How well the model reads each club

The same log that scores the model on the accuracy page, grouped by club. Every prediction here was written down before kick-off and scored once the match finished; nothing is added afterwards. A club's domestic season and its European ties are ranked separately and need 5 settled predictions each, which makes 617 rows across 572 clubs.

Some clubs are here twice, and that is the point. A domestic league and a European tie are not the same test of a model. The league is the same twenty or so sides over and over, on form we have watched all season; a European draw throws two leagues together on very little directly comparable evidence, which is a weakness we set out on the methodology page. Averaging the two into one figure hid it, and hid something plainer as well: CSKA Sofia's 15 predictions are 9 in the First League and 6 in the UEFA Europa League, not 15 rounds of one competition. So the two are counted apart, 45 clubs hold a row for each, and the line under the club name always says which one you are reading.

31 rows cover both at once. Those are the clubs — Europe's early exits, mostly — that have not reached 5 settled predictions in either half on its own, and dropping them to keep the rule tidy would have taken 572 clubs down to 541. They say and Europe in their label instead.

The order is the plain hit rate — how often the model called the right result, home win, draw or away win, in that club's matches. Nothing is smoothed or weighted on the way: whoever we have got right most often is top, whoever we have got wrong most often is last.

Which is exactly why the sample column matters. At 5 predictions the margin of error on a hit rate is around 22 percentage points, so the very top and the very bottom of this table are partly luck and will move. Read the number of predictions next to every figure. The middle of the table, where most sit near 50.0%, is the honest picture of this model.

Every club we have scored

All 617 of them, best-called first. The last column is the part that is about the model rather than about the football: how much ground it gained on a forecast that knew nothing and split every match three ways. A row can sit high on the hit rate and low on that one — a side or a league whose matches are easy to call flatters us — and where the two disagree, the last column is the one describing us.

The same question, by competition

A club is a small sample by nature — a league gathers ten times as many predictions in the same weeks, which is why that table uses a bar of 25. How well the model reads each competition is the same record at the other level.

A statistical estimate, not betting advice. Nothing here predicts a future match; it records how the model has done so far.