000, processes = 7, seed.active = 10,
+ copula = c("normal", "t"))
+ copula = c("clayton", "gumbel", "frank", "joe", "amh"))
+ copula = c("galambos", "huslerReiss", "tawn", "tev", "fgm", "plackett"))
To evaluate the gained objects of class "gofCOP", one can manually inspect the resulting p-values
and look closer at the performances and differences between the single tests and the corresponding
hybrids. However, the easiest and most informative way is to visualize the p-values, which is done
using the plot function.
R> plot(cop_1)
singleTests hyb2 hyb3 hyb4 hyb5 hyb6 hyb7 hyb8 hyb9 hyb10 hyb11 hyb12
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
normal
t
Figure 3: p-values of the hybrid tests for the data from Figure 2 for elliptical copulae.

R> plot(cop_2)
singleTests hyb2 hyb3 hyb4 hyb5 hyb6 hyb7 hyb8 hyb9 hyb10 hyb11 hyb12 hyb13 hyb14 hyb15
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
clayton
gumbel
frank
joe
Figure 4: p-values of the hybrid tests for the data from Figure 2 for Archimedean copulae.
R> plot(cop_3)
singleTests hyb2 hyb3 hyb4 hyb5
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
galambos
huslerRe
tev
plackett
galambos
huslerRe
tev
plackett
galambos
huslerRe
tev
plackett
galambos
huslerRe
tev
plackett
galambos
huslerRe
tev
plackett
Figure 5: p-values of the hybrid tests for the data from Figure 2 for EV, FGM, and Plackett copulae.
Interpreting Figures 3, 4, and 5 clearly shows the ability of the tests to detect the true copula. The
column singleTests in Figure 4 indicates that the Clayton copula is appropriate. The decision is
supported by the higher-order hybrid tests, as all p-values except for the Clayton copula become 0,
strongly rejecting the H0 -hypothesis in these cases. Notice that similar to the introductory example in
Section 2.4, the AMG, Tawn, and FGM are automatically excluded, which is why they do not appear in
the plots. Having such a result at hand, the user can proceed with the modeling task with the selected
copula.
Validating the model for the time estimation
In the next step, we validate the assumption of using a linear model for estimating the computation
time in gofCheckTime. We have chosen the gofKendallKS test as a representative for the group of single
bootstrapping tests and gofKernel, as the test having a double bootstrapping procedure. Both tests are
available for all copulae in the bivariate case. For gofKendallKS, we measured the computation times
for 12 copulae, varying numbers of bootstrap loops (M) and sample sizes (n) of the underlying dataset,
which is simulated from a normal copula with Kendall’s τ = 0.1. This value is selected because it falls
within the attainable interval of Kendall’s τ for all copulae, see Table A.1. For gofKernel, we fixed
n = 100 and investigated the situation for 12 copulae, different M, and different sample sizes MJ of
the internal bootstrap. The results are shown in Figures 6, 7, and 8. The t-EV copula is not included in
these illustrations due to its tremendous computation time, which can exceed the one of the t-Copula
even for small sample sizes by a factor of 10 and higher. However, similar properties in the behavior of
the computation time depending on the number of bootstrapping loops can be found. All calculations
were performed without parallelization using an Intel Core i7-4712MQ CPU with 2.3 GHz on a 64-Bit
Windows 10 system.

For the gofKendallKS test, the computation time increases linearly with the number of bootstrap-
ping loops M, while the t-Copula is generally the most time-demanding of the considered copulae.
This holds for all the analyzed sample sizes. A similar observation can be made for the gofKernel
test. Here, a rapid increase in computation time is expected if both M and MJ increase. However,
following Figures 7 and 8, this is not the case, and a linear dependency is justifiable. Therefore, the
package implements for gofKernel a linear model with M and MJ being independent variables.
n = 100 n = 500
copula copula
150
normal 400 normal
Time in s
Time in s
t t
100
clayton clayton
gumbel 200 gumbel
50 frank frank
joe joe
0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
M M
n = 1000 n = 1500
3000
copula 1500 copula
normal normal
Time in s
Time in s
2000 t 1000 t
clayton clayton
gumbel gumbel
1000 500
frank frank
joe joe
0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
M M
n = 100 n = 500
75 copula 200 copula

amh amh
150
Time in s
Time in s
galambos galambos
50
huslerReiss huslerReiss
100
tawn tawn
25 fgm fgm
50
plackett plackett
0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
M M
n = 1000 n = 1500
400
copula 600 copula
amh amh
300
Time in s
Time in s
galambos galambos
400
huslerReiss huslerReiss
200
tawn tawn
fgm 200 fgm
100
plackett plackett
0 0
0 500 1000 1500 2000 0 500 1000 1500 2000
M M
Figure 6: Computation times of gofKendallKS for different copulae, sample sizes n, and number of
bootstrapping loops M.

