GLMs for A-Share Stock Selection

Article BigQuant

Summary

This research summary examines generalized linear models as ways to combine stock factor exposures into forecasts of future returns. Its workflow includes extracting features and labels, preprocessing features, forming training sets, and retraining on a rolling basis. At each month end, the model generates forecasts for the next period; the combined score can be tested as a factor and used to construct portfolios with different industry-neutrality choices. The report evaluates models using backtests and test-set information coefficients or classification accuracy.

Reported sensitivity analyses favor rolling windows of 12–24 months, retaining more principal components, and different sample selections depending on the benchmark. Ridge, Lasso, and elastic net reportedly do not improve meaningfully on linear regression. Logistic regression and stochastic gradient descent classifiers outperform regression in the reported tests, with the hinge-loss classifier strongest; the summary gives favorable risk-adjusted metrics for one industry-neutral portfolio. These are historical results, not guarantees. The authors suggest binary return labels may reduce noise while discarding information, and warn that the models may stop working as market conditions change.

Key ideas

The study frames multi-factor stock selection as predicting future returns from current factor exposures.
Its monthly workflow preprocesses inputs, trains on rolling samples, and evaluates predictions and portfolio backtests.
Reported results vary with rolling-window length, principal-component retention, and sample selection.
Regularized regressions perform similarly to linear regression in the reported tests.
Classification models, especially the hinge-loss SGD model, reportedly outperform linear regression, but the findings are historical and may not persist.

Generalized Linear Models for China A-Share Multi-Factor Stock Selection

Summary

Key ideas

Tags