Model Comparison

5 models trained on the same data. Champion selected by lowest MAE on holdout set.

Performance Comparison

5 models predict BOC1 price levels. MAE measures prediction error. Directional accuracy measures true next-day direction calls. Click headers to sort.

Residual Diagnostics

Champion model: loading

Feature Importance

XGBoost native gain-based importance

Learning Curves

Training vs validation MAE as function of training set size

Walk-Forward Performance by Fold

Does the model break down in specific periods?

Model Selection & Hyperparameters

Why Ridge?

Ridge regression was selected as champion for lowest MAE (2.69 c/lb) on the holdout set. Despite XGBoost's ability to capture nonlinear interactions, Ridge's L2 regularization prevented overfitting on this relatively small feature set (5 commodity prices). XGBoost's MAE was 2.5x worse — it memorizes noise in the training data rather than learning generalizable patterns.

Selection criterion: lowest MAE on 20% temporal holdout (no shuffle). Walk-forward validation on 5 folds confirms generalization.

XGBoost Hyperparameters