Bidirectional layer for time series forecasting

In the case of time series forecasting, using windowed data, why does using a Bidirectional layer make sense considering that we are forecasting the future based on past observations? Isn’t standard LSTM more appropriate as we only consider information from one direction?

Bidirectional layer processes the inputs from both directions and so improves the learning.

But for time series forecasting, what’s the input from the future side?

There exist 2 unidirectional RNN layers. One of them will process the inputs from start of sequence and the other will process the input from end of sequence and their outputs will be merged via concatenation.

  1. See call arguments here
  2. Watch this lecture. Don’t forget to take note of the disadvantages of bidirectiaonl rnn (need for the entire sequence to be known in advance).