Five analysts searched seventeen seasons of NFL closing lines for anything the market leaves on the table. The best three ideas were frozen, then tested once on three seasons nobody had looked at. All three made money there, +11.3u on 255 bets between them, but none by enough to rule out luck.
Each made money in the seasons it was found in and in the sealed ones. Each also shrank by half or more once it left the data it was chosen on, and every sealed interval still includes losing. That is what “survived” means here: worth watching, not worth betting yet.
If a team missed the spread by 28+ points last week, bet on it this week. If it beat the spread by 28+, bet against it. Regular season, closing spread.
The line over-corrects for the most recent extreme result. The same rule using the game two weeks back did nothing (+0.7%), which is what a real over-reaction effect should look like.
Why it isn’t proven: Fires about 16 times a season. Lockbox CI is wide; the red team expected a true edge of +5 to +7%, not the +16% found in discovery.
Tease favourites of −7.5 to −8.5 down to −1.5/−2.5, and underdogs of +1.5 to +2.5 up to +7.5/+8.5, through both 3 and 7, only when the game total is 49 or lower. 2-team tickets.
NFL margins pile up on 3 and 7, so six points that cross both buy more than six points elsewhere. Low totals mean fewer points and stickier key numbers.
Why it isn’t proven: The whole edge is the ticket price: it pays at −120 and loses at the −135 most books now charge. It also fires only about 33 tickets a season, and 2024 lost.
Bet the under at the closing total when kickoff wind is 12 mph or more in an open-air stadium.
The closing total drops about a point for strong wind, and each extra mph has cost about 0.2 points of scoring in every era since 2006, so the close under-adjusts.
Why it isn’t proven: Uses wind measured at kickoff, not the forecast you could bet off. All of its lockbox profit came in 2023; it lost in 2024 and 2025.
Every game we published carried both sides, the better number at each book. A pair only profits if the score lands between the two numbers, which needs about 8% of pairs. 1 of 60 did. These are execution notes, not bets, and graded as bets they cost the vig.
The shadow model behind no public number. The market’s own price forecast games far better than it did, which is why it never left the admin page.
The NFL closing line is very hard to beat with public information. Rest, travel, starting-QB and coaching changes and last week’s form are already in the price. What survived are structural quirks: how margins pile up on 3 and 7, how far a line over-corrects after one freak result, and how much wind is worth.
None is strong enough to call a winner, and this dataset has now had its last look. The only thing that can promote one is new seasons, graded at prices we actually recorded. Until then the cheapest edge anyone can take is still the one on the bet sheet: the better number at the other book.
Closing lines and prices: nflverse. Flat 1-unit bets at the consensus closing price. Intervals are 95%, bootstrapped by week. Formula files, hashes and the referee run are committed with this page (reports/nfl-thesis-2026).