Abstract
In this paper, we propose a goodness-of-fit test, named pseudo “in-and-out-of-likelihood” (PIOL) ratio test, to check for misspecification in semi-parametric copula models for univariate time series. The proposed test extends the idea of the IOS test by Presnell and Boos (J Am Stat Assoc 99:216–227, 2004) and PIOS test by Zhang et al. (J Econom, 193:215–233, 2016), which are problematic for direct application to univariate time series. The PIOL test provides an integrated framework for both independent data and time series data. In addition, an approximation method is implemented to alleviate the computational burden of calculating the test statistics. Asymptotic properties of the proposed test statistics are discussed. The finite-sample performance is examined through simulation studies. We also demonstrate the proposed method through the analysis of a time series of daily transactions of Apple trade.
Similar content being viewed by others
References
Banerjee A, Dolado JJ, Galbraith JW, Hendry D (1993) Co-integration, error correction, and the econometric analysis of non-stationary data. Oxford, New York
Beare BK (2010) Copulas and temporal dependence. Econometrica 78:395–410
Beare BK (2012) Archimedean copulas and temporal dependence. Econom Theory 28:1165–1185
Beare BK, Seo J (2015) Vine copula specifications for stationary multivariate Markov chains. J Time Ser Anal 36:228–246
Berg D, Bakken H (2005) A goodness-of-fit test for copulae based on the probability integral transform. Technical Report SAMBA/41/05, Norsk REgnesentral, Oslo
Brent P (2002) Algorithms for minimization without derivatives. Dover Publications, New York
Breymann W, Dias A, Embrechts P (2003) Dependence structures for multivariate high-frequency data in finance. Quant Financ 3:1–14
Chen X, Fan Y (2006a) Estimation and model selection of semiparametric copula-based multivariate dynamic models under copula misspecification. J Econom 135:125–154
Chen X, Fan Y (2006b) Estimation of copula-based semiparametric time series models. J Econom 130:307–335
Chen X, Wu WB, Yi Y (2009) Efficient estimation of copula-based semiparametric markov models. Ann Stat 37:4214–4253
Clayton D (1978) A model for association in bivariate life tables and its application in epidemiological studies of familial tendency in chronic disease incidence. Biometrika 65:141–151
Craiu R, Craiu M (2008) On the choice of parametric families of copulas. Adv Appl Stat 10:25–40
Darsow WF, Nguyen B, Olsen ET (1992) Copulas and Markov processes. Ill J Math 36:600–642
DasGupta A (2008) Asymptotic theory of statistics and probability. Springer, New York
de la Peña VH, Ibragimov R, Sharakhmetov S (2006) Characterizations of joint distributions, copulas, information, dependence and decoupling, with applications to time series. Lect Notes Monogr Ser 49:183–209
De Luca G, Zuccolotto P (2017) Dynamic tail dependence clustering of financial time series. Stat Pap 58:641–657
Dehling H (2006) Limit theorems for dependent U-statistics. In: Bertail P, Soulier P, Doukhan P (eds) Dependence in probability and statistics, vol 187. Lecture notes in statistics. Springer, New York, pp 65–86
Demarta S, McNeil A (2005) The t copula and related copulas. Int Stat Rev 73:111–129
Dobrić J, Schmid F (2007) A goodness of fit test for copulas based on Rosenblatt’s transformation. Comput Stat Data Anal 51:4633–4642
Domowitz I, White H (1982) Misspecified models with dependent observations. J Econom 20:35–58
Durante F (2009) Construction of non-exchangeable bivariate distribution functions. Stat Pap 50:383–391
Emura T, Pan CH (2017) Parametric likelihood inference and goodness-of-fit for dependently left-truncated data, a copula-based approach. Stat Pap. https://doi.org/10.1007/s00362-017-0947-z
