Elo and Poisson Models, Explained in Plain Language
Elo and Poisson models are the two most established statistical approaches for turning football data into match probabilities: the Poisson/Dixon-Coles model treats goals as a counting process and produces a full score matrix, while the Elo rating maintains a single strength number that updates after every result. This guide explains the plain math behind both, and how they operate as two of the three core probability engines inside 11Stat's 12-engine analysis. It is football data analytics, not betting advice.
Why Poisson? Goals are a counting process
A goal in football is a rare event spread across 90 minutes: an average match produces roughly 2.6-2.9 goals, and each attack ends in a goal with only a small probability. That structure is the textbook case for the Poisson distribution. If you know a team's goal expectancy per match (λ, lambda), Poisson hands you the complete set of probabilities for scoring 0, 1, 2, 3... goals.
λ is not a fixed constant; it is derived from attacking strength, the opponent's defensive strength, the home/away effect and the league's scoring baseline. At 11Stat this expectancy is fed by xG (expected goals) data: chance quality, not the scoreboard, sets λ.
The Dixon-Coles correction: independence breaks at low scores
A pure Poisson model counts the two teams' goals as independent of each other; real data disagrees precisely at low scores. Draws like 0-0 and 1-1, and narrow results like 1-0 and 0-1, occur at different frequencies than independence predicts — a leading team slows the game down, and at level scores both sides manage risk. In the low-score region, goals are not fully independent.
The 1997 Dixon-Coles correction repairs exactly that corner: a small dependence parameter (ρ, rho) re-weights only the cells with 0 or 1 goals and leaves the rest of the matrix untouched. The model also down-weights old matches with exponential time decay, so a squad's form from two seasons ago cannot outvote its form from last month. Both terms are defined in the glossary.
From goal rates to a score matrix: 1X2, over/under and BTTS from one source
Combine the two teams' λ values with the Dixon-Coles adjustment and you get the score matrix: a probability for every exact score from 0-0 upward. That matrix is a single source of truth — the mass on either side of the diagonal yields the 1X2 win probabilities, the mass on either side of a total-goals line yields over/under, and the cells where both teams score at least once sum to the both-teams-to-score probability.
Because they all come from the same matrix, these percentages cannot contradict each other; that internal consistency is the most concrete advantage a probability model holds over hand-made predictions. And a percentage always carries uncertainty: a 60% probability means the outcome is expected in roughly 60 of 100 similar scenarios — the other 40 are part of the probability, not a failure of it.
Elo ratings: strength that updates with every result — plus home advantage
The Elo rating, adapted to football from chess, is a results-based strength measure. Every team carries one number; before kickoff the rating gap converts into an expected result, and after the match the difference between the actual and expected result — scaled by a K-factor — is added to the rating. A favourite beating a weak side gains almost nothing; a team producing an upset jumps.
Home advantage enters the model as bonus rating points — its size varies by league, but in typical implementations it shifts win probability by several percentage points in an otherwise even matchup. Recent form is captured by weighting the latest matches more heavily; how that signal is constructed is covered on the team form index page.
Where each model is strong, and where it is blind
The two models look at the same question through different windows:
- Poisson/Dixon-Coles feeds on goal production: it is naturally strong for score, total-goals and both-teams-to-score questions, but it is sensitive to squad turnover and thin data.
- Elo/form learns from results only: it is robust to noise and consistent across leagues, but it never sees how a match was won and suffers a cold-start problem early in a new season.
When both engines point the same way, confidence rises; when they diverge, that is not a malfunction but information — usually a sign that the scoreboard and the underlying production disagree. In that case 11Stat lowers the confidence score instead of amplifying the signal.
Inside 11Stat: two of the three core probability engines
In 11Stat's analysis pipeline every match passes through 12 engines, three of which are core probability engines: the Poisson/Dixon-Coles goal engine, the Elo/form engine, and the advanced bivariate engine. The two models on this page are the first two legs of that core; the third models low-score dependence in a more flexible way.
Engine outputs are never published raw: a calibration layer first aligns the probabilities with historical performance, a consensus rule then counts how many independent engines point the same way, and weak data quality dampens the signal. The full validation process is documented on the methodology page. The output is an auditable probability measurement, not a promise.
Frequently Asked Questions
Does the Poisson or Elo model guarantee the result?
No, it does not. No statistical model can guarantee the outcome of a single match; a model only produces a probability distribution. An outcome given 65% probability fails to happen in roughly 35 of 100 comparable matches — that is not an error, it is the definition of probability. 11Stat's outputs are data analytics, not a promise of results.
Is a high model probability a sure bet?
No — there is no such thing as a sure bet, and 11Stat does not use that framing. Even the most probable outcome fails at a meaningful frequency, and single-match variance is always large. What the platform shows instead is the probability, the engine consensus and the data quality, so the reader can see the uncertainty for themselves.
Why Poisson instead of a normal distribution?
Goal counts are non-negative small integers: 0, 1, 2, 3... The normal distribution is continuous and symmetric; in a match averaging 2.6 goals it would assign probability to nonsense values like minus one goal. Poisson is built for integer counts and fits the real distribution of goals in low-average, rare-event settings remarkably well.
What does the Dixon-Coles rho (ρ) actually do?
Rho adjusts the probabilities of only four scores — 0-0, 1-0, 0-1 and 1-1 — by a small factor. Pure Poisson is systematically off in that low-score region because the two teams' goals are not independent there. Rho is typically a small value and leaves the rest of the score matrix completely untouched.
How much is home advantage worth in Elo?
It varies by league and era; in typical implementations the bonus added to the home side translates into a shift of several percentage points of win probability in an even matchup. Data from matches played without crowds showed the advantage is not constant, which is why 11Stat estimates the parameter per league from data and validates it in simulated backtests.
Which is better, Elo or Poisson?
They are complements, not rivals. Poisson/Dixon-Coles produces a natural score matrix for goal-based questions; Elo delivers a results-based strength signal that is robust to noise. In simulated backtests neither is consistently superior on its own, which is exactly why 11Stat runs both — together with the advanced bivariate engine — inside one consensus.