Figure 7: Computation times of gofKernel for different copulae, number of bootstrapping loops M,
and internal bootstrap sample size MJ.
Figure 8: Computation times of gofKernel for different copulae, number of bootstrapping loops M,
and internal bootstrap sample size MJ.
Application
Cryptocurrency market
We intend to demonstrate the functionality of the gofCopula package and show the empirical pro-
cedure as described in Section 2.5.1 on a real-world example from the market of cryptocurrencies.
To account for the relevant steps in a realistic application study, we split the procedure into Data
Investigation and Goodness-of-Fit testing.
Data investigation
We have chosen Bitcoin (BTC) and Litecoin (LTC) for our analysis. The objective is to detect which
copula is appropriate to model the dependence structure between BTC-LTC and check whether the
copula changes over the years. For that purpose, we use the volatility-adjusted log-returns of the
currencies in the time span from 2015 to 2018. The volatility correction was performed by fitting a

GARCH(1,1) process to each time series for each year separately in order to extract their standardized
residuals. These are included in the package as CryptoCurrencies, whereas each element of the list
contains the data for a particular year. In order to gain a visual impression beforehand, we plotted
the data with margins transformed to standard normal, leading to Figure 9. A strong dependency
between both cryptocurrencies is visible, especially in the year 2018. Based on these residual diagrams,
it is possible to take a guess which copula is the most adequate for the given situation. For 2015, one
could possibly argue that the elliptical shape of a normal copula is present, while in 2016 and 2018,
the shapes are more similar to the one of a t-Copula. Finally, for the year 2017, Figure 9 shows a
comparable plot to Figure 2 from the simulated example in Section 2.5.1, indicating a Clayton copula
might be present. However, these visual impressions are to a certain degree subjective and need to be
backed up by the GoF tests. Ideally, the test results would match our plot-based guesses.
R> data("CryptoCurrencies", package = "gofCopula")
R> years = as.character(2015:2018)
R> for(i in years){

+ x1 = CryptoCurrencies[[i]][,1]
+ x2 = CryptoCurrencies[[i]][,2]
+ n = length(x1)
+ plot(qnorm(cbind(ecdf(x1)(x1), ecdf(x2)(x2)) * n / (n + 1)), col = "blue3",
+ pch = 19, cex.lab = 1.25, main = i, xlab = "Bitcoin", ylab = "Litecoin")
+ }
2015 2016
3
2
2
1
1
Litecoin
Litecoin
0
0
−1
−1
−2
−2
−3
−2 −1 0 1 2 −3 −2 −1 0 1 2 3
Bitcoin Bitcoin
2017 2018
2
2
1
1
Litecoin
Litecoin
0
0
−1
−1
−2
−2
−2 −1 0 1 2 −2 −1 0 1 2
Bitcoin Bitcoin
Figure 9: Residual plots for BTC-LTC with margins transformed to standard normal.
Goodness-of-fit testing
In this example, the focus in testing is on the most popular copula models in practice: normal, t,
Clayton, Gumbel, and Frank copulae. To get the highest testing power, we include all tests, which
are available for all five copulae. Thus, following Table 1, each test in the package except the ones
based on the transformation for Archimedean copulae (see Section 2.3.4) is computed. Additionally, all

possible hybrid tests are considered. We use the function gof while setting the bootstrap parameters
M = 100 and MJ = 1000. We specify the number of cores for the parallelization to processes = 7.
For replicability, we set seed.active = 1:101 and apply the non-parametric margin transformation
by default.
R> copulae = c("normal", "t", "clayton", "gumbel", "frank")
R> BTC_LTC_15 = gof(x = CryptoCurrencies[["2015"]], copula = copulae, M = 100,

+ MJ = 1000, processes = 7, seed.active = 1:101)
After finishing the calculations, we proceed by plotting the received objects of class "gofCOP". For a
detailed explanation about the information contained in the gofCopula pirateplots, please see Section
2.4.2 and Phillips (2017).
R> plot(BTC_LTC_15)
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
Figure 10: p-values of the single and hybrid tests for BTC-LTC in the year 2015.
R> plot(BTC_LTC_16)
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t

