Create a preprocessing step that limits the size of the training window to the `n` most recent observations per location.