Penalized Linear Regression for Multi-Factor Stock Selection

Summary

This study applies Lasso, adaptive Lasso, and Elastic Net to select equity factors and combine them into stock-selection signals. Its motivation is that correlated predictors can make ordinary multivariate regression unstable. The reported experiments show that penalizing coefficients does not by itself remove multicollinearity: with 165 candidate factors, Lasso retained fewer than 30 in some fits, yet multicollinearity still appeared in more than 30% of test-period samples. The authors then combine penalized regression with a procedure for screening highly correlated factors.

For the resulting modified Elastic Net signal, the summary reports an information coefficient of 0.135, a t-statistic of 17.1, a 61.7% annualized long-short return, and a Sharpe ratio of 5.3, with an average of 25.3 factors selected per period. Training within the CSI 300, CSI 500, and CSI 800 universes is also reported to improve stability; the CSI 300-trained signal has a 0.075 information coefficient and 25.8% annualized long-short return within that universe. These are reported historical results, and the text does not provide enough methodological detail here to assess costs, validation design, or portability.

Key ideas

• Lasso, adaptive Lasso, and Elastic Net are used to select factors and build composite stock-selection signals.
• Coefficient penalties can reduce predictor influence, but the reported Lasso results still show frequent multicollinearity.
• The study adds a screen for highly correlated factors to modify the penalized regression process.
• It reports results for the modified Elastic Net signal and finds that training within narrower Chinese stock universes can improve reported performance stability.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.