It combines ECMWF's open forecasts, corrected for each airport, with WT-1, a neural network trained on 26 years of hourly reports from 6,047 airports. An independent project, free to use, with no ads and no accounts.
The claim
Over a year held out from training, GateWeather's next-day high was off by 0.92°C on average, 8% less than the best other forecast tested (ECMWF AIFS, corrected, 1.00°C) and 37% less than ECMWF's raw IFS forecast (1.47°C).
319 airports, 358 days (1 November 2025 to 24 October 2026), every forecast issued at 18:00 local the evening before from what was public then. Scored against the official US highs instead, at 90 airports, it was off by 1.05°C against 1.13°C for corrected ECMWF AIFS.
Please keep these caveats with it
- Self-reported. Every forecast and measurement is public on the test data page, so anyone can rescore it; the model itself is not open.
- A test run after the fact, not the live record. The combination's weights were fitted on alternating weeks of that year; fitting them on earlier months only gives about the same accuracy. The live forecast has run since 27 October 2026.
- Scored on hourly airport reports, not the official maximum and minimum (what the high means).
- One day ahead, at airports only. Not for flight planning or safety decisions.
Figures
- Compare: every system's error, with 95% ranges, against ECMWF, NOAA GFS, US guidance (MOS and the National Blend of Models) and DWD's MOSMIX.
- The test data: every forecast and measurement of the held-out year, CC BY 4.0.
- How it works and Patterns, for background.
Images
Free to use with credit "GateWeather":
- Chart: next-day high, average error (1200 × 630)
- Share image (1200 × 630)
- Logo with name and mark (SVG)
Credits
Based on data and products of the European Centre for Medium-Range Weather Forecasts (ECMWF), CC BY 4.0, and data from NOAA and Deutscher Wetterdienst. Not an official forecast of any weather service. All credits.
Contact
hello@gateweather.com · x.com/analyzestripes. Per-airport breakdowns on request.