Model vs market

Before kickoff, our model publishes its probabilities; at the same moment, the betting market sets its closing price. After full time, both are graded on what actually happened (log-loss — lower is better). The market is the judge, never an input — and we publish the verdict either way. How the ratings behind the model work.

The market leads across 90 graded matches. That gap is the mission — every model improvement ships only if it closes it, and this page will say so the day it does.

A 3-way close is a regulation-time instrument — a knockout Tie contract settles on the 90-minute scoreline — so every row is graded on the 90-minute outcome wherever it can be derived from the goal record, and on the full-time outcome (labelled 'full') where it cannot. The market and every engine on a row always grade against the SAME outcome: one basis per row, never mixed. 96 of 98 recorded quotes grade on the 90-minute outcome.

The headline pooled numbers weight one MATCH once: a fixture quoted on two venues is collapsed to a single per-match log-loss before pooling. The per-venue table keeps every row as recorded.

By venue

Small samples stay labeled as such — a venue with a handful of matches proves nothing yet, in either direction.

Every graded match

The probability each grader gave the result that actually happened — higher was better that day.

World Cup 2026 · 98 graded quotes of 126 recorded market closes (15 matched no played fixture, 13 matched one no engine covers); unmatched closes are counted, never guessed onto a match. Club competitions join as the season's fixtures grade.

The model record

Every engine logs its probabilities before kickoff — immutably — and is graded the moment the match resolves. This table grows on its own, ten minutes at a time. It grades on the full-time result; the headline section above grades every engine against the market on the shared 90-minute basis — two different questions, each labelled.

A prediction is logged only when both teams carry a rated history — a fixture we cannot honestly price logs nothing, and that refusal is part of the record. Once a logged match has a market close of its own, it joins the graded table above.