Our model against the market: the first graded weeks

We log every forecast before kickoff, grade it against what happened, and publish the running score — including the parts where we're losing.

This site runs forecasting engines that estimate, before kickoff, how likely each side is to win. Prediction markets do the same thing with money. The obvious question — the one we built the public scoreboard to answer — is: who reads football better? We grade both the same way, with a statistic called log-loss: every forecast is scored against what actually happened, and a lower number means sharper forecasts. No cherry-picking; every logged match counts, and the ledger is immutable once a match kicks off.

The score so far

Honestly: the market is ahead where most of our matches have graded. On the venue with our deepest coverage — 90 graded matches at the time of writing — the market's closing prices score 0.829, our Dixon-Coles engine 0.844, our Elo engine 0.883. Closing prices aggregate everything every trader knows a minute before kickoff; beating them consistently is one of the hardest problems in sports forecasting, and we publish the gap rather than hide it.

On the second venue we record, the early picture inverts: across its first 8 graded matches, both of our engines scored ahead of that market's close — 0.934 and 0.957 against 0.989. Eight matches is far too small a sample to declare anything, and we say so on the scoreboard itself, next to the denominators. Small honest numbers now, so that the large numbers later mean something.

Why publish while losing

Because the alternative is worthless. A forecasting record that only appears once it looks good is marketing, not measurement. The scoreboard states its grading basis (results after ninety minutes, matched to the market's own question), discloses how many market closes matched no played fixture, and grades every engine on the same matches. When our model earns a lead, the same page that documents today's deficit will document it — with the same denominators attached.