Investable and Interpretable Machine Learning for Equities

Abstract

We propose three principles for evaluating the practical efficacy of machine learning for stock selection, and we compare the performance of various models and investment goals using this framework. The first principle is investability. To this end, we focus on portfolios formed from highly liquid US stocks, and we calibrate models to require a reasonable amount of trading. The second principle is interpretability. Investors must understand a model’s output well enough to trust it and extract some general insight from it. To this end, we choose a concise set of predictor variables, and we apply a novel method called the Model Fingerprint to reveal the linear, nonlinear, and interaction effects that drive a model’s predictions. The third principle is that a model’s predictions should be interesting, by which we mean they should convincingly outperform simpler models. To this end, we evaluate performance out-of-sample compared to linear regressions. In addition to these three principles, we also consider the important role people play by imparting domain knowledge and preferences to a model. We argue that adjusting the prediction goal is one of the most powerful ways to do this. We test random forest, boosted trees and neural network models for multiple calibrations which we conclude are investable, interpretable, and interesting.