Poisson vs Dixon-Coles: Where the Difference Actually Matters
For most scorelines, the independent Poisson model and Dixon-Coles agree to within a fraction of a percent. The entire difference lives in four cells of the scoreline grid: 0–0, 1–0, 0–1, 1–1. Here's exactly how big the gap is, using a typical mid-table clash (λH = 1.45, λA = 1.15).
| Scoreline | Independent Poisson | Dixon-Coles (ρ=−0.13) | Difference |
|---|---|---|---|
| 0–0 | 7.4% | 9.0% | +1.6pp |
| 1–0 | 10.8% | 9.2% | −1.6pp |
| 0–1 | 8.5% | 7.0% | −1.5pp |
| 1–1 | 12.4% | 14.0% | +1.6pp |
| 2–1 (control) | 9.0% | 9.0% | 0.0pp |
The pattern is systematic: draws get more likely, 1–0/0–1 get less likely, and everything else is untouched. Aggregated to 1X2, the draw probability rises by about 3 percentage points — from 26.0% to 29.3% in this example.
Why this matters more than it looks
Three percentage points on the draw doesn't sound like much until you convert to odds. A 26.0% draw is fair at 3.84; a 29.3% draw is fair at 3.42. If the bookmaker's model includes the correction and yours doesn't, you're systematically overestimating the value of draw bets — or underestimating it, depending on which side you're on. Over hundreds of bets, that leak compounds.
When Poisson alone is fine
For totals markets (over/under 2.5) the correction barely matters: the shifts in the four cells largely cancel out in the total-goals sum. A plain Poisson is adequate for totals. It's specifically the draw — and correct-score bets on 0–0/1–1 — where Dixon-Coles earns its keep.
Try it yourself
Our match predictor lets you toggle between independent Poisson and Dixon-Coles on the same inputs, so you can see the draw shift live. The methodology page documents where the model still fails — because the correction fixes the average, not every match.
Related reading: How the Dixon-Coles Model Predicts Soccer Scores · How to Read Our Match Predictions