Dixon-Coles Model explained for betting

In one sentence

Dixon-Coles is a Poisson goals model that rates each team's attack and defence, adjusts the four lowest scorelines, and down-weights older matches.

How it works

Mark Dixon and Stuart Coles published the model in 1997, aiming it squarely at beating football betting markets. It starts from the standard setup: each team has an attack rating and a defence rating, plus a home advantage, and those combine to give each side's expected goals.

They then noticed that plain Poisson gets 0-0, 1-0, 0-1 and 1-1 wrong: real matches finish 0-0 and 1-1 more often than predicted. So the model multiplies just those four scorelines by a correction factor controlled by one parameter, ρ (rho). Every other score is left alone, and the corrections are built so the probabilities still add up to 100%.

The second idea is time decay. A result from last month tells you more about a team than one from two seasons ago, so each past match is weighted by how long ago it happened.

The maths

P(X=x, Y=y) = \tau_{\lambda,\mu}(x,y)\,\frac{\lambda^{x}e^{-\lambda}}{x!}\,\frac{\mu^{y}e^{-\mu}}{y!}

\tau(0,0)=1-\lambda\mu\rho,\quad \tau(0,1)=1+\lambda\rho,\quad \tau(1,0)=1+\mu\rho,\quad \tau(1,1)=1-\rho

\lambda = \alpha_{\text{home}}\,\beta_{\text{away}}\,\gamma, \qquad \mu = \alpha_{\text{away}}\,\beta_{\text{home}}

λ and μ are home and away expected goals.
α is a team's attack strength, β its defensive weakness, γ the home advantage.
τ is the correction, equal to 1 for every score other than the four listed.
ρ controls the correction; a negative ρ adds 0-0 and 1-1 and removes 1-0 and 0-1.
Time decay multiplies each match's contribution by e raised to (−ξ × days ago), where ξ sets how fast old results fade.

In words: start with Poisson, tweak the low scores, and trust recent form more.

Worked betting example

Match Odds and Correct Score, illustrative inputs. Home expected goals λ = 1.3, away μ = 1.0, and ρ = −0.1.

Independent Poisson: P(0-0) 10.03%, P(1-1) 13.03%, P(1-0) 13.03%, P(0-1) 10.03%.
Corrections: 0-0 × 1.13, 0-1 × 0.87, 1-0 × 0.90, 1-1 × 1.10.
Adjusted: P(0-0) 11.33%, P(1-1) 14.34%, P(1-0) 11.73%, P(0-1) 8.72%. The full grid still sums to 100%.
P(draw) rises from 27.96% to 30.57%. Fair draw odds fall from 3.58 to 3.27.
The exchange shows 3.40 on the draw. Backing £10 under plain Poisson: 0.2796 × £24 − 0.7204 × £10 ≈ −£0.49. Under Dixon-Coles: 0.3057 × £24 − 0.6943 × £10 ≈ +£0.39, which shrinks to about +£0.25 after 2% commission.

The correction flips the verdict, but the edge after commission is only about 2.5% of stake. That is typical.

Where it's good

Match odds, draw, correct score and low totals (under 1.5, under 2.5) in football.
Lower leagues where the market is less sharp and data is still plentiful.
A well-understood baseline to benchmark fancier models against.
Updating ratings weekly with the time decay doing the forgetting for you.

Limitations and pitfalls

Published in 1997 and now widely copied. The major Betfair football markets already price in what it knows.
Results-based ratings react slowly to injuries, transfers and manager changes. Expected goals inputs help.
ρ is one number for the whole league; the true low-score effect may differ by team style.
The decay rate ξ is chosen by trial and error, which invites overfitting to your backtest.
Promoted teams and the start of a season are weak spots: little relevant data, big uncertainty.
It ignores in-match dynamics such as red cards and game state.

How to build it

Python: penaltyblog has a ready Dixon-Coles implementation; or write the likelihood and optimise with scipy.optimize.
Data: results with dates for two or more seasons; expected goals improves ratings.
Tip: choose ξ by walk-forward testing on log loss, never by in-sample fit.

Poisson: the core engine.
Bivariate Poisson: a different way to handle correlated scores.
Expected goals: better inputs than raw goals.
Model vs market: the test any Dixon-Coles build must pass.