Mixture Pair-Copula-Constructions

We propose the use of convex combinations of parametric copulas as pair-copulas in high-dimensional vine copula models. By doing so, we circumvent the error-prone need to choose and estimate a parametric copula for each paircopula in a vine model. We show in simulations that our proposed model fits the dependence structure in a given data sample significantly better than a competing benchmark. In our empirical study on the models' accuracy for forecasting the Value-at-Risk of financial portfolios, we show that our proposed mixture pair-copula construction yields significantly better results in backtesting while the benchmark overestimates portfolio risk