How forecasts are made and scored

The forecast

Before each coupon round, a probability triple — 1 / X / 2 — is recorded for every match on the coupon, along with which signs are actually marked. That timestamp must precede kickoff; a forecast recorded after the fact is not a forecast, and the build fails if one ever is.

Ranked Probability Score

RPS is the headline metric, not raw accuracy. Accuracy rewards always backing the favourite and depends entirely on which fixtures happened to be on that week's coupon — it is close to meaningless on a small sample. RPS instead scores the full probability triple against what actually happened, and — because the outcomes 1 / X / 2 have a real order — it punishes a confident home call that finished as an away win harder than one that finished a draw. That is the right behaviour for football.

There is no external benchmark folded into this figure. RPS is tracked as a self-measure of forecast quality over time — this season against last month, not this site against the betting market.

Draw watch

Tracked on its own because it is a failure this project has already made in production: a coupon that excluded the draw on several soft favourites, in a round that went on to produce five draws. The expected draw count is the sum of predicted draw probabilities; a result within two standard deviations of that expectation is noise, not a bias worth correcting for.

What this site will not do

• Present a forecast as betting advice.
• Backfill a prediction after a match has kicked off.
• Hide a match that was never forecast, or fake a pending score.
• Store a derived figure — every number here is computed at build time from the underlying match data, so it cannot drift out of sync with what produced it.