I am applying the N Beats Model of the pytorch-forecasting package on a traffic dataset. I am doing single step prediction with a context length of 5. Now the prediction is unfortunately slightly shifted from the original data. For me this looks pretty much like a wrong hyperparameter. I set batch_size to 25, context_length to 5 and prediction_length to 1. This is a plot plot of concatenated single step predictions vs the actual data Why is my plot shifted and how to prevent this? Continue reading...
