LASSO Regression on Asset Pricing Factor Selection

This paper investigates the problem of redundant factors in asset-pricing models. The excessive inclusion of factors may result in model overspecification and multicollinearity, thereby increasing estimation uncertainty and reducing the accuracy of asset-return predictions. To address this issue, the study proposes a factor-selection procedure based on the least absolute shrinkage and selection operator (LASSO) regression model. The LASSO objective function imposes a penalty on model complexity, discouraging the inclusion of redundant factors and shrinking the coefficients of irrelevant or weakly informative factors toward zero. The proposed procedure was evaluated using monthly data from the Chinese A-share market and successfully selected seven factors from a set of 29 commonly used factors. Nevertheless, in predicting the absolute levels of individual stocks' excess returns for the following month, the model did not significantly outperform the naïve historical-mean benchmark, as indicated by an out-of-sample R-squared value slightly below zero. However, the positive mean rank information coefficient (Rank IC) suggests that the LASSO model retains some ability to identify cross-sectional differences in expected stock returns. Therefore, although its point forecasts remain limited, the model may still provide useful information for stock ranking, portfolio construction, and investment decision-making.