This paper pioneers the use of multi-class machine learning classifiers for cross-sectional stock return prediction, framing the problem as a classification task rather than the traditional regression. The authors employ Artificial Neural Networks, Random Forest, and Gradient Boosting Trees to assign stocks to three categories: outperformers (top decile), underperformers (bottom decile), and midperformers. The resulting long-short portfolios, which buy predicted outperformers and sell underperformers, deliver exceptional out-of-sample performance with annual Sharpe ratios of 1.67 (value-weighted) and 3.35 (equal-weighted), and annual alphas ranging from 29% to 48%. These results are robust to controlling for machine learning regressions and focusing on large-cap stocks, and they demonstrate that classification models provide complementary information to regression-based approaches.

A key innovation is the introduction of 'machine uncertainty,' measured as the information entropy of predicted probabilities across return states. This metric quantifies the model's assessment of information scarcity. Using Fama-MacBeth regressions, the authors show that higher machine uncertainty significantly reduces prediction correctness: a one-bit increase leads to a 9% decline in correct predictions. Moreover, machine uncertainty negatively predicts future stock returns in the cross-section, with a per-bit increase associated with a 1.2% decrease in annual returns. This effect is primarily driven by past underperformers, consistent with theories of slow information diffusion and investor inattention. The study also finds that the negative impact of machine uncertainty on prediction accuracy is amplified when it aligns with human proxies of information uncertainty, such as analyst forecast dispersion.

The paper contributes to the machine learning and asset pricing literature by demonstrating the viability and economic value of classification-based return prediction, and by providing a novel measure of machine uncertainty that sheds light on the information-return relationship. The findings suggest that machine learning classifiers can capture nonlinear patterns and state-dependent dynamics that regressions may miss, offering practical implications for portfolio management and risk assessment.