HN Explorer

Ask HN: Should I try building weather prediction models using machine learning?

4 points by mthoms ↗ HN
Is it realistic for a developer new to machine learning to attempt building weather prediction models with "off the shelf" open source tools? What about hardware? Is it possible to get decent results with modest hardware? (for example, a small cluster of AWS instances).

8 comments

[ 3.7 ms ] story [ 27.0 ms ] thread
harterrt
Depends on what you're trying to forecast and what you want to use the forecast for. If you're just curious, go for it. If you're planning on trading NG futures, well... go for it, but expect to be disappointed.

I'd bet the biggest hurdle you'll face is getting quality data.

mthoms
Good point - I should have clarified that I'm only interested in short term predictions at the moment (1-3 days).
mindcrime
I'd bet the biggest hurdle you'll face is getting quality data.

I guess it depends on what you're looking for, but at least some weather/climate/earth-surface stuff can be grabbed right off the air from the various satellites. Some of that stuff is beamed down unencrypted and anybody with the right receiver setup (which can be as simple as an RTL SDR dongle, AFAIK) can grab the data and use it.

http://www.rtl-sdr.com/rtl-sdr-tutorial-receiving-noaa-weath...

http://noaasis.noaa.gov/NOAASIS/pubs/Users_Guide-Building_Re...

You can also download a lot of data here, although it presumably wouldn't be as timely as getting it direct from the satellite.

https://www.ncdc.noaa.gov/data-access/satellite-data/satelli...

mindcrime
I don't know, but it sounds like it would be fun to try. Other than that, all I can add is that I saw this come across my inbox a while back, and it seems like something you might find interesting:

https://climate.apache.org/

mthoms
Thanks for that, it looks interesting.
gesman
Weather prediction is an unsolvable task to begin with.
mthoms
I don't understand your comment. Since 100% accuracy is unachievable we shouldn't try at all?

Wouldn't a prediction with 80% accuracy still be enormously useful in most use cases?