Maher’s 1982 Poisson Model for Football Scores
Michael Maher’s 1982 paper established the attack–defence Poisson framework that still underlies many football score models. Instead of treating every team as if it had the same scoring distribution, the model gives teams separate attacking and defensive strengths, combines them with home advantage and estimates the parameters from match results.
Original 1982 paper: Michael J. Maher, Modelling association football scores, Statistica Neerlandica, 36(3), 109–118. DOI: 10.1111/j.1467-9574.1982.tb00782.x . ImpliedScore links to the publisher/DOI record and does not host or reproduce the paper PDF.
- Home and away goals are modelled as Poisson variables.
- The expected goals depend on one team’s attack and the opponent’s defensive weakness.
- Maher tested whether teams needed separate home and away strength parameters.
- The selected independent model used common home advantage rather than team-specific home effects.
- The independent model fitted reasonably well, but a bivariate Poisson extension improved the fit.
| Element | Maher’s selected Model 2 |
|---|---|
| Home scoring mean | αiβj |
| Away scoring mean | k²αjβi |
| Team effects | One attack and one defensive-weakness parameter per team |
| Home advantage | Common factor for all teams |
| Base dependence | Independent home and away scores |
| Extension | Bivariate Poisson with correlation ρ ≈ 0.2 |
| Data | Four English divisions × three seasons (12 datasets) |
ρ notation warning: Maher’s ρ in the bivariate-Poisson extension is a positive score-correlation parameter. It is not the Dixon–Coles ρ used in the four-cell τ correction, and the two parameters are defined differently and are not numerically comparable.
Why Maher’s model mattered
Earlier studies had compared pooled football goal counts with a single probability distribution. Those pooled counts appeared more dispersed than a Poisson distribution, leading researchers to prefer a negative binomial model.
Maher challenged the pooling assumption. A strong attack facing a weak defence should not have the same expected number of goals as a weak attack facing a strong defence. Once each match is allowed to have its own mean, much of the apparent failure of Poisson can disappear.
The central contribution was team heterogeneity. Maher did not claim that every football match has one universal Poisson mean. He modelled the mean as a function of the two teams involved.
This distinction created the basic architecture used by later independent Poisson, Dixon–Coles, bivariate Poisson and dynamic team-strength models.
Why pooled football scores can look non-Poisson
Suppose goals are conditionally Poisson once the teams and match context are known:
If Λ changes from match to match, the variance of the pooled data is:
The first term is ordinary Poisson variance. The second comes from differences between teams and fixtures. Therefore, pooled variance can exceed the pooled mean even when every match is conditionally Poisson.
If half the fixtures have Λ = 0.8 and half have Λ = 2.0, the pooled mean is 1.4. The pooled variance is 1.4 + 0.36 = 1.76. The extra 0.36 comes from mixing different scoring rates, not necessarily from a failure of the conditional Poisson assumption.
A gamma mixture of Poisson rates produces a negative binomial distribution exactly. Maher’s point was more general: a negative-binomial-like aggregate pattern can arise because fixtures have different means.
The original attack–defence model
Let team i play at home against team j. In Maher’s most general starting specification:
The two goal counts were initially assumed independent. The parameters had the following interpretation:
This gives four raw parameters per team: home attack, away attack, home defence and away defence. Maher did not simply assume that all four sets were necessary. He tested a hierarchy of nested models.
| Level in the hierarchy | What varies by team? | Question being tested |
|---|---|---|
| Model 0 | Nothing | Can all teams be treated as equal? |
| Models 1A / 1B | Attack only or defence only | Is one dimension of team quality sufficient? |
| Model 2 | Attack and defence | Do teams need separate attack and defence ratings? |
| Models 3C / 3D | Partial home/away separation | Does one type of strength need a venue-specific parameter? |
| Model 4 | Home/away attack and defence separately | Is the fully separated model justified? |
Likelihood-ratio tests showed that team-specific attack and defence were important, but separate home and away strength systems were not consistently supported. Maher therefore selected the more compact Model 2.
