Cross-Validated Stepwise Feature Selection with Permutation Tests

Summary

The article describes an enhanced stepwise feature-selection method for supervised learning. Instead of retaining only one current feature set, it keeps several high-performing subsets and tests candidate additions against them, aiming to capture useful feature combinations without exhaustive search. Cross-validation scores guide selection and can trigger early stopping when adding predictors no longer improves performance. The procedure also uses shuffled-target replications to estimate the chance of apparent performance and of an apparent gain from a newly added feature.

An MQL5 implementation is presented as a model-agnostic base class that can be adapted to different learning methods, with an example application to regression. The text explains tuning choices such as retained candidate count, fold count, predictor limits, and permutation replications, but the provided excerpt gives no detailed numerical benchmark. Cross-validation and permutation checks can reduce some overfitting risks, yet results still depend on the data, model, validation design, and computational budget; feature selection itself does not establish out-of-sample trading value.

Key ideas

• The method retains multiple promising feature subsets to explore combinations beyond a single greedy path.
• Cross-validation performance is used to compare subsets and support early stopping.
• Shuffled-target replications estimate whether observed performance or an incremental gain could arise by chance.
• The implementation is designed to accommodate different supervised learning models.
• Selected predictors and their apparent significance depend on validation design and do not by themselves prove trading usefulness.

Tags

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