A basic Poisson model is elegant, but it has a well-known blind spot: the smallest scorelines. Matches ending 0-0, 1-0, 0-1 or 1-1 are the most common in football, so errors there matter more than anywhere else.
What goes wrong
The simple model treats the two teams' goals as completely independent. In real matches they are not quite so. Once a team scores first, the other side may open up, or the leader may sit back and defend. Tight games also behave differently from open ones: two cautious teams can cancel each other out in a way that a pair of independent distributions does not capture well.
When analysts compared simple Poisson predictions with real results, they noticed a pattern. The model tended to misjudge how often matches finish with those low scores, and as a result it could misjudge how often matches end in a draw.
The Dixon-Coles correction
In 1997, Mark Dixon and Stuart Coles proposed a modest, clever adjustment. Rather than rebuilding the whole model, they introduced a small correction factor that applies only to the four low scorelines: 0-0, 1-0, 0-1 and 1-1. For those four cells of the scoreline grid, the probability is nudged up or down depending on the teams' expected goals. Every other scoreline is left as it was, and the probabilities are kept consistent so that they still add up to 100%.
The idea is not to claim that goals are mysterious or magical, only to admit that independence is a simplification and to repair it where it hurts most. Predicta applies this correction when it builds its scoreline grid.
Why it matters for draws
Draws come from the diagonal of the grid: 0-0, 1-1, 2-2 and so on. Because 0-0 and 1-1 dominate that diagonal, any adjustment to them has a direct effect on the estimated chance of a draw. Without a correction, a simple model can understate or overstate how often tight matches end level.
Even with the fix, draws remain the hardest of the three results to call. They are not tied to either team being stronger, and a single late goal can turn one into a win. That is not a flaw in the maths; it is the nature of a low-scoring sport.
Key takeaways
- Simple Poisson models treat goals as independent, which fits low scores imperfectly.
- Dixon and Coles (1997) added a correction for 0-0, 1-0, 0-1 and 1-1 only.
- This adjustment directly improves the estimated probability of draws.
- Draws stay the most difficult result to estimate, even with the correction.