Does the latest version of the released model support multivariate prediction?
I would like to inquire if the latest model is based on Timer-XL? Does it support multi-variable prediction? I have checked the Github repositories of Large-Time-Series-Model and Timer-XL, but there seems to be no clear explanation.And it also seems that both repositories point to the same huggingface model.
The current version is built on the Timer-XL structure (e.g. TimeAttention and RoPE), but this checkpoint is univariate pre-trained since the details of downstream tasks (e.g. the variate number) is agnostic. To make multivariate prediction, we currently recommend the channel-independence approach during inference.
If you're interested in building domain-specific large models (e.g., the variate number can be fixed), we recommend Timer-XL for explicit multivariate pre-training. The example pipeline is conducted on the ERA5-MS on our codebase.