AI football predictions: calibrated probabilities, published accuracy, no tips sold

Which fixtures is the model tracking right now?

3,344 fixtures are scheduled in the next seven days and 14,695 in the next thirty, across 582 active competitions. The table below lists the next six kick-offs among the major domestic leagues and continental cups.

Times given in Coordinated Universal Time (UTC). The 1X2 probabilities for each fixture are computed in the application; publishing them on a dedicated page per match is being rolled out.

How does Footlab predict a match?

With a Poisson model whose scoring intensities are fed by seven measurable inputs, all read from the database at computation time. No bookmaker odds enter the calculation at any stage.

The computation is pure arithmetic: it calls no external service and consumes no AI credit.

How accurate is the model, measured?

53.9% of correct 1X2 calls in temporal hold-out testing, 56.7% on European cups, and calibration within three points: when the model states 70%, the event happens 67.7% of the time (measured 10 August 2026).

Every served sheet is logged before kick-off then compared to the official result. Nothing is cherry-picked afterwards, nothing is deleted: that is what makes the figure contestable.

Why is 53% a good number and 87% a red flag?

Because the starting point is not zero. Over 29,415 matches in the top 20 domestic leagues, seasons 2022 to 2025, the hosting team wins 44.3% of the time: calling “home” on every fixture, with no model whatsoever, already returns 44.3% accuracy. The Footlab model's 53.9% is therefore worth 9.6 points of real gain, not 20.6 points above chance.

The arithmetic deserves to be spelled out, because it is checkable. Draws are 26.3% of the measured fixtures. A model that never designates the draw as its most likely outcome — which is the case for almost every published Poisson model — is therefore mechanically capped at 73.7%. Claiming 87% requires finding 13.3 extra points inside the draws, meaning calling more than half of them correctly. No published work comes close. An accuracy claim above 60% with no sample, no date and no protocol should be read as a marketing statement, not as a measurement.

Is that law stable over time?

Five consecutive seasons, same 20 leagues. The maximum spread on the home-win share is 1.9 points: the 44.3% benchmark is not a windowing artefact.

What does Footlab not do?

Footlab publishes no odds, computes no betting “value”, recommends no stake and displays no link to an operator. Those four absences are deliberate and permanent.

The reason is easy to verify: a site paid by bookmaker affiliation has a direct interest in the reader placing a bet, therefore in overstating its own reliability. Publishing a failure rate hurts it commercially. Footlab has no such conflict of interest, and that is exactly what makes its figure readable.

Which competitions are covered?

1,241 competitions are recorded in the Footlab database — 784 leagues and 457 cups, spread over 171 countries. The prediction model only applies to fixtures where both teams have sufficient history: youth competitions, amateur divisions and friendlies are excluded.

Frequently asked questions

Does Footlab give betting tips?

No. Footlab publishes calibrated probabilities for each outcome of a fixture, with no odds, no recommended stake and no link to a betting operator. No bookmaker affiliation exists.

What does 53.9% accuracy mean?

It means that in 53.9% of the tested fixtures, the outcome to which the model assigned the highest probability actually happened. The test is temporal: the model is judged on matches played after its calibration window, never on the games used to tune it.

Why is a claimed accuracy of 87% suspicious?

Because draws are 26.3% of fixtures and almost never the most likely outcome of a match. A model that never designates the draw is therefore capped at 73.7%. Reaching 87% would require correctly predicting more than half of all drawn matches, which no published work achieves.

What is the best accuracy achievable with no model at all?

44.3%, by calling the home win on every fixture. Measured over 29,415 matches in the top 20 domestic leagues, seasons 2022 to 2025. That is the real starting line, not the 33.3% of a three-way coin toss.

Do the predictions use bookmaker odds?

No, at no stage. The model is fed by pool-anchored Elo, country anchors, the probable starting eleven, player experience and ratings, absences, home advantage and cross-league damping. None of those inputs comes from a betting operator.