How the Poisson model estimates football score probabilitiesGuides · Football · Stats explained

How the Poisson model estimates football score probabilities

Key points

  • The Poisson model converts expected goals into individual score probabilities
  • Add scorelines to derive win, draw and loss probabilities
  • Independent scoring makes low-score draws such as 0-0 and 1-1 less reliable

1Why football suits a Poisson model

Goals are scarce in football, and each attack resembles one of many low-probability opportunities that may or may not produce a goal. Statisticians often use the Poisson distribution for exactly this kind of question: how many times will an event occur within a fixed period? Once you know a team's average expected goals for a match, the distribution gives its probability of scoring 0, 1, 2 and so on.

That makes Poisson a natural starting point for many football probability models, and it can be built from nothing more exotic than goals scored and conceded. The classic football application is Maher's 1982 paper: assign every team an attacking strength and a defensive strength, then feed those values into a Poisson distribution.

2Step one: measure attacking and defensive strength

Begin with the league baseline. Suppose home teams have averaged 1.5 goals per match this season and away teams 1.2. Then compare an individual team with that baseline:

  • Home attacking strength = the home team's home goals per match ÷ the league's home goals per match
  • Away defensive strength = the away team's away goals conceded per match ÷ the league's away goals conceded per match (which equals the league's home goals per match)

A value above 1 means better-than-average attack, or worse-than-average defence; below 1 means the reverse. Calculate the away team's attack and the home team's defence in the same way, using the league's away scoring average as the reference.

3Step two: calculate expected goals

Multiply three figures to obtain the expected-goals rate for the match:

Home expected goals = home attacking strength × away defensive strength × league home goals per match

For example, take a home attacking strength of 1.3, an away defensive strength of 1.1 and a league home average of 1.5. The home expectation is 1.3 × 1.1 × 1.5 ≈ 2.1. Run the equivalent calculation for the visitors and suppose it returns 0.9.

4Step three: turn the rates into score probabilities

Given expected goals λ, the probability of scoring k goals is e^(−λ) × λ^k ÷ k!. With λ = 2.1, the home team's chance of scoring 0 is about 12%, 1 about 26%, 2 about 27% and 3 about 19%.

Multiply the home and away probabilities to get the chance of each exact score, such as 2-0, 2-1 or 1-1. Add every scoreline where the home side has more goals to find the home-win probability; add the level scores for the draw; the rest make up the away-win probability. The same grid can answer other questions, including the chance of more than 2.5 total goals or of both teams scoring.

5This is the introductory version

The team-by-team division above is designed to make the mechanics easy to follow. Academic models such as Maher and Dixon-Coles fit every match in a season simultaneously, estimating every club's attacking and defensive strengths together instead of calculating clubs in isolation. That produces more stable estimates, but the chain of logic remains the same: strengths determine expected goals, and expected goals determine score probabilities.

6Where the model falls short

  • It assumes the teams score independently. Real matches change with the score: a trailing side pushes forward while a leader may sit deeper. Maher found the independent model already worked quite well, though with a small systematic bias. Dixon and Coles later added corrections for low scores—0-0, 1-0, 0-1 and 1-1—so those are precisely the results where the unadjusted model is least trustworthy.
  • It does not know about injuries, rotation or a crowded schedule. The inputs reflect only past scoring and conceding.
  • Early-season samples are tiny. After only three or four matches, strength estimates can swing wildly. Analysts often blend in the previous season or give recent matches extra weight.

Treat Poisson as a data-led reference point, not a prophecy. League tables provide the goals-for and goals-against record—for example, the Premier League table—while our xG guide explains how to judge the quality of the chances behind those goals.

7FAQ

What is the Dixon-Coles correction?

Dixon and Coles introduced the refinement in 1997. A basic Poisson model treats each team's scoring as independent, but observed frequencies for 0-0, 1-0, 0-1 and 1-1 do not fit that assumption perfectly, so their model applies a correction factor. It also allows team strength to change over time; their fit used English league and cup matches from 1992 to 1995.

Is the independence assumption badly wrong?

No. Maher's study found an independent Poisson model was already reasonably accurate, with only a modest systematic bias. A bivariate version that allowed some relationship between the two teams' scoring—with a correlation coefficient of about 0.2—improved the fit further. Independence is therefore a useful simplification, not a perfect description.

Why is even the most likely exact score only around ten percent?

Football offers many plausible scorelines, so the probability is spread across outcomes such as 1-0, 1-1 and 2-1. In this article's example, with home expected goals of 2.1 and away expected goals of 0.9, even the most likely score, 2-0, is only about 11%. The model is better at comparing which side is likelier to win than at calling the exact score.

What can a model do when the season is only a few games old?

Dixon and Coles let team strength vary over time, giving recent results more relevance. A common practical solution is to blend in last season's record, then reduce that prior data as the new season supplies enough matches of its own.

8Sources

Illustrated portrait of A-He

About the author

A-He

Editor-in-chief

Editorial pen name

A-He is the newsroom's editor-in-chief, covering cross-sport topics, tools, rules explainers and esports. She starts with source quality and the current rulebook, then asks which numbers actually answer the reader's question, making definitions, limits and exceptions easy to compare across sports.

View profile and articles ›

‹ Back to guides · Glossary