I am building a multivariate time series prediction. I want to train and use the network with fixed length series of n events. I know that I could use a RNN for this. What I do not understand though is why this should perform any better than a regular deep network, e.g. with a 1d convolutions. I understand that LSTMs and GRUs have a hidden state. However, if I always use a fixed length series of events, this hidden state should be lost after each prediction. I guess I could keep it, but I am not interested in this because I only want to take the given n events into consideration. So, if the hidden state is lost after each prediction, I do not see why a RNN should help at all. Is not this just limiting my network from learning to detect even more powerful patterns?

Bonus: Apart from TCNs (Temporal Convolutional Networks), are there other archtectures which I should explore?

Edit: To use a regular network with a 3d tensor, I would just reshape it first.

  • 1
    $\begingroup$ The point in using models for time series data is their ability to model autocorrelation. The recursive factor in RNNs is what makes it a native tool for "correcting" autocorrelation and applying autoregression. If other neural network architectures are able for such a feat, then they're potential tools for time series analysis. What is better and what is worse can only be determined a posteriori. $\endgroup$ – Digio Sep 6 '18 at 10:00
  • $\begingroup$ @Digio I see, thank you. Is there anything known about what makes a network autocorrelating/-regressive? $\endgroup$ – Joel Richard Sep 6 '18 at 18:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.