← Blog · Methodology Series

Bayesian Poisson Models for Football Score Prediction

Published: June 2026 · 7 min read

At the heart of FootballMoney's prediction engine is a Bayesian hierarchical Poisson model — the same class of models that Dixon and Coles introduced to football analytics in 1997. With 53.0% accuracy across 15,292 matches, it remains our strongest single predictor.

1. Why Poisson for Football?

Football goals are rare, discrete, non-negative events that occur at a roughly constant average rate over a fixed 90-minute period. The Poisson distribution is the natural choice: it models the probability of observing exactly k goals given an expected rate λ.

For a match between Team A at home and Team B away, the expected goal rates are:

The key insight: each team has an attack strength (how many goals they score above average) and a defense strength (how many they concede below average). The expected goals in a match is the product of the attacking team's attack strength and the defending team's defensive weakness.

2. The Bayesian Advantage: Priors

A pure Poisson model would estimate attack/defense parameters from data alone. But early in a season — or for newly promoted teams — data is sparse. This is where the Bayesian approach shines: we don't start from zero.

Instead, we place prior distributions on each parameter. For a newly promoted Championship side, the prior says: "You're probably about one standard deviation below the Premier League average in attack, with moderate uncertainty." The model starts skeptical and gets more confident as match data accumulates.

3. Hierarchical Structure

Our model is hierarchical: parameters aren't estimated independently for each team. Instead, league-level hyperparameters define the distribution from which individual team parameters are drawn. This provides partial pooling — teams "borrow" information from the league average when their own data is thin.

4. From Goals to Win/Draw/Loss Probabilities

Once we have λ_home and λ_away, we can compute the probability of every possible scoreline:

Then sum over all scorelines:

• P(home win) = sum of P(i,j) for all i > j
• P(draw) = sum of P(i,j) for all i = j
• P(away win) = sum of P(i,j) for all i < j

We cap i and j at 10 goals each (101 scorelines) since the probability of >10 goals is vanishingly small (< 0.01%).

5. The Low-Scoring Correlation Fix

A naive assumption is that home and away goals are independent — but real football shows a slight negative correlation at low scores. 0-0 and 1-1 draws occur more often than the independent Poisson model predicts, while 1-0 and 0-1 results are slightly less frequent.

We apply the Dixon-Coles correction: a parameter ρ that adjusts the probability of low-scoring outcomes (0-0, 1-0, 0-1, 1-1) to match empirical frequencies. This correction improves Brier score by approximately 0.005 — small but meaningful over thousands of predictions.

6. Why Ensemble Beats Single Models

Our Bayesian Poisson achieves 53.0% accuracy — but the ensemble (48.7%) has better calibration (Brier 0.190 vs 0.205). Why?

The Bayesian Poisson model is optimistic — it produces sharp, confident probability estimates that are right more often but wrong more confidently when wrong. The Gradient Boosted model (GB-RF) is more conservative, spreading probability across all three outcomes. Averaging them produces the best of both worlds: well-calibrated probabilities that neither over- nor under-estimate certainty.

For value betting, calibration matters more than accuracy. A model that says 55% and is right 55% of the time is far more useful than one that says 70% and is right 55% of the time — because the first enables accurate EV calculation and the second doesn't.