Emura T, Lin CW, Wang W (2010) A goodness-of-fit test for Archimedean copula models in the presence of right censoring. Comput Stat Data Anal 54:3033–43
Emura T, Long TH, Sun LH (2017) R routines for performing estimation and statistical process control under copula-based time series models. Commun Stat Simul Comput 46:3067–87
Fermanian JD (2005) Goodness-of-fit tests for copulas. J Multivar Anal 95:119–152
Genest C (1987) Frank’s family of bivariate distributions. Biometrika 82:543–552
Genest C, Rivest LP (1993) Statistical inference procedures for bivariate Archimedean copulas. J Am Stat Assoc 88:1034–1043
Genest C, Favre AC (2007) Everything you always wanted to know about copula modeling but were afraid to ask. J Hydrol Eng 12:347–368
Genest C, Rémillard B, Beaudoin D (2009) Goodness-of-fit tests for copulas: a review and a power study. Insur Math Econ 44:199–213
Gilbert P, Varadhan R (2019) numDeriv: accurate numerical derivatives. R package version. https://CRAN.R-project.org/package=numDeriv
Hofert M, Kojadinovic I, Maechler M, Yan J (2018) copula: multivariate dependence with copulas. R package version. https://CRAN.R-project.org/package=copula
Huang W, Prokhorov A (2014) A goodness-of-fit test for copulas. Econom Rev 33:751–771
Huang XW, Emura T (2019) Model diagnostic procedures for copula-based Markov chain models for statistical process control. Commun Stat Simul Comput. https://doi.org/10.1080/03610918.2019.1602647
Ibragimov R (2009) Copula-based characterizations for higher order Markov processes. Econom Theory 25:819–846
Ibragimov R, Lentzas G (2017) Copulas and long memory. SSRN Electron J. https://doi.org/10.2139/ssrn.1185822
Joe H (1997) Multivariate models and dependence concepts. Chapmann & Hall, London
Kendall M (1938) A new measure of rank correlation. Biometrika 30:81–89
Kim B, Lee S (2013) Robust estimation for copula parameter in SCOMDY models. J Time Ser Anal 34:302–314
Longla, Martial (2015) On mixtures of copulas and mixing coefficients. J Multivar Anal 139:259–265
Malevergne Y, Sornette D (2003) Testing the Gaussian copula hypothesis for financial assets dependences. Quant Financ 3:231–250
Meyn SP, Tweedie RL (1993) Markov chains and stochastic stability. Springer, London
Nelder J, Mead R (1965) A simplex method for function minimization. Comput J 7:308–313
Nelsen RB (2006) An introduction to copulas. Springer, New York
Panchenko V (2005) Goodness-of-fit test for copulas. Phys A Stat Mech Appl 355:176–182
Patton AJ (2011) Copula methods for forecasting multivariate time series. In: Elliott G, Granger C, Timmermann A (eds) Handbook of economic forecasting, vol 2. Elsevier, Oxford
Presnell B, Boos DD (2004) The IOS test for model misspecification. J Am Stat Assoc 99:216–227
Quessy JF (2016) Consistent nonparametric tests for detecting gradual changes in the marginals and the copula of multivariate time series. Stat Pap. https://doi.org/10.1007/s00362-016-0846-8
Rémillard B, Papageorgiou N, Soustra F (2012) Copula-based semiparametric models for multivariate time series. J Multivar Anal 110:30–42
Scaillet O (2007) Kernel based goodness-of-fit tests for copulas with fixed smoothing parameters. J Multivar Anal 98:533–543
ŞErban M, Brockwell A, Lehoczky J, Srivastava S (2007) Modelling the dynamic dependence structure in multivariate financial time series. J Time Ser Anal 28:763–782
Smith MS (2015) Copula modelling of dependence in multivariate time series. Int J Forecast 31:815–833
Trapletti A, Hornik K (2019) tseries: time series analysis and computational finance. R package version 0.10-47. https://CRAN.R-project.org/package=tseries
Wang W, Wells M (2000) Model selection and semiparametric inference for bivariate failure-time data. J Am Stat Assoc 449:62–72
White H (1982) Maximum likelihood estimation of misspecified models. Econometrica 50:1–25
