Our method, step by step
Every forecast on AtmoAtlas is produced by our own engine. It takes several weather models, keeps the strongest one for each place and lead time, corrects what the measurements allow us to correct, then checks itself against real weather stations every night. Below: what it does, and the results of that check.
Choosing a model for each place and each hour
For temperature, humidity, pressure and wind, the global base is ECMWF-AIFS, the machine-learning model of the European Centre for Medium-Range Weather Forecasts. Its rainfall is left aside on purpose: in our own tests it smoothed showers into drizzle that never falls.
Over Europe, ICON-EU from the German weather service (7 km cells) takes over for the first five days, for all fields. NOAA GFS (about 25 km) fills the three-hour steps between the other models and provides days 11 to 14. In the United States, our API also uses HRRR (3 km) for the first two days.
Once merged, the result goes through consistency rules: gusts are never weaker than the mean wind, the daily summary is built from the hourly values, and past hours are never shown as forecast.
What observations correct
Every hour we collect METAR weather reports from about 1125 airports. Near a station (up to 500 km), the engine measures the systematic error of its recent pressure forecasts and removes it, within limits: never more than 5 hPa, and nothing below 0.3 hPa so that random noise is left alone.
The same correction applied to temperature and wind was tested by leaving each of 1,066 stations out in turn. It degraded the forecasts at every distance, so it is not used: a correction is switched on only if the measurements show it helps.
Rain in the next hour
The rain indicator relies on our radar picture, rebuilt every ten minutes from open official radar networks: EUMETNET OPERA for 24 European countries, the German weather service (DWD) for Germany and its surroundings, and NOAA MRMS for the United States. Around Germany we take the DWD radar forecast as is; elsewhere under radar coverage we measure how the rain moved between the two latest images and extend that motion over the next hour. Where no open radar exists, the answer comes from our hour-by-hour rain forecast (ICON-D2, ICON-EU, GFS).
Reading the confidence labels
Each of the 14 days is tagged. High confidence until day 7, where models agree and verify well; medium until day 10; low afterwards, where days 11 to 14 are an outlook and are shown as one.
Seasonal normals and felt temperature
The “usual for the season” value next to each day comes from our processing of the ERA5 reanalysis over 1991–2020, the official reference period: 180 samples per month for every point of the globe, with about 0.2 to 0.3 °C of residual uncertainty.
Felt temperature has no world standard. We compute the recognised indices (humidex, heat index, apparent temperature, wind chill) in one place and say which one is shown.
How accurate we were, recomputed every night
Each night the forecasts are compared with what the 1125 stations observed over the last 20 days (59,976 forecast/observation pairs for temperature). The table gives the mean absolute error of our combined forecast and, for comparison, of GFS alone, by lead time. Last update: 2026-10-29 03:20 UTC.
| Lead time | Temperature (°C) | Wind (km/h) | Pressure (hPa) |
|---|---|---|---|
| 0 h | 1.29 · GFS only 1.79 | 5.26 · GFS only 5.52 | 1.35 · GFS only 1.41 |
| 24 h | 1.41 · GFS only 1.88 | 5.43 · GFS only 5.68 | 1.49 · GFS only 1.58 |
| 48 h | 1.47 · GFS only 1.95 | 5.60 · GFS only 5.87 | 1.66 · GFS only 1.81 |
| 72 h | 1.58 · GFS only 2.04 | 5.80 · GFS only 6.19 | 1.99 · GFS only 2.20 |
| 120 h | 1.78 · GFS only 2.18 | 6.16 · GFS only 6.64 | 2.50 · GFS only 2.82 |
AtmoAtlas · Mean absolute error = average distance between forecast and measurement, in the unit of each column; smaller is better. 0 h is the analysis, 120 h is five days ahead.
Where the data comes from
ECMWF open data (AIFS, AIFS-ENS), NOAA NCEP (GFS, HRRR, GFS-Wave), Deutscher Wetterdienst (ICON-EU, ICON-D2, radar), EUMETNET OPERA and NOAA MRMS (radar), aviationweather.gov (METAR), Copernicus CAMS (air quality, pollen), Copernicus ERA5 (normals, past weather) and NASA JPL MUR (sea temperature). Every source allows commercial reuse with attribution; we do not take forecasts from other weather websites.
What we cover
Forecasts for 220,018 places worldwide in English and French; the same engine serves the website and the AtmoAtlas API.