This article compares the forecasting ability of the recently proposed Realized GARCH model with that
of the standard GARCH models that use only the daily returns, and the other time series models based
on the realized measures of volatility. Each model is used for forecasting the conditional variance of 16
international stock indices, for a sample period of about 14 years. We find that the relative forecasting
performance of the Realized GARCH and EGARCH models is sensitive to the choice of the loss criterion.
With the realized measures, the exponentially weighted moving average model generally outperforms
the Realized GARCH model in out-of-sample forecasts. This result is robust across different volatility
regimes and loss criteria