Skip to content

Unified Temporal-Tabular Foundation Architectures

Home / Open Questions / Unified Temporal-Tabular Foundation Architectures

Background: Time series foundation models often rely on distinct architectural paradigms for modeling temporal dynamics versus in-context covariate regression. Achieving robust integration of both capabilities is essential for models deployed in scenarios with rich exogenous information.

Question / Future Work: Developing a unified foundation model architecture that effectively synthesizes explicit temporal dependency modeling (e.g., autoregressive forecasting) with robust in-context tabular regression capabilities remains an unresolved challenge for comprehensive covariate-aware forecasting.

Why It Matters: Current time series foundation models typically excel at either univariate temporal modeling or feature-target regression, creating a performance gap in domains where both patterns are crucial for accuracy.

Evidence: In particular, a foundation architecture that combines explicit temporal dependency modeling, as in Chronos-2, with in-context regression capabilities inspired by TabPFN-TS, could represent an interesting direction for covariate-aware time series foundation models.

Metadata & Links