The estimation window, I guess, is not inheritant of the function; instead, it depends on how you implement the forecasting process. You can apply the expanding and rolling estimation window in the below manner:
- In order to implement the expanding window, you continuously need to add new data to your training set as they become available, retraining or updating your model with the entire dataset up to the current time point.
- In order to implement the rolling window, you need to maintain a fixed-size dataset by removing the oldest data point and adding the newest one, retraining or updating your model with this subset of data.
I hope this clears your doubt :)
