SVM Classification and Stock Selection

Article SuperMind

Summary

The article explains support vector machines (SVMs) as classifiers that seek a separating boundary between classes. It introduces the geometric intuition of mapping observations into a higher-dimensional feature space, where a nonlinear boundary in the original space may become linear. For cases that remain imperfectly separable, it describes slack variables as a way to allow margin violations while penalizing errors. The kernel trick computes inner products in the transformed space through a kernel function, avoiding explicit construction of potentially large feature vectors; kernel choice remains an important modeling decision.

For a stock-selection example, the article labels stocks by whether their prior-month return was positive or negative, standardizes financial features, trains a classifier, and buys stocks predicted to rise at the next rebalance. It reports that the SVM lagged decision-tree and random-forest approaches in the described backtest, with relative outperformance over the benchmark appearing only later in the period. The account calls this a preliminary strategy and notes that parameter tuning and feature research are important; it gives no detailed performance statistics or controls for look-ahead bias and transaction costs.

Key ideas

SVMs classify observations by finding a boundary with a wide margin between classes.
Higher-dimensional mappings can make nonlinear class boundaries separable by a linear boundary.
Slack variables permit margin violations and penalize misclassified or borderline observations.
Kernel functions avoid explicitly computing high-dimensional feature mappings, but kernel selection matters.
The stock-selection example uses lagged financial features and prior-month return direction as its training label.
The reported backtest is preliminary and identifies tuning as a limitation.

Support Vector Machines for Classification and Stock Selection

Summary

Key ideas

Tags