Wooldridge JM (1994) Estimation and inference for dependent processes. Handb Econom 4:2639–2738
Wong SF, Tong H, Siu TK, Lu Z (2017) A new multivariate nonlinear time series model for portfolio risk measurement: the threshold copula-based TAR approach. J Time Ser Anal 38:243–265
Yan J (2007) Enjoy the joy of copulas: with a package copula. J Stat Softw 21:1–21
Zhang S, Okhrin O, Zhou QM, Song P (2016) Goodness-of-fit test for specification of semiparametric copula dependence models. J Econom 193:215–233
Zhang S, Song P, Shi D, Zhou QM (2012) Information ratio test for model misspecification on parametric structures in stochastic diffusion models. Comput Stat Data Anal 56:3975–3987
Zhou QM, Song P, Thompson ME (2012) Information ratio test for model misspecification in quasi-likelihood inference. J Am Stat Assoc 107:205–213
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendix
Appendix
1.1 Appendix I: Proof of theorems in Section 3
Proof of Theorem 1
First, by condition (A1), applying law of large number for stationary ergodic time series (DasGupta 2008), we have
By conditions (A2), we have
Since \(\Vert {\hat{\theta }}-\theta ^*\Vert {\mathop {\rightarrow }\limits ^{pr}} 0 \) and \(\Vert G_n - G^*\Vert _{\infty } {\mathop {\rightarrow }\limits ^{pr}} 0\), as \(n\rightarrow \infty \), we have
Following similar arguments as above, we also can prove that
It follows from Assumption (A3) and Slutsky’s Theorem that
\(\square \)
To prove Theorem 2, we need the following Lemma 1, which gives the expansion of the estimated negative sensitivity matrix \(S_n\) and estimated variability matrix \(V_n\). Let \(S_n({\hat{U}}_{t-1}, {\hat{U}}_{t}; {\hat{\theta }})_{ij}\) denote the (i, j)th element of \(S_n\), and \(V_n({\hat{U}}_{t-1}, {\hat{U}}_{t}; {\hat{\theta }})_{ij}\) denote the (i, j)th element of \(V_n\).
Lemma 1: Under the conditions (A2)–(A3) and (B1)–(B3), we have
and
where \(C^1_{ij} \triangleq \mathbb {E}_{0}\left[ \frac{\partial \ddot{\ell }\left\{ U_0,U_1 ; \theta ^{*}\right\} _{i j}}{\partial \theta ^{T}}\right] \), \(C^2_{ij} \triangleq 2\mathbb {E}_{0}\left[ \dot{\ell }( U_{t-1},U_{t}, \theta ^*)_i\frac{\partial {\dot{\ell }}\left\{ U_0,U_1 ; \theta ^{*}\right\} _{j}}{\partial \theta ^{T}}\right] \), and
with
Proof of Lemma 1
First, we prove Eq. (8). Denote the pathwise derivative of \(\ddot{\ell }( G(Y_{t-1}), G(Y_{t}); \theta ^*) \) w.r.t. G in the direction of dG as \(\nabla \ddot{\ell }( G(Y_{t-1}),G(Y_{t}); \theta ^*)[dG]\), where \(dG = G-G^*\). Following similar arguments of Proposition 4.3 in Chen and Fan (2006b), we have
Expanding \(\ddot{\ell }( U_{t-1},U_{t};\theta ^*)\) around \(\theta ^*\) and \(G^*\) in the direction dG, we obtain
By Proposition 4.3 in Chen and Fan (2006b), we have
By condition (B3), employing similar arguments in Theorem 1, we have
By proposition 4.3 in Chen and Fan (2006b), we obtain
Hence, we have
Applying similar arguments as above, we can prove
\(\square \)
Proof of Theorem 2
Under the null hypothesis of the copula model being correctly specified, the Bartlett identity holds, hence, we have \({{\mathcal {S}}} = {{\mathcal {V}}}\), i.e. \({{\mathcal {S}}}^{-1}{{\mathcal {V}}} =I_p\). The test statistic R given in Eq. (7) can be represented as follows:
By Lemma 1, we have
and
Thus, we can write \(\sqrt{n}\left( R-p\right) = \frac{1}{\sqrt{n}}\sum \limits _{t = 1}^{n} h_R(U_{t-1},U_{t};\theta ^*)\), where
Hence, by condition (B3) and central limit theorem for time series, we obtain
where asymptotic variance \(\sigma _R^2 = Var\left[ h_R(U_{t-1},U_{t};\theta ^*)\right] \), which can be computed consistently via the parametric bootstrap method in Sect. 4.2. \(\square \)
To prove Theorem 3, we need the following Lemma 2, which states the limiting difference between the parameter estimates \({\hat{\theta }}\) and \({\hat{\theta }}_{-t}\) or \(\tilde{\theta }_{-t}\).