R> plot(BTC_LTC_17)
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
frank
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
R> plot(BTC_LTC_18)
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
Figures 10 to 13 show the resulting p-values in the form of "gofCOP"-plots for all the considered
years. Following the usual approach in practice, we select the copula corresponding to the highest
p-values. For the year 2015, we see that the t-copula is favored by the tests, as all remaining p-values
become 0 with increasing hybrid testing size; see Figure 10. This rejects our initial visual guess that
a normal copula might be appropriate. Continuing with 2016, we see our visual opinion solidified,
as the plot suggests using a t-copula to capture the dependence structure. We can even see that the
p-values converge to 1 for the t-copula when we consider the hybrid testing orders 9, 10, 11, and 12 .
Figure 12 is in line with our visual impression as well. We see that the tests favor a Clayton copula,
while the other copula models are rejected by the higher-order hybrid tests. Finally, for 2018 Figure
13 gives for the t-copula the highest p-values, although the difference to the p-values of the Clayton
copula is not too large. Therefore, in three out of four years, the results from gofCopula matched our
visual impressions from the residual plots.
Summarizing, the following conclusions can be drawn from this analysis:
• Generally, the dominant copulae in describing the dependency between the volatility-adjusted
log-returns of BTC-LTC is the t-copula. Following the test results, the year 2017 is an exception,
as the dependence structure shifted towards a Clayton copula. This observation reflects the
change in the market during the year 2017, as many investors got attracted by cryptocurrencies
in this phase. Due to the developed hype, both the prices of BTC and LTC drastically increased,
resulting in a modified underlying dependence structure between the two currencies.
• The hybrid tests are able to stabilize the results of the single tests, as they clearly selected for
2015, 2016, and 2018 the t-copula and in 2017 the Clayton copula. Therefore it is recommendable

to take hybrid tests into account in order to use the package adequately and get the highest
testing power.
Stock return data
As a second real-world example, we analyze the volatility-adjusted stock log-returns of Citigroup (C)
and the Bank of America (BoA) in the time span from 2004 to 2012. The procedure is, again, splitted
into Data Investigation and Goodness-of-Fit testing.
Data investigation
This data was analyzed by Zhang et al. (2016), and we are expanding their procedure and consider the
same copulae and tests as in the example in Section 2.6.1. The volatility correction was performed
similarly in terms of fitting a GARCH(1,1) process, and the resulting data is included in gofCopula in
the list Banks. Note that in this section, we focus on the years 2004 and 2007, while the results of the
other years are given in Appendix B. We start by visualizing the residuals with margins transformed
to standard normal.
R> data("Banks", package = "gofCopula")
R> for(i in c("2004", "2007")){

+ x1 = Banks[[i]][,1]
+ x2 = Banks[[i]][,2]
+ n = length(x1)
+ plot(qnorm(cbind(ecdf(x1)(x1), ecdf(x2)(x2)) * n / (n + 1)), col = "blue3",
+ pch = 19, cex.lab = 1.25, main = i, xlab = "Citigroup", ylab = "Bank of America")
+ }
2004 2007
2
2
Bank of America
Bank of America
1
1
0
0
−1
−1
−2
−2
−2 −1 0 1 2 −2 −1 0 1 2
Citigroup Citigroup
Figure 14: Residual plots for C/BoA in 2004 and 2007 with standard normal margins.
Analyzing the shape of the data in Figure 14 for 2004, one may argue that the elliptical normal
copula is present, while in 2007, a t-copula is possibly more appropriate. To check these assumptions,
we proceed with the GoF testing.
Goodness-of-fit testing
We set M = 100 and MJ = 1000 as bootstrap parameters, parallelize via processes = 7, and set
seed.active = 1:101 for reproducibility. Further, we implicitly keep the default margins = "ranks"
to perform the margin transformation nonparametrically.

R> copulae = c("normal", "t", "clayton", "gumbel", "frank")
R> C_BoA_04 = gof(x = Banks[["2004"]], copula = copulae, M = 100, MJ = 1000,

+ processes = 7, seed.active = 1:101)
R> C_BoA_07 = gof(x = Banks[["2007"]], copula = copulae, M = 100, MJ = 1000,
+ processes = 7, seed.active = 1:101)
Following these calculations, we continue to plot the results, leading to Figures 15 and 16. For a
detailed explanation about the information contained in the gofCopula pirateplots, please see Section
2.4.2 and Phillips (2017). Interpreting these "gofCOP"-plots of the p-values, the tests propose for 2004
indeed a normal copula (and a Frank one, which is radially symmetric), although a t-copula is a valid
assumption as well. Compared to 2004, the p-values for the normal copula definitely decreased in 2007
and converged slowly to 0 with increasing hybrid testing size. The decision goes clearly in favor of the
t-copula, which is in line with our original guess. Evaluating the results from Appendix B leads to
similar conclusions as in Section 2.6.1. The hybrid tests are relatively stable and match in the majority
of the cases the visual impressions from the residual plots. The proper copula seems to be the t-copula
in most of the years, although in 2004 and 2009, the normal copula is a reasonable assumption. The
hybrid tests are able to stabilize the selection of the copula.
R> plot(C_BoA_04)
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
Figure 15: p-values of the C/BoA data for 2004. The column hyb12 of the t-copula is empty, as the
p-value of the test gofWhite could not be computed due to instability in the test statistics. For a
detailed description of this phenomenon, see Nagler et al. (2019).
R> plot(C_BoA_07)
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
0
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
Figure 16: p-values of the C/BoA data for 2007.

