How confidence is calculated

We deliberately avoid the word "confidence" as a headline number, because a single percentage dressed up as certainty is exactly the kind of overclaim this site exists to avoid. Here is what actually drives how strongly the engine backs a fixture.

The process gate (engine score)

Each fixture gets a 0–100 engine score built additively from concrete factors: favourite strength, agreement between bookmakers, expected value versus the sharp devigged price, agreement between the model and the market, recent form and league context. A negative model-vs-market edge is a veto. The score is a gate — it decides whether a fixture is worth publishing — not a promise about the result.

Model confidence = sample size

The model's own confidence grows with how much relevant data it has for the teams involved, and that weight scales its influence on the score. Thin data means the model speaks more quietly, not more loudly.

Calibration is the honest test

The only meaningful check on probabilities is calibration: do the fixtures we price at 60% actually win about 60% of the time? We publish that curve on the track record. If our 70%s won 55% of the time, the number would be wrong and the page would show it.