Backtesting

The calibration and evaluation layer (Brier score, log loss, ECE, reliability bins) is implemented and unit-tested in the prediction engine.

• Time-based train/test splits and walk-forward validation runner — requires historical match data (set MOCK_DATA=false with SPORTMONKS_API_KEY)
• Stored model versions, training periods and metrics (model_versions / model_metrics tables)
• Future matches are never shuffled into training; splits are strictly chronological to avoid leakage