Comparing Naive Bayes, LDA, and QDA for Chinese Multi-Factor Stock Selection

Summary

This research summary compares Naive Bayes, linear discriminant analysis (LDA), and quadratic discriminant analysis (QDA) as supervised methods for multi-factor stock selection. The described process extracts and preprocesses features and labels, retrains monthly, and uses time-ordered cross-validation so validation observations follow training observations rather than introducing future data. The models estimate next-period upside probabilities, which are assessed using classification accuracy, AUC, and portfolio backtests. The summary says longer training periods improved predictive results.

Reported industry-neutral backtests show Naive Bayes performed relatively well within the CSI 300 and CSI 500 universes, while LDA led in the broader A-share universe. LDA also generally achieved higher test accuracy and AUC. These are results as reported in the summary; the underlying report is linked but not reproduced here, and the excerpt gives limited detail about sample period, costs, and implementation assumptions. It notes that Gaussian Naive Bayes assumptions may be unrealistic despite reported robustness, and discusses factor correlations as consistent with LDA's assumptions.

Key ideas

• The study compares three supervised classification methods for multi-factor equity selection.
• Monthly retraining and time-ordered cross-validation are used to avoid training on future observations.
• The reported results favor Naive Bayes within major index universes and LDA in the broader A-share universe.
• LDA generally leads the alternatives on classification accuracy and AUC in the reported tests.
• The summary provides limited information about costs, test period, and implementation details.

Tags

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