In July 2026 our cup section looked bad — rows of missed picks in UEFA qualifiers. Instead of shrugging, we audited every historical qualifier we hold: across 1,150 graded games, our over/under calls landed 51.4%. A coin flip.
The reason is structural. Qualifier pairings cross wildly different leagues, and the model's anchor — the average of the two clubs' domestic goal rates — is systematically wrong there: pairings from two high-scoring leagues (say, an Icelandic side against a Kazakh one) produced an anchor near 61% while the games actually landed around 54%. Qualifier football is cagier than either club's domestic league: two-legged stakes, unfamiliar opponents, long travel, mid-pre-season fitness.
We shipped two responses. First, a phase-aware recalibration fitted on 2,931 anchored UEFA games and validated on held-out seasons: qualifier anchors get squeezed, autumn league-phase anchors (which run 4+ points hot the other way) get boosted, and knockout anchors — which barely correlate with domestic rates at all — get flattened. Second, and more honestly: we no longer make picks on qualifier rounds at all. The percentages stay visible for information; our name stays off them.
Method: walk-forward testing — every prediction is made using only information available before kick-off, then graded against what happened. Studies run June-July 2026 on the site's own dataset. Questions: contact us.