Complete Correlation Modeling for TSFMs

Background: Time series foundation models are typically trained using channel-independent modeling, focusing on temporal dependencies while often neglecting complex inter-channel relationships.

Question / Future Work: Investigating the true extent to which more comprehensive modeling of dynamic (DCorr), heterogeneous (HCorr), and partial (PCorr) correlations can further enhance the generalization and predictive performance of Time Series Foundation Models (TSFMs) beyond the capabilities demonstrated by the proposed CoRA adapter.