The Cross-Section of Stock Returns: Rank-Based, Linear, and Non-Linear Models (slides) by Matéo Molinaro :: SSRN

This study empirically compares rank-based, linear, and non-linear models for cross-sectional stock return prediction and factor portfolio construction in the U.S. equity market. Using fundamental and market data for Russell 1000 constituents from Bloomberg, we construct over 40 firm-level characteristics, standardized and winsorized to reduce outlier effects.  We evaluate three main approaches: Model-free strategies: univariate sorts and sequential filtering on individual characteristics. Linear models: LASSO, Ridge, and Elastic Net regressions. Non-linear models: Random Forests and Neural Networks. Portfolios are formed monthly by selecting the top decile of stocks ranked by the chosen characteristic or predicted return, with equal weighting and no transaction costs. Performance is measured out-of-sample using the Information Ratio (IR) and Excess Sharpe Ratio (eSR). Results indicate that model-free approaches, particularly quality and profitability-based factors, achieve strong economic performance with simplicity and interpretability. Sequential filtering slightly outperforms simple sorts. Among predictive models, LASSO and Random Forest show the highest risk-adjusted returns, although predictive R² values remain low and rank correlation with realized returns drives performance. We discuss the limitations of the short backtest period for model-based strategies, the potential of rank-based loss functions to improve predictive ordering, and directions for future research in cross-sectional asset pricing using machine learning.