Tomorrow Weather Upington
Mixed agreement. The models are split. The shape of the day is clearer than the exact number.
Hot with clear sky, gusting to 36 km/h.
Mixed agreement. The models are split. The shape of the day is clearer than the exact number.
Hot with clear sky, gusting to 36 km/h.
Each coloured line is one independent forecast model. The white line is their average, weighted so that the models which have actually been accurate here count for more and with each model's excess bias already removed, which is why it does not sit exactly in the middle of the band. The grey band is the disagreement between them: a narrow band means the forecast is trustworthy, a wide one means the shape will hold but the numbers will not. Average spread across the 9 days: 4.2°C.
| Model | Source | Mean error | Error | Bias |
|---|---|---|---|---|
| ICON | German weather service | 1.12°C | -0.66°C | |
| GEM | Canadian model | 1.13°C | -0.57°C | |
| UKMO | UK Met Office | 1.56°C | -1.45°C | |
| GFS | US national weather service | 1.82°C | -1.37°C | |
| ECMWF | European centre | 1.95°C | -1.87°C | |
| AROME | Meteo-France | 2.25°C | -2.01°C |
Station distance: 5 km. Scored against 173 days of real hourly observations at Upington, comparing each model's forecast of the daily maximum with what the airport actually measured. The models are queried at the station's own coordinates for this comparison, not the city centre, so the score describes the same piece of ground the measurements came from. Mean error is the average size of the mistake; bias shows whether the model runs warm or cold. Best here is ICON at 1.12°C. The headline average is weighted by these errors, so a model that has been wrong by three degrees here carries a fraction of the vote of one that has been wrong by half a degree. Rainfall is a plain mean, because one dry model should not be allowed to cancel a wet one. Each model's number also has its own long-run bias taken off before averaging, but only the part beyond one degree: here that is AROME, which has run 2.0°C cold at this station, so its number is lifted by 1.01°C before anything is averaged. A bias under a degree is left alone, because over a six month record it cannot be told apart from ordinary luck, and correcting noise adds error instead of removing it. No model is far enough behind to be dropped at this station, so all six count towards the average.