A Poisson model estimates an expected average number of goals for each team, then derives a probability distribution across every possible score — not a single score. It assumes goals happen independently of each other at a constant rate, two assumptions football only approximately respects. xG (expected goals) feeds this kind of model as an input, never as an output. The resulting distribution stays an estimate to check against an odd, which itself carries a margin and the market's opinion.
The essentials in a few seconds
A Poisson model estimates an average number of goals per team, then derives a probability for every possible score. xG often feeds this kind of model upstream, as an input. The output is a distribution — never a single score, never a prediction.
Three ideas carry the whole page:
- the model assumes, it doesn't observe — independence of goals and a constant pace are approximations, not laws of football;
- xG goes into the model, it doesn't come out of it — it's a measure of chance quality, not a prediction;
- a distribution isn't an odd — comparing the two means first converting the odd into a probability, margin included.
"A statistical model gives you the exact score of the match."
A Poisson/xG model gives every possible score a probability. In football, even the most likely score in a realistic distribution rarely exceeds 15 to 20% probability: the vast majority of conceivable scores share the rest between them.
What is a Poisson model applied to football?
A Poisson model assigns each team an expected average number of goals for a match, then calculates the probability of scoring 0, 1, 2, 3 or more goals from that average alone. Combining the two distributions, one per team, gives a probability for every exact score — 1-0, 2-1, 0-0 — and, by summing them, a probability for every outcome of the 1X2 or Over/Under market.
The principle traces back to Maher (1982), the first to formalise this approach for football and to test its statistical validity on real league data. His conclusion already held both sides of the story: an independent Poisson model reasonably describes the observed distribution of scores, with identifiable systematic gaps rather than mere noise.
The model stays simple to describe, which explains its longevity: two goal averages are enough to produce a complete distribution across every outcome of a match. Its simplicity is also its limit — it rests on assumptions worth knowing before reading a result produced by this kind of model.
What does Poisson's law assume?
Two assumptions carry the whole model: independence between goals, and a constant scoring rate across the entire match. Neither is strictly true in football — the question is how far off they are, and what that gap changes.
| Model assumption | What it assumes | What the data show |
|---|---|---|
| Independence of goals | A team's goal count doesn't depend on the other team's | A weak but measured correlation between the two — close to 0.2 according to Maher (1982) |
| Constant pace | The probability of scoring is the same in minute 5 as in minute 85 | Pace varies with the current score (a losing team attacks more, a leading team can ease off) |
| Fixed team strength | Attacking/defensive strength doesn't change during the match | Red cards, injuries and tactical changes shift real strength mid-match |
| Context-independent scores | Each match is treated in isolation | Schedule, stakes and fatigue affect real pace, uncaptured by the goal average alone |
The correlation Maher measured between the two teams' goals isn't zero, but it stays weak — around 0.2. That's exactly the gap Dixon and Coles (1997) set out to correct fifteen years later, adding a dependence parameter that adjusts the probability of close scores (0-0, 1-0, 0-1, 1-1), the ones an independent Poisson represents least accurately. Their work has a second, less often cited contribution: they tested it directly against bookmakers' odds, to check whether the corrected model identified exploitable price gaps on the English market of the mid-1990s.
An assumption and a fact aren't the same thing. A model assumption is a simplifying choice, made to keep the calculation tractable. A fact is what the data show once the model is checked against them. Using a Poisson/xG model well starts with knowing which of its assumptions carries the most gap on the match you're looking at.
xG, an input to the model
Expected goals (xG) measure the quality of chances created, from their position, their build-up and the context of the action. Used as an input, they build a more stable goal average than a simple average of goals scored, which is sensitive to finishing swings from one match to the next.
That use places xG precisely in the chain: it feeds the input parameter of a Poisson model, it's never its output. A team can post a high xG over several consecutive matches without the model that uses it producing a certain win — xG informs an average, it removes none of the assumptions attached to the model that then uses it.
The full football angle — xG, form, in-game indicators, and why football resists modelling so much — is covered in detail by the guide dedicated to AI applied to football.
xG, form data and models applied to football →
Checking the distribution against the market
A probability distribution from a model and a bookmaker's odd answer two different questions, and confusing them means comparing a calculation to a price. The odd factors in an estimate, but also a margin and the exposure tied to stakes already received — two elements no Poisson model captures.
To compare the two on a common basis, you need to convert the odd into an implied probability, then place the model's probability on the same outcome. If the model's estimate exceeds what the odd requires, the gap can be read as a fair odd different from the displayed price — a signal to question, never a certainty, since the gap can just as easily reveal a weakness in the model as an inefficiency in the market.
That's exactly the question Dixon and Coles were asking in 1997: does a better-specified model identify price gaps that survive contact with a real market? Their answer, positive on their original sample, doesn't generalise automatically — a betting market adjusts, and a gap identifiable in one season can close the next as other participants exploit it.
