Maher (1982) and Dixon-Coles (1997): Football Scoring Models in Plain Words

Maher's football model connects team attack and defence to expected goals. Dixon-Coles adds a local dependence correction and gives older results less weight when estimating team strengths. Both produce probabilities across possible scores, rather than one certain result.

This introduction explains how the ideas fit together and how ImpliedScore uses them. The important distinction is between estimating goal rates from historical results and fitting them to current betting odds.

Try the Poisson football calculator →

In his 1982 paper, Maher examined Poisson models with team-specific attacking and defensive parameters. A strong attack facing a weak defence should have a different scoring expectation from the same attack facing a strong defence. Treating every fixture as if it shared one average loses that distinction.

A compact attack–defence model combines each team's attack, the opponent's defensive weakness and a common home effect. A larger defensive-weakness value means more goals conceded; some implementations use the opposite convention, so the label matters.

Once the home and away goal rates are known, independent Poisson multiplies the two goal-count probabilities to form each score cell:

P(home h, away a) = Poisson(h; λhome) × Poisson(a; λaway)

Maher found a reasonably good independent-Poisson fit in his data and also investigated a correlated bivariate extension. He did not establish independence as a universal law of football. The detailed Maher guide covers the original specification, estimation and paper results.

The 1997 Dixon-Coles model modifies the independent matrix at exactly four scores: 0–0, 1–0, 0–1 and 1–1. Its correction function, tau (τ), is controlled by a dependence parameter, rho (ρ).

With the standard sign convention used here, negative rho increases 0–0 and 1–1 while decreasing 1–0 and 0–1. Positive rho reverses that movement; zero recovers independence. Each multiplier must remain non-negative, which constrains rho according to the goal rates. Rho is not a general correlation coefficient.

This is a targeted modelling choice. It does not mean every competition has too few predicted draws, or that all dependence in football is confined to these four cells. The Dixon-Coles guide and correction calculator give the full tau formulas and valid parameter range.

Take home expected goals of 1.60, away expected goals of 1.10, and an illustrative valid ρ = −0.10. Keep both goal rates fixed while applying the correction.

The correction adds about 1.18 percentage points to each of the two low draws and removes the same amount from each 1–0 or 0–1 win. The changes cancel overall, preserving total probability and each team's goal distribution.

Probabilities are calculated before rounding. This is a mathematical illustration, not a forecast for a named match or a recommended rho for every league.

It changes 1X2 probabilities: in this example, total draw probability rises by about 2.37 percentage points, while home-win and away-win probabilities each fall by 1.18 points. Because both win probabilities shift equally at fixed rates, their relative order stays the same. The draw's position among the three outcomes can change.

BTTS Yes rises through the 1–1 cell. Over/Under 2.5 stays unchanged at fixed goal rates: all four corrected scores are Under 2.5, and their adjustments sum to zero. Lower lines such as 0.5 and 1.5 can change.

Scores such as 2–1 and 2–2 receive no direct correction. If a calculator refits the goal rates after changing model settings, those other scores and Over/Under 2.5 can move indirectly. See totals and BTTS from the score matrix for the calculation details.

When team strengths are estimated from past results, older matches may be less informative about current teams. Dixon-Coles uses exponential weighting to reduce their contribution to the fitting objective as they age.

This changes how the parameters are estimated; it is separate from adjusting the four cells of a particular match. A calculator that accepts two goal rates and rho implements the correction, but does not thereby fit the full historical team model. The decay setting also requires validation rather than an assumed universal value.

ImpliedScore starts with supplied 1X2 odds and optional Over/Under 2.5 odds. It first removes margin within each market, then fits home and away goal rates to the resulting probability targets. Removing margin does not itself calculate expected goals.

In Dixon-Coles mode, the correction is included while those rates are fitted, as documented in the site methodology. The site uses supplied or heuristic preset rho values; this is a different process from estimating team attack, defence and rho together from historical results.

The outputs describe what the supplied odds imply under the chosen model. They are not shot-based xG or independent evidence that the market is wrong. Historical time weighting is not part of this odds-inversion step. Input prices may reflect current information, but the calculator does not independently verify team news.

For a walkthrough from supplied odds to the displayed results, use the correct score calculator guide.

That depends on what it predicts on later matches. Compare forecasts made with the same available information, fit parameters without using future outcomes and evaluate the full probability distribution. A closer fit to the training sample is insufficient.

The Poisson, Dixon-Coles and bivariate Poisson comparison explains their different dependence structures. For evaluating predictions, start with calibration and the Brier score. Adding rho does not automatically improve accuracy or create a profitable betting strategy.