Conclusion
This paper introduces a gofCopula package that provides maximum flexibility in performing statistical
Goodness-of-Fit tests for copulae. The package provides an interface for 16 most popular GoF tests for
13 copulae with automatic estimation of margins via different techniques. The user is not limited to
the implemented tests as self-defined test statistics functions can be easily embedded via a function
provided in the package. As the computation of p-values relies on a parametric bootstrap, efficient
and user-friendly parallelization is available. During the bootstrapping procedure, all tests inform the
user about the progress of the calculations as well as the estimated time until the results are available.
Additionally, gofCopula allows for the replication of said results. The package offers intelligible and
interpretable visualization of the results of the hybrid tests that strengthen the overall test power. The
flexibility and the usefulness of the tests are shown via a simulation and two empirical studies in
economic sciences. In a nutshell, the broad range of tests, the comprehensive combination of methods,
and an informative user-interface make gofCopula a fire-and-forget package providing flexibility in
testing for the proper copula.
Acknowledgements
The authors are grateful to Shulin Zhang, Qian Zhou, Peter Song, and Sören Pannier for helpful
discussions and to Oliver Scaillet for the code of the version of his test used in Zhang et al. (2016) that is
being adapted in this package. Financial support from NUS FRC grant R-146-000-298-114 “Augmented
machine learning and network analysis with applications to cryptocurrencies and blockchains” as well
as CityU Start-Up Grant 7200680 “Textual Analysis in Digital Markets” is gratefully acknowledged
by Simon Trimborn. Ostap Okhrin thankfully received financial support from RFBR, project number
20-04-60158.
Bibliography
P. Barbe, C. Genest, K. Ghoudi, and B. Rémillard. On Kendalls’s process. Journal of Multivariate Analysis,
58:197–229, 1996. URL https://doi.org/10.1006/jmva.1996.0048. [p471]
W. Breymann, A. Dias, and P. Embrechts. Dependence structures for multivariate high-frequency data
in finance. Quantitative Finance, 1:1–14, 2003. URL https://doi.org/10.1080/713666155. [p472]
S. X. Chen and T. Huang. Nonparametric estimation of copula functions for dependence modeling. The
Canadian Journal of Statistics, 35(2):265–282, 2007. URL https://doi.org/10.1002/cjs.5550350205.
[p470]
X. Chen and Y. Fan. Estimation and model selection of semiparametric copula-based multivariate
dynamic models under copula misspecification. Journal of Econometrics, 135(1–2):125–154, 2006. URL
https://doi.org/10.1016/j.jeconom.2005.07.027. [p470]
X. Chen, Y. Fan, and V. Tsyrennikov. Efficient estimation of semiparametric multivariate copula models.
Journal of the American Statistical Association, 101(475):1228–1240, 2006. URL https://doi.org/10.
1198/016214506000000311. [p470]
G. Csárdi and R. FitzJohn. progress: Terminal Progress Bars, 2019. URL https://CRAN.R-project.org/
package=progress. R package version 1.2.2. [p469]
P. De Valpine. The common sense of p values. Ecology, 95(3):617–621, 2014. URL https://doi.org/10.
1890/13-1271.1. [p467]
S. Demarta and A. J. McNeil. The t-copula and related copulas. International Statistical Review, 73(1):
111–129, 2005. URL https://doi.org/10.1111/j.1751-5823.2005.tb00254.x. [p469, 495]
J. Dobrić and F. Schmid. A goodness of fit test for copulas based on Rosenblatt’s transformation.
Computational Statistics & Data Analysis, 51(9):4633–4642, 2007. URL https://doi.org/10.1016/j.
csda.2006.08.012. [p472]
D. Dokuzoğlu and V. Purutçuoğlu. Comprehensive analyses of gaussian graphical model under

different biological networks. Acta Physica Polonica, A., 132(3), 2017. URL https://doi.org/10.
12693/APhysPolA.132.1106. [p467]

