An Economic Regularization for Cross-Sectional Return Prediction

Thomas M. Menino Convention & Exhibition Center 

Cross-sectional stock return prediction is a high-dimensional problem characterized by weak signals, strong collinearity, and heavy-tailed outcomes. In such settings, model performance often depends less on flexibility than on how regularization is imposed. This paper studies the role of economically motivated structure as a regularization device, treating thirteen economic themes as exogenous organizing principles used exclusively to discipline shrinkage rather than to identify new predictive mechanisms. We estimate an additive spline model with a theme-structured Sparse Group LASSO that applies shrinkage both across and within these themes. Using a strict rolling out-of-sample forecasting design on U.S. equities from 2013--2023, conventional forecast metrics such as out-of-sample R^2 and MAE ratios provide limited separation across structured and unstructured penalties. In contrast, structured regularization consistently yields stronger long--short decile portfolio performance --- measured by Sharpe ratio and cumulative return --- where portfolios are formed by sorting stocks into deciles based on predicted returns, analogous to standard anomaly portfolio construction. Theme-level exposures are also more coherent under structured penalties, with results robust to excluding the January 2021 meme-stock episode. The findings suggest that in weak-signal environments, economic structure adds value by disciplining regularization and shaping how predictive weight is allocated --- not by improving statistical forecast accuracy.

High-dimensional data and Regularization

Model interpretability / Statistical learning

Sparse Group LASSO

Financial economics (Asset Pricing)

Portfolio construction and performance

Additive models / splines 

Business and Economic Statistics Section