How the GoalPredicts Prediction Model Works
An overview of the statistical model behind GoalPredicts predictions: Poisson goal estimates, team strengths, recency weighting and known limitations.
By GoalPredicts Editorial Team · Reviewed
Expected goals from team strengths
Our baseline model estimates how many goals each team is expected to score in a match. It starts from the competition's average goals for home and away sides, which captures home advantage, and adjusts them by each team's attacking strength and their opponent's defensive strength.
Strengths are measured relative to the competition average, so a team with an attack rating of 1.2 has scored about 20% more than an average side would against the same opponents. Ratings are adjusted for opponent quality, so scoring against strong defences counts for more than scoring against weak ones.
From expected goals to probabilities
The model treats each team's goals as following a Poisson distribution, a standard statistical model for counts of independent events. Combining the two teams' distributions gives the probability of every possible scoreline, and adding up the relevant scorelines gives the probability of each outcome: home win, draw, away win, over or under a goal line, and both teams scoring.
This approach was described for football by M. J. Maher in 1982 and remains a common baseline because it is transparent and easy to check.
Recency and small samples
Recent matches carry more weight than old ones; a match from six months ago counts for half as much as one played yesterday. Teams with only a few matches in our data are pulled towards the league average, so a lucky or unlucky start doesn't produce extreme estimates.
If a team has played too few matches in the competition, the model does not produce a prediction at all. Newly promoted sides, for example, get no model predictions until they have played enough games.
No look-ahead
A prediction only uses matches that finished before it was generated. When we test the model on past seasons, each historical match is predicted using only the matches played before it, the same way it would have been in real time.
Known limitations
The model does not know about injuries, suspensions, line-ups, motivation, weather or fixture congestion. It assumes the two teams' goals are independent, which slightly underestimates low-scoring draws. It models each competition separately, so cup and continental matches between teams from different leagues are not covered.
Model output is reviewed by an editor before publication, and any manual change keeps the original model output on record. Predictions remain estimates and can be wrong.