Shouldn't the training data be preprocessed to transform time values to a relative coordinate system? I would expect a time-series training pipeline to contain something like:
1. Define {N = context duration, M = forecast duration} upfront
2. Select some time value T
3. Extract historical data whose timestamps lie in time interval (T, T+N+M)
4. Transform timestamp values to (-N, M) interval by subtracting T+N from each timestamp
5. Append timestamp-transformed data to training data
6. Goto 2
Or are you saying that people don't want to define N and M upfront?