Where the numbers come from

Nothing on this site is invented or estimated by us. The forecasts are the output of six national weather services, and the accuracy scores are computed against real measurements from airport weather stations. This page is both the credit and the licence notice, because the open licence under which the data is published requires us to say where it came from and what we did to it.

Forecast data

Forecast data is obtained through Open-Meteo, which publishes it under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). The underlying models, and the services that produce them, are:

• ICON, the German weather service, Deutscher Wetterdienst.
• ECMWF, the European Centre for Medium-Range Weather Forecasts.
• GFS, the National Weather Service of the United States, through NOAA and NCEP.
• AROME, Meteo-France, the French national meteorological service.
• UKMO, the Met Office of the United Kingdom.
• GEM, the Global Environmental Multiscale model of Environment and Climate Change Canada.

Measurements

The scores that decide how much each model's vote is worth are calculated from real hourly observations of what the nearest airport weather station actually measured. Those observations come from the METAR archive published by the Iowa Environmental Mesonet at Iowa State University, which archives reports originating from the United States National Weather Service. The station used for each location is named on that location's page, together with how far it is from the town.

What each model predicted on those same past days

To score a model we need both sides: what it said would happen, and what happened. What it said comes from Open-Meteo's historical forecast archive, which republishes the stored output of each model for past dates. So the scoring compares a model's own past output against the station measurement for the same day, and the difference is the score you see on the page.

What we changed, which the licence requires us to state

• We average the six models rather than publishing one of them.
• We weight each model's contribution by how accurate it has been at the nearest station over the past six months, so a model that has been badly wrong there counts for less.
• We remove long run bias from a model where it exceeds one degree at that station, and we say so on the page.
• We publish the spread, which is the gap between the highest and lowest model, rather than hiding it behind a single number.
• We round temperatures for display.

We do not alter the observations. Where our numbers differ from a provider's own published figure, it is because of the averaging and weighting described above.

If you are a data provider

If you produce any of this data and something here is wrong, credited incorrectly, or used in a way you object to, write to hello@seoninjas.co.za and we will correct it or remove it.