Review Stephan Rasp
I would like to thank the authors for their complete description and readable code. The authors adapt the method of Scheuerer et al. 2020 to predict terciles based on the variable to be predicted. The method, a CNN, is trained only for a limited geographical region over Eastern Europe but applied globally after training. The train/valid split is well executed and no suspicion of overfitting is warranted. Therefore, this submission receives a pass from me.
Here are some further questions for the authors:
-
Why did you choose to train only for the limited region? Was it a computational problem? Would you expect your method to perform better if it had been trained globally? How does this work with the basis functions? Aren’t they location specific?
-
How well does the method perform on the limited region trained? Is it much better there than elsewhere?
Best, Stephan Rasp