F. Durante and C. Sempi. Copula theory: An introduction. In P. Jaworski, F. Durante, W. K. Härdle,

and T. Rychlik, editors, Copula Theory and Its Applications, pages 3–31. Springer, Berlin, Heidelberg,
2010. URL https://doi.org/10.1007/978-3-642-12465-5_1. [p467]
Y. Fang and L. Madsen. Modified gaussian pseudo-copula: Applications in insurance and finance.
Insurance: Mathematics and Economics, 53(1):292–301, 2013. URL https://doi.org/10.1016/j.
insmatheco.2013.05.009. [p467]
J.-D. Fermanian. Goodness-of-fit tests for copulas. Journal of Multivariate Analysis, 95(1):119–152, 2005.
URL https://doi.org/10.1016/j.jmva.2004.07.004. [p471]
J.-D. Fermanian and O. Scaillet. Nonparametric estimation of copulas for time series. Journal of Risk, 5:
25–54, 2003. URL https://doi.org/10.21314/JOR.2003.082. [p470]
P. Gaensler and W. Stute. Seminar on Empirical Processes. Springer Basel AG, Boca Raton, 1987. URL
https://doi.org/10.1007/978-3-0348-6269-1. [p470]
C. Genest and J. Nešlehová. Copulas and copula models. Wiley StatsRef: Statistics Reference Online,
2014. URL https://doi.org/10.1002/9781118445112.stat07523. [p467]
C. Genest and B. Rèmillard. Validity of the parametric bootstrap for goodness-of-fit testing in semi-
parametric models. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 6(44):1096–1127,
2008. URL https://doi.org/10.1214/07-AIHP148. [p471]
C. Genest and L.-P. Rivest. Statistical inference procedures for bivariate Archimedean copulas. Journal of
the American Statistical Association, 88(3):1034–1043, 1993. URL https://doi.org/10.1080/01621459.
1993.10476372. [p469, 471]
C. Genest, K. Ghoudi, and L.-P. Rivest. A semi-parametric estimation procedure of dependence

parameters in multivariate families of distributions. Biometrika, 82(3):543–552, 1995. URL https:
//doi.org/10.1093/biomet/82.3.543. [p469, 470]
C. Genest, J.-F. Quessy, and B. Rémillard. Goodness-of-fit procedures for copula models based on
the probability integral transformation. Scandinavian Journal of Statistics, 33:337–366, 2006. URL
https://doi.org/10.1111/j.1467-9469.2006.00470.x. [p471]
C. Genest, B. Rémillard, and D. Beaudoin. Goodness-of-fit tests for copulas: A review and a power
study. Insurance: Mathematics and Economics, 44:199–213, 2009. URL https://doi.org/10.1016/j.
insmatheco.2007.10.005. [p467, 470, 471, 472, 474]
T. Gneiting and A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal
of the American Statistical Association, 102(477):359–378, 2007. URL https://doi.org/10.1198/
016214506000001437. [p467]
M. Gomes, R. Radice, J. Camarena Brenes, and G. Marra. Copula selection models for non-gaussian
outcomes that are missing not at random. Statistics in medicine, 38(3):480–496, 2019. URL https:
//doi.org/10.1002/sim.7988. [p467]
J. Größer and O. Okhrin. Copulae: An overview and recent developments. Wiley Interdisciplinary
Reviews: Computational Statistics, page e1557, 2021. URL https://doi.org/10.1002/wics.1557.
[p467]
C. Hering and M. Hofert. Goodness-of-fit tests for archimedean copulas in high dimensions. In K. Glau,
M. Scherer, and R. Zagst, editors, Innovations in Quantitative Risk Management, pages 357–373, Cham,
2015. Springer International Publishing. URL https://doi.org/10.1007/978-3-319-09114-3_21.
[p472]
M. Hofert, I. Kojadinovic, M. Maechler, and J. Yan. copula: Multivariate Dependence with Copulas, 2020.
URL https://CRAN.R-project.org/package=copula. R package version 1.0-0. [p467]
K. Huang, L. Dai, M. Yao, Y. Fan, and X. Kong. Modelling dependence between traffic noise and traffic
flow through an entropy-copula method. Journal of Environmental Informatics, 29(2), 2017. URL
https://doi.org/10.3808/jei.201500302. [p467]
W. Huang and A. Prokhorov. A goodness-of-fit test for copulas. Econometric Reviews, 33(7):751–771,
2014. URL https://doi.org/10.1080/07474938.2012.690692. [p473]
H. Joe. Multivariate Models and Dependence Concepts. Chapman & Hall, London, 1997. [p469, 470]

