Higher order conditional moment dynamics and forecasting value-at-risk (in Russian)

We empirically investigate the possibilities for enhancing value-at-risk predictions by explicit modelling conditional higher order moment dynamics of financial returns. Using one-day-ahead VaR forecasts for 5 highly liquid constituents of the S&P500 index from different industrial sectors, we compare performances of the benchmark GARCH model with skewed generalized Student's innovations with a set of models allowing for time-varying asymmetry and kurtosis such as ARCD-type models with normal inverse gaussian and skewed generalized Student's errors. As predictive accuracy tests we exploit both the scoring rules for left tail forecasts and likelihood-ratio tests for correct (un)conditional quantile forecasts. We also propose a parsimonious ARCD model with the skewed generalized error distribution for innovations, asymmetric power ARCH for volatility and autoregressive dynamics for skewness and kurtosis related parameters which is shown to perform not worse than the aforementioned models in terms of VaR prediction accuracy, while being computationally less demanding.