Review Jesper Dramsch
This contribution adapts a neural network from Scheuerer et al. 2020 "Using Artificial Neural Networks for Generating Probabilistic Subseasonal Precipitation Forecasts over California". The models are trained independently on the two variables t2m and tp and the corresponding lead times, amounting to four models. The models are trained to directly predict the terciles using categorical cross-entropy.
Finally, the models are trained on 5 different random seeds and averaged and then smoothed using a Gaussian filter. The data is trained on the years 2000-2019 there is no sign of overfitting to 2020.
Curiously, the training was done over a limited area over Easter Europe and then applied to the global prediction. My questions arising from this choice are similar to those of #3.
I want to thank the authors for the documentation, links to the paper, and commented code.