Modelling multivariate Time series with LSTM, the y of the TS are on >50% consist of zeros. the same is true for features
I use loss = 'mean_squared_error', optimizer =Adam, and make grid search for other parameters.
as an output - I receive either constant (e.g. set(y_pred) = -0.0345), or the model
throws an error saying it encountered NAN or infinity.
I experimented with the TS, and it appeared that once the TS is more variable (less than 40% zeros) i start getting meaningful results. My questions are: 1. Can LSTM produce meaningful prediction for TS with zeros at the level of >50% 2. Will it help if I experiment with the optimizer?
A reference to a paper/example will be highly appreciated.