Abstract

The aim of this paper is to compare the performance of the daily nonlinear support vector machines, the new semi-parametric tool for regression estimation, heterogeneous autoregressive (SVM-HAR)-ARCH type models based on the daily realized volatility (which uses intraday returns) with the performance of the classical HAR-ARCH type models by using different innovation distribution when the one-day ahead value-at-risk (VaR) is to be computed. The daily realized volatility is calculated using 5-, 15-min and optimally sampled intraday returns for Nikkei 225 index. This paper shows that the particular hybrid SVM-HAR-ARCH type model provides better performance when 15-min intraday returns are used. This paper also shows that the models based on a long memory skewed student distribution provide the better performance of one-day ahead value-at-risk forecasts.

Key words: Value-at-risk, HAR-RV model, nonlinear support vector machine-HAR-RV model, ARCH type models, Skewed student distribution, high frequency Nikkei 225 data.