We tested if a generic model as LSTM can learn time series and its latent state. We found that we need to model the full distribution of the time series; and if the latent state is low-variance, we need to feed surrogate to the latent state to the fitting procedure.

https://github.com/jli05/CS229-TimeSeries-LSTM