Methodology
This page explains, end to end, how SharpFooty Analytics turns raw results and odds into a published, gradeable call. Nothing here is a black box — the discipline is that the calibrated statistical model and the market own the number, and AI is only ever allowed to explain, challenge or veto, never to set a price.
1. The model
A Dixon-Coles bivariate-Poisson goal model estimates each team's attack and defence strength, blended with ELO ratings, and retrained continuously on a growing archive of real match results. For any fixture it produces expected goals per team (λ), a full scoreline matrix, and 1X2 / over-under / BTTS probabilities. Model confidence grows with sample size and weights the model's influence accordingly.
2. The market
Real bookmaker prices are captured and timestamped from a licensed odds feed. The bookmaker margin is removed (devigging), anchored to a sharp book where one is held, to produce a margin-free "fair" probability for each outcome. The model's number is priced against this fair line — value is the gap between the best available price and the fair line, stated as a fact about a price.
3. The AI layer (bounded, measured)
Grounded AI briefs gather real team news (injuries, rotation, motivation) and can apply a small, bounded adjustment to expected goals; adversarial critic passes stress-test each near-bet fixture. Every AI and market output becomes a named, point-in-time signal that is graded like any pick. A signal cannot influence a live decision until it has beaten the baseline on a meaningful sample — and even then it can only veto or nudge, never invent a bet.
4. Grading
Every published call is settled against the real result and the closing price. We record closing-line value (did the market move toward our price after we flagged it?) and the win/loss outcome. The ledger is public and permanent — see the track record.
In short
- Model estimates probabilities from results; market sets the fair line.
- AI explains and challenges, within hard bounds, and is graded before it counts.
- Every call is timestamped and judged against the close — win or lose.
- We lead with calibration and CLV, never profit claims.