Factor Investing with Classification-Based Supervised Machine Learning

There are two types of supervised machine learning (SML): regression and classification. In this study, the authors propose classification-based machine learning algorithms for factor investing with artificial neural networks in which the cross section of stock returns is grouped into five categories: strong buy, buy, neutral, sell, and strong sell. Their empirical out-of-sample results demonstrate some advantages of classification-based machine learning relative to regression-based learning in which the actual stock returns denote the response variable. The classification-based models also deliver slight outperformance relative to the ordinary least squares model, although the outperformance is not statistically significant. Furthermore, the out-of-sample results show that “deep” learning with multilayers of neuron layers cannot outperform a less sophisticated “shallow” learning for both classification-based and regression-based SML algorithms. Their findings suggest that market noise, common in the financial markets, during the training process overwhelms the nonlinear association uncovered in the machine learning process; and the classification of the cross section of stock returns may have reduced some of the noise.