How AI football predictions work

Two layers: the calculator and the narrator

The calculator is a statistical model. It reads strength ratings, recent form and home advantage, turns them into expected goals for each side, and produces a probability for every possible scoreline. It cannot be talked into anything — it only knows maths.

The narrator is the language model. It reads the model's output plus real fixtures, standings and public news headlines, then explains the numbers in plain language — in English or Chinese. It is strictly forbidden from inventing data: if the model did not compute it, the AI does not say it.

The data behind the numbers

Real data only: fixtures, results and standings come from public data sources, and news is limited to public headlines. Nothing is simulated, and the site shows no odds and no betting recommendations — probabilities are the product.

From ratings to a probability

Elo-style strength ratings built from historical results are refreshed per league; recent form and home advantage adjust each side's expected goals; a Dixon-Coles adjusted Poisson model converts those into the full scoreline distribution, which sums into win/draw/loss, over/under, both-teams-to-score and exact-score probabilities.

Why a chat, not a static tips table

A table ends at its columns. In a chat you can follow up: why so high on the draw? What changed after the midweek game? How does the table look if they win? The answer always cites the same underlying numbers, so the conversation stays grounded.

How to recognise a serious prediction service

Three questions separate the serious from the noisy: Is the method published? Are predictions locked before kick-off and graded afterwards, misses included? Are small samples flagged honestly? Footalytics answers yes to all three — the methodology is public, official picks are graded against real scores, and the accuracy dashboard flags anything under 30 matches.