Goodness-of-fit test for specification of semiparametric copula dependence models

This paper concerns goodness-of-fit tests for semiparametric copula models. Our contribution is two-fold: we first propose a new test constructed via the comparison between “in-sample” and “out-of-sample” pseudo-likelihoods. Under the null hypothesis that the copula model is correctly specified, we show that the proposed test statistic converges in probability to a constant equal to the dimension of the parameter space. We establish the asymptotic normality and investigate the local power of the test. We also extend the proposed test to the specification test of a class of multivariate time series models, and propose a new bootstrap procedure to establish the finite-sample null distribution, which is shown to have better control of type I error than the commonly used bootstrap. Secondly, we introduce a Bonferroni-based hybrid mechanism to combine several test statistics, which yields a useful test. This hybrid method is particularly appealing when there exists no single dominant optimal test. We conduct comprehensive simulation experiments to compare the proposed new test and hybrid approach with two of the best “blanket” tests in the literature. For illustration, we apply the proposed tests to analyze two real datasets.