Which weather model is most accurate in New York?
New York, New York · daily high temperature · trailing 30 days · refreshed at most every 6 hours
Every day, The Spread archives what each major weather model predicted for New York at 1-, 3-, and 7-day leads, then scores those forecasts against ERA5 reanalysis once the day has happened. This page covers the trailing 30 days (2026-07-06 to 2026-08-05), for daily high temperature.
One day out, GEM (Canada) has been most accurate (typical miss 1.6°F); at three days, UKMO (UK) (2.15°F); at seven days, ECMWF AIFS (AI) (3.26°F).
The largest standing bias belongs to GEM (Canada), which has run cool by about 5.15°F at 3-day lead over this window.
The Spread's own bias-corrected blend scored a 1.2°F typical miss at one day. Its corrections come from the same window it is scored on (in-sample), so read it as a ceiling, not a promise.
The scoreboard
Typical miss (MAE) with the signed bias in parentheses — positive bias means the model ran warm. Days = scored days behind the 1-day number.
| Model | 1-day | 3-day | 7-day | Days |
|---|---|---|---|---|
| GEM (Canada) | 1.6° (-0.1°) | 5.3° (-5.2°) | 4.6° (-4.3°) | 30 |
| DWD ICON | 1.7° (-0.6°) | 2.3° (-0.7°) | — | 30 |
| UKMO (UK) | 1.9° (+0.3°) | 2.1° (+0.4°) | — | 30 |
| ECMWF IFS | 1.9° (-0.3°) | 2.2° (+0.1°) | 4.8° (+1.4°) | 30 |
| NOAA GFS | 2.2° (+0.3°) | 2.6° (-1.1°) | 4.7° (+0.5°) | 30 |
| NOAA AIGFS (AI) | 2.3° (-2.0°) | 3.0° (-2.7°) | 4.3° (-1.9°) | 30 |
| ECMWF AIFS (AI) | 2.5° (-2.4°) | 3.1° (-2.7°) | 3.3° (-1.7°) | 30 |
| NBM (US blend) | 6.0° (-5.9°) | 6.3° (-6.0°) | 5.1° (-3.9°) | 30 |
| Blended (bias-corrected) | 1.2° (-0.0°) | 1.7° (-0.0°) | 2.8° (+0.0°) | 30 |
Method, honestly
Forecasts come from each model's actual archived runs via Open-Meteo's Previous Runs API — what the model said at the time, not a reconstruction. Truth is ERA5 reanalysis daily highs. Rankings cover the models themselves; the two consensus products at the bottom of the table — NBM, which is itself a blend of models, and The Spread's own bias-corrected blend — are shown for comparison, never ranked or named "most accurate" here. The blend's corrections come from the same trailing window it is scored on (in-sample), so treat its score as a ceiling. A rolling 30-day window means these numbers move — that is the point.
Receipts — was the forecast right?
One page per settled date: what every model said as the day approached, what happened, and who was closest.
- Mon Aug 3 7 days out the models said 77–86; it hit 83. ECMWF led (off 0.5°); GEM ran 5° cold.
- Sun Aug 2 6 days out the models said 76–82; it hit 82. AIGFS led (off 0.1°); ICON ran 2° cold.
- Sat Aug 1 5 days out the models said 79–87; it hit 85. AIFS led (off 0.0°); ICON ran 3° cold.
- Fri Jul 31 4 days out the models said 79–87; it hit 84. ICON led (off 0.2°).
- Thu Jul 30 3 days out the models said 76–80; it hit 77. AIFS led (off 0.1°); GEM ran 2° hot.
Other cities
Atlanta, Georgia Boston, Massachusetts Buffalo, New York Charlotte, North Carolina Chicago, Illinois Dallas, Texas Denver, Colorado Houston, Texas Los Angeles, California Miami, Florida Minneapolis, Minnesota New Orleans, Louisiana Oklahoma City, Oklahoma Orlando, Florida Phoenix, Arizona Salt Lake City, Utah San Francisco, California Seattle, Washington Washington, District of Columbia
The Spread — the forecast spread — is in closed beta. These verification pages are public while the rest is gated.