If you have ever wondered how prediction sites turn two teams' scoring averages into a full set of scoreline probabilities, the answer is usually the Poisson distribution. It is a statistical tool for modelling how many times an event happens in a fixed period, and it fits football goals surprisingly well: goals are relatively rare, roughly independent events, and they arrive at a rate that varies by team. With just a few inputs, Poisson lets you estimate the chance of a 1-0 win, a 2-2 draw, or any other scoreline, and from there the chance of a home win, a draw, or an away win.

The core idea is simple. If a team scores an average of 1.5 goals per match, Poisson gives you the probability it scores exactly zero, one, two, or three goals in a given match. The distribution is skewed: zero and one are common, four or more are rare, and the shape is fully determined by that single average. To predict a match, you estimate each team's expected goals, often called lambda, then combine the two distributions. Multiply the probability the home team scores two by the probability the away team scores one, and you have the probability of a 2-1 scoreline. Sum all scorelines where the home team scores more, and you have the home win probability.

The model is only as good as its lambda values, and this is where most beginners go wrong. Using raw season scoring averages ignores who those goals were scored against. A team averaging two goals a game after facing the league's weakest defences is not as potent as the number suggests. The standard fix is to split scoring into attack strength and defence weakness: divide a team's goals scored by the league average, do the same for goals conceded, then combine them. Home advantage gets its own adjustment, since home teams historically score more than away teams.

Expected goals data, if available, improves the inputs further. A team creating high-quality chances but finishing poorly will usually regress toward its expected goals over time, so modelling on expected goals rather than actual goals can spot value the raw numbers hide. Whichever input you use, apply it consistently and validate against past matches. A Poisson model should be tested on matches it did not learn from, with its predicted probabilities compared against actual outcomes, before anyone trusts it with a stake.

Poisson assumes goals are independent, but football is not quite that tidy. Teams change tactics when leading or trailing: a side that goes two goals up often sits back, reducing further scoring, while a trailing side pushes forward and leaves gaps. This makes very low-scoring draws slightly more common and very high-scoring games slightly less common than pure Poisson predicts. Researchers often correct for this with adjustments for low scorelines, or by moving to more flexible models such as the bivariate Poisson, which allows the two teams' scoring to correlate.

Poisson also assumes scoring rates stay constant within a match, ignoring red cards, injuries, and momentum shifts. None of this makes the model useless. For pre-match markets, a well-built Poisson model remains one of the best starting points in football prediction, precisely because it is transparent: you can inspect every assumption, adjust it, and see the effect. Use it as a foundation, compare its probabilities with the market, and treat large disagreements as prompts for investigation, not automatic bets. The model points; your judgement, and the rest of your evidence, decides.

Before staking on a Poisson model, backtest it properly. Take a full past season, estimate expected goals from data available before each match, generate scoreline probabilities, and compare the predicted home, draw, and away probabilities with actual outcomes. Check calibration by bucket, look at the Brier score, and compare your probabilities with de-vigged market odds to see where you disagree. Large, systematic disagreements are worth investigating: sometimes the model is wrong, but sometimes it has spotted something the market undervalues, such as a team's improving underlying numbers.

Keep the model updated as the season progresses. Early-season ratings are noisy because few matches have been played, so blend early estimates with prior-season data, weighting recent matches more heavily as the sample grows. By mid-season, the model should stand on current form alone. A Poisson model that is never refreshed will drift away from reality as squads change and tactics evolve, and its predictions will quietly worsen. Maintenance is unglamorous, but it is what separates a model from a one-off spreadsheet.