Forecasting stock return volatility: Realized volatility-type or duration-based estimators

In this paper, we study the predictive performance of two kinds of volatility estimators: the realized volatility (RV) type and duration-based ones. This is motivated by the theoretical and empirical support for these distinct estimators. We use intraday data for 218 component stocks of the CSI 300 index in the Chinese equity market from 2010–2019 and perform in- and out-of-sample 1-, 5-, and 22-day ahead volatility forecasts from combinations of volatility models and these estimators. We show that, although empirically more efficient with the US data, the duration-based estimators fail to compete statistically, or in terms of economic value, with RV-type ones in the Chinese market. We perform a comprehensive set of simulations to rationalize these results and show that duration-based estimators underperform as they cannot handle the occasional heightened level of volatility in the Chinese market.