Tabular foundation models can effectively forecast multiple correlated time series by converting the problem into independent scalar predictions, enabling zero-shot forecasting that captures inter-channel dependencies better than treating each series independently.
This paper shows how to use tabular foundation models (like TabPFN) to forecast multiple time series variables together, rather than treating each variable separately. By reformulating multivariate forecasting as individual scalar prediction problems, the method works zero-shot without retraining, and captures relationships between different time series channels.