Realized GARCH models: Simpler is better

Within the framework of Realized GARCH (RealGARCH), different RealGARCH variants have been proposed for volatility forecasting. The question remains unknown that which RealGARCH variant is more efficient. This paper compares three RealGARCH variants including the log-linear RealGARCH, the RealEGARCH and the GARCH@CARR. A comprehensive empirical study is performed on a stock index and 28 individual stocks, and the results show that the GRACH@CARR model outperforms the other two. Given that GARCH@CARR is more parsimonious in its specification, this finding is consistent with the principle of parsimony that models of simple structure usually provide better forecasts than the complex ones.