Our Accuracy Record — The Receipts
The short version — every accuracy figure we publish is a walk-forward backtest against seasons the engine had never seen, reproduced from our repository before publication — and we print the misses too.
79% of six-gameweek forecasts on target
Across a full season the engine had NEVER seen, 79% of our six-gameweek player forecasts landed within an average of two points a gameweek. Walk-forward: the engine only ever saw data available before each gameweek, then we scored it against what actually happened.
8.5 of 11 starters called
Across a two-season backtest, our predicted line-ups named an average of 8.5 of the 11 actual starters — before official team news.
0.79 rank correlation on team strength
Replaying a full season, our team ratings correlated 0.79 with the final league table.
84% of the player pool on real shot data
84% of projected players are backed by three seasons of real expected-goals data. For the rest — newly promoted or newly signed — we say so and project conservatively rather than guessing.
What we do NOT claim
We do not publish a price-change predictor yet: ours stays switched off until it proves at least 90% right on 150+ real price moves. A tool that is not provably good does not ship — that is the standing bar, and it is why one of our analyst desks is still empty.
Backtests are not guarantees. Football has variance; a projection is a probability statement, not a promise. What we promise is honest measurement: the same walk-forward method, misses included, every time.