MODEL EXPLAINER Β· 8 min read

How the Dixon-Coles Model Predicts Soccer Scores

In 1997, Mark Dixon and Stuart Coles published a short paper that quietly became the foundation of modern soccer modeling. Their insight was simple: the standard Poisson model β€” the obvious starting point for predicting goals β€” gets low scores systematically wrong. And low scores are where soccer lives.

The Poisson starting point

Model each team's goals as an independent Poisson distribution. If the home team averages 1.5 expected goals and the away team 1.1, the probability of a 1–0 home win is just P(home=1) Γ— P(away=0). Clean, elegant β€” and wrong in a specific, measurable way.

When Dixon and Coles compared Poisson predictions against thousands of English league matches, they found the model underestimated 0–0, 1–0, 0–1 and 1–1 results while slightly overestimating others. Teams don't play like independent random-number generators at low scores: a team leading 1–0 late in the game often stops attacking, and a team trailing 0–1 throws everyone forward. The two goal counts are dependent exactly where it matters most.

The one-parameter fix

Their solution was a correction factor, Ο„ (tau), applied only to the four low scorelines, controlled by a single parameter ρ (rho):

P(0–0) Γ— (1 βˆ’ Ξ»HΒ·Ξ»A·ρ)
P(1–0) Γ— (1 + Ξ»A·ρ)
P(0–1) Γ— (1 + Ξ»H·ρ)
P(1–1) Γ— (1 βˆ’ ρ)

With ρ negative (around βˆ’0.13 across most leagues), the correction boosts 0–0 and 1–1 while trimming 1–0 and 0–1 β€” matching what actually happens on the pitch. One parameter, fitted once per league, and the model's draw probabilities suddenly line up with reality.

The practical takeaway: any soccer model that doesn't correct for low-score dependence will systematically misprice draws β€” usually by 2–4 percentage points. That's the difference between a fair price of 3.40 and 3.10 on the draw, which is real money over a season.

How we use it at GoalMath

Our match predictor runs the full Dixon-Coles calculation over every scoreline from 0–0 to 12–12, then aggregates into 1X2, over/under and BTTS probabilities. Expected goals come from team attack/defense strengths relative to league averages:

Ξ»H = home attack Γ— away defense Γ— league avg home goals
Ξ»A = away attack Γ— home defense Γ— league avg away goals

We fix ρ at βˆ’0.13 as the default β€” the value that has held up across leagues and seasons in published research β€” and we show our calibration record on the methodology page, including where the model fails.

What Dixon-Coles can't do

A model is a starting point for thinking, not a substitute for it. The math tells you what usually happens; the news tells you whether "usually" applies today.

Related reading: Poisson vs Dixon-Coles: Where the Difference Actually Matters Β· Expected Goals (xG) Explained: What the Number Really Means