The heterogeneous autoregressive (HAR) models are used in modeling high frequency multipower realized volatility of the
S&P 500 index. Extended from the standard realized volatility, the multipower realized volatility representations have
the advantage of handling the possible abrupt jumps by smoothing the consecutive volatility. In order to accommodate
clustering volatility and asymmetric of multipower realized volatility, the HAR model is extended by the threshold
autoregressive conditional heteroscedastic (GJR-GARCH) component. In addition, the innovations of the multipower realized
volatility are characterized by the skewed student-t distributions. The extended model provides the best performing insample
and out-of-sample forecast evaluations.