H. Joe. Asymptotic efficiency of the two-stage estimation method for copula-based models. Journal
of Multivariate Analysis, 94(2):401–419, 2005. URL https://doi.org/10.1016/j.jmva.2004.06.003.
[p470]
H. Joe and D. Kurowicka. Dependence Modeling: Vine Copula Handbook. World Scientific, 2010. URL
https://doi.org/10.1142/7699. [p467]
M. Jouini and R. Clemen. Copula models for aggregating expert opinions. Operations Research, 3(44):
444–457, 1996. URL https://doi.org/10.1287/opre.44.3.444. [p471]
S. Konigorski, Y. E. Yilmaz, and S. B. Bull. Bivariate genetic association analysis of systolic and diastolic
blood pressure by copula models. BMC Proceedings, 8(1):S72, 2014. URL https://doi.org/10.1186/
1753-6561-8-S1-S72. [p467]
O. Kuss, A. Hoyer, and A. Solms. Meta-analysis for diagnostic accuracy studies: a new statistical
model using beta-binomial distributions and bivariate copulas. Statistics in medicine, 33(1):17–30,
2014. URL https://doi.org/10.1002/sim.5909. [p467]
Z. Liu, S. Guo, L. Xiong, and C.-Y. Xu. Hydrological uncertainty processor based on a copula function.
Hydrological sciences journal, 63(1):74–86, 2018. URL https://doi.org/10.1080/02626667.2017.
1410278. [p467]
X. Ma, S. Luan, B. Du, and B. Yu. Spatial copula model for imputing traffic flow data from remote
microwave sensors. Sensors, 17(10):2160, 2017. URL https://doi.org/10.3390/s17102160. [p467]
F. Michiels and A. De Schepper. A copula test space model how to avoid the wrong copula choice.
Kybernetika, 44(6):864–878, 2008. URL http://hdl.handle.net/10338.dmlcz/135896. [p495]
T. Nagler, U. Schepsmeier, J. Stoeber, E. C. Brechmann, B. Graeler, and T. Erhardt. VineCopula: Statistical

Inference of Vine Copulas, 2019. URL https://CRAN.R-project.org/package=VineCopula. R package
version 2.3.0. [p467, 473, 490, 496, 497, 498]
D. Oakes. Multivariate survival distributions. Journal of Nonparametric Statistics, 3(3-4):343–354, 1994.

URL https://doi.org/10.1080/10485259408832593. [p470]
D. H. Oh and A. J. Patton. Time-varying systemic risk: Evidence from a dynamic copula model of cds
spreads. Journal of Business & Economic Statistics, 36(2):181–195, 2018. URL https://doi.org/10.
1080/07350015.2016.1177535. [p467]
N. Phillips. yarrr: A Companion to the e-Book “YaRrr!: The Pirate’s Guide to R”, 2017. URL https:
//CRAN.R-project.org/package=yarrr. R package version 0.1.5. [p469, 477, 487, 490]
J.-D. F. D. Radulovic and M. Wegkamp. Weak convergence of empirical copula processes. Bernoulli, 10
(5):847–860, 2004. URL https://doi.org/10.3150/bj/1099579158. [p470]
B. Remillard and J.-F. Plante. TwoCop: Nonparametric test of equality between two copulas, 2012. URL
https://CRAN.R-project.org/package=TwoCop. R package version 1.0. [p467]
B. Remillard and O. Scaillet. Testing for equality between two copulas. Journal of Multivariate Analysis,
100:377–386, 2009. URL https://doi.org/10.1016/j.jmva.2008.05.004. [p467]
M. Rosenblatt. Remarks on a multivariate transformation. Annals of Mathematical Statistics, 23(3):

470–472, 1952. URL https://doi.org/10.1214/aoms/1177729394. [p471, 472]
I. D. L. Salvatierra and A. J. Patton. Dynamic copula models and high frequency data. Journal
of Empirical Finance, 30:120–135, 2015. URL https://doi.org/10.1016/j.jempfin.2014.11.008.
[p467]
O. Scaillet. Kernel-based goodness-of-fit tests for copulas with fixed smoothing parameters. Journal
of Multivariate Analysis, 98(3):533–543, 2007. URL https://doi.org/10.1016/j.jmva.2006.05.006.
[p467, 473, 474]
U. Schepsmeier. Efficient information based goodness-of-fit tests for vine copula models with fixed
margins: A comprehensive review. Journal of Multivariate Analysis, 138:34–52, 2015. URL https:
//doi.org/10.1016/j.jmva.2015.01.001. [p473]
P. Shi, X. Feng, and J.-P. Boucher. Multilevel modeling of insurance claims using copulas. The Annals of
Applied Statistics, 10(2):834–863, 2016. URL https://doi.org/10.1214/16-AOAS914. [p467]

J. H. Shih and T. A. Louis. Inferences on the association parameter in copula models for bivariate
survival data. Biometrics, 51(4):1384–1399, 1995. URL https://doi.org/10.2307/2533269. [p470]
A. Sklar. Fonctions de répartition à n dimension et leurs marges. Publications de l’Institut de Statistique

de l’Université de Paris, 8:299–231, 1959. [p467]
D. Valle and D. Kaplan. Quantifying the impacts of dams on riverine hydrology under non-stationary
conditions using incomplete data and gaussian copula models. Science of The Total Environment, 677:
599–611, 2019. URL https://doi.org/10.1016/j.scitotenv.2019.04.377. [p467]
W. Wang and M. Wells. Model selection and semiparametric inference for bivariate failure-time data.
Journal of the American Statistical Association, 95(449):62–76, 2000. URL https://doi.org/10.1080/
01621459.2000.10473899. [p471]
H. White. Maximum likelihood estimation of misspecified models. Econometrica: Journal of the

