Predicting Equity Trading Volumes

Abstract

We construct a model for daily, stock-individual trading volumes to forecast the number of shares traded on a one-day horizon. The model combines short term trend, long term trend and exogenous effects, such as earnings announcements and holidays. Kalman filter techniques are used for the evaluation and the estimation of the model. Cross-sectional correlation between stocks' volume series is taken into account. As our large dataset covers the entire MSCI World Index universe, we run into computational issues in the estimation phase. We therefore use a combination of quasi maximum likelihood (QML) estimation and efficient recursive computation. Test statistics for the (QML) parameters are derived, to see if different countries, sectors or market cap groups have different trading volume dynamics. Test results show that regional differences are most significant. Results on the forecasting accuracy show that our model significantly outperforms models widely used in the industry, such as average daily volume (ADV). Furthermore, the results are robust in the cross-section as well as in the time dimension. We use cost functions to translate our volume forecasts to forecasts of trade impact costs. Confidence intervals are constructed and tested, for single stocks as well as for trade baskets. Finally, we use Thomson Reuters news sentiment data to show the relation between stock-related news items and trading volume.