Forecasting volatility with outliers in Realized GARCH models

The Realized generalized autoregressive conditional heteroskedasticity (GARCH) model proposed by Hansen is often applied to forecast volatility in high‐frequency financial data. It is frequently found, however, that the distribution of the estimated residuals from Realized GARCH models has peak fat‐tail characteristics. Considering this feature may be a result of neglected additive outliers (AOs) and innovative outliers (IOs), this paper proposes the Realized GARCH model with additive outlier and innovative outlier (Realized GARCH‐AI model) for forecasting volatility. This model can detect and correct abnormal returns and realized volatility by estimating the coefficients of volatility models and calculating the outlier test statistics. In the process of simulation, this paper considers different outlier cases in the GARCH model and the Realized GARCH model, and evaluates the performance of the proposed procedure through the accuracy of parameter estimation under different critical values. We find that the critical values will affect the results of outlier detection and correction. When the value is in a suitable range, the proposed procedure based on high‐frequency data can obtain unbiased parameter estimation, and the estimation result is close to those of the intervention model containing outlier information. Finally, we use the MCS test proposed by Hansen et al. (Econometrica, 2011, 79(2), 453–497) to study the volatility prediction accuracy, value at risk, and expected shortfall prediction ability of the new model. The empirical analysis demonstrates that the proposed model produces better prediction effects than the traditional Realized GARCH model.