Predictorous

A football model that shows its work.

A probability and a fair price for every fixture in the big five. Every call is recorded before kickoff and graded afterwards — including the wrong ones.

Data from 2026-11-20

Anyone can publish predictions. The test is whether the stated probabilities are honest: when this model says 45%, does it happen about 45% of the time? Each row is a band of confidence — the pale bar is what it said, the solid bar is what actually happened. Only calls committed before kickoff count here; the 146 fixtures of the opening rounds, rebuilt after the fact, are shown apart.

30-40%n=116
said
35.3%
landed
31.9%
40-50%n=103
said
44.8%
landed
51.5%
50-60%n=74
said
54.4%
landed
64.9%
60-70%n=48
said
64.1%
landed
70.8%
70-80%n=14too thin to read
said
74.0%
landed
100.0%

Buckets are the model's stated probability for the outcome it called. A model that is confident and wrong shows up here as a solid bar far short of its pale one — which is why this is the first thing on the page rather than a methodology footnote. The buckets cannot show a bias in one outcome's level, so here is the one that exists: the model has said 27% draw on average and 22.3% have drawn. The full breakdown, and why, is on the track record.

This round, league by league: the model's probability for each result, the outcome it calls, its most likely scoreline, and the fair price for each result — what it is worth. Above that price there is value in it; below it there is not. We never see your book, so the comparison is yours to make.

round 11 · 11-21 → 11-23

Percentages are the model's probability; the smaller number is the fair price (1 ÷ probability) — on a phone it appears under each percentage once you switch it on above. The called outcome is shaded. Scorelines are the most likely result given that call, not a prediction of the exact score. Played fixtures stay listed until the whole round is done. Data as of 2026-11-20.