Wind unders on forecast wind — findings (2026-10-17)

Rules: PREREG.md (written before any forecast data was pulled). Data: GFS MOS (MAV) from the IEM archive, one station per stadium (mos/chosen.txt), 2006–2025 outdoor/open-roof US games (n=3,925, forecast found for all). Odds: nflverse closing total and under_odds (2% missing → −110).

Primary rule — PASSES

Bet the under when the latest MOS run ≥6 h before kickoff forecasts ≥15 mph at kickoff.

(The identical half records are a coincidence; season-by-season records differ.) At flat −110: +7.8%. Under odds at close median −106.

Secondary

• Forecast vs observed game wind: r = 0.705.
• Thresholds: 12 mph +10.2% (n=990, both halves CI > 0); 18 mph +9.1% (n=247); 20 mph +8.0% (n=128).
• Day-before run (≥24 h): +6.3% (n=413), CI spans 0. Forecast skill drops with lead time.
• Observed wind ≥15 on the same games: +9.2% (n=461). The forecast does as well as observed wind.
• Forecast ≥15 but observed <15: 94-65, +12.7% at −110 (n=161) — the effect is not biasbiasWhether the misses lean consistently one way. A projection can have a good average error size but still be biased if it is almost always too high. Bias is often the more fixable problem." data-def="Accidentally using information that was not available at the time. It makes predictions look brilliant and is the single most common way a backtest fools you.">look-aheadlook-ahead biasAccidentally using information that was not available at the time. It makes predictions look brilliant and is the single most common way a backtest fools you..
• All outdoor games: −0.9% (no general under bias).
• By forecast band (ROI / closing total): 0–5 −2.5% / 44.3; 5–10 −7.5% / 44.0; 10–15 +5.8% / 43.8; 15–20 +10.4% / 42.8; 20+ +8.0% / 41.7. The market lowers totals for wind but not enough, and calm games lean over: the line looks too flat across wind.
• No single stadium drives it (leave-one-out worst: +5.5% without PHI). By month: Sep −0.7%, Oct +4.3%, Nov +13.5%, Dec +11.3%, Jan +1.0%.

Caveats

1. Not an independent discovery. The idea came from the 2026-10 25-heuristic screen, which used observed wind on these same games (1 of 25 survived). The forecast (r=0.7) removes the look-ahead objection but not the selection: this is a replication on overlapping data, not a held-out test. The 2026-08 reanalysis study found 50.0% unders on a held-out n=44.
2. The pooled CI lower bound is +0.3%; neither half is significant alone.
3. Volume is ~20 bets a season, so confirming 56% vs a 52.4% break-even takes years.
4. Closing totals and closing under odds: a bettor has to be able to get the close.

The clean test is forward: 2026 onward, flagged before kickoff from live MOS.