Social media sentiment, model uncertainty, and volatility forecasting

Abstract

Many economic indicators including consumer confidence indices used to forecast volatility or macroeconomic outcomes, are published with a considerable time lag. To obtain a timelier measure of consumer sentiment many central bank and economic researchers are turning towards using state-of-the-art text sentiment analysis tools. We examine if there are benefits for forecasting volatility from (i) incorporating a sentiment measure derived using deep learning from Twitter messages at the 1-min level, and (ii) acknowledging specification uncertainty of the lag index in the heterogeneous autoregression (HAR) model. We present evidence from an out of sample forecasting exercise that suggests including social media sentiment can significantly improve the forecasting accuracy of a popular volatility index, particularly in short time horizons. Further, our results document large gains in predictive accuracy from a newly proposed estimator that allows for model uncertainty in the specification of the lag index when using a HAR estimator.