Determine whether the model selectively uses covariates

Determine whether the t0 forecasting model uses every covariate that carries information and ignores covariates that carry none.

Background

The paper evaluates the forecasting improvement obtained by supplying covariates on fev-bench, but notes that the benchmark does not distinguish a model’s ability to exploit informative covariates from its general univariate forecasting accuracy. Consequently, the available evaluation does not establish whether t0 consistently identifies useful covariates or filters out irrelevant ones.

The authors identify two unresolved questions: whether the model uses every informative covariate and whether it ignores covariates with no predictive signal. They state that answering these questions requires benchmarks specifically designed to measure covariate-selection skill.

References

Two questions therefore stay open. Does the model use every covariate that carries information? Does it ignore the ones that carry none? Answering these questions calls for benchmarks built for the purpose.

— $t_0$: A Time-Series Foundation Model for Forecasting with Context  (2609.24559 - Meyer et al., 21 Sep 2026) in Section 6, “Limitations,” paragraph “Extend covariate lift studies and benchmark”