Modelling volatility in turbulent times: a Robust Realized GARCH framework

Standard Realized GARCH models are sensitive to outliers, which can distort volatility persistence estimates. We propose a robust Realized GARCH framework to address this issue in high-frequency financial time series. By incorporating a bounded influence function for the innovation terms, we dampen the impact of extreme observations in both the return and measurement equations. This approach provides a parsimonious and computationally tractable solution that preserves the information content of realized measures while filtering out noise. Using Realized Kernel estimates derived from intraday S&P 500 data, we evaluate the model's performance against standard specifications via quasi maximum likelihood estimation. We demonstrate that our specification achieves superior statistical fit, particularly during turbulent periods. To illustrate the model's practical utility, we use it to disentangle genuine volatility clusters from microstructure anomalies during the Global Financial Crisis and the COVID-19 pandemic. Notably, the robust framework correctly identifies the largest volatility shock of the 2007-2009 crisis (February 27, 2007) as a technical trading glitch, distinguishing it from fundamental economic news.