This week it’s raining in Switzerland. It’s actually raining buckets. So instead of enjoying the summer, we are discussing the newly released error indications, a.k.a. „probabilistic weather forecasts“, in the official weather apps of the Swiss government agency MeteoSwiss. Which, by the way, now also pushes AlertSwiss emergency notifications, so you better get yourself set up in these uncertain times. Oh, and need I mention the data is partly open? Let’s #hackweatherdata
Bild: Ein Blick hinter die «Kulissen» der Lokalprognosen von MeteoSchweiz, 11. Mai 2021
MeteoSwiss introduce numerical forecast models in their website, a standard feature of weather forecasts. Their information is for the moment only available in Ger/Fre/Ita, so I’m going to try to translate some of this here for my own curiosity and general interest. In short, despite a lot of progress in extrapolating the climate ‚in silico‘, some processes are still not adequately represented by data models - a weather process is often not represented using „physical laws“, but in a simplified way using „surrogate“ models and data.
There is a very readable introduction to this in English by Chuck Doswell and Harold Brooks at the US National Severe Storms Laboratory. Of course, even the best models are only as good as your measurements - which, at MeteoSwiss, are really precisely calibrated but still not infallible. Knowing how much your interpretations deviate from the data is very important, something we like to forget…
So how does this actually work, like, if you wanted to do the math? The MeteoSwiss documentation helpfully explains:
From the ensemble scenarios, a median scenario can be calculated - for a given place and time - for each meteorological parameter, which by definition is the one for which half of the ensemble forecast scenarios give a higher value and the other half a lower value. To quantify the most extreme values predicted by these scenarios, the highest 10 per cent of the maximum values and the lowest 10 per cent of the minimum values are used in each case. Thus, in mathematical statistician jargon, the value below which the lowest 10 per cent lie is called the 10 per cent quantile (Q10) and the value above which the highest 10 per cent lie is called the 90 per cent quantile (Q90). (The median is thus nothing other than the 50-% quantile). (meteoswiss.admin.ch)
While this approach has long ago become part of the work of forecasters and integrated into their services (browse the Climate bulletins and key terms page for more on communicating probabilistically), most people still love to complain when the forecast does not match the situation. So the information about how much certainty there is in the forecast needs to flow into the graphical representation, especially of local forecasts. Hence, error bars!
In our social media discussion, we explored the MeteoSwiss app’s new visualizations, how they differ from the desktop version, and what we think about the usability and impact of all this on everyday users. I look forward to see more comments on the practice and impact of visualizing uncertainty.
Many thanks to Adina Renner, Ralph Straumann and Joël Fisler for precipitating (ha ha) today’s post. Stay safe, don’t read too much into the graphs, and let’s get inspired by this incredibly agile frog to go #hackweatherdata