Where is the weather actually predictable?
Everyone complains the forecast is wrong. Almost nobody checks. We scored real forecasts — what the model said one through seven days ahead — against what the weather station went on to measure.
14 stations · 2024-02-01 to 2026-07-23. Coverage grows as the verification backfill runs.
A 7-day forecast for MIAMI INTL AP misses by 2.4°F on average — better than a next-day forecast for MARYLAND SCI CTR, which misses by 3.4°F. Forecast quality is not one number. It depends enormously on where you are.
How fast forecasts decay
Average miss (°F) on the daily high, across all verified stations, by how many days ahead the forecast was made.
Ranked by long-range accuracy
| # | Station | 1-day | 7-day | Decay | Within 3°F |
|---|---|---|---|---|---|
| 1 | MIAMI INTL AP, FL | 2.0° | 2.4° | +0.3° | 73% |
| 2 | TAMPA INTL AP, FL | 2.6° | 3.2° | +0.6° | 58% |
| 3 | NEW ORLEANS AUDUBON, LA | 4.2° | 4.5° | +0.3° | 44% |
| 4 | BIRMINGHAM AP, AL | 3.0° | 4.8° | +1.8° | 41% |
| 5 | ATLANTA HARTSFIELD-JACKSON INT, GA | 3.0° | 4.9° | +1.9° | 40% |
| 6 | JACKSONVILLE NAS, FL | 2.6° | 5.2° | +2.5° | 36% |
| 7 | CHARLOTTE DOUGLAS AP, NC | 2.8° | 5.2° | +2.5° | 38% |
| 8 | PITTSBURGH ALLEGHENY CO AP, PA | 2.6° | 5.6° | +3.0° | 38% |
| 9 | PHILA INTL AP, PA | 2.7° | 5.7° | +3.0° | 37% |
| 10 | NASHVILLE INTL AP, TN | 3.3° | 5.7° | +2.4° | 37% |
| 11 | NY CITY CNTRL PARK, NY | 2.8° | 5.9° | +3.1° | 35% |
| 12 | RICHMOND INTL AP, VA | 2.9° | 5.9° | +3.1° | 35% |
| 13 | BOSTON, MA | 2.7° | 6.1° | +3.3° | 36% |
| 14 | MARYLAND SCI CTR, MD | 3.4° | 6.1° | +2.7° | 33% |
Lower is better. “Decay” is how much accuracy is lost between a next-day and a 7-day forecast — the truest measure of predictability, since a place can be hard to forecast at every range without being unpredictable.
Method. Forecasts come from an archive of what the model predicted 1–7 days before each target date. Truth comes from the NOAA station itself — not a reanalysis grid — which is the part that makes these numbers meaningful rather than a model grading its own homework.
Why only airport stations. We score exclusively at NWS first-order sites (ASOS), which record midnight-to-midnight local standard time. Cooperative-observer stations read their instruments at whatever hour suits the volunteer, often 7am or 5pm, so scoring a calendar-day forecast against them would measure the observation schedule rather than the forecast.
Want the climatological view instead — what a date is like when no forecast exists? Beyond the Forecast.