Model Laboratory

Per-match model comparison (Poisson, Monte Carlo, Bayesian live, Ensemble) is live on every match page under the Models tab. The ensemble weights are configurable in services/prediction-engine and are deliberately not hard-coded into business logic.

• Gradient boosting (CatBoost/LightGBM) — requires historical training data; install the library and it integrates automatically, the ensemble renormalises without it
• Dixon-Coles fitting on real historical results — requires a data provider subscription
• Calibration dashboards (reliability diagrams, Brier, log loss, ECE) — implemented in football_intelligence/calibration.py; needs completed-match outcomes to populate
• Per-competition and per-match-phase calibration buckets — schema implemented, data pending