0

I'm making a CNN model for time series forecasting and towards evaluating the Model I did a time series cross validation that shows in pic What I did that was by every iteration I train my model and test it and save the resulting error and at the end I calculate the average error.Do I miss something? enter image description here

image reference:https://medium.com/next-gen-machine-learning/types-of-cross-validation-in-machine-learning-8bd33bf3e12f

Izo
  • 1
  • 1
  • Nope you are good. Keep going. – usεr11852 Jul 25 '20 at 15:23
  • thanks for the feedback – Izo Jul 25 '20 at 15:26
  • @usεr11852 actually, if Izo is using CNN for time series forecasting, then they don't need to use time series cross validation, normal CV is also applicable. – Skander H. Jul 25 '20 at 16:12
  • why I don't need it? can you please explain – Izo Jul 25 '20 at 16:24
  • The standard text book response to your question is that what you are doing is correct. However, there are relatively recent results that show that normal CV is applicable to time series when you use ML models (as opposed to traditional models like ARIMA or Holt-Winters) - see [here for details](https://stats.stackexchange.com/a/375031/89649) – Skander H. Jul 25 '20 at 16:51
  • @SkanderH. Thank you for pointing that out it is a good point for the OP to follow. What you suggest requires reformatting the original data. Given that the OP does not reformat the data beforehand, what is currently described, is correct. Also it will greatly simplify life because the same TS-CV schema will be used throughout when comparing with "standard methods" (e.g ARIMA). – usεr11852 Jul 25 '20 at 19:12
  • Thanks a lot for helping – Izo Jul 25 '20 at 19:34
  • @usεr11852 except that the OP has to perform that reformatting anyway if they want to use a Deep Learning model. – Skander H. Jul 25 '20 at 21:22

0 Answers0