The power of news data in forecasting tail risk: evidence from China

This study investigates whether the inclusion of news information can help predict tail risk in the Chinese market. To quantify information from news, we develop an information volume proxy and two sentiment proxies. We employ the asymmetric GJR-GARCH model, along with different specifications for the error distributions and for the exogenous variables in the conditional volatility dynamics. It results in 24 models, including 18 models with three kinds of exogenous variables and six baseline models without exogenous variables. Subsequently, we assess the extent to which the 18 models with exogenous variables improve value-at-risk (VaR) forecasts for various Chinese industries compared to the six baseline models, utilizing both the backtesting and Model Confidence Set (MCS) procedures. The findings indicate that incorporating information volume as exogenous variables significantly enhances the VaR predicting abilities of the models, whereas incorporating sentiment variables does not. Moreover, we consider three different methods for combining VaR predictions and find that employing the MCS procedure to filter models first, and subsequently assigning weights to the selected models based on the reciprocal of their MCS ranking values, yields the best performance among the three combination approaches.