How these projections work
Two numbers per fixture: the goals a club is expected to score, and its chance of keeping a clean sheet. Where the bookmakers have priced the gameweek, the clean sheet figure is theirs. Beyond that it is ours.
The bookmakers first
Bookmakers price every match with money at stake, which makes their numbers about as close to a right answer as football gets — but only a week or two ahead. For those gameweeks we publish their clean sheet percentage directly, margin removed, rather than a model of the same thing. No model of ours improves on a direct quote.
Hover any clean sheet number to see which one you are looking at.
Then our own, for the rest of the season
Every club gets two ratings — how well it scores, and how well it stops other people scoring. A fixture puts one club's attack against the other's defence, adds home advantage worth about a third of a goal, and comes out with expected goals. That goes through a Poisson distribution, which is the standard way of turning "about 1.4 goals" into "so what are the odds of exactly none?"
Refitted against the bookmakers as the season goes
Poisson alone is a little too kind to defenders, so the number carries a correction. That correction is refitted every few days against the bookmakers' own clean sheet percentages for the gameweeks they have priced — twenty fresh comparisons a week, accumulating all season. Across the season so far it tracks them at a correlation of 0.95, and it is what makes the far-out gameweeks continue the market's view rather than stepping away from it at the point the pricing runs out.
Fitting against live prices replaced a correction measured on last season's finished results. That one was worth six percentage points, but it had been measured against each match's realised goals and was then applied to a projection — so it corrected twice, and published an average clean sheet chance of 21% when football produces about 25%. Checked against this season's actual results, the version fitted on live prices is the better of the two.
What we don't take from them
Expected goals stays ours throughout. The odds pages carry match result prices but no goals market, and working goals back out of result odds alone leans entirely on Poisson's weakest part — it under-counts draws — so the total comes out low by construction. A knowingly biased number is worse than a stale one. Ours run within about a tenth of a goal of what those prices imply, which is the check we keep on the ratings rather than a figure we overwrite them with.
None of it is a promise, mind. A 35% clean sheet still fails two times in three.