Football prediction models are everywhere now. Pundits quote expected goals, betting sites publish win probabilities, and fans argue about which model is most accurate. Yet few people outside the data teams that build them understand what these models actually do. At their core, they are not fortune-tellers. They are structured ways of turning past matches into probabilities for future ones, and they come with limits that matter.

The most famous building block is expected goals, usually shortened to xG. Every shot is given a value between zero and one, representing how often a shot from that position, with that type of assist and under those defensive conditions, would be scored. A shot from six yards out in open play has a high xG. A hopeful effort from thirty yards has a low one. Add up the xG of all a team's shots and you get a measure of the quality of chances it created, separate from whether those chances happened to go in.

The second building block is a team rating. The best-known approach is similar to the rating system used in chess. When a team beats a strong opponent, its rating rises more than when it beats a weak one. When it loses to a weak opponent, its rating falls sharply. Over many matches, these ratings settle into a reasonable ranking of team strength, adjusted for home advantage and, in more advanced models, for the margin of the result.

Once a model has ratings for both teams, it converts them into probabilities for a home win, a draw and an away win. It does this by simulating the match many thousands of times, or by using a statistical formula that maps the rating difference to a likely scoreline distribution. The output is not a single prediction like a 2-1 home win. It is a set of probabilities, for example a 45 percent chance of a home win, a 27 percent chance of a draw and a 28 percent chance of an away win. That distinction is crucial. A model that says a team has a 60 percent chance of winning will still be wrong four times out of ten, and that is not a failure.

Models have real limits, and honest builders admit them. First, they struggle with new information that is not in the data yet: a key player returning from injury, a manager changing formation, a team playing a weakened side before a cup final. The model sees the team's past results, not the team sheet for tonight. Second, they struggle with rare events. Red cards, penalties and goalkeeper errors are hard to predict and can swing a match that the model rated as comfortable. Third, models inherit the biases of their inputs.

For fans, the practical lesson is to read probabilities, not certainties. When a model gives a team a 70 percent chance, that is a strong favourite, not a guarantee. When it gives a match a near-even split, that is a genuine coin flip, even if one team is more famous. The value of a model is in the long run: it helps you see which teams are overperforming their results, which are underperforming, and where the underlying numbers disagree with the league table. Used well, AI football prediction models are not oracles. They are lenses that make the game clearer, provided you remember what they can and cannot see.