Quantile Machine Learning and the Cross-Section of Stock Returns

We investigate the predictability of future stock return quantiles using machine learning models trained on firm characteristics and macroeconomic variables, and find that multi-task neural networks dominate linear, tree-based, and feed-forward neural network models. We introduce a quantile-based risk premium measure, robust to outliers and heteroskedasticity, and demonstrate that it delivers significant predictive and economic gains for investors, outperforming leading machine learning techniques. We construct machine learning skewness and volatility measures, and find a strong positive relationship between conditional skewness and average returns. We develop a portfolio strategy that incorporates machine learning Sharpe ratio forecasts, leading to significant economic gains.