Determine the cause of covariate-related performance degradation in time series foundation models

Determine whether the performance degradations observed when time series foundation models use known-future covariates arise from uninformative covariates themselves or from the numerical representation used to encode those covariates.

Background

The paper discusses prior evidence that covariate integration can improve time series foundation model forecasts, while noting that such gains are not universal. In particular, reported results for t₀-beta show performance degradation on some forecasting tasks with known-future covariates.

The unresolved issue is whether these degradations reflect the substantive lack of predictive information in the covariates or instead result from how the covariates are numerically represented and supplied to the model. Resolving this distinction would clarify whether poor performance should be addressed through covariate selection or through improved encoding strategies.

References

Whether these degradations stem from uninformative covariates themselves or from their representation cannot be decided from the reported results.

— On the Encoding of Calendar Features in Covariate-Informed Time Series Foundation Models  (2610.13676 - Kreusel et al., 9 Oct 2026) in Section 2, Related Work and Foundations, paragraph “Time Series Foundation Models and Covariates”