Dissecting Momentum: We Need to Go Deeper

Abstract

Cross-sectional predictability of returns by past prices, or momentum, is a lasting market anomaly. Previous research reports numerous ways to measure momentum and establishes a multitude of factors predicting its performance. The emerging machine learning asset pricing literature further identifies price-based firm characteristics as major predictors of returns. I investigate predictive power of a broad set of price-based variables over various time horizons in a deep learning framework and document rich non-linear structure in impact of these variables on expected returns in the US equity market. The magnitude and sign of the impact exhibit substantial time variation and are modulated by interaction effects among the variables. The degree of non-linearity in expected returns varies over time and is highest in distressed markets. Incorporating insights from the literature on time-varying, market state-dependent momentum risks and momentum crashes helps to improve out-of-sample performance of neural network portfolios, especially with respect to the downside risk -- investment strategies built on predictions of the deep learning model actively exploit the non-linearities and interaction effects, generating high and statistically significant returns with a robust risk profile and their performance virtually uncorrelated with the established risk factors including momentum. Lastly, I make a case for adoption of automated hyperparameter optimization techniques as an important component of disciplined research in financial machine learning.

Keywords: Momentum, Return Predictability, Time-Varying Risk, Machine Learning, Deep Learning, Neural Networks, Bayesian Optimization

JEL Classification: G11, G12, G17, C45