Methodology
The short version: real data in, our own model in the middle, and a public, permanent record of every number that comes out. Open about results, careful with the recipe.
The data
Three decades of English football: 30,920 matches since 1993, with detail on how each one was actually played, not just the final score. Bookmaker odds are never a model input — we only look at the market afterwards, to check ourselves against it.
The model
Built entirely in-house. It estimates how strong every club really is right now, reading underlying performance rather than headlines, and turns that into a probability for each result. The exact method is ours and stays that way — what we publish instead is every output, graded, so you can judge the model by the only thing that matters.
The test we must pass
Before any market appears on this site or in the Discord bot, it has to prove itself on matches the model was never tuned on: probabilities close to the market's best, and true-to-life in every probability band — when the model says 60%, reality must agree about 60% of the time. Markets that fail stay hidden until a version of the model earns them.
The record
Probabilities freeze into a permanent record 24 hours before kick-off, stamped with the model version that made them. After full time, each one is graded against the result and the market's number for the same match. Postponements are set aside with a reason, never rewritten. We cannot quietly edit our history, and that is the point.
What the model cannot see
Line-ups, injuries, rotation and motivation are not in our data today. That information is most of the market's remaining edge over us, and it is why our numbers are deliberately more cautious than bookmaker prices on heavy favourites. We say this out loud because trust is the product.