METHODOLOGY
How does a match forecast work?
Our forecast is a Dixon-Coles statistical model (Dixon and Coles, 1997): every team has an attacking strength and a defensive strength, all teams share a single home advantage, and the expected goals of the two teams are derived from these values. Adding up the probability of every exact score, up to 6 goals each, gives the probabilities of a home win, a draw and an away win.
Frequently asked questions
What does time decay mean?
Older matches count for less: each match is weighted by exp(−ξ × days), where days is the time since the match and ξ is 0.001. With this value the weight halves in about 693 days. ξ is a fixed value, not estimated together with the team strengths, because estimating it jointly would lead to ignoring everything except the latest match.
What data is it calculated from?
From the results of finished matches in all the seasons in our data; with too few matches the model is not fitted. It uses only goals scored: xG, line-ups, injuries and suspensions do not enter.
Why is there no forecast for some teams?
The model assigns a strength only to clubs that appear in the historical data. A club with no Serie A history in our data, such as a newly promoted side, has no estimated strength and so no Dixon-Coles forecast.
Limits and missing data
A probability is not a prediction of what will happen: it says how plausible each outcome is according to the model.