univariate implementation of PatchTST

[koukouer]

Description

I have a question regarding the implementation of PatchTST. While the paper suggests that it supports multivariate time series, I've noticed that the code utilizes the BaseWindows class, which seems to default to univariate prediction.
Could you please clarify why the implementation is univariate?

Use case

No response

[marcopeix]

Hello! The paper mentions channel independence, which is what we mean by "univariate" and "multivariate". Models like NHITS and PatchTST can forecast multiple time series, but they will be predicted independently due to channel independence. The relationship between series is not learned by the model. As such, the implementation follows the paper.

However, a "multivariate" model, like TSMixer, iTransformer, learn interdependencies between all series in the dataset.

So, all models can forecast multiple time series, but in a different way. "Univariate" models will apply channel independence, so forecast each separately, "multivariate" models learn the relationship between all series.

This is also explained in our table of forecasting models. You can see the univariate/multivariate explanation, and also see which model supports exogenous features.