Reading a prediction
Each match has a bar split into three: the chance of a home win, a draw and an away win. Wider means more likely. After the match, the part that actually happened stays bright and the rest fades.
A 70% favourite should still lose or draw about 3 times in 10. That is not the prediction being "wrong": football is unpredictable, and good predictions are honest about it.
How often are we right?
Over 12 past seasons, here is how often each method's most likely result was what actually happened:
Why isn't it higher? About 1 match in 4 ends in a draw, and a draw is almost never the single most likely result, so nobody calls those in advance. Upsets do the rest. That is why even the bookmakers, with up-to-date knowledge of injuries and team news, top out at around 55%.
The different ways to predict a match
Goals model
Our main prediction. Every team gets an attack rating and a defence rating from its results, with recent matches counting most. From these the model works out how many goals each side is likely to score, then the chance of every scoreline, and so of a home win, draw or away win.
Elo ratings
Our second model, invented to rank chess players and now used in FIFA's world rankings. Each team has one rating; after every match the winner takes points from the loser, more for a surprise win. Simple, and surprisingly good.
The bookmakers
Betting odds turned into chances (odds of 2.00 mean about 50%). The market knows about injuries and team news, so it is the benchmark to beat. You can compare it with our models in the Data lab.
Corners, cards and shots
The same idea as the goals model: each team gets a rating for corners, cards and shots won and conceded. These stats are much more random than results, so our numbers only beat the league average by a little (see the Data lab).
Season simulation
For the predicted table and the Market, we play out the rest of the season 10,000 times using our predictions. A team's title chance is simply how often it finished first.
What the models don't know: injuries, suspensions, new signings, a new coach, or a team resting players for a cup final. They only learn from results, so they catch up after a few matches. The bookmakers know straight away, which is a big part of why they are slightly more accurate.
Current team ratings
| # | Team | Elo | Attack | Defence |
|---|
Our prediction record
Before each round, our predictions are saved to a public record on GitHub and never edited afterwards. GitHub keeps a history of every change, so anyone can check what we predicted and when it was published.
Want the numbers?
The Data lab puts our prediction side by side with Elo and the bookmakers for every match, and has the full accuracy results season by season.