How Football Prediction Models Work

Every prediction starts with two numbers: how many goals each side is expected to score in this specific fixture. Those numbers come from each team's attacking output, the defensive quality of the opponent, and an adjustment for home advantage, which is worth roughly a fifth of a goal in most top divisions.

A team that averages 1.8 goals per game against average defences will be marked down against a strong one and up against a weak one. The output is not a prediction of the scoreline; it is the average of all the scorelines the fixture could produce.

Goals in football arrive at a roughly constant rate, which makes the Poisson distribution a good approximation for how many a team scores. Applying it to both expected-goals figures produces a grid of every plausible scoreline with a probability attached.

Everything else is arithmetic on that grid. Summing the cells where the home side scores more gives the home win probability; summing cells where both numbers exceed zero gives BTTS; summing cells totalling three or more gives over 2.5 goals. Because all markets come from one grid, they never contradict each other.

A statistical model does not know that a first-choice striker limped off in training, that a manager is resting players before a cup tie, or that a storm has turned the pitch to mud. Those factors move real matches and sit outside the data.

This is why probabilities should be treated as a baseline rather than a verdict. Use them to understand the shape of a fixture, then adjust for information the model has no access to.

Hit rate alone is a poor measure: always backing the favourite produces a respectable-looking number without any skill. Better measures compare the probability issued against the outcome that occurred, such as the Brier score, where lower is better and a well-calibrated model beats naive baselines over hundreds of matches.

Calibration matters more than confidence. If a model says 70% and those selections land close to 70% of the time across a large sample, it is doing its job, even when individual predictions miss.