Question: How Effectively Does TFT Deal with Nonstationarity?

After reading through your paper and applying TFT to model COVID-19 trends across US states, I am still unclear about how effectively TFT deals with nonstationarity. My trained model appears to be doing as well as can be expected, given how aberrant the pandemic trends have been in the US; however, I'm wondering if TFT necessitates stationarity transformations in this case or if the sequential attention is able to learn to deal with this on its own fairly effectively (given that obviously the more feature engineering that is done can help improve model performance but then again part of the appeal of TFT is that one can avoid much of that work through the model).

Thanks for your thoughts. Much appreciated!

No repository file, test, or entry point is named. Start with the referenced TFT paper and the model's treatment of nonstationarity; a useful resolution would explain whether stationarity transformations are required or whether sequential attention handles the issue.

Written by the indexing model from the issue text.