it looks promising as a foundation model for transfer learning

Thanks for the insight! We presented our work on transferability of different architectures at the International Symposium of Forecasting. We found that most architectures can work reasonably well (nhits, tft, pathtst, lstm, tcn), outperforming autoarimas and other baselines trained directly in the target dataset.

very cool, any chance there’s a recording of that? also curious if this was done on global models

@Cristian (Nixtla) Would is be interested in your insights 🙏