Elastic Net Regression with Coordinate Descent for Trading Models

Summary

The article describes elastic net regression as a way to regularize linear predictive models used in trading. Its penalty blends the lasso’s absolute-coefficient penalty, which can suppress predictors, with ridge regression’s squared-coefficient penalty, which shrinks coefficients while retaining predictors. Alpha sets the balance between these penalty types, while lambda controls regularization strength. The implementation uses coordinate descent, optimizing one variable at a time, first to select lambda and then to estimate model coefficients.

The author outlines an MQL5 class that standardizes inputs, trains models, and can use covariance updates, alongside a moving-average-based predictive strategy example. Flat arrays represent matrices to support cross-platform compatibility. The article includes an out-of-sample return figure for the example, but the excerpt does not provide enough experimental context to assess its reliability or generalizability. Regularization can reduce overfitting and help identify less useful predictors, but the result still depends on data selection, validation, and model settings; the presented example is not evidence of robust live performance.

Key ideas

• Elastic net combines lasso and ridge penalties to regularize linear models.
• Alpha controls the balance between absolute and squared coefficient penalties, while lambda determines regularization strength.
• Coordinate descent optimizes one parameter at a time and is used for both lambda selection and coefficient fitting.
• The MQL5 implementation standardizes data and can use covariance updates to speed computation.
• An example applies the method to moving-average predictors, but its reported out-of-sample result lacks enough context to establish robustness.

Tags

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