1

I am using neural networks for time series prediction and am completely new to the process. I have 6 covariates in total. Since I am doing time series prediction, I have decided to use several lags of each variable while also including lags of my dependent variable. In total, I have a data frame with 49 possible variables to use.

I literally do not have the time to go through every combination of possible variables and parameters of the network so I have decided on some feature selection processes. I have two questions.

First, if I suspect the underlying data generating process of my dependent variable to be non-linear, does using feature selection methods that are linear, like PCA, hinder the ultimate outcome?

Second, is it appropriate to use non-neural network feature selection methods to find features to then use in the final neural network?

Thanks in advance for any advice/answers!

user111417
  • 91
  • 1
  • 8
  • This is only somewhat related to your second question, but I think it illustrates the kind of thinking that you'll have to do to decide whether feature selection is portable between two different modeling approaches. http://stats.stackexchange.com/questions/164048/can-random-forest-be-used-for-feature-selection-in-multiple-linear-regression/164068#164068 – Sycorax Oct 31 '16 at 19:51

0 Answers0