A modern form of Maher’s selected model
In Maher’s original notation, Model 2 linked the away parameters to the home parameters through one constant: δi = kαi and γi = kβi. The home scoring mean stayed αiβj, while the away mean became:
Relative attacking and defensive strengths are therefore identical at home and away, and every team’s scoring power is reduced by the same factor k² when playing away — the origin of the modern common home-advantage factor.
The same structure can be written in a modern log-linear form, which is an equivalent reparameterization rather than Maher’s original notation. Let ai be team i’s attacking effect, di its defensive-weakness effect, h the common home effect and c the league scoring baseline:
Equivalently, on the multiplicative scale:
where B = ec is the baseline, H = eh is home advantage, A is attacking strength and D is defensive weakness. A value above 1 increases expected goals; a value below 1 decreases them.
Defence needs a sign convention. In this page, a larger D means a weaker defence because it raises the opponent’s expected goals. Some software instead defines defensive strength, where larger values mean better defence. The formulas then require a minus sign or division.
Once λ and μ are known, any exact-score probability follows from the independent product:
The matrix can then be summed to obtain home win, draw, away win, totals, BTTS and other derived markets. See Poisson distribution in football for the mechanics of the score matrix.
How the parameters are estimated
For match m, let the observed score be (xm, ym) and the modelled means be (λm, μm). Under independent Poisson scoring, the log-likelihood is:
Maximum-likelihood estimation chooses the attack, defence and home parameters that maximize this expression. Maher described iterative estimation, including Newton–Raphson and an alternating scheme that repeatedly updates attack and defence parameters.
Why an identification constraint is required
The model is not unique without normalization. Multiplying every attack strength by a constant and dividing every defensive-weakness parameter by the same constant leaves all products unchanged — and because the modern form also carries a separate baseline c, a second, independent rescaling exists: the shifts ai → ai + s and di → di + t are invisible whenever the baseline absorbs them as c → c − s − t.
Two normalization constraints are therefore normally required, for example:
Equivalent alternatives include:
- normalizing the geometric means of the attack and defence multipliers to 1;
- fixing both the attack and the defence parameter of one reference team;
- fixing the mean attack effect and the mean defence effect.
Fixing only one parameter — a single reference team’s attack alone, for instance — removes one degree of freedom but leaves the other rescaling unconstrained. Maher’s original multiplicative parameterization needed only the single constraint Σαi = Σβi because it carried no separate baseline; adding c in the modern form is what creates the second freedom. Different valid constraints change the numerical scale of the ratings, but not the fitted λ and μ values.
Worked example
Consider a hypothetical match with these multiplicative parameters:
| Parameter | Value | Meaning |
|---|---|---|
| League baseline B | 1.25 | Baseline expected goals before team and venue effects |
| Home factor H | 1.22 | Home scoring rate is multiplied by 1.22 |
| Home attack Ai | 1.18 | 18% above the normalized attack level |
| Away defence weakness Dj | 1.10 | Allows 10% more than the normalized defence |
| Away attack Aj | 0.95 | 5% below the normalized attack level |
| Home defence weakness Di | 0.90 | Allows 10% fewer than the normalized defence |
The expected goals are:
Display values are rounded to two decimals. The probabilities and fair odds below use the unrounded products λ = 1.97945 and μ = 1.06875.
The resulting independent-Poisson outputs include:
| Output | Probability | Fair decimal odds |
|---|---|---|
| Home win | 58.50% | 1.71 |
| Draw | 21.52% | 4.65 |
| Away win | 19.98% | 5.01 |
| Over 2.5 | 58.75% | 1.70 |
| Both teams to score | 56.59% | 1.77 |
The highest individual score probabilities are close together:
| Score | Probability |
|---|---|
| 1-1 | 10.04% |
| 2-1 | 9.93% |
| 1-0 | 9.39% |
| 2-0 | 9.29% |
The most likely score is not a high-confidence prediction. In this example, 1-1 ranks first but still has only a 10.04% probability. The model is a distribution over many outcomes, not a deterministic score forecast.
