Over / Under

The Poisson Model for Football Goals

How the Poisson distribution turns scoring averages into Over/Under, 1X2, BTTS and correct-score probabilities — the model behind the calculator.

3 min de lecture#poisson#goals#over under#expected goals#score matrix#value bet

Introduction

The Poisson distribution is the workhorse of football goal modelling. Because goals are relatively rare events spread over 90 minutes, the number a team scores is well approximated by a Poisson process. This guide shows how to turn simple scoring and conceding averages into probabilities for Over/Under, 1X2, both-teams-to-score and exact scores — the exact model behind the Poisson calculator.

It is a transparent, first-principles approach: no black box, just a distribution and a few strength ratios.

1. The Poisson distribution

For a team expected to score λ goals, the probability of scoring exactly k is given by the Poisson probability mass function:

P(X = k) = \frac{\lambda^{k}\, e^{-\lambda}}{k!}

λ (lambda) is the mean number of goals. The whole model reduces to estimating a sensible λ for each side.

2. Expected goals from attack and defense

λ is built from three ingredients: the league scoring baseline, the team’s attacking strength and the opponent’s defensive strength. Strengths are ratios to the league average:

\text{Att}_{H} = \frac{\text{home goals scored}}{\bar{g}_{H}}, \qquad \text{Def}_{A} = \frac{\text{away goals conceded}}{\bar{g}_{H}}

Expected goals for each side then combine the baseline with the relevant strengths:

\lambda_{H} = \bar{g}_{H} \cdot \text{Att}_{H} \cdot \text{Def}_{A}, \qquad \lambda_{A} = \bar{g}_{A} \cdot \text{Att}_{A} \cdot \text{Def}_{H}

where ḡ_H and ḡ_A are the league-average home and away goals per game. A team scoring twice the league home average has an attack strength of 2.

3. The score matrix

Assuming the two teams’ goal counts are independent, the probability of an exact scoreline i–j is the product of two Poisson probabilities:

P(H=i,\, A=j) = \frac{\lambda_{H}^{\,i}\, e^{-\lambda_{H}}}{i!} \cdot \frac{\lambda_{A}^{\,j}\, e^{-\lambda_{A}}}{j!}

Building this matrix for i, j = 0, 1, 2, … (up to about 10 each captures nearly all the probability) gives the full distribution of results, from which every market follows.

4. Deriving the markets

Every market is a sum over the relevant cells of the score matrix. Writing Pᵢⱼ = P(H=i, A=j), the match result (1X2) is:

P(1) = \sum_{i>j} P_{ij}, \qquad P(X) = \sum_{i=j} P_{ij}, \qquad P(2) = \sum_{i<j} P_{ij}

Over/Under a goal line L (e.g. 2.5) and both-teams-to-score:

P(\text{Over } L) = 1 - \sum_{i+j < L} P_{ij}, \qquad P(\text{BTTS}) = \sum_{i\ge 1}\sum_{j\ge 1} P_{ij}

Because the sum of two independent Poisson variables is itself Poisson, total goals follow Poisson(λ_H + λ_A), so Over/Under can also be read directly from a single total-goals Poisson:

\text{Total goals} \sim \text{Poisson}(\lambda_{H} + \lambda_{A})

5. From probability to value (EV)

To find value, compare the model probability P̂ with the odds. Odds O imply a probability 1/O; the expected value of a unit stake is:

EV = \hat{P} \times O - 1

A positive EV means the odds are longer than the model’s fair price (1/P̂) — a potential value bet, never a guarantee on any single wager.

6. Limitations and refinements

The basic model assumes independence between the teams’ scores and treats λ as exact. Reality departs from this, especially for low scores (0–0, 1–1), which occur slightly more often than independence predicts. Common refinements:

Independence: real matches show mild dependence at low scores — the Dixon–Coles correction adjusts the 0–0, 1–0, 0–1 and 1–1 cells with a factor τ.
Static λ: form, injuries and match importance are not captured unless folded into the inputs.
Expected goals (xG): using xG instead of raw goals often yields a more stable λ.
Time decay: weighting recent matches more heavily keeps ratings current.

Conclusion

The Poisson model is the clearest bridge from simple averages to a full set of match probabilities. With just league baselines and each team’s scoring and conceding rates, it produces coherent Over/Under, 1X2, BTTS and correct-score estimates.

Treat its outputs as probabilities to inform judgement — refine the inputs (use xG, recent form, a Dixon–Coles correction) for a sharper λ, and always analyse within responsible limits.

References

Maher, M. J. (1982). Modelling association football scores. Statistica Neerlandica, 36(3), 109–118.
Dixon, M. J., & Coles, S. G. (1997). Modelling association football scores and inefficiencies in the football betting market. JRSS C, 46(2), 265–280.
Karlis, D., & Ntzoufras, I. (2003). Analysis of sports data by using bivariate Poisson models. JRSS D, 52(3), 381–393.
Baio, G., & Blangiardo, M. (2010). Bayesian hierarchical model for the prediction of football results. Journal of Applied Statistics, 37(2), 253–264.

References gathered as editorial background only; the text above is original to G10Tips.