Forecasting volatility with empirical similarity and Google Trends

This paper proposes an empirical similarity approach to forecast weekly volatility by using search engine data as a measure of investors attention to the stock market index. Our model is assumption free with respect to the underlying process of investors attention and significantly outperforms conventional time-series models in an out-of-sample forecasting framework. We find that especially in high-volatility market phases prediction accuracy increases together with investor attention. The practical implications for risk management are highlighted in a Value-at-Risk forecasting exercise, where our model produces significantly more accurate forecasts while requiring less capital due to fewer overpredictions.