Guide · updated 2026-10-05
The Dixon–Coles model, explained
The Dixon–Coles model gives every team an attack and a defence rating, turns them into expected goals for each side, and from those builds a probability for every possible score. Adding up the scores gives home, draw and away probabilities.
From ratings to expected goals
Each side's goals are modelled as a Poisson count. The home side's expected goals rise with its attack rating, the away side's defensive weakness and a league-wide home advantage; the away side's expected goals work the same way without the home term.
The low-score correction
Dixon and Coles added a term that nudges the probabilities of 0–0, 1–0, 0–1 and 1–1 up or down, because real football produces these results at rates a pure Poisson model misses. Setting the term to zero gives back the independent Poisson model.
How our version is set up
- Training data: Premier League full-time results since 2005-06, with only matches finished before the prediction.
- Time weighting: a one-year half-life, so last season counts half as much as this one.
- Shrinkage: ratings are pulled towards average when a team has little data, and newly promoted clubs start below average, as promoted sides historically have.
- Player availability: expected goals are adjusted for injured, suspended and doubtful players listed by the Fantasy Premier League feed.
Full details are on the methodology page.
Example: Arsenal vs Leeds United
The model expects 1.93 goals for Arsenal and 0.77 for Leeds United. Its most likely single score is 2–0 at 12.5%, and summing every scoreline gives Arsenal 63.8%, draw 22.7% and Leeds United 13.4%. See the full match page.
How good is it?
On 1,141 held-out matches the model's log loss was 0.991, against 0.969 for de-vigged closing odds. Bookmakers' closing odds have been more accurate than the model; the accuracy guide explains the comparison.
Common questions
What is the Dixon–Coles model?
A football score model published by Mark Dixon and Stuart Coles in 1997. It treats each team's goals as Poisson counts driven by attack and defence ratings and home advantage, with a correction for low scores such as 0–0 and 1–1.
Why not a plain Poisson model?
Independent Poisson counts tend to misprice low-scoring results. The Dixon–Coles correction adjusts the probabilities of 0–0, 1–0, 0–1 and 1–1, which matters because those are among the most common scores.
How often is the model refitted?
Ratings are refitted every time the site updates, using only matches that had already finished. Its settings are checked once a month on held-out matches.
More guides
Probabilities are not guarantees and nothing here is betting advice. Responsible gambling.