Lemma 2 : Under the conditions (B3), we have
Proof of Lemma 2
First, we prove Eq. (10). By Eq. (3), \({\hat{\theta }}_{-t}\) solves the following equation
Expanding \({\dot{\ell }}({\hat{U}}_{s-1},{\hat{U}}_{s};{\hat{\theta }}_{-t})\) around \({\hat{\theta }}\) leads to
where \({\check{\theta }}_{-t}\) lies between \({\hat{\theta }}\) and \({\hat{\theta }}_{-t}\). It follows that
By conditions (B3),
In addition, by condition (B3), using the similar arguments in the proof of Theorem 1, we can show
Hence,
To approximate \({\hat{\theta }}_{-t}\), shown in Eq. (5), we use \({\tilde{\theta }}_{-t}\) which is defined as
By the similar arguments above, we also can show that
\(\square \)
Proof of Theorem 3
By definition, \(T=\sum _{t = 1}^{n}\ell ({\hat{U}}_{t-1},{\hat{U}}_{t}; {\hat{\theta }}) - \sum _{t = 1}^{n}\ell ({\hat{U}}_{t-1},{\hat{U}}_{t}; {\hat{\theta }}_{-t})\). Expanding \(\ell ({\hat{U}}_{t-1},{\hat{U}}_{t}; {\hat{\theta }}_{-t})\) around \({\hat{\theta }}\) leads to
where \({\check{\theta }}_{-t}\) lies between \(\theta \) and \({\hat{\theta }}_{-t}\). Lemma 1 implies that
In summary, we prove that \(T - R = O_p(n^{-1})\). Similarly, we can prove \({\tilde{T}} - R = O_p(n^{-1})\). \(\square \)
1.2 Appendix II: Copula families
For the sake of self-containedness, below we present some brief descriptions of the four copula families: Clayton, Frank, Gaussian and Student’s t, considered in the simulation studies. Figure S11 shows the scatter plots of 1000 observations generated from the four copula families with Kendall’s \(\tau = 0.5\). In addition, we provide the explicit expression of \(C_{2 | 1}^{-1}\left( x | u_1, \theta \right) \), which is used to generate the pseudo-observations of a univariate time series from the assumed copula. Note that this function is the inverse of the function \(C_{2 | 1}\left( u_2 | u_1, \theta \right) \) described in Sect. 4.1, which is the conditional cdf of \(U_{t}\) given \(U_{t-1}=u\).
-
1.
Clayton copula (Clayton 1978)
$$\begin{aligned} C(u_1,u_2; \theta ) = \left( u_1^{-\theta }+u_2^{-\theta }-1\right) ^{-1/\theta }, \end{aligned}$$where the dependence coefficient \(\theta \ge 0\), which corresponds to Kendall’s \(\tau = \theta /(\theta +2)\). This copula represents lower tail dependency and is of particular interest in modeling the value-at-risk data. In addition, \(C_{2 | 1}^{-1}\left( x | u_{1} ; \theta \right) =\left[ \left( x^{-\theta /(1+\theta )}-1\right) u_{1}^{-\theta }+1\right] ^{-1 / \theta }\).
