Techniques that penalize model complexity to prevent overfitting. Common regularizers (L1, L2, dropout) add a penalty term to the loss for large parameter values. In trading ML, L1 (Lasso) or L2 (Ridge) regularization can shrink less important feature weights, effectively performing feature selection. For example, a linear regression with L2 regularization will reduce sensitivity to outlier price spikes. Regularization improves generalization: it forces the model to learn only the strongest signals (useful trends) and ignore noise in the market data.

Back to Robots Glossary Index