A Study of Internet Search Volume's Contribution to Day-Ahead Volatility Forecasts

Abstract

We investigate whether day-ahead forecasts of individual stocks' volatility can be improved with Internet search volume data. We extend Heterogenous Autoregressive models of Realized Volatility (HAR-RV models) with past search volume data, and evaluate these models' forecasting performance. We find that short term search volume can improve forecasting performance for a subset of the companies in our sample. The improvement is greater if we isolate idiosyncratic volatility components for each company, from volatility components that can be explained by the market. This decomposition itself yields a significant improvement of forecasting performance. Using Google Trends' "company" filter and "investing" filter in place of simple search volume also increases forecasting performance, as does combining (averaging) forecasts based on different filters. Utilizing both a decomposition and a combination of forecasts based on filtered search volume extensions, we improve single stock volatility forecasts by an average of $4.9\%$ over the standard HAR-RV model. Additionally, we find that aggregate search volume of multiple companies can improve forecasts of market volatility.