Econometric Society, 50:1–25, 1982. URL https://doi.org/10.2307/1912526. [p473, 474]
F. Wu, E. Valdez, and M. Sherris. Simulating from exchangeable archimedean copulas. Communications
in Statistics—Simulation and Computation, 36(5):1019–1034, 2007. URL https://doi.org/10.1080/
03610910701539781. [p472]
S. Zhang, O. Okhrin, Q. M. Zhou, and P. X.-K. Song. Goodness-of-fit test for specification of
semiparametric copula dependence models. Journal of Econometrics, 193(1):215–233, 2016. URL
https://doi.org/10.1016/j.jeconom.2016.02.017. [p467, 468, 474, 489, 491]

Appendix A
Table A.1 contains parameter ranges, the bivariate cumulative distribution function (CDF), and possible
values of Kendall’s τ for the copulae available in gofCopula.
Copula θ∈ C ( u1 , u2 ) τ∈
R Φ −1 ( u 1 ) R Φ −1 ( u 2 ) n
2θst−s2 −t2
o
Normal [−1, 1] √1 exp dsdt [−1, 1]
−∞ −∞ 2π 1−θ 2 2(1− θ 2 )
o − ν +2
[−1, 1] R t−ν 1 (u1 ) R t−ν 1 (u2 )
√1
n
s2 +t2 −2θst 2
t −∞ −∞ 1+ ν (1− θ 2 )
dsdt [−1, 1]
ν>0 2π 1−θ 2
n o− 1
[−1, ∞)\{0} max(u1−θ + u2−θ − 1, 0)
θ
Clayton [−1, 1]
o
1 θ
n 1
Gumbel [1, ∞) exp − (− log u1 ) θ + (− log u2 ) θ [0, 1]
h i
{exp(−θu1 )−1}{exp(−θu2 )−1}
Frank (−∞, ∞)\{0} − 1θ log 1 + exp(−θ )−1
[−1, 1]
1
[1, ∞)

Joe 1 − (1 − u1 ) θ + (1 − u2 ) θ − (1 − u1 ) θ (1 − u2 ) θ θ
[0, 1]
5−8 log 2
AMH [−1, 1] u1 u2 [ 3 , 13 ]
1−θ (1−u1 )(1−u2 ) ≈ [−0.1817, 0.3333]

− 1θ
[0, ∞) (− log u1 )−θ + (− log u2 )−θ

Galambos u1 u2 exp [0, 1]
n o
log u log u
Husler-Reiss [0, ∞) exp log(u1 )Φ( 1θ + 12 θ log log u1 ) + log(u2 )Φ( 1θ + 21 θ log log u2 )
2
[0, 1]
1
q
8 arctan( 1)
3

Tawn [0, 1] u1 u2 exp
log u log u
−θ log u11+log u22 [0, √ − 2]
3
≈ [0, 0.4184]
[−1, 1]
t-EV see Demarta and McNeil (2005) [0, 1]
ν>0
FGM [−1, 1] u1 u2 + u1 u2 θ (1 − u1 )(1 − u2 ) [− 29 , 29 ]

q
{1+(θ −1)(u1 +u2 )}− {1+(θ −1)(u1 +u2 )}2 −4u1 u2 θ (θ −1)
Plackett (0, ∞) 2( θ −1)
[−1, 1]
Table A.1: This table is mainly based on Michiels and De Schepper (2008). AMH abbreviates Ali-
Mikhail-Haq. FGM stands for Farlie-Gumbel-Morgenstern. Φ is the CDF of the univariate standard
normal distribution and Φ−1 its inverse. t− 1
ν is the inverse CDF of the univariate t-distribution with ν
degrees of freedom. The expression of the CDF for the t-EV is complex due to the construction via a
Pickands dependence function, which is why we do not explicitly list it. The given parameterization
of the tawn copula is based on the one implemented in the package copula.
Appendix B
This section is devoted to the results of Section 2.6.2.

p−value p−value p−value
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
normal normal normal

t t t
clayton clayton clayton
gumbel gumbel gumbel
frank frank frank
singleTests
singleTests
singleTests

t t t
hyb2
hyb2
hyb2

frank frank frank
t t t
C ONTRIBUTED R ESEARCH A RTICLES
hyb3
hyb3
hyb3

frank frank frank
t t t
hyb4
hyb4
hyb4

frank frank frank
The R Journal Vol. 13/1, June 2021

t t t
hyb5
hyb5
hyb5

frank frank frank
t t t
hyb6
hyb6
hyb6

frank frank frank
t t t
hyb7
hyb7
hyb7

frank frank frank
t t t
hyb8
hyb8
hyb8

frank frank frank

t t t
hyb9
hyb9
hyb9

frank frank frank
Figure B.2: p-values of the C/BoA data for 2006.

