I test the Sagemaker AWS solution for RNN: deepAR.
Previously I used sklearn for this and obviously I cleaned the data to avoid highly correlated time series (KBest, PCA, etc.)
Do you think it is needed here, or RNN will handle internally correlated series.
PS. This is typically the case with blackbox commercial solutions. They look fantastic, but when it comes to details you have nowhere to consult.
$\begingroup$
$\endgroup$
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
I wouldn't do much to the data before using as input to DeepAR. I would just make sure the timeseries I'm using as input have some variance (are not stationary for long).
DeepAR uses a simple yet efficient sampling technique to handle timeseries of different velocity.
DeepAR's paper state some of the capabilities the architecture should handle. https://arxiv.org/abs/1704.04110
In any case, it is always a responsible thing to try the simple first and use that as a baseline.
$\begingroup$
$\endgroup$
I am not a specialist but I would say that you only have to use the right format for your data. Since NNs have the ability to handle high dimensional features it will not harm your model if those features are correlated.
