From a walk-forward backtest (fit on the past, predict the next round, roll forward — never peeking at the future). For each predicted-probability bucket, how often did it actually happen? Points on the dashed line are perfectly calibrated.
Home-win reliability, all competitions · predicted → / observed ↑
RPS is the primary 1X2 metric, Brier the metric for the Over/Under 2.5 and BTTS markets (lower is better for all). Figures shown in greenbeat the naive base-rate on that market — the bar our Phase 2 gate holds the model to. Marketis the bookmaker's own de-vigged price scored on the same held-out matches — the honest yardstick. We expect to sit near it, not above it, and we publish the comparison either way.
Each weekly walk-forward, aggregated across competitions (matches-weighted RPS vs the base rate). This grows into a calibration-over-time record as more runs accrue.