Modelling temporal dependence of realized variances with vines

Models for realized volatility that take the specific form of temporal dependence into account are proposed. Current popular methods use the idea of mixed frequencies for forecasting realized volatility, but neglect the potential non-linear and non-monotonic temporal dependence. The proposed approach utilizes vine copulas to mimic different memory properties. HAR, MIDAS and bivariate copulas, which can be seen as special cases of the suggested modeling framework, are chosen as benchmarks. All models are evaluated within an extensive empirical study both in- and out-of-sample and their forecasting ability is compared statistically. The results suggest that one specific vine copula construction is significantly superior over the considered benchmarks in modeling time dependencies of realized volatilities.