Score-driven models for realized volatility

This paper sets up a statistical framework for modeling realized volatility (RV) using Dynamic Conditional Score (DCS) model. It first shows how, dataset on stock indices, preliminary analysis of RV, based fitting linear Gaussian model to its logarithm, suggests the use two component dynamic specification. also indicates departure from normality, weekly pattern in data and presence heteroscedasticity. Fitting DCS specification with leverage day week effect is then carried out directly RV Generalized Beta Second Kind (GB2) conditional distribution or, equivalently, logarithm an Exponential (EGB2) distribution. The forecasting performance this model, without heteroscedasticity, compared that Heterogeneous Autoregression (HAR), some extensions it other models. Overall there clear gain GB2-DCS even when HAR uses additional information, such as semi-variance. When aim forecast tail behavior, fat-tailed GB2 performs much better than models thin-tailed distributions. A further exercise exchange rates compare quarticity. Again information offers no advantage. are transparent, provide comprehensive description properties difficult beat forecasting.