Abstract

Corrected Dynamic Conditional Correlation (cDCC) model has been largely used to model multivariate returns of financial time series. For large dimensions, composite likelihood method has been used to estimate the model. We consider several issues related to this method. The main conclusion are: i) the likelihood function can become very flat for moderate dimension; ii) we can obtain gain in terms of mean square error by increasing the number of series permutations until the MSE is dominated by the bias; and iii) Comparing two ways to consider the permutation, the log-likelihood as the sum of likelihood of different permutations and the estimator as a function of the estimates of each permutation, none of than showed to dominate the other. Finally, we study the robustness of the estimator to outliers and found out that none of them are robust.