Odds and probability
How Football Prediction Models Work: Poisson, Elo, xG and the Market
How football prediction models turn data into probabilities: Poisson goals, Dixon and Coles, Elo ratings, expected goals and market prices, with limits.
Give a home side an expected 1.6 goals and the visitors 1.1, and a simple model can tell you the chance of a 1-1 draw: about 11.8 percent. The chance of 0-0 comes out near 6.7 percent. Add up every scoreline where the home team scores more and the home win sits close to 49 percent, the draw about 25, the away win about 26. Those three numbers come from two inputs and one well-known formula, and they are the starting point for most serious football prediction models.
Models do not see the future. What they offer is consistency: the same evidence always produces the same estimate, which makes it possible to test whether the estimates were any good.
Poisson Goal Models: From Averages to Scorelines
The Poisson distribution describes how often a rare event happens in a fixed window when the events arrive more or less independently. Goals in a football match fit that description reasonably well. Feed it an average and it returns the probability of each count.
| Goals | Home side (average 1.6) | Away side (average 1.1) |
|---|---|---|
| 0 | 20.2% | 33.3% |
| 1 | 32.3% | 36.6% |
| 2 | 25.8% | 20.1% |
| 3 | 13.8% | 7.4% |
| 4 or more | 7.9% | 2.6% |
Multiply a home row by an away row and you get that scoreline’s probability, assuming the two sides score independently. The 1-1 figure in the opening paragraph is simply 32.3 percent times 36.6 percent. The same grid feeds the goals markets too; the page on over/under lines and both teams to score uses it to price a 2.5 goals line.
The hard part is the averages. Mike Maher’s 1982 paper proposed giving each team an attacking strength and a defensive strength, estimated from past results, plus a home advantage term. A team’s expected goals in a match is then its attack, multiplied by the opponent’s defensive weakness, multiplied by home advantage where it applies. With twenty clubs and a season of results, those strengths can be fitted quite well.
- Fit team strengthsAn attacking and a defensive strength per club from past results, plus a home advantage term.
- Set the averagesAttack times the opponent's defensive weakness, times home advantage where it applies.
- Count the goalsPoisson turns each average into a probability for 0, 1, 2 and more goals.
- Build the gridMultiply a home row by an away row for every scoreline, then add them into home, draw and away.
The Dixon and Coles refinement
Mark Dixon and Stuart Coles published an improved version in 1997. They noticed that a plain Poisson model gets the lowest scores slightly wrong, particularly 0-0 and 1-1, which happen more often than independence predicts. Their adjustment corrects those four low scorelines. Their second change mattered just as much: older matches get less weight than recent ones, so a team that changes manager in November is not judged mainly on its form from the previous spring.
- 1982Mike Maher gives each team an attacking and a defensive strength.
- 1997Dixon and Coles correct the low scorelines and weight recent matches more.
- 2010sOpta and other data companies bring expected goals into mainstream coverage.
- NowMany goal models use xG instead of raw goals for attack and defence strengths.
Elo Ratings: One Number Per Team
Arpad Elo designed his rating system for chess, and football borrowed it because it is simple and adapts quickly. Each team has a single number. Before a match, the difference between the two ratings gives an expected result; after the match, both ratings move by an amount proportional to how surprising the outcome was.
A 100-point gap gives the stronger side an expected score of about 0.64 under the standard formula, where a win counts 1 and a draw 0.5. Beat a much weaker side and your rating barely moves. Lose to them and it drops sharply. Football versions add home advantage and often a factor for the margin of victory. The Club Elo project publishes ratings for European clubs going back decades, and national-team Elo lists exist as well.
Elo’s strength is that it needs only results. Its weakness is the same: it knows nothing about who played, how many chances were created or whether a 1-0 win was a smash-and-grab. It also gives an expected score rather than separate win, draw and loss chances, so turning it into a three-way forecast needs an extra step.
Expected Goals: Measuring Chances Instead of Scores
Expected goals, usually shortened to xG, assigns each shot a probability of becoming a goal based on where it was taken, how, and in what situation. A penalty is worth roughly three-quarters of a goal; a speculative effort from 30 yards might be worth 0.03. Add up a team’s shots and you get the goals an average finisher would have scored from those chances.
