Comment by cryptoz | Hacker News Reader

Comment by cryptoz | Hacker News Reader

While this article mentions that the GFS model was more accurate than the others (ECMWF & NAM), it neglects to give the backstory as to why it was more accurate and less trusted. For years the GFS has been made fun of because of its poor performance when compared to the Euro - and it just got the major upgrades it needed this month! New supercomputers, new high-res models, all went online just before the storm arrived. So, very few trusted the model even though it was in the best spot to get everything right.

In an unexpected move, The Weather Channel was much more accurate than the US NWS / AccuWeather / etc because of their use of the GFS as a trusted model. More info in this article: http://www.alternet.org/environment/why-almost-everyone-got-...

Edit: And of course, it's my opinion that all the models are currently suffering from a lack of input data. The "...and what should be done" part of the headline, in my opinion, should be answered with, "everyone download PressureNet and contribute your phone's sensor data to weather models", so that we can better predict difficult quantities like heavy snowfall and severe thunderstorms: https://play.google.com/store/apps/details?id=ca.cumulonimbu... ;)

Do you know of any research about the accuracy of forecasts as a function of sensor data input? I.e. what kind of an improvement will we see if some number of Android users install PressureNet?

Here is a recently published paper by Cliff Mass that is the first attempt at forecast experiments with smartphone pressure data: http://journals.ametsoc.org/doi/full/10.1175/BAMS-D-13-00188...

The early results are good. There will be an improvement [1]. How much of an improvement? We don't know yet. Cliff thinks it could be a revolution for some types of forecasts, but we don't have the density of sensors yet to know for sure.

For some comparisons, we sent cliff about 20,000 measurements per hour for his experiment in that paper. We're now delivering about 200,000 per hour to researchers, and that's not nearly enough. Our aim is for 2M per hour and I hope to reach that in the next 2-3 months. Around 1-2M per hour is probably sufficient to provide the "revolution" in accuracy that Cliff predicting.

I should also note that 1-2M per hour is small. We should be able to get closer to 1B per hour, but it'll take a while to ramp up to that kind of scale (that'll be like, every smartphone + watch + car that has a barometer).

The improvements will be slow and steady until we get massive scale and are able to run our models in real time. Until then, it's tough to guess how good the improvements will be.

[1] In the linked paper above, I believe the results were a reduction in root mean square error of about 1deg C for a 3-hour temperature forecast in the Pacific Northwest.

Do you know of any mobile apps that use GFS as a source?

Using the chrome app you can go to

http://weather.rap.ucar.edu/model/?model=gfs

I never tried it on my phone until a couple seconds ago. It looks pretty good!

You can create a shortcut to that URL on android, presumably legacy phones can do the same.

Reply on news.ycombinator.com