The Over/Under market, which bears directly on a total number of goals, is where a Poisson distribution meets a price most naturally.
The totals market, and how it settles →
What a Poisson/xG model allows — and what it doesn't
Three levels never to confuse.
- What it lets you do — combine two goal averages into a complete distribution across every possible score, and place each market outcome on that distribution.
- What it lets you estimate — a probability per score or outcome, with a margin of error that depends directly on how well the independence and constant-pace assumptions hold on the match at hand.
- What it never lets you conclude — the score of the match, nor even the most likely outcome as a settled result. The review by Bunker and Susnjak, already cited by the site's AI football guide, notes that football remains one of the hardest sports to model precisely because of its low goal count.
One practical consequence of this last limit: judging a model's quality on a single match makes no statistical sense. You need a track record, and how much data it takes to become interpretable depends on identifiable parameters.
What the amount of data needed depends on →
Responsible gambling — a model doesn't remove the risk. Sports betting carries a risk of financial loss and, according to the Autorité nationale des jeux (the French gambling regulator), poses the highest individual risk of problem gambling among regulated gambling activities. A probability distribution produced by a model makes no bet risk-free. Set a budget, don't chase your losses, and use the limit-setting or self-exclusion tools available to you. Learn more about responsible gambling.
Where OddScore stands relative to these models
OddScore doesn't build a score model and doesn't publish a probability distribution. The platform collects the odds of several bookmakers, converts them into implied probabilities, strips out the margin built into each price and tracks how they move up to kick-off.
What this means in practice: a Poisson/xG model and a market reading answer two related but distinct questions — one estimates a distribution from goal averages, the other reads what several operators collectively think about an outcome, price by price. The two can inform each other; neither replaces the other, and neither is a prediction.
The role an odd plays within a broader analysis — as one data point among others, never as an answer — is covered in detail by the guide linking odds and predictions.
Fold an odd into an analysis, without copying it →
Dig into the market
Odds movements are only part of the story. Here are the next topics to read.
Sports betting predictions: how to analyse before betting
The full cluster guide: data, odds, reliability, biases and the role of AI.
Read the guideFootball predictions and artificial intelligence
xG, form data and models applied to football.
See the football applicationImplied probability: converting it from an odd
The formula to go from an odd to its implied probability.
Convert an oddFair odds: from an estimated probability to an odd
Going from a model probability to a fair odd.
Calculate a fair oddOver/Under: the totals market
The market this kind of distribution most directly serves to evaluate.
See the marketThe limits of AI predictions
Incomplete data, black boxes, overconfidence in a precise number.
See the limitsHow many bets before judging a performance?
What the amount of data needed depends on.
See the methodFrequently asked questions
What is a Poisson model applied to football?
It's a statistical model that estimates an expected average number of goals for each team, then calculates the probability of scoring 0, 1, 2, 3 or more goals from that average alone. It produces a distribution of scores, not a single score.
Are xG and Poisson's law the same thing?
No. xG measures the quality of chances created during a match already played or in progress. Poisson's law is a probabilistic model that can use a goal average (derived from xG or not) as an input parameter, before a match.
Why doesn't football follow an exact Poisson distribution?
Because its two central assumptions — independence of goals and a constant scoring rate over 90 minutes — are only approximate. Goals within the same match are slightly correlated, and the pace of play shifts with the current score.
Can a Poisson/xG model predict an exact score?
No. It assigns a probability to every possible score, and even the most likely score remains a minority outcome within the distribution. A model that produced a single score would stop being a probabilistic model.
How do you compare a statistical model with a bookmaker's odd?
By converting the odd into an implied probability, then comparing that probability with the one the model produces for the same outcome. Any gap should first cast doubt on the model before it casts doubt on the market.
Do bookmakers use this kind of model?
Variants of the Poisson model are among the historical tools of sports pricing, alongside other methods. They don't alone explain a displayed price, which also factors in the operator's margin and exposure.
Does OddScore provide predictions?
No. OddScore compares the odds of several bookmakers, strips out the margin and tracks how they move. The platform publishes no betting selections and offers no staking advice.
Sources & methodology
This page draws on the two founding works on Poisson-based football score modelling, on the academic review of team-sport outcome prediction already cited by the site's AI football guide, and on the French regulatory framework published by the Autorité nationale des jeux.
- Reuse the independent Poisson model's assumptions as formalised by Maher (1982), without simplifying them.
- Present the correlation correction introduced by Dixon and Coles (1997) as the academic answer to the independence limit, not as a minor technical detail.
- Treat xG as one input among others, never as a model output or a result in itself.
- Explicitly distinguish a probability distribution produced by a model from a market odd, which factors in a margin and commercial exposure.