Behind the data
Every model with a section here, written out: the terms, the constants, the thresholds, and the measured results. Referee Effect has no section because that surface is still being built.
Sample size travels with the claim. A rate without an n attached is not evidence, so every rate on the site carries the count it came from. Where a split is inside the noise, it is drawn muted rather than left to look like a finding.
Nothing is tuned against its own backtest. The fatigue model’s constants were set by reasoning about the physical effect rather than fitted to the win rates this site publishes. One has since moved: the altitude multiplier was raised on 2026-08-02 to match altitude’s measured size against a back-to-back on final margin — a different target from these win rates, and recorded as such. A model fitted to maximise its own reported accuracy would report whatever accuracy it was asked for.
Limits are published beside results. Each section ends with what its model cannot see. Those lists are the most useful part of this reference: they are the conditions under which the number on the product page is wrong.
Data spans 41 seasons, 1985-86 to the present. Not every field reaches back that far — see Data & limits for which seasons carry what.