t t t
hyb10
hyb10
hyb10

frank frank frank
t t t
hyb11
hyb11
hyb11

frank frank frank
t t t
hyb12
hyb12
hyb12

frank frank frank
ISSN 2073-4859
the p-value of the test gofWhite could not be computed due to instability in the test statistics. For a
Figure B.3: p-values of the C/BoA data for 2008. The column hyb12 of the t-copula is empty, as
496
p−value p−value p−value
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

t t t
frank frank frank
singleTests
singleTests
singleTests

t t t
hyb2
hyb2
hyb2

frank frank frank
t t t
C ONTRIBUTED R ESEARCH A RTICLES
hyb3
hyb3
hyb3

frank frank frank
t t t
hyb4
hyb4
hyb4

frank frank frank
The R Journal Vol. 13/1, June 2021

t t t
hyb5
hyb5
hyb5

frank frank frank
t t t
hyb6
hyb6
hyb6

frank frank frank
t t t
hyb7
hyb7
hyb7

frank frank frank
t t t
hyb8
hyb8
hyb8

frank frank frank

t t t
hyb9
hyb9
hyb9

frank frank frank

t t t
hyb10
hyb10
hyb10

frank frank frank
t t t
hyb11
hyb11
hyb11

frank frank frank
t t t
hyb12
hyb12
hyb12

frank frank frank
ISSN 2073-4859
497
1
0.9
0.8
0.7
0.6
p−value
0.5
0.4
0.3
0.2
0.1
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
t
clayton
gumbel
frank
normal
Ostap Okhrin
Chair of Econometrics and Statistics, esp. in Traffic Sciences
Institute of Transportation Economics
Faculty of Transportation
Technische Universität Dresden
Würzburger Street 35, 01187 Dresden
Germany
[email protected]
Simon Trimborn
Department of Management Sciences
and School of Data Science
City University of Hong Kong
7-268, Lau Ming Wai Academic Building
Hong Kong
[email protected]
Martin Waltz
Chair of Econometrics and Statistics, esp. in Traffic Sciences
Institute of Transportation Economics
Faculty of Transportation
Technische Universität Dresden
Würzburger Street 35, 01187 Dresden
Germany
[email protected]

RJ 2021 060

Uploaded by

Copyright:

Available Formats

RJ 2021 060

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

RJ 2021 060

Uploaded by

Copyright:

Available Formats

C ONTRIBUTED R ESEARCH A RTICLES 467

gofCopula: Goodness-of-Fit Tests for

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Consider a d-dimensional random vector X = { X1 , . . . , Xd } with corresponding marginal distributions

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

(θ̂, α̂1 , . . . , α̂d )⊤ = argmax L(X , θ, α1 , . . . , αd ), (1)

Goodness-of-fit tests for copulae

Empirical copula process

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

process can be defined:

Cd (ui |u1 , . . . , ui−1 ) = P{Ui ≤ ui |U1 = u1 . . . Ui−1 = ui−1 }

The Rosenblatt transform (c.f. Rosenblatt, 1952) is then defined as follows.

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Transformation for Archimedean copulae

Definition 2 Hering and Hofert’s transformation of an Archimedean copula C of dimension d ≥ 2 with

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Kernel density estimator

Using this approach, the testing problem can be formulated as follows:

H0W : S(θ ) = V(θ ) vs. H1W : S(θ ) ̸= V(θ ).

TW = n{d(θ̂ )}⊤ Â−1 d(θ̂ ),

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Bootstrapping test statistics

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Functionality of the R-package

The margins will be estimated as: ranks

user system elapsed

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Plotting gofCOP class objects

R> plot(result, copula = c("clayton", "joe", "plackett"), hybrid = c(1, 3, 5))

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Hybrid testing and further functionality

The margins will be estimated as: ranks

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Please use the functions gofGetHybrid() and gofOutputHybrid() for

The functions are:

R> gofCustomTest(copula = "normal", x = IndexReturns2D, M = 10,

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

[1] "gofArchmChisq" "gofArchmGamma" "gofArchmSnB"

[1] "normal" "t" "clayton" "gumbel" "frank" "joe"

The margins will be estimated as: ranks

The margins will be estimated as: ranks

Please use the functions gofGetHybrid() and gofOutputHybrid() for

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Computational time and parallelization

Test normal t clayton gumbel frank

gofKendallCvM 200.33 530.22 166.85 249.86 167.81

gofKendallCvM 106.19 288.86 104.29 140.25 94.69

Empirical process of using the package

R> par(mfrow = c(1,2))

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

+ xlab = expression(x[1]), ylab = expression(x[2]))

0.0 0.2 0.4 0.6 0.8 1.0 −3 −2 −1 0 1 2 3

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

Validating the model for the time estimation

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859

75 copula 200 copula

The R Journal Vol. 13/1, June 2021 ISSN 2073-4859