-
2.
Frank copula (Genest 1987)
$$\begin{aligned} C(u_1,u_2; \theta ) = -\theta ^{-1}\log \left( 1+ \frac{[e^{-\theta u_1}-1][e^{-\theta u_2}-1]}{e^{-\theta }-1}\right) , \end{aligned}$$(12)where \(\theta \ne 0\) and the corresponding Kendall’s tau is \(\tau = 1-4(D(\theta ) -1)/\theta \), \(D(\theta ) = \frac{1}{\theta }\int _{0}^{\theta }\frac{t}{e^t-1}dt\) for \(\theta >0\) and \(D(\theta ) = \frac{1}{\theta }\int _{0}^{\theta }\frac{t}{e^t-1}dt + 0.5\theta \) for \(\theta <0\). In addition, \(C_{2 | 1}^{-1}\left( x | u_{1} ; \theta \right) = - \theta \log \left[ 1-\frac{1-e^{-\theta }}{1+e^{-\theta u_1}(1/x-1)}\right] \).
-
3.
The bivariate Gaussian copula is defined as
$$\begin{aligned} C(u_1,u_2;\theta ) = \Phi _2(\Phi ^{-1}(u_1), \Phi ^{-1}(u_2); \rho ), \end{aligned}$$where \(\Phi (x)\) is the standard univariate Gaussian distribution, \(\Phi _2(x_1,x_2; \rho )\) is the bivariate Gaussian distribution with zeros mean and unit variance, and \(\rho \in [-1,1]\) is the correlation coefficient. Here, the copula parameter \(\theta =\rho \), corresponding to Kendall’s \(\tau = \frac{2}{\pi }\arcsin (\theta )\). In addition, \(C_{2 | 1}^{-1}\left( x | u_{1} ; \theta \right) = \Phi \left( \theta \Phi ^{-1}(u_1)+\sqrt{1-\theta ^2}x\right) \).
-
4.
The Student’s t copula is a two-parameter copula, which is given by
$$\begin{aligned} C(u_1,u_2; \nu ; \theta )= & {} \int _{-\infty }^{t_{\nu }^{-1}(u_1)}\int _{-\infty }^{t_{\nu }^{-1}(u_2)}\frac{1}{\sqrt{2\pi (1-\rho ^2)^{1/2}}}\\&\times \,\left( 1+\frac{x^2-2\rho xy + y^2}{\nu (1-\rho ^2)}\right) ^{-\nu /2-1}dxdy, \end{aligned}$$where the copula parameter \(\theta =(\rho ,\nu )\) with the correlation coefficient \(\rho \) and the d.f. \(\nu \). The Kendall’s \(\tau = \frac{2}{\pi }\arcsin (\rho )\). Student’s t copula has received increasing attention in the modeling of multivariate financial time series. Many empirical studies have shown that the Student’s t copulas fit better than Gaussian copulas since the former better captures the phenomenon of dependent extreme values, which is often observed in the financial asset return data (Breymann et al. 2003; Demarta and McNeil 2005). In addition, \(C_{2 | 1}^{-1}\left( x | u_{1} ; \theta \right) = F_{t_{\nu }}(\rho F^{-1}_{t_\nu }(u_1)+\sigma x)\), where \(F_{t_v}\) is the cdf of a t random variable with d.f. \(\nu \), and \(\sigma =\sqrt{\left[ \nu +F^{-1}_{t_\nu }(u_1)^2\right] (1-\rho ^2)/(\nu +1)}\).
Rights and permissions
About this article
Cite this article
Zhang, S., Zhou, Q.M. & Lin, H. Goodness-of-fit test of copula functions for semi-parametric univariate time series models. Stat Papers 62, 1697–1721 (2021). https://doi.org/10.1007/s00362-019-01153-4
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00362-019-01153-4