Data companies such as Opta brought xG into mainstream coverage over the 2010s, and it is now standard in Premier League broadcasts. For prediction its value is in stability. Scorelines swing on deflections and goalkeeping errors; chance quality is steadier, so a team’s recent xG for and against tends to forecast its next matches better than its recent goals do. That is why many modern goal models replace raw goals with xG when estimating attack and defence strengths.
One honest caveat: xG data for smaller competitions is patchier. Coverage of the NPFL or the Kenyan Premier League is far thinner than for Europe’s top divisions, and a model built for one data-rich league can mislead badly when carried to another without that data.
The Market as a Model
A betting market is itself a forecast, built from the views of everyone willing to put money behind an opinion. Strip out the bookmaker margin, as described in the walk-through of odds and implied probability, and the remaining percentages are often more accurate than any single statistical model, especially close to kick-off when late team news has been absorbed.
A model that matches the closing price closely has learned something real; a model that differs wildly is more often wrong than brilliant.
For model builders this is useful and humbling at once. Useful, because the market makes a strong benchmark. Humbling, because beating it consistently is rare. A model that matches the closing price closely has learned something real; a model that differs wildly is more often wrong than brilliant.
Strengths and Limits Side by Side
Each approach covers a gap in the others. Poisson models produce full scoreline grids but depend on good strength estimates. Elo adapts quickly with minimal data but sees only results. Expected goals measures underlying performance but needs detailed shot data. Market prices fold in information no public dataset holds, but they include margin and can be swayed by popular clubs.
Poisson
- Full scoreline grids
- Depends on good strength estimates
Elo
- Adapts quickly, needs only results
- Blind to who played and chances made
Expected goals
- Measures underlying performance
- Needs detailed shot data
Market price
- Holds information no public dataset has
- Includes margin, swayed by popular clubs
None of them handles the things that happen off the pitch well: a squad hit by illness, a club in financial trouble, a side with nothing to play for in May. That is where human judgement tends to add or subtract the most, and where it most needs to be written down and checked.
How to tell whether any of this works in practice is the subject of the guide to measuring prediction accuracy, which covers the Brier score, calibration and why a few hundred matches are needed before a record means much. The homepage explainer on direct win forecasts shows the same probabilities applied to a single fixture.
A model that rated every match 46 percent home, 26 draw and 28 away would be roughly calibrated in many leagues and useless for picking out any particular game. Try comparing your own forecasts with that flat baseline first. If they cannot beat it over a season, the problem lies in the inputs rather than the formula.
FAQ
Your questions, answered.
What is the most accurate football prediction model?
No single model wins everywhere. In big leagues, the closing market price is usually the hardest benchmark, because it already absorbs team news and the views of many informed people. Well-built Poisson or Dixon-Coles models, strengthened with expected goals data, come close in many competitions. Elo ratings are simpler and surprisingly competitive on match results, though they say nothing about scorelines.
Why do football models use the Poisson distribution?
Goals are rare, roughly independent events spread across ninety minutes, which is the situation the Poisson distribution describes well. Give it an average, say 1.4 goals for a team in a given match, and it returns the chance of that team scoring zero, one, two or more. Pairing two such distributions produces probabilities for every scoreline.
What did Dixon and Coles add to the Poisson model?
Their 1997 paper made two changes. It adjusted the probabilities of the low scores 0-0, 1-0, 0-1 and 1-1, which a basic Poisson model tends to misjudge, and it gave recent matches more weight than older ones so that team strengths could change through a season. Both ideas remain standard in modern goal models.
Can expected goals predict the next result?
Expected goals is better at describing how well a team has been playing than raw scores are, because it strips out much of the luck in finishing and goalkeeping. That makes it a useful input for the next few matches. It does not predict a single result on its own, and it misses things such as injuries, tactics against a particular opponent or a change of manager.
Do bookmakers use prediction models?
Prices for major football markets are widely understood to start from statistical models and then move with the money and information that arrive before kick-off. That is one reason the price close to the start is hard to beat. The exact methods are private, so outside forecasters judge their own models by comparing them with those prices over many matches.
How should a home-made model be tested?
Build it on one block of matches, then score its forecasts on later matches it has never seen, using the Brier score or log loss. Compare the result with a simple benchmark such as the league's average home, draw and away rates, and with the market price. A model that only looks good on the data it was built from has learned noise.