Kernel Support Vector Machines for Chinese A-Share Stock Selection

Summary

The report compares linear, polynomial, Gaussian, and sigmoid support vector machine classifiers, along with support vector regression, for multi-factor stock selection. Its workflow extracts features and labels, preprocesses inputs, trains and tunes models through cross-validation, then evaluates predictions on an out-of-sample set. Monthly predictions are used to build industry-neutral portfolios from CSI 300 stocks, CSI 500 stocks, and the broader A-share universe.

The reported tests generally favor the Gaussian-kernel SVM over the other kernels and linear regression on predictive accuracy, AUC, and portfolio returns or information ratios. The report also finds classification-based SVM strategies outperform support vector regression in backtests. However, SVM does not show a clear advantage in maximum drawdown. These findings come from the report's historical sample and portfolio setups; they do not establish that the results will persist in other periods or markets.

Key ideas

• The study compares several SVM kernel choices and support vector regression for multi-factor stock selection.
• Its workflow combines feature preprocessing, in-sample training, cross-validation, and out-of-sample evaluation.
• Gaussian-kernel SVMs generally lead the tested alternatives in prediction metrics and portfolio performance.
• The reported SVM strategies outperform linear regression on returns and information ratios, but not clearly on maximum drawdown.
• The reported historical backtests do not guarantee future performance.

Tags

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