If you have ever wondered how a football model gives a percentage chance for a home win, a draw or a 1-1 scoreline, there is a good chance the answer starts with the Dixon Coles model. It was published by statisticians Mark Dixon and Stuart Coles in 1997, and it is still one of the most common starting points for football prediction today.
This guide walks through the idea step by step, without any heavy maths.
Start with goals, not results
Most beginners try to predict results directly: will it be a home win, a draw or an away win? The Dixon Coles approach goes one level deeper. It predicts how many goals each team is likely to score, and then works out the results from that.
This is useful because once you know the chance of every possible scoreline, you can answer lots of questions at once. Home win, draw, over 2.5 goals, both teams to score and correct score all come from the same set of numbers.
The Poisson distribution
Goals in football are fairly rare, and they arrive in a way that is close to random once you know how strong the teams are. That pattern fits a well known bit of statistics called the Poisson distribution.
You give Poisson one number, the average number of goals you expect a team to score in this match, and it tells you the chance of that team scoring 0, 1, 2, 3 or more goals.
For example, if a team is expected to score 1.5 goals, Poisson says:
- 0 goals: about 22%
- 1 goal: about 33%
- 2 goals: about 25%
- 3 or more: the remaining 20% or so
So the big question becomes: how do we work out that expected goals number for each side?
Attack and defence ratings
The model gives every team two ratings:
- Attack rating: how good they are at scoring compared with an average team.
- Defence rating: how much they tend to concede compared with an average team.
An attack rating above 1 means better than average at scoring. A defence rating below 1 means harder than average to score against.
These ratings are not guessed. The model looks back at past results and finds the set of ratings that best explains the scores that actually happened. A team that keeps scoring lots against strong defences ends up with a high attack rating. A team that keeps shipping goals to weak attacks ends up with a poor defence rating.
To get the expected goals for a match, you roughly multiply together the scoring team's attack rating, the opponent's defence rating and a baseline for how many goals a typical team scores.
Home advantage
Teams tend to do better at home. The model handles this with a single home advantage factor that nudges up the home side's expected goals. Like the team ratings, this factor is estimated from past results rather than picked by hand.
A worked example
Here is a made up example to show how the pieces fit. The numbers are for illustration only.
Say the model, after combining attack, defence and home advantage, expects:
- Home team: 1.5 goals
- Away team: 1.1 goals
Using Poisson for each side, the chance of some scorelines is found by multiplying the two teams' chances together:
- 0-0: 22% × 33% = about 7.4%
- 1-0: 33% × 33% = about 11.2%
- 1-1: 33% × 37% = about 12.3%
- 2-1: 25% × 37% = about 9.2%
Do this for every realistic scoreline, add up all the ones where the home side wins, and you have the home win chance. Do the same for draws and away wins, and you have a full prediction.
The low scoring adjustment
This is the part that made the Dixon Coles paper famous.
A plain Poisson model treats the two teams' goals as completely separate. In real football, that is not quite true for low scoring games. Dixon and Coles found that simple Poisson tends to get the chances of 0-0, 1-0, 0-1 and 1-1 slightly wrong, with draws like 0-0 and 1-1 a little more common than it suggests.
Think about how a match feels at 0-0 or 1-1 late on. Teams may settle for a point. A side leading 1-0 may sit back. Those tactical choices link the two teams' scoring in a way basic Poisson ignores.
So the model adds a small correction that only touches those four low scorelines. Everything else stays as normal Poisson. It is a simple fix, but it makes draw and under goals predictions noticeably more realistic.
Recent matches count more
Dixon and Coles also suggested giving recent matches more weight than old ones. A team's form from last month says more about them now than a result from two seasons ago. Squads change, managers change and players age, so older results are gradually faded out rather than counted equally.
What the model does not know
The Dixon Coles model is a solid base, but it only sees past scores. On its own it does not know about:
- Injuries, suspensions or rotated line ups
- A new manager or a big signing
- Whether a team created lots of chances but was unlucky in front of goal
- Fixture congestion or a team resting players for a cup game
That is why many modern models feed in extra information, such as expected goals from shot data, alongside or instead of raw scorelines. It is also why a model's numbers should be treated as an estimate, not a certainty.
Why this matters for you
When you see a model say a team has a 45% chance of winning, that number has usually come from something like the process above: ratings built from past results, turned into expected goals, turned into scoreline chances, then added up.
Understanding this helps you ask better questions. Is the model using recent enough data? Does it know about the key striker who is out? Is the edge it shows big enough to matter once you allow for the bookmaker's margin?
A model can help you think clearly about a match, but no model can tell you what will actually happen. Football will always have plenty of surprises.
A note on responsible gambling
If you use predictions like these for betting, please keep it fun. You must be 18 or over, only stake what you can afford to lose, and if gambling stops feeling like entertainment, visit BeGambleAware.org for free, confidential support.