search

DAMO-DI-ML / NeurIPS2023-One-Fits-All

The official code for "One Fits All: Power General Time Series Analysis by Pretrained LM (NeurIPS 2023 Spotlight)"
446 stars 62 forks source link

Long-term Forecasting Target #41

Open jp-guo opened 4 months ago

jp-guo commented 4 months ago

Thanks for your great work!

In the ETT datasets, the target shoule be the oil temperature (OL), but in the repo I find that OFA seems to predict all 7 features in a single-to-single manner. Also the feature category is not part of the input. So OFA seems to hypothesize that all 7 features are from the same distribution. Is this the same setting with other baseline methods? I find most papers claim they treat OL as the only target feature.

tianzhou2011 commented 4 months ago

This approach, often referred to as channel-independent, is utilized for multivariable forecasting and has been demonstrated by several previous studies to help mitigate overfitting. Meanwhile, the paper you're examining may focus on univariate forecasting, where operational loss (OL) is the sole objective.