Rewind · week 3, 2026

Atlanta 35, Green Bay 14: what we missed

We had Green Bay at 69%. Atlanta won by 21. The empty tank angle we flagged landed, but the bigger miss was ours: Michael Penix was back and the model priced a backup-quarterback offence.

25 Sept 2026 · 3 min read

35-14Atlanta win at Lambeau

We had Green Bay to win at 69%, by 7.3 points. The market had them by 4.5. Atlanta won 35-14. That is a miss by 28 points, and it is worth being clear about why.

GB 69%Our callby 7.3

GB -4.5Marketde-vigged 70%

ATL +21Result35-14

The angle we flagged

Before kickoff we flagged the empty tank: teams playing a short week straight after an overtime game are 8-29 against the spread since 1999. Green Bay went to overtime against the Jets on Sunday. It landed. We tested it properly afterwards: it improved the model in every held-out season, but not by enough to pass our bar, so for now it is a flag on game pages, not a model input.

The bigger miss was ours

Atlanta's first two games were played by backups, Cooper Rush and J. Strand, with Tua Tagovailoa and Michael Penix both out. That is why their offence ranked 32nd. On Thursday Penix was back. Our model saw a backup-quarterback offence, because it has a rule for a starting quarterback being ruled out and no rule for one coming back. That is on us, not bad luck.

What we are doing about it

A fix that rates an offence on the returning starter's own recent games, not the backups', is being tested on four seasons the model has never seen. It goes in only if it improves the predictions, and we will publish the result either way. Separately, a change to how fast the model trusts this season's form passed its test and went live today.

Sources: nflverse play-by-play and schedules 1999-2026; NFL injury reports; our published pre-kickoff run, 24 Sept 2026. Every number here comes straight from our pipeline.