Local Predictability in High Dimensions

Abstract

We propose a dynamic subset forecast combination approach to investigate local predictability of aggregate stock returns in high dimensions. Our method accommodates a vast number of heterogeneous predictive signals and is designed to pick up structural changes. Based on more than 12,000 indicators, we find that text-based indicators have provided valuable signals for daily aggregate stock returns over the last two decades, a period where the predictive signals of economic indicators have weakened.

Keywords: Aggregate Stock Returns, Dynamic Subset Forecast Combination, Textual Data, Estimation Error

JEL Classification: G12; G17; C10