How we measure forecast accuracy

Thursday, 15 October · Zefita

Everyone promises “the most accurate forecast”. We measure it and publish the result, city by city.

How. At 06:10 UTC we save each model’s forecast (ICON, ECMWF, GFS, UKMO, MET Norway) for the next 5 days. The next day we read the real observations from the nearby airport weather station (METAR messages, public domain) and compute: the mean error of the daily high and low, the share of days within 2 °C, whether rain was predicted correctly (yes/no) and the wind speed error. The method follows the World Meteorological Organization guide (WMO/TD 1023).

The baseline. We also compare each model with “persistence” — the naive forecast “tomorrow like today”. A model that does not beat persistence is not worth the money.

When scores appear. After 30 days of data for temperature and 90 for rain. Until then the page honestly says “not enough data”. Scores are recomputed daily and never adjusted by hand.