The Anatomy of Machine Learning-Based Portfolio Performance

Abstract

Asset return predictability is routinely assessed by economic value: based on a set of predictors, out-of-sample return forecasts are generated—increasingly via “black box” machine learning models—which serve as inputs for portfolio construction, and performance metrics are computed over an evaluation period. We develop a methodology based on Shapley values—the Shapley-based portfolio performance contribution (SPPC)—to directly estimate the contributions of individual or groups of predictors to a performance metric. We illustrate the SPPC in an empirical application measuring the economic value of cross-sectional stock return predictability using a large number of firm characteristics and machine learning.

Keywords: Asset return predictability, Machine learning, Out-of-sample forecast, Economic value, Shapley value, XGBoost, Firm characteristics

JEL Classification: C53, C55, C58, G11, G17