That said, this is a point of contention that even the authors of the paper allude to:

"The most principled approach is thus to build ancestry models in which source and ‘outgroup/reference’ populations are older than, or at least contemporary with, the target genome or group that we are trying to model. However, this has been challenging, due to the limited statistical power offered by the thousands-fold lower sample sizes and reduced sequence quality of ancient genomes"