$t_0$: A Time-Series Foundation Model for Forecasting with Context
Lucas Meyer, Claudio Sole, Huikan Xiang, Nicolas Li, Lucas Franceschino, Arnau Quera-Bofarull, Maarten P. Scholl, Joachim Fainberg, et al.
cs.LG · cs.AI
Quick verdict
Vendor report on open-weights TSFMs t0-alpha/beta; t0-beta ranks 3rd on GIFT-Eval/fev-bench; thorough evals but deep analysis covers only the weaker model.
Summary
The paper introduces t0, a family of open-weights time-series foundation models (t0-alpha, 102M parameters; t0-beta, 256M) that forecast probabilistically while conditioning on target history, past covariates, and known-future covariates without task-specific retraining. The transformer alternates attention along time and across variates, with asymmetric variate attention preventing future-information leakage, and emits monotone-by-construction quantiles. Pretraining combines a rebalanced GIