Mosaics of Predictability

Abstract

Existing studies on asset return predictability focus on aggregate performance. We examine the oft-overlooked grouped heterogeneity in return predictability across different assets and macroeconomic regimes. A novel tree-based asset clustering methodology is introduced to partition the panel of asset-return observations according to return predictability, using high-dimensional asset characteristics and aggregate time series predictors. When implemented on U.S. equities over the past five decades, we find that some characteristics-managed (dollar trading volumes, unexpected earnings, earnings-to-price, and cashflow-to-price) and/or macro-based (dividend yield and default yield) clusters are more predictable, resulting in a heterogeneous predictive model with outperformance. Finally, less predictable clusters generally exhibit lower risk-adjusted investment performance, revealing an important empirical link between return predictability and trading profitability.