Predictions

For the result and the main goals lines we start from the betting market's own prices, with the bookmaker margin removed. Across several seasons of testing, no model of ours beat those prices on these markets, so we don't pretend to. We turn them into the expected goals of each team and read every other goals line from there: over 1.5 and 3.5, both teams to score, each team's goals, and the first and second half.

When a match has no price yet, a model built from results alone takes over: each team's attack and defence, with recent games counting more, and thin records pulled towards the league average. Those matches are marked Model only and never get a VALUE flag.

Here we build the prediction ourselves, from what is known before kickoff: each team's recent numbers for and against (newer games weigh more, and a team with few games is pulled towards the league average), how one-sided the match is expected to be, and how many goals it should have. Corners are modelled as a match total rather than two separate teams, because one side's corners come at the other's expense.

Cards use the same inputs plus the referee: how many cards he shows per match, weighed by how many matches we have seen him in. We add him as soon as the appointment is announced. We count only cards shown to players who took part in the match, and a second yellow counts as a yellow and a red.

Each model was tested season by season, predicting every match with a version trained only on the seasons before it — never on the games it was being tested on. A market is shown only if it beat simple alternatives such as the league average and the average of each team's last ten games. Markets that didn't pass are held back; shots on target in Norway and Sweden are one of them, because the history mixes two ways of counting shots.

Fair odds are 1 ÷ probability: the price at which a bet would break even in the long run. When a real price is higher than our fair odds, the difference is the expected value (EV).

We call it VALUE only when the price is at least 3% above fair, we rate the pick at 35% or more, the match is within 72 hours (earlier prices move too much), and the price is in line with the rest of the market — a single price far above all the others is usually a mistake, not a chance. For corners and shots on target we often hold only one price; there the bar is 5%, anything above 25% is ignored as a likely stale line, and the price is marked 1 price.

Most likely lists the model's most confident call per match, at 70% or more. Trends are streaks that held in at least 80% of a team's league games this season, once it has played six. Next to a pick you see how often it happened in each team's games this season, and in the league meetings between the two.

Limited data marks a match where a team has fewer than five games in our history — a promoted club, early in the season. Until a team has five games this season, its facts also include last season (or the second tier), and we say so.

Injuries and suspensions are shown on every match page, but the model doesn't yet adjust for who is missing: the prices it starts from do, partly, for result and goals. Nothing on these pages knows about a late change after the last update time shown at the top.

Every prediction is saved before kickoff and graded after the final whistle; nothing is edited afterwards. The Results page shows how each probability band actually came in. Predictions stay in Beta until each kind of market has at least 200 graded picks.

This is statistical analysis for adults (18+). We don't accept bets, we don't work for a bookmaker, and a probability is not a promise about any single match. If betting stops being fun, stop.