Any weather app can look accurate if it sets its own exam. So we wrote ours down first — the rules we score ourselves by, and the places where we refuse to claim anything at all.
Every warning is checked against ground observations, not against our own earlier forecast. A forecast that agrees with itself is not accurate. It is only consistent.
Mountains are not interchangeable. What predicts a storm over an Alpine ridge can be wrong over the Caucasus, and badly wrong over a monsoon valley. We measure range by range, and where we have not measured yet, the app says so instead of pretending. An unmarked warning in an unmeasured region would be the most dishonest thing this app could do.
When we publish how often GRAT catches a storm, the false alarms will be printed beside it, in the same size. Those two numbers move against each other: anyone can push one up by letting the other go. Shown alone, either is a sales figure rather than a measurement.
“How often does it catch a storm” has no answer until you say what counts as a storm. Thunder heard at a weather station, a storm cloud overhead, a strike recorded within ten kilometres — the same app scores differently against each. We will always say which one a number was measured against, because quoting the flattering one is the oldest trick in this field.
We are rebuilding the yardstick. While measuring GRAT against observation records, we found that those records disagree with each other often enough to matter — hours logged as storms with no lightning anywhere near, and hours with lightning that were never logged at all.
Until that is sorted out, any figure we printed here would be measured with a bent ruler. So there is no figure here. When the measurement is sound, it goes on this page — including the parts that do not flatter us.
We do not publish the recipe: which models are blended, how the thresholds are chosen, how the regions are drawn. That work is the product. What we do publish is everything you can hold us to — what it caught, what it missed, and where we have not looked.
GRAT is built on open weather data published by national meteorological services and open-data projects, among them Open-Meteo, Deutscher Wetterdienst, NOAA, Metéo-France, MET Norway, the UK Met Office, the Japan Meteorological Agency, Environment and Climate Change Canada, MeteoAlarm and RainViewer. Parts of that data are published under the Creative Commons Attribution 4.0 licence; this note is that attribution. Naming who publishes the data is not the same as describing how GRAT uses it — see the section above.
Checking our own accuracy needs a second kind of data: a record of what the weather actually did. For that we use lightning observed from space — the Lightning Imager on EUMETSAT’s Meteosat Third Generation over Europe, Africa and the Middle East, and NOAA’s Geostationary Lightning Mapper over the Americas — together with cloud-top temperatures from those satellites and from the Japan Meteorological Agency’s Himawari-9. MTG Lightning Imager data: © EUMETSAT, used under CC BY 4.0. Lightning is a harder thing to argue with than a written weather report, which is why we went looking for it. It does not cover everywhere: there is no comparable lightning observation over Japan, New Zealand or the Himalaya, and we say so rather than quietly scoring those places as if there were.
A day you actually remember on the mountain, where the app was wrong, is worth more to this project than a download. Send it: support@gratweather.com.