Reproduce “$t_0$: A Time-Series Foundation Model for Forecasting with Context” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lucas Meyer, Claudio Sole, Huikan Xiang, Nicolas Li, Lucas Franceschino, Arnau Quera-Bofarull, Maarten P. Scholl, Joachim Fainberg and 1 more
t0 introduces an open-weight, zero-shot time-series foundation model built to forecast targets alongside past and known-future covariates. Its alternating temporal and cross-variate attention improves covariate-informed forecasting, with t0-beta reaching 0.4738 CRPS on GIFT-Eval, 46.7% skill on fev-bench, and roughly 38% lower MAE than a lagged-price baseline in an ERCOT electricity-price evaluation.
Key insights
- Covariates provide the largest gains when their relationship with the target is operationally meaningful and temporally aligned: known-future inputs improved skill by 6.3 percentage points across 30 tasks, although 11 tasks became worse when covariates were added.
- The model is reasonably calibrated in its central prediction ranges, but its native 0.1–0.9 quantiles underrepresent extreme tails; post-hoc tail interpolation helps, and t0-beta expands native output levels to 0.01–0.99.
- Performance degrades under missing history and long-horizon rollout, and t0-beta’s stronger benchmark results cannot be cleanly attributed to scale alone because the report does not provide a controlled scaling study.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.