Research
When the projection and the weather disagree
Our other studies ask what weather does to scoring. This one asks a different question: does the weather read add anything once a scoring projection already exists? We built a simple, strictly forward-looking projection for 21,001 NFL, college football and MLB games, split each game by whether the projection and the weather read pointed the same way, and graded one bet per game against the real closing total.
The short answer
The result that repeated, in both football codes, is negative: when a projection pointed one way and a clear weather read pointed the other, taking the projection side lost badly and consistently — NFL 38.1% over 63 games, College Football 44.2% over 209, with the 95% confidence interval sitting entirely below the 52.4% break-even at -110. Weather disagreement behaves like a veto, not a pick.
The agreeing buckets trended positive in every sport — most interestingly in MLB, where they went 180-141 (56.1% over 333 graded games) — but no bucket in any sport clears our evidence bar, and we label the small ones as unusable rather than hiding them. This is not a betting system, and we will not pretend otherwise.
How we tested it
- A simple non-weather projection was built strictly walk-forward: each game projected only from games played before it (trailing team form plus league average, shrunk together). No hindsight, no season-end ratings.
- The weather read comes from the published football and MLB weather coefficients, applied as a scoring shift with a neutral dead zone — light weather is treated as no read at all.
- One bet per game, always the projection's side, graded against the real closing total at -110 (break-even 52.4%). Pushes are excluded from win% and ROI.
- Every bucket carries a Wilson 95% confidence interval. Only a lower bound above 52.4% counts as evidence of an edge; only an upper bound below it counts as a reliable loser.
NFL
1,140 graded games · model edge 3.0+ points, weather read 2.0+ points either way
| Bucket | Games | W-L-P | Win% | ROI | 95% CI |
|---|---|---|---|---|---|
| Confirmed — projection and weather agree | 12 | 8-4-0 | 66.7% | +27.3% | 39.1% to 86.2% |
| Conflict — projection against the weather | 63 | 24-39-0 | 38.1% | -27.3% | 27.1% to 50.4% |
| Model-only — no clear weather read | 335 | 169-160-6 | 51.4% | -1.9% | 46.0% to 56.7% |
College Football
6,652 graded games · model edge 4.0+ points, weather read 2.0+ points either way
| Bucket | Games | W-L-P | Win% | ROI | 95% CI |
|---|---|---|---|---|---|
| Confirmed — projection and weather agree | 47 | 25-21-1 | 54.3% | +3.8% | 40.2% to 67.8% |
| Conflict — projection against the weather | 209 | 91-115-3 | 44.2% | -15.7% | 37.6% to 51.0% |
| Model-only — no clear weather read | 2,928 | 1,525-1,372-31 | 52.6% | +0.5% | 50.8% to 54.5% |
MLB
13,209 graded games · model edge 0.4+ runs, weather read 0.8+ runs either way
| Bucket | Games | W-L-P | Win% | ROI | 95% CI |
|---|---|---|---|---|---|
| Confirmed — projection and weather agree | 333 | 180-141-12 | 56.1% | +7.0% | 50.6% to 61.4% |
| Conflict — projection against the weather | 832 | 406-398-28 | 50.5% | -3.6% | 47.0% to 53.9% |
| Model-only — no clear weather read | 7,854 | 3,865-3,781-208 | 50.5% | -3.5% | 49.4% to 51.7% |
Break-even at -110 is 52.4%. Bold conclusions need the confidence bounds, not the raw percentages — see the caveats below before quoting any row.
What it means when you check a game
We cannot tell you your handicap is right — nobody honestly can. What the data says a weather check is for is the other direction: it cannot pick your bet, but it can flag when the weather argues with it. In both football codes, the games where bettors took a projection against a clear weather read are the most consistently losing behaviour we have measured anywhere on this site. That is a seatbelt, not an engine.
It is also not a law of nature. In MLB the conflict bucket was flat (50.5%), so football and baseball clearly behave differently here. Sport-specific market behaviour matters, which is why we test each sport separately rather than assuming transfer.
Honest caveats
- This is retrospective. "It held in the past" is not "it will hold next week"; a locked forward test is running before any of this informs anything user-facing.
- No bucket in any sport clears the evidence bar: no confidence lower bound sits above 52.4%. There is no validated edge here.
- The agreeing buckets are underpowered: NFL n=12 and College Football n=47 are far below our 40-game minimum once split further. MLB n=333 at 56.1% is the most interesting sample, but its confidence interval still includes break-even.
- The baseline projection is intentionally simple and transparent (recent form plus league average, shrunk), so these results speak to the weather check, not to model skill — the agreeing buckets may be carrying the weather signal rather than model skill.
Where this meets the site
Every game page pairs the conditions with context you can weigh yourself — start with the NFL weather board, then read why weather matters for how we think a forecast should be used, and the forecast accuracy page for how our own forecasts grade against what actually happened.
OverUnderWeather Research Desk
Independent data analysis on public sports and weather records. Studies match historical game outcomes to archived game-window weather, apply pre-set sample and significance rules, and publish null results when that is what the data show. Editorial policy and corrections log.
Cite this research
Free to republish with a visible credit and a link to this page. Ready-made citations below — click to copy.
Frequently asked
Does the weather read add value beyond a scoring projection?
The tested evidence is one-directional. In both football codes, bets that took the projection against a clear weather read lost reliably (NFL 38.1% over 63 games, College Football 44.2% over 209), with the 95% confidence interval entirely below break-even. That makes weather disagreement a usable veto. The opposite direction — agreement meaning a better bet — trended positive everywhere but no bucket clears our evidence bar yet, so we do not sell it as a green light.
Is this a betting system?
No. Nothing here is a validated edge: no bucket in any sport has a confidence lower bound above the 52.4% break-even at -110. What repeated was the failure case — ignoring a clear weather read is where bettors got hurt. Treat the weather check as a seatbelt, not an engine.
Why publish numbers that do not clear your evidence bar?
Because the one result that did repeat cleanly is the negative one, and it supports the core idea behind checking the forecast before betting. We publish the caveats alongside it, label every underpowered bucket as such, and a locked forward test is running before any of this informs anything user-facing.
Which projection did you use — your best model?
No. The baseline is deliberately simple and transparent (recent form plus league average, shrunk together, built only from games already played). That makes the results a statement about the weather check, not about model skill — and it means the agreement buckets may be carrying the weather signal rather than model skill.