What Maher found in the 1982 paper
Maher analysed 12 datasets: the four English Football League divisions across three seasons, 1971–72 to 1973–74.
Attack and defence both mattered
Likelihood-ratio comparisons strongly supported team-specific attacking and defensive parameters. The fully separated home/away parameter system was not consistently justified, leading Maher to adopt the more compact structure with common home advantage.
The conditional Poisson fit was reasonably good, not perfect
Across the 12 datasets, Maher performed separate goodness-of-fit tests for home and away goal frequencies. Nineteen of the 24 tests were non-significant at the 5% level. This was evidence that a team-adjusted Poisson model described marginal goal counts reasonably well.
Non-rejection in 19 of 24 in-sample tests is not proof that the conditional distribution is Poisson, and five tests did reject it. A non-significant result alone cannot establish whether the model is adequate or a test has limited power. In Maher’s pooled comparison, when the observed and expected frequencies of all twelve datasets were combined, the χ² statistics of 16.2 for home scores and 28.8 for away scores led to clear rejection of the model. Small, directionally consistent residuals accumulate.
Small systematic differences remained. In the pooled comparison across the datasets, the fitted model tended to overestimate zero and high goal counts and underestimate some one- and two-goal frequencies. Maher described the conditional distribution as slightly narrower than Poisson.
Independence was the weaker assumption
The independent model underestimated draws in the goal-difference distribution. Maher therefore tested a bivariate Poisson construction in which the two teams share a common Poisson component:
Here, U, V and W are independent Poisson variables and the shared component W induces positive covariance. Maher parameterized the construction so that the correlation, rather than the shared-component mean, was approximately constant across fixtures. For a match with means λ and μ:
Then Cov(X, Y) = q and Corr(X, Y) = ρ for every fixture, which is why one value of ρ could be applied across matches with different means. The construction requires 0 ≤ q ≤ min(λ, μ); otherwise a component mean would turn negative. This differs from modern presentations that treat the shared-component mean itself as one constant parameter, where the implied correlation then varies with the marginal means.
The extension was not a full joint re-estimation. Using the attack, defence and home parameters already fitted under the independent model, Maher tried a range of common correlation values and checked the fit of the score-difference distribution. A value around 0.2 improved it substantially, and the same value was applied to all twelve datasets.
A common oversimplification: Maher’s paper is often described only as an independent Poisson model. Independence was the main model he developed and tested, but the paper also proposed and evaluated a bivariate Poisson extension.
Maher, Dixon–Coles and bivariate Poisson
| Model | Goal rates | Dependence treatment | Main distinction |
|---|---|---|---|
| Maher independent Poisson | Team attack × opponent defence × home effect | None | Established the team-strength framework |
| Maher bivariate extension | Same team-strength framework | Shared Poisson component | Introduces positive covariance across the matrix |
| Dixon–Coles | Closely related attack–defence structure | Local τ adjustment at 0-0, 1-0, 0-1 and 1-1 | Adds low-score correction and time weighting |
| Modern bivariate Poisson | Can include dynamic and Bayesian team effects | Explicit covariance parameter | Generalizes dependence and parameter estimation |
Dixon and Coles explicitly built on Maher’s attack–defence structure. Their 1997 model added a correction for four low scorelines and downweighted older matches. See the Dixon–Coles model explained.
Later work by Karlis and Ntzoufras developed bivariate Poisson methods more systematically, including models for sports data with a shared latent scoring component. The comparison of football score models separates that construction from the local Dixon-Coles correction.
How this differs from ImpliedScore
Maher’s model estimates team strengths from historical results. ImpliedScore starts from supplied betting-market prices and fits λ and μ to approximate margin-free 1X2 and optional Over/Under 2.5 probabilities. With both markets supplied, three independent constraints generally cannot be matched exactly by two rates when the dependence setting is fixed.
| Component | Maher model | ImpliedScore |
|---|---|---|
| Input | Historical match scores | Supplied 1X2 odds and optional Over/Under 2.5 odds |
| Team attack and defence | Estimated explicitly | Already reflected indirectly in market prices |
| Home advantage | Estimated from results | Embedded in the odds |
| Output | Model-generated match probabilities | Market-implied score probabilities |
| Main use | Rating and forecasting teams | Decoding the probability structure of odds |
Both approaches end with a score matrix, but they reach λ and μ from different evidence. ImpliedScore does not estimate attack and defence separately; it treats current market prices as a compressed estimate that may reflect injuries, form, tactics, venue and other information — imperfectly, and with no guarantee that the odds are objectively correct.
Convert 1X2 and Over/Under odds into market-implied λ values and a complete score matrix.
Open the football score calculatorFor the optimization process used by the site, see the ImpliedScore methodology.
Limitations of the Maher framework
Team strengths are treated as fixed within the fitted period
Managers, players, tactics and injuries change. A static parameter estimated over a long window can lag behind current team quality.
Independent Poisson ignores game interaction
A goal changes incentives. Leading teams may defend more deeply, trailing teams may take greater risks and red cards alter both scoring rates. Maher’s own bivariate extension was an aggregate model for dependence between final scores. It did not model score-state or event-driven intensity changes themselves, and fitting covariance does not establish its causal mechanism.
Poisson imposes equality of conditional mean and variance
After controlling for teams, football scores can still show residual underdispersion or overdispersion. A good aggregate fit does not prove every fixture follows the same mechanism.
Home advantage need not be constant
The selected model used a common home effect, but travel, stadium, crowd, altitude and team style can make venue effects heterogeneous.
Parameter count grows with the league
Each team requires at least an attack and defence parameter. Promoted teams, sparse schedules and cross-league matches create estimation problems without regularization or hierarchical priors.
Good fit does not imply a betting edge
Maximum likelihood finds parameters that explain the sample. Profitability requires out-of-sample probabilities that outperform market prices after margin, limits and execution costs.
Implementation checklist
- Define clearly whether defence means strength or weakness.
- Use one normalization for Maher’s original scale, or two for the log-linear form with a separate baseline.
- Separate training and test periods.
- Check calibration and log loss, not only correct-score hits.
- Compare static, time-weighted and dynamic strength estimates.
- Test independence rather than assuming it.
- Do not confuse a fitted historical model with market-implied λ.
Original paper and canonical citation
Canonical citation. Maher, M. J. (1982). Modelling association football scores. Statistica Neerlandica, 36(3), 109–118. https://doi.org/10.1111/j.1467-9574.1982.tb00782.x
| Author | Michael J. Maher |
| Journal | Statistica Neerlandica (Wiley) |
| Published | September 1982 · Volume 36, Issue 3 · Pages 109–118 |
| DOI | 10.1111/j.1467-9574.1982.tb00782.x |
What the paper introduced
Earlier work had rejected a simple pooled Poisson description of football scores in favour of the negative binomial without fully accounting for differences in team quality. Maher re-examined the Poisson approach with team-specific attack and defence parameters estimated by maximum likelihood, compared a hierarchy of model specifications, and showed that a team-adjusted independent-Poisson model describes observed score frequencies reasonably well. He also showed that independence is not exact: a bivariate Poisson extension with positive score correlation improved the fit, an issue later addressed more locally by Dixon and Coles (1997).
How ImpliedScore relates to it
ImpliedScore uses the same independent-Poisson score-matrix foundation, but it obtains the goal rates from different evidence. Maher estimated team attack and defence strengths from historical results; ImpliedScore infers λ Home and λ Away from bookmaker prices after margin removal. The site therefore uses Maher's probabilistic foundation without presenting its market-implied λ values as a historical attack-defence fit.
We link to the publisher/DOI record and do not host the paper itself.
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 . Journal of the Royal Statistical Society: Series C, 46(2), 265–280.
- Karlis, D. & Ntzoufras, I. (2003). Analysis of Sports Data by Using Bivariate Poisson Models . The Statistician, 52(3), 381–393.