How our numbers are made

Every chance on this site is an estimate. This page says where it comes from, how a pick is chosen, and what the method can't do. A chance of 75% still loses about one time in four.

What goes in

Fixtures, results, league tables, injuries and suspensions, top scorers and bookmaker prices, all from API-Football. We scan each night at about 23:15 UK time for the next day's matches, and nothing after that moment is used. Line-ups aren't known then, so they play no part.

How a chance is made

Each team gets an attacking and a defensive rating from its goals scored and conceded this season and last, at home and away separately. Early in a season the ratings lean on last season; as games are played, this season takes over. Multiplying the ratings with the league's average goals gives each side's expected goals.

Small, capped adjustments follow: recent form (at most 3%), past meetings between the two teams (at most 4%), how many goals both teams' recent games had (at most 10%), and missing players, chiefly a missing top scorer.

From the expected goals, a standard statistical model gives the chance of every scoreline, and the scorelines add up to the chance of each result: home win, draw, away win, over 1.5 and 2.5 goals, both teams to score. A last correction, fitted on thousands of past matches, adjusts where the raw model tends to be too confident or not confident enough.

Which number is the Win chance

For home win, away win and win with money back on a draw, where 3 or more bookmakers price the match, the Win chance is the average of their prices with their profit margin removed. Tested on past matches, the bookmakers judged match results better than our model, so on these markets we show theirs. Everywhere else, and for every goals market, it is our model's chance. Every pick shows both numbers where both exist.

How a pick is chosen

One pick per match at most: the market with the highest Win chance, if it is 75% or more.
Where the bookmakers price a market, our model and theirs must be within 10 points of each other. If they disagree by more, there is no pick: one of us is likely wrong.
Some high bands are held back until our own published picks show they are judged well: both teams score at 60% or more, and over 1.5 goals at 90% or more.
Only leagues on a list we review are picked from, and only matches still to kick off.
Every pick has one sentence explaining it, written from the stored statistics.
The day's top picks are the 20 with the highest Win chance, chosen once when the day is published. They don't change as matches start.

Once published, a pick is never changed.

How results are judged

Each pick is frozen at kick-off, as published. After the match it is marked won or lost from the final score. A draw on a “money back on a draw” pick is money back, a match called off is void, and a match no licensed result source covers is marked unverified. None of those counts as won or lost. If a result has to be corrected, the correction is added with its reason and the original stays.

We don't publish an overall hit rate yet. A few weeks of picks is too few to judge a method by, and the rules changed while the record was young. Each day's picks and results are public from the start of our record; see yesterday's results. We check our own picks against what happened before trusting any band, and need at least 30 days of them first.

What the method can't do

Bookmakers usually know more. Their prices already reflect the likely result, so a likely pick is not a cheap one.
Our model rates strong favourites lower than the bookmakers do. Until that is fixed, the 10-point rule blocks most home and away win picks, so most picks are goals markets.
The effect of missing players can't yet be checked against past matches: there is no history of injuries to test it on.
Teams early in a season, new to a league, or with little data lean on last season or on league averages. The match page shows how much data there is.
Weather, referees, travel, tired squads, manager changes and motivation are not in the model. Some appear on the match page as context.

Chances are estimates, not guarantees.