FREL Feature Weighting with Regularized Energy-Based Learning

Summary

The article describes FREL, a feature weighting and selection method based on regularized energy-based learning. It frames prediction as weighted nearest-neighbor classification: candidate predictors receive weights, and Manhattan distance determines how strongly nearby training observations influence a classification. An energy function scores input configurations, while a per-sample loss penalizes cases where incorrect configurations appear too similar in energy to correct ones. Optimization then finds weights intended to emphasize more relevant predictors.

The MQL5 implementation accepts a matrix with observations in rows, predictors in columns, and the target in the final column. It initializes weights and minimizes the objective using Powell’s method, with optional regularization and bootstrap settings. The article demonstrates the approach on synthetic data and reports that the computation is slow; smaller bootstrap samples improved execution speed, while parallel or GPU implementation is suggested as future work. Predictors should have similar scales, and the demonstration does not establish that selected features improve out-of-sample trading performance.

Key ideas

• FREL assigns predictor weights by optimizing a regularized energy-based loss.
• Its nearest-neighbor component uses weighted Manhattan distances to compare observations.
• The method is intended for datasets with candidate predictors and a single target, with predictors on similar scales.
• The MQL5 example uses Powell’s method and allows bootstrap sampling.
• The synthetic demonstration is computationally costly and does not demonstrate live or out-of-sample trading performance.

Tags

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