What im familiar with is: univariate: input is one time series, output one time series. Multivariate: input is many time series (like features or channels) and output either multi or univariate signal (could be forecasting the input, or predicting some other signal(s))

In neuralforecast, it's more about channel dependence and independence. In the case of FEDformer, there is no channel dependence, so it's "univariate" in that sense. However, you can still pass multivariate datasets (like ETTh) and get predictions for all series. Only those predictions will not consider the inter-dependence of each series on one another. Take another model, like TSMixer, where channel dependence is inherently part of the model due to the mixing layers. Here, the model is implemented as "multivariate". I hope this clarifies thigs a bit!

Got it! Though I do believe this is not the right naming convention