Every weekend, thousands of football predictions are produced not by pundits but by software. Some come from clubs' analytics departments, some from academic researchers and many from hobbyists running models on a laptop. They all promise the same thing: a more objective view of what is likely to happen on the pitch. But what is the "AI" actually doing? This guide explains it step by step, without equations.
1. A prediction is a probability, not a verdict
A good model never says "Team A will win". It says something like this:
| Outcome | Model probability |
|---|---|
| Home win | 48% |
| Draw | 27% |
| Away win | 25% |
The numbers above are an illustration, not a real match.
A 48% favourite still fails to win more often than it wins. When people say "the model was wrong" after a single upset, they usually misunderstand what the model claimed. The real test is whether, over hundreds of matches, events the model rated at 48% happen roughly 48% of the time.
2. What data goes in
Most football models are built from a handful of data families:
- Results and goals. The oldest and simplest input: who scored how many, home and away.
- Chance quality. Expected goals (xG) estimates how likely each shot was to become a goal, based on its position, angle and type. It tells a model whether a team's results are backed by the chances it creates.
- Team strength ratings. Systems such as Elo ratings update a single number for each team after every match.
- Context. Home advantage, rest days between matches, travel and sometimes confirmed line-ups and injuries.
- Market information. Some models also use betting odds as an input, because odds aggregate the opinions of many people.
3. The classic approach: counting goals
One of the most widely used ideas treats goals as rare, random events. If you estimate how many goals each team is expected to score in a match, the Poisson distribution turns those averages into the probability of every possible scoreline: 0–0, 1–0, 2–1 and so on. Add up all the scorelines where the home team scores more, and you have a home-win probability.
The expected goals for each side usually come from an attack strength and a defence strength for every team, plus a home-advantage factor. Statisticians Mark Dixon and Stuart Coles refined this approach in an influential 1997 paper, adjusting for the fact that low scores such as 0–0 and 1–1 happen slightly more often than the basic model expects.
4. Where machine learning comes in
Machine-learning models, such as gradient-boosted trees or neural networks, do not assume that goals follow a particular formula. Instead, they are shown thousands of past matches, each described by dozens of features — recent xG, rest days, ratings and so on — together with the result. The algorithm learns which combinations of features tend to come before which outcomes.
The advantage is flexibility: the model can pick up interactions a human might miss. The risk is overfitting — learning the noise of past seasons instead of real patterns. Football has relatively few matches per season and a lot of randomness, so careful testing on data the model has never seen is essential.
5. Why combining models works
No single model captures everything. A goals-based model may understand scorelines well but ignore line-up news; a machine-learning model may react to form but overreact to short streaks. A common technique is to combine several models into a consensus, averaging their probabilities or weighting them by past accuracy.
The idea is the same as asking several independent experts and taking the average: individual errors partly cancel out. In forecasting research, combined forecasts frequently match or beat the best individual model, which is why many serious prediction systems are ensembles rather than single models.
6. How a model is judged
Because predictions are probabilities, a model cannot be judged on "right or wrong". Two common checks:
- Brier score — a scoring rule: the average squared difference between the predicted probability and what happened. Lower is better.
- Calibration — grouping predictions (for example, everything rated 60–70%) and checking how often those events actually happened.
A well-calibrated model is honest about its uncertainty, which is often more valuable than a model that is occasionally spectacular.
7. What AI still struggles with
- Late team news. A key player ruled out an hour before kick-off can change a match more than a season of data.
- Motivation. End-of-season matches with nothing at stake, cup rotation or a new manager's first game are hard to quantify.
- Small samples. Promoted teams or newly assembled squads have little relevant history.
- Pure randomness. Deflections, red cards and goalkeeping errors happen. Football is one of the lowest-scoring major sports, so a single moment decides many matches.
The takeaway: AI does not see the future. It turns large amounts of data into honest probabilities, and its real value shows up over many matches, not one.
See how a rating system reacts to results with our Elo rating calculator.
Further reading
- Poisson distribution — Wikipedia
- Elo rating system — Wikipedia
- World Football Elo Ratings — Wikipedia
- Brier score — Wikipedia
The AI Sports team writes about data, models and technology in sport.