Testing SVM Classifiers Against Majority-Class Baselines in FX

Summary

The article introduces support vector machines for classifying next-day currency returns. It explains maximum-margin boundaries, soft margins that allow some classification errors, and the kernel trick for representing nonlinear boundaries in a higher-dimensional feature space. It then describes a model using currency price and lagged-return features, a degree-three kernel, and grid search with three-fold cross-validation to select parameters.

The reported comparison finds that neither linear regression nor the SVM reliably beats a 50% random-guess benchmark across currencies. Repeated random train-test splits show occasional SVM advantages, including a stated average directional accuracy of 54% for USD/JPY. However, comparison with a baseline that always predicts the most frequent training-set class shows that much of the apparent accuracy comes from class imbalance; the SVM adds only about two percentage points of nonlinear predictive information. The article notes that this information is difficult to interpret and offers no evidence that the results translate into profitable trading after costs or across other periods.

Key ideas

• SVM classifiers choose a decision boundary that balances a wide margin against penalties for misclassified observations.
• Kernel functions allow nonlinear class boundaries without explicitly computing a high-dimensional feature mapping.
• The study tests next-day currency direction using price and lagged-return features, tuning the model with cross-validation.
• Accuracy should be compared with a majority-class baseline because class imbalance can make naive benchmarks misleading.
• The reported SVM improvement is modest and difficult to interpret, and directional accuracy alone does not establish trading profitability.

Tags

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