The method
No magic.A pipeline you can audit.
Every probability on GrabbleStats comes out of the same six steps — and each step exists to keep the number honest. Here is the whole machine, with nothing hidden.
A decade of history, professionally sourced
Results, lineups, events and odds since 2015, from a professional data provider — refreshed every morning and during every live match.
Poisson models, fitted per league
Each team carries an attack and a defence strength, fitted on its own league's history with recent matches weighing more, plus home advantage. Goals models of the Dixon-Coles family — the workhorse of serious football modelling, not a black box.
Measured on seasons the model never saw
Refreshed every week, and always tested out-of-sample: a model is scored on more than a thousand matches it was never fitted on, before anything reaches you. New models climb through shadow, canary and partial stages — nothing is promoted the day it is born.
Confronted with the market, margin removed
Every reading is set against the bookmaker consensus with the margin stripped out — the only fair comparison. A gap is a disagreement worth examining, never a promise.
The AI explains. It never computes.
Language models write the context around a number — injuries, form, what could mislead us. They are structurally unable to produce a probability: the pipeline does not accept numbers from them.
Judged in public, fed back into the product
Kick-off freezes each claim; the final score judges it; the record stays on display. That record then governs the product: fragile leagues are flagged and capped, disproved markets are excluded, daily proposals prefer where announcements have held.
One number measures all of this: the Brier score — the average squared gap between what we announced and what happened. 0 would be perfect; 0.25 is a coin flip. We publish ours league by league and market by market, including where it is bad.
What the method refuses
An unwritten model says "not written".
No number without a fitted model
Cards, player props: where the fitted model does not exist yet, the product says so in plain words instead of showing a plausible figure.
No probability rewritten after kick-off
The number you saw before the match is the number that gets judged. Live probabilities are computed separately and labelled as such.
No verdict on thin evidence
Under twenty judged analyses, a league's gap is sample noise — the product refuses to call it good or fragile, and says how many it has.
No promise beyond the number
Neither accuracy, nor certainty, nor gain: a calibrated 60% does not happen four times out of ten, by definition. The product states this everywhere the number appears.
See the machine run.Open the app.
Next: the difference · straight answers