Why LSTM doesn't predict low and high values in regression?

Question

I made a stack of bidirectional LSTM layers following by 4 Dense/Dropout layers with a swish activation function in order to predict a continuous value between 0 and 2. I co