Title: Vine based modeling of multivariate volatility time-series
Authors: Nicole Barthel - Technische Universitaet Muenchen (Germany) [presenting]
Claudia Czado - Technische Universitaet Muenchen (Germany)
Yarema Okhrin - Universitaet Augsburg (Germany)
Abstract: Studying realized volatility based on high-frequency data is of particular interest in asset pricing, portfolio management and evaluation of risk. We propose an approach for dynamic modeling and forecasting of realized correlation matrices that allows for model parsimony and automatically guarantees positive definiteness of the forecast. We use the one-to-one relationship between a correlation matrix and its associated set of partial correlations corresponding to any regular vine specification. Being algebraically independent the partial correlations of the vine do not need to satisfy any algebraic constraint such as positive definiteness. We present several selection methods to choose, among the large number of vines, the vine structure best capturing the characteristics of the multivariate time-series of the correlation parameters. The individual partial correlation time-series are studied using ARFIMA and HAR models to account for long-memory behavior. The dependence between assets is flexibly modeled using vine copulas that allow for nonlinearity and asymmetric tail behavior. Correlation point forecasts are obtained as nonlinear transformation of the forecasts for the partial correlation vine. The usefulness of the methodology is investigated in a one-day ahead forecasting framework comparing existing methods based on a model confidence set approach.