How it works: from data to alert

1. We download the models to our own serverEvery 20 minutes we check for a new run from the ECMWF (the European centre: its classic IFS model, its artificial intelligence model AIFS and the 50 ENS scenarios), the American GFS and the German ICON-EU. We add the French ARPEGE and the British UKMO. Each run is downloaded once and served to everyone already processed, which is why the dashboard is fast.
2. We combine themWe calculate the consensus of the six models: the median, the range between the most optimistic and the most pessimistic, and how many exceed a threshold. If they agree, confidence is high; if they disagree, we tell you instead of hiding it.
3. We correct their quirksEvery day we compare what each model predicted with what the airport stations measured. That teaches us each model's bias in each place and how much weight it deserves. Near those stations, this correction has cut tomorrow's temperature error by 13.5 % compared with a plain average of the models.
4. We make our own forecast: the Atlas ForecastIn Catalonia we go further: we combine the six models with what the nearly 190 XEMA stations (Meteocat) measure. Using months of data, the system has learned how wrong each model tends to be by time of day, season and place, and it corrects itself with the error the nearby stations have seen in recent days. Tested on stations it had never seen, it gets the temperature for the next 24 hours 15 % closer than the median of the models (and 12 % closer than the best of them), and its rain probability has half the error of climatology. In the simple view it tells you which models weigh most and why. With the nearly 860 AEMET stations, it will switch on province by province across the rest of Spain as it proves it beats the models.
5. We cross-check with what is happeningRain radar and its 90-minute extrapolation, the European MTG satellite, lightning, stations and the official AEMET warnings for Spain's 233 warning zones.
6. We detect what mattersCut-off lows, depressions, thunderstorms, heavy rain, gusts, heat and frost are detected automatically, and your alerts are checked against every new run.
7. We tell you, and we measure whether we were rightIn whichever view you choose, always with the source, the time and the uncertainty. And we keep score: the dashboard publishes each model's error against 32 airports over the last 30 days.

What it is not