Model accuracy & calibration

Every prediction is graded against the real result — nothing cherry-picked. This is the model's live, out-of-sample record, as of 3 Nov 2026 — distinct from the historical backtest.

51%
1,313/2,571 picks correct
49%
average confidence
58%
Over/Under 2.5
58%
Both teams to score

Over 2,571 graded picks the model was, on average, 49% confident and right 51% of the time — a excellent calibration match (gap 1.8 pts, ECE 2.1%). Last 90 days: 51% over 2,565 picks.

Calibration — predicted vs actual

Each dot is a confidence band. If the model is honest, dots sit on the dashed line: a band we called ~70% should win ~70% of the time. Dot size = number of predictions.

40 40 50 50 60 60 70 70 80 80 90 90 100 100 33–40% band: predicted 38%, actual 38% (542 picks) 40–50% band: predicted 44%, actual 46% (1020 picks) 50–60% band: predicted 54%, actual 55% (576 picks) 60–70% band: predicted 64%, actual 71% (288 picks) 70–80% band: predicted 74%, actual 86% (116 picks) 80–90% band: predicted 83%, actual 65% (26 picks) 90–100% band: predicted 91%, actual 100% (3 picks) Predicted probability (%) Actual hit rate (%)
Confidence bandAvg predictedActualPredictions
33–40%38%38%542
40–50%44%46%1020
50–60%54%55%576
60–70%64%71%288
70–80%74%86%116
80–90%83%65%26
90–100%91%100%3

Want the deeper numbers? See the historical backtest (walk-forward across seasons) and the verifiable ledger — every day's picks are sealed with a hash before kickoff, so this record can't be edited after the fact.

Statistical estimates graded against real results. Past accuracy doesn't guarantee future results. For information and entertainment only. 18+.