Are news important to predict large losses?

In this paper we investigate the impact of news to predict extreme financial returns using high frequency data. We consider several model specifications differing for the dynamic property of the underlying stochastic process as well as for the innovation process. Since news are essentially qualitative measures, they are firstly transformed into quantitative measures which are subsequently introduced as exogenous regressors into the conditional volatility dynamics. Three basic sentiment indexes are constructed starting from three list of words defined by historical market news response and by a discriminant analysis. Models are evaluated in terms of their predictive accuracy to forecast out-of-sample Value-at-Risk of the STOXX Europe 600 sectors at different confidence levels using several statistic tests and the Model Confidence Set procedure of Hansen et al. (2011). Since the Hansen's procedure usually delivers a set of models having the same VaR predictive ability, we propose a new forecasting combination technique that dynamically weights the VaR predictions obtained by the models belonging to the optimal final set. Our results confirms that the inclusion of exogenous information as well as the right specification of the returns' conditional distribution significantly decrease the number of actual versus expected VaR violations towards one, as this is especially true for higher confidence levels.