Time Series Forecasting with RNNs

I am playing around with the Kaggle competition Web Traffic Time and wanted to practice with more LSTMs for forecasting. I just finished fast.ai lesson 7. Should I continue with part 2 fast.ai lessons (perhaps skip to a particular lesson) OR spend the next two weeks (Sept 10 deadline) with other examples first?

I have looked on the forums here and found this article on the forums about time series and moved over to this blog post by Dr. Jason Brownlee. It seems that I am on the right path, but my numbers are training extremely slowly for even 1 row when changing to a much larger dataset. A notebook of Dr. Browniee’s code is here.

Jeremy discusses timeseries some more in the last lesson of part 2. So watching it might give you some mroe tips on how to proceed.

1 Like

Thanks! That is very helpful, and I will check that out tomorrow morning!

I feel I am getting closer to a different article discussing forecasting. Unfortunately, I am running into many problems such as:

  1. The modeling is super slow for 1 row. It will not be scaleable for the large dataset.
  2. The errors are incredibly inaccurate for predicting 30 days out. It seems like values just explode out there.
  3. Many of the layers from Lesson 6 I have tried porting over to no success.

You can see my code here.

Thanks for the great advice! Two weeks and a week being locked in for hurricane Harvey brought me some lovely predictions!

I am still confused by the cats and contin section found [here] (https://www.youtube.com/watch?v=1-NYPQw5THU&feature=youtu.be&t=53m54s). How do I choose which values go where? Currently, I put them together like this:

Is there another resource explaining these? I matched the Rossman data with weight the best I could, but these are still just guesses.