Hi! I have a question about the NBEATS Implementation in Neural Forecast. Specifically, how were the default parameters decided? Are these the parameters used in the original paper?

Hi @Nakul Upadhya. While some parameters are based on the original paper, we reduced the total size of the architecture (for example less number of blocks) since the original paper was evaluated on a large dataset with several thousand time series. We wanted to have a smaller (and faster) model as default.

Thanks for the help. Were these the parameters used for the results reported in the NHITS paper then? Appendix D in the paper only mentions the HP space for Autoformer, Informer, ARIMA, Prophet, and NHITS, but nothing about NBEATS.

No, for the NHITS paper we created a small hyperparameter search space and select the optimal parameters based on the validation loss.