'MS' forecasting task downperforming 'S'

[ClaudiaShu]

Hi, thank you for your work.

I'm testing your code with ETTh1 dataset using both multivariate predict univariate task (MS) and univariate predict univariate (S). If I understand the process correctly, MS is to forecast the future OT with an input of [feature1, feature2, ..., featureN, OT], while the S task is to forecast the future OT based on the observed OT. However, the MSE of the MS task is larger than that of the S task.

My question is, why is the performance worse with more features provided?

Thank you!

[ANTAGNIST]

Besides informer, using lstm will meet the same questions.