Maher (1982) — Modelling Association Football Scores

Summary

M.J. Maher's 1982 paper "Modelling Association Football Scores" (Statistica Neerlandica, Vol. 36, pp. 109–118) is the foundational academic work establishing the independent Poisson model for football match prediction. Maher proposed that the number of goals scored by each team in a football match follows an independent Poisson process, with expected goals (attack strength × defense weakness) derived from a baseline rate adjusted for team-specific attacking and defensive capabilities.

This paper introduced what is sometimes called the "BP model" (Birnbaum–Poisson or Basic Poisson). Maher tested the model against English Football League data and found that the independent Poisson model provides a reasonably accurate description of football scores, though a bivariate Poisson model with modest correlation (ρ ≈ 0.2) between home and away goals improves the fit slightly. This correlation is the conceptual precursor to the Dixon–Coles rho correction.

Key Concepts

  • Independent Poisson assumption: Home and away goals are modeled as independent Poisson random variables, each with its own λ (lambda) rate parameter
  • Attack/defense strength: Each team has an attack parameter and a defense parameter; λ_home = baseline_home × attack_home × defense_away
  • Maximum likelihood estimation: Parameters are estimated via MLE from historical match data
  • Bivariate Poisson extension: Maher noted that adding a correlation term (ρ ≈ 0.2) between home and away goals improves fit — this is the conceptual bridge to Dixon–Coles
  • Goodness-of-fit testing: Chi-square tests applied to assess Poisson fit; found to be adequate but not perfect

Formulas

Poisson rate for team i playing at home:
$$\lambda_i^{home} = \mu \cdot \alpha_i \cdot \beta_j$$

Where:
- μ = overall mean goals per home team per match
- α_i = attack strength of team i
- β_j = defense weakness of opponent j

For away team:
$$\lambda_j^{away} = \mu_{away} \cdot \alpha_j^{away} \cdot \beta_i^{home}$$

Constraint for parameter identification:
$$\prod_i \alpha_i = 1, \quad \prod_i \beta_i = 1$$

(one team's attack parameter normalized to 1, etc.)

Bivariate Poisson (Maher's extension):
$$P(X=h, Y=a) = \sum_{k=0}^{\min(h,a)} P_{corr}(k) \cdot P_{Poisson}(h-k; \lambda_1) \cdot P_{Poisson}(a-k; \lambda_2)$$

Where P_corr(k) is the correlation mass function controlling joint low-scoring outcomes.

Notes

  • Maher 1982 is the canonical citation for Poisson-based football models — any academic paper on sports prediction cites this
  • The paper's bivariate Poisson correlation finding (ρ ≈ 0.2) directly motivated Dixon and Coles' 1997 correction for low-scoring draws
  • The existing poisson-distribution.md note covers the mechanics well; this source adds the academic grounding and the bivariate extension
  • The model was estimated on English Football League 1992–1993 data (premier league data era) — basic structure applies universally to football
  • Key insight: Maher found that the Poisson model systematically under-predicts 0-0 and 1-1 draws, motivating subsequent corrections