I tried to use LSTM to predict one-step ahead of a time series.
But after training, it seems the only thing that LSTM learned is to shift the curve one step ahead and modify the old value very little ( o - ground truth, x - predicted value of the same time step)
Does anyone know how this happens and what could be done to improve the prediction behavior?
I split the data as training and test parts. Here are the whole results of test data segment.