In the Python TSFresh package it is possible to use tsfresh.utilities.dataframe_functions.roll_time_series() and tsfresh.utilities.dataframe_functions.make_forecasting_frame() to help preprocess data for time series forecasting, according to https://tsfresh.readthedocs.io/en/latest/text/forecasting.html.
Specifically, how would the output from rolling extract_features differ from just using extract_features on the full training data set:
- Which types of features generated would be different and how?
- When would roll_time_series be preferred over using the vanilla extract_features directly? Why?
https://github.com/blue-yonder/tsfresh/blob/main/notebooks/05%20Timeseries%20Forecasting.ipynb https://github.com/blue-yonder/tsfresh/blob/main/notebooks/advanced/05%20Timeseries%20Forecasting%20(multiple%20ids).ipynb
Give more examples of the rolling extra_features being used. However, neither show why the rolling version was used - and the documentation does not appear to explain the differences either.