The Poisson model behind football predictions

Why Poisson fits football

Goals are rare, independent-ish events with a stable average — exactly what the Poisson distribution describes. Given a team's expected goals (say 1.6), Poisson gives the probability of it scoring 0, 1, 2, 3… goals. Morley's 1958 analysis of World Cup scoring and a long line of research since found the fit good enough that this remains the standard starting point for match modelling.

From two teams to a scoreline matrix

Estimate each side's expected goals — Footalytics uses Elo-style strength ratings, recent form and home advantage — then apply Poisson to both and multiply: P(home scores i) × P(away scores j) fills the matrix cell for scoreline i–j. Sum cells and you get 1X2, over/under 2.5, BTTS and exact-score probabilities from one consistent object.

The Dixon-Coles correction

Raw Poisson under-rates low-score draws (0–0, 1–1) and over-rates some blowouts, and treats results long ago the same as last week's. Dixon and Coles's 1997 fix adjusts the joint distribution near low scores and time-weights the input data. It is the standard refinement in academic match modelling — and the version Footalytics runs.

What the model doesn't know

Line-ups, injuries confirmed an hour before kick-off, weather, motivation swings and manager changes are only partly reflected in the ratings. The model is a strong prior, not an oracle — which is why the AI layer can layer current context on top when you ask, and why every official pick is graded in public rather than trusted blindly.

See it on a real match

Open today's fixtures and ask about any match: the answer shows the probabilities, the top scorelines and the reasoning — the same distribution this guide describes, with the noise stripped away.