Model Validation — current model: 2025-2026 Season

Held-out backtest results for the current production model (version ridge_logistic_v1, fit for the 2025-2026 Season) against recent completed seasons it never trained on - each season's ratings and prediction models are refit using only information available before that season's own tournament, then scored against what actually happened. This is a read-only summary; the interactive hyperparameter-sweep tool lives in Administration. For a specific season's actual, game-by-game result (not this aggregate backtest), see Bracket Results.

See the Data Dictionary for what each of these metrics actually measures - lower Brier score / log loss / margin MAE is better; higher accuracy is better. Precomputed once per model fit, not recomputed on this page load.