FROM SOURCES TO FORECAST
How we calculate forecasts
A weather forecast, a rain radar outlook and air quality each have their own sources. Here you can follow how we turn them into information for your location and what happens when data are missing.
The path to a forecast
- Weather models and radar
- Assembling model outputs and a separate rain outlook
- A result for a location and time
Weather models form the foundation
The public weather forecast uses model data downloaded directly from DWD (ICON-D2 and ICON-EU) and ECMWF (IFS). These organisations calculate the weather models; we assemble their outputs for the selected location.
The hourly forecast starts with ICON-D2 and can continue with ICON-EU. The longer daily outlook uses IFS. Each part keeps its source and model-run time.
When the latest run is unusable, we try suitable backup runs from the direct model data. Without usable data, the affected information or forecast is not shown; Open-Meteo does not fill the gap.
Models and calculation details
Separately, we evaluate combinations of models against station observations. Older experiments obtained some model inputs through Open-Meteo. Open-Meteo also remains a background comparison forecast; these experiments do not describe the public forecast shown above.
In these station-combination experiments, temperature and wind speed use a weighted mean after bias correction. Precipitation uses a weighted wet-or-dry vote and a weighted median of wet inputs, without additive bias correction. This method alone does not demonstrate higher accuracy for the public forecast.
Radar tracks the movement of rain
We estimate precipitation intensity from CHMI radar frames. We track how the precipitation field moved between frames and project that movement forward. This calculation does not itself predict the formation of a new shower.
The radar outlook runs separately from the weather-model combination. Comparing it with later radar frames answers a different question from comparing a forecast with rain measured at a station.
Radar method and horizon
We convert radar reflectivity into a rain-rate estimate, estimate motion with Lucas–Kanade optical flow and move the field using semi-Lagrangian extrapolation.
- Time step
- 5 minutes
- Horizon
- Up to 2 hours
Air quality, pollen and UV come from CAMS
For air quality, pollen and UV we use model data from the Copernicus Atmosphere Monitoring Service (CAMS), downloaded directly through the Atmosphere Data Store (ADS). These are model estimates, not sensor measurements at your address.
We show only values supported by available CAMS inputs. Open-Meteo does not fill gaps in these data.
We derive the air-quality index, pollen bands and UV index from these inputs. They remain separate from the weather forecast.
Celestial positions and weather are separate calculations
We calculate the positions of the Sun, Moon and planets for the selected location and time. Stars come from a catalogue. The sky view shows a calculation, not a live image.
We assess observing conditions separately using available cloud cover and visibility from direct weather-model data. A conditions category does not express a probability of seeing a particular object.
Astronomical sources
Astronomy Engine calculates celestial positions. Stars come from the BSC5P catalogue.
Calculation time is not validity time
Forecast time tells you when a value applies. The model-run timestamp identifies the origin of the combined forecast, while a radar frame has its own timestamp. Response generation time alone does not say that every input has just been produced.
The sky distinguishes the selected viewing time, response generation time and the timestamp of weather conditions.
How we handle missing data
If neither the direct forecast nor a suitable backup run is usable, we do not show a forecast. When no usable air-quality data are available, the weather forecast can remain available without air data.
Unavailable radar does not provide a valid radar outlook. We do not turn incomplete sky conditions into favourable ones. A missing value is not zero.
How we check forecasts
We pair forecasts with subsequent observations for the same station, time and variable. We use CHMI observations and aviation METAR reports. Separately, we compare the radar outlook with later radar frames.
What the error metrics mean
- MAE
- Mean size of a temperature or wind-speed error, regardless of sign.
- RMSE
- A temperature or wind error metric that gives larger deviations more weight.
- Bias
- Mean signed error showing whether forecasts overestimate or underestimate values.
- POD / FAR / CSI
- For precipitation we count hits, misses and false alarms. POD describes detection, FAR false alarms, and CSI accounts for both misses and false alarms.