How Our AI Match Predictions Work (and What They Can't Do)

Every match on this site carries an AI prediction: win/draw/loss probabilities, a predicted scoreline, a confidence rating and the players most likely to decide the game. Prediction content on the internet is usually a black box, so this page explains exactly how ours is made — what goes in, what comes out, and what it can and cannot tell you.

What goes in

Our match data is synced from professional football data feeds — the same fixture, result, lineup and player-statistics data used across the industry. For each prediction, the model considers:

• FIFA rankings of both teams, as a baseline strength signal;
• Tournament form — results, goals scored and conceded in the current competition, pulled from our own standings and results database;
• Squad quality — the players available to each side, including per-match statistics we collect for every player at the tournament;
• Match context — the round, what's at stake, and knockout dynamics like extra time and penalties.

How the prediction is generated

We use Claude, a large language model built by Anthropic, prompted with the structured data above and asked to produce a strict, machine-readable assessment: probabilities for each outcome, a most-likely scoreline, a confidence score, an upset likelihood, and short written reasoning. We store the model's full output — including its reasoning — and display it on each match's prediction page, so you can always see why the model leans the way it does, not just a percentage.

Predictions are not static. They are generated before kickoff, then regenerated at kickoff, at halftime, and after goals — so the probabilities you see during a live match reflect the current score and time remaining, not a stale pre-match guess.

What the model gets wrong

Football is famously the lowest-scoring major sport, which makes single matches genuinely hard to predict — for humans, statistical models and language models alike. A side given a 65% chance still loses about one time in three; that is what 65% means. Penalty shootouts are close to coin flips no matter what any model says, a point this tournament's round of 32 made emphatically when Germany and the Netherlands both went out on penalties as heavy pre-match favorites.

We are building a public accuracy tracker so the model's record — outcomes called correctly, exact scorelines, calibration of its probabilities — is visible for anyone to audit. That work continues into the 2026-27 club season, where we'll publish predictions for every matchweek and grade them in the open.

What our predictions are not

• Not betting advice. Our predictions are editorial and entertainment content. We are not a bookmaker, we do not sell tips, and nothing on this site should be the basis for a wager.
• Not insider information.The model knows what the data feeds know. It does not know about a dressing-room argument or a knock in training that hasn't been reported.
• Not a substitute for watching. If the model were always right, nobody would need to play the matches — and Paraguay vs Germany is why they do.

Where to see it in action

Every match page links to its prediction — for example, any fixture on the fixtures page — and the tournament predictions pageaggregates the model's title odds and Golden Boot forecast, refreshed as results come in. Fan votes on each match page are kept separate from the AI's numbers, so you can see where the crowd and the model disagree.

Questions about the methodology are welcome — see our contact page. For how we produce written content like this article, see our editorial policy.