I have forecast dataset containing multiple multivariate time series that are not independent from each other. A state in one of the the series in time "t" can influence the state in another in time "t+n".

I was initially thinking in doing a LSTM model for each time series, but I ran in the problem of how to make a LSTM speak with another, is that possible? How could I do that?

I also thought in transforming the multiple series in a single one, but the goal is to make a forecast of the class of each series and I don't think this way I could have the classification of each one as an output, right?

Are there any other solutions? Could I make a CNN-LSTM work for something like this?

  • $\begingroup$ There certaintly are fancier methods, but a vector autoregression (VAR) seems like a reasonable starting point. $\endgroup$ Commented Apr 6, 2022 at 8:39

1 Answer 1


For time series classification tasks it's usual to build one model trained on all the samples in your training set, rather than a separate model for each time series. The great multivariate time series classification bake off paper provides a summary and comparison of the latest multivariate classification algorithms, including some CNN (InceptionTime, ResNet, both available in the sktime-dl python package) and CNN/LSTM (TapNet) ones.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.