Receipts

Every prediction here was made out of sample — the model was refitted repeatedly and only ever saw matches played before the one it was predicting. That is what makes them forecasts rather than hindsight.

Two are kept per match, win or lose: the result call (who wins) and the model's most confident selection of any kind. Neither is chosen after the fact, which is the point — a page showing only the winners would look like a record and work like an advertisement.

A hit rate on its own means nothing. Backing the home side wins a lot of football matches without predicting anything, so the baseline is shown beside every rate. The gap between them is the only part that is evidence of skill.

For the same reason the list defaults to against the grain first. Sorting by confidence puts a 100% call on a 97–39 mismatch at the top, which is correct and proves nothing — anyone would call it. A confident call on the side the baseline does not back is the one worth showing.

A market family that claims 30% and delivers 30% is working perfectly, and one that claims 90% and delivers 74% is broken — even though the second has the far higher hit rate. What matters is the gap between what the model claimed and what happened, so that is what this is sorted on.