I am trying to use a multivariate time series to predict the future for one of the channels and use the other channels as context features. As far as I can tell all of the huggingface models either expect access to the context at inference time or to also produce a multivariate prediction. I then tried to just condense down the multivariate output to just a single channel by using a wrapper around TimeSeriesTransformerForPrediction. But it seems the forward method does not return any actual predictions and using generate would be inefficient.
So my question is: Is there any intended way to perform multi- to single channel time series forecasting within the HF ecosystem?
Also I realize that this whole Idea only makes sense if the context features are used in a meaningful way, so that each channels prediction depends on all values in the other channels. I am not sure which, if any, of the models do this so I would appreciate any insight on that.
1 Like
For questions that are logically difficult, you’re probably better off asking them on the general channel of the HF Discord.
There are a lot of active users there who are LLM-related.
So it seems the PatchTSMixer model can do what’s needed via the prediction_channel_indices argument. I’ll try using that.
1 Like
This topic was automatically closed 12 hours after the last reply. New replies are no longer allowed.