How AI Ball works out its numbers

Short answer

AI Ball turns information available before kick-off into win/draw/loss probabilities, a likely range of scorelines and a confidence level. Every finished match is then marked right or wrong on the open record page.

What the estimates are built from

Published pre-match signals for the three outcomes and for the number of goals; the two sides' results, including their head-to-head meetings and recent matches; and availability — injuries, suspensions and how crowded each side's fixture list is.

What does not go in

Shot-by-shot data — so “expected goals” on AI Ball is not shot-based xG — line-ups, player tracking, weather and news. A page that explains a number should also say what that number cannot know.

The methods

Poisson model (Maher, 1982)

A standard way to model counts of rare events; in football, the number of goals a side scores.

Dixon–Coles adjustment (1997)

A published correction to the Poisson scoreline model. It fixes the known under-count of 0-0, 1-0, 0-1 and 1-1, and lets recent matches weigh more than old ones. The most likely scorelines and the total-goals distribution come from this score matrix.

Elo-style rating

A strength rating that moves after each result, by more when the result was less expected. It sets each side's goal expectation before the score matrix is built.

Shin's method (1993)

A published technique for removing the systematic bias from a set of publicly available pre-match figures before they are read as probabilities, so that they add to one.

Weighted multi-path fusion

Several estimates of the same match are combined with fixed weights rather than used one at a time, so no single input decides the outcome on its own. The blended result is then nudged for injuries and fixture congestion, and the draw is lifted slightly when a match is finely balanced and low-scoring.

Calibration

Calibration asks a narrower question than accuracy: when the model says 70%, does it happen about 70% of the time? A model can be well calibrated and still uncertain, and it can be right often while badly calibrated. The bands on the open record are the calibration check.

Why six models

The six modes use the same ingredients with different weights: Balanced (the default, and the only one used on the open record), Market, Data, Change, Conservative and Value. Comparing modes on a match page shows how much a conclusion depends on those choices.

What “confidence” means

Confidence is not the chance of being right. It measures how far apart the three probabilities are and how many of the model's views point the same way, then maps that through a table fitted on past matches. The open record shows whether higher-confidence matches were in fact right more often.

How you can check it

The open record lists every finished match since 27 August 2026. The model's view is locked between 45 and 120 minutes before kick-off and cannot be edited or deleted afterwards; records from before the lock existed are labelled as such. Each match is marked against the final result, next to two baselines: always siding with the market's favourite, and picking at random (33.3%). Bands with fewer than 30 matches show counts, not percentages.

Limits

These are statistical estimates from pre-match information. Football carries a large random component that no model removes, and draws are especially hard to call. Nothing on this site is advice, and no outcome is guaranteed.

References

• Maher (1982), Statistica Neerlandica 36(3)
• Dixon & Coles (1997), Applied Statistics 46(2)
• Shin (1993), The Economic Journal 103
• Brier (1950), Monthly Weather Review 78(1)
• Elo (1978), The Rating of Chessplayers, Past and Present