This message was deleted.

Hi @marah othman, We use by default in the RNN-Based methods (RNN, GRU, LSTM, TCN,...) all the available history for the series at the time of the prediction y[:t]. If you want to reduce the inference time you can use the

input_size

, to make the inputs of fixed size y[t-input_size:t].

do we treat the problem of non stationary time series? in the library . and what about this