Can a machine learn from behavioral biases? Evidence from stock return predictability of deep learning models

Abstract

We examine how the return predictability of deep learning models varies with stocks’ vulnerability to investors’ behavioral biases. Using an extensive list of anomaly variables, we find that the long-short strategy of buying (shorting) stocks with high (low) deep learning signals generates greater returns for stocks more vulnerable to behavioral biases, i.e., small, young, unprofitable, volatile, non-dividend-paying, close-to-default, and lottery-like stocks. This performance of deep learning models for speculative stocks becomes pronounced when investor sentiment is high, and when new information is delivered through earnings announcements. Moreover, our nonlinear deep learning signals are negatively associated with analysts’ earnings forecast error especially for speculative stocks, implying that analysts’ forecasts are too low for speculative stocks with high deep learning signals. These results suggest that deep learning models with nonlinear structures are useful for capturing mispricing induced by behavioral biases.