Comment by kqr | Hacker News Reader

Comment by kqr | Hacker News Reader

One thing I'm missing from virtually every weather service is access to ensemble forecasts. Meteorologists internally deal in ensemble forecasts, but never, as far as I know, publish them. All public data has been condensed down to a single forecast.

The ensemble forecast would -- I imagine -- give me an idea of the internal variability of the forecast for a single hour/day, which I could feed into my personal loss function and determine more accurately how to prepare to minimise discomfort.

As a concrete example, a single forecast of 17 °C can correspond to many ensemble forecasts, including

- 90 % range 11--20 °C

- 90 % range 18--22 °C

and these two mean very different things for how I'd dress and what extra clothes I'd bring.

It's even more important e.g. when planning a long drive. If the 90 % range includes heavy downpour, it might be worth rescheduling or adding buffer time to the trip, even if the most likely single forecast only predicts drizzle.

What would it take to get access to this data?

The ECMWF provide meteograms which show the ensemble spread for their global forecasts: https://www.ecmwf.int/en/forecasts/charts/web/classical_mete.... That webpage has a lot is great visualisations

Ooh, thats pretty much exactly what I've been looking for. Cheers!

I think you want to check out the Norwegian Meteorological Institute, they have a open free API with heaps of data, i think they have what you want. It powers yr.no which is used worldwide. https://api.met.no/

They have an endpoint to give you weather per hour now and going forward for any location in the world

Including ensemble forecasts into Open-Meteo is already on my backlog. And yes, this will return those 90% confidence intervals.

Forecast variability is difficult to express. For example precipitation probability describes how likely rain is given the forecasted atmospheric conditions. Instead some people assume it is raining in 10% of an area.

Making use of probabilities it also tricky. If you run an ice cream shop, with high operation cost and there is a 50% probability of rain, do you open with the chance of barely anyone coming or keep closed to save cost?

Ensemble forecasts also take up a lot of disk space, because each model ensemble is stored individually.

You could store the standard deviation as well as the mean (or median) of the ensemble. That would maximally double your data volume (but would add quite a bit less with compression). The UK Met Office have done same nice work into how to optimally combine the severity and risk of an event into an impact measure

Based just on the probability of rain, yes, that would be hard. But if I ran an ice cream shop, I would keep a simple record of the relationship between rain amounts and customers, pad it with a fat safety margin and the rest is just plugging in the numbers from the forecast.

didn't dark sky used to show these uncertainty ranges?

Some providers do it only for long-term forecasts, but not daily or hourly. Dark Sky might have been one of those.

Reply on news.ycombinator.com