How we are actually doing

Where the ground is lost

Is a stated 30% really 30%?

Over/under 2.5 goals

Fewer matches than the table above, and a reader is owed the reason. Every fixture is graded from the first entry that published it — one fixture, one file, so anyone checking by hand knows exactly where to look — and entries sealed before this market was added carry no probability for it. The Premier League therefore joins the totals record in September rather than August. That is a choice about how auditable we want the rule to be, not a law of nature. The reference points differ too — guessing 1/3-1/3-1/3 on a result scores 1.0986, guessing a coin flip on a half-goal line scores 0.6931 — so averaging the two would produce a number that means nothing.

Two groups, counted apart

The headline above pools every division. That is the honest total and it stays, but a pooled log loss moves when the mix of divisions moves, whether or not anything about the model has — so the pooled line on its own would let a composition change look like a result. On 17 September 2026 twelve thinner divisions were added to the eleven this site had been pricing since 28 August. These rows keep them apart, each with its own interval.

This split was decided and published before a single match in the new group had been graded — before one had been sealed. The proof is in the ledger, which nobody can rewrite: the last entry sealed before the decision is 2026-09-17, and no entry on or before that date names any of the twelve. Group membership is frozen. A division added later joins neither group and gets its own line, so neither of these can ever be reshaped into the one that reads better.

Which division sits in which group is listed on the method page and set in config.py, in the repository, where the change is dated like everything else here. A group with no graded matches says so rather than borrowing the other group's number.

Division by division

The same question asked 23 times. A model that is genuinely close to the market should be close in most of them; an edge that lives in one division and nowhere else is usually a lucky autumn, not an edge. A division's log loss appears here only once it has 20 graded matches. Below that the gap's own margin of error is several times the gap, so the number would say nothing except which way a coin landed — and the division where it landed our way is the one that would get quoted.

Per-division rows are small samples for a long time, and even the ones shown carry an interval wider than the gap they report. The headline at the top of this page is the one with the most matches behind it, and it needs a season too. "Pending" counts sealed predictions still waiting on a result or a closing price.

Week by week

Coverage

What is missing from the table above

Every other table on this page starts from the ledger and asks how the predictions did. This one starts from the matches that were actually played and asks whether the ledger has them at all — the only direction that can see a match we never sealed. A prediction that does not exist cannot appear in a table of predictions, so without this section a division whose fixture feed goes dark simply contributes fewer rows and the headline never flinches. Counted from the same results files the grading uses, against the same ledger, joined the same way.

599 of 649 matches played since each division entered the chain were sealed before kickoff — 92.3%. 38 were never sealed at all, and never can be: the ledger is append-only and a prediction published after kickoff is worse than no prediction.A further 12 were sealed but carry a club name that does not yet join to the results file — those are scored retroactively the moment the spelling is added, without anything being rewritten.

“In the chain since” is the first date the ledger can prove the division was being asked about. Before 6 September an entry recorded only the divisions that sealed, not the ones that were asked and answered nothing, so for the older divisions that date is a lower bound — these percentages can therefore only flatter us, never the reverse.

New markets

BTTS and Asian handicap

What this page will never do