Factor Investing: Hierarchical Ensemble Learning
29 Pages Posted: 6 Feb 2019
Date Written: January 31, 2019
We present a Bayesian hierarchical framework for both cross-sectional and time-series return prediction. Our approach builds on a market-timing predictive system that jointly allows for time-varying coefficients driven by fundamental characteristics. With a Bayesian formulation for ensemble learning, we examine the joint predictability as well as portfolio efficiency via predictive distribution. In the empirical analysis of asset-sector allocation, our hierarchical ensemble learning portfolio achieves 500% cumulative returns in the period 1998-2017, and outperforms most workhorse benchmarks as well as the passive investing index. Our Bayesian inference for model selection identifies useful macro predictors (long-term yield, inflation, and stock market variance) and asset characteristics (dividend yield, accrual, and gross profit). Using the selected model for predicting sector evolution, an equally weighted long-short portfolio on winners over losers achieves a 46% Sharpe ratio with a significant Jensen’s alpha. Finally, we explore an underexploited connection between classical Bayesian forecasting and modern ensemble learning.
Keywords: Hierarchical Model, Firm Characteristics, Market Timing, Portfolio Efficiency, Return Predictability, Risk Anomalies, Seemingly Unrelated Regressions
JEL Classification: C1, G1
Suggested Citation: Suggested Citation