SVM Stock Selection with Factor Features and Kernel Comparisons

Summary

This article introduces support vector machines for classification and regression, then applies them to A-share stock selection. It explains the maximum-margin principle for linear SVMs, slack variables for imperfectly separable observations, and kernel functions that make nonlinear classification possible. It compares linear, polynomial, sigmoid, and Gaussian or RBF kernels, and discusses how the penalty parameter C and gamma influence model fit, computational burden, and overfitting. Feature scaling and outlier handling are part of the described workflow.

The example builds features from 18 factors and labels stocks according to whether their forward five-day return falls in the top 5%. It trains on one historical interval, predicts a later period, and describes a weekly equal-weight portfolio of 50 predicted stocks with a minimum five-day holding period. The article reports that linear and RBF kernels beat the polynomial and sigmoid versions in its comparison, with RBF described as more stable and having an 8.7% maximum drawdown. These are results from the stated example, not evidence of general predictive power. The page also warns that its material targets an older platform version and identifies choices such as factors, labels, and parameters for further work.

Key ideas

• SVM classification seeks a wide separating margin, while slack variables allow limited errors in imperfect data.
• Kernel methods extend SVMs to nonlinear decision boundaries by using inner products in transformed feature spaces.
• The kernel choice and the parameters C and gamma influence fit, speed, and overfitting risk.
• The example uses 18 factor features and labels the top 5% of forward five-day stock returns as positive.
• In the reported backtest, linear and RBF kernels outperformed polynomial and sigmoid variants, with RBF showing the smallest stated maximum drawdown.
• The results depend on the example’s historical data and portfolio rules, and the article is for an older platform version.

Tags

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