Empirical Asset Pricing via Machine Learning: The Role of Research Design Choices

Abstract

We explore the impact of research design choices on the profitability of machine learning (ML) investment strategies. Specifically, we consider eight choices, including training window length, data filters and portfolio construction methods, across seven ML models. Based on 5376 portfolios, we find that design choices cause significant variation in strategy returns. The nonstandard errors (NSEs) of ML strategies are up to five times the standard errors (SEs) and remain large even after controlling for high-impact decisions such as eliminating micro-caps and using value-weighted portfolios. Nonetheless, ML still generates significant returns for about one-third of the portfolios even after transaction costs.