Cross-Validation Techniques for Sports Data

## Beyond Train/Test Split Cross-validation improves on a single train/test split by using the data more efficiently — training and testing on multiple splits and averaging the results. ## K-Fold Cross-Validation Standard k-fold: 1. Split data into k equal folds (e.g. k=5) 2. For each fold: train on the other k−1 folds, test on this fold 3. Average test performance across all k folds **Problem for time-series data:** Standard k-fold randomly assigns matches to folds, which allows future data to inform predictions of past matches — a form of lookahead bias. ## Time-Series Cross-Validation For sports data, use time-blocked cross-validation: - Block 1: Train on season 1, test on first half of season 2 - Block 2: Train on seasons 1–2, test on second half of season 2 - Block 3: Train on seasons 1–2, test on season 3 - Continue... This preserves the temporal ordering and simulates real deployment conditions. ## The Leave-One-Match-Out Test Extreme version: for each match, train on all other matches and predict this one. Most computationally expensive but most rigorous. For computational efficiency: leave-one-season-out is more practical and nearly as informative. ## Nested Cross-Validation for Hyperparameter Tuning When you are also selecting model hyperparameters (e.g. decay rate, regularisation strength): use nested cross-validation. - Outer loop: splits for model evaluation - Inner loop: splits for hyperparameter selection within the training set This prevents "double-dipping" on the test set for both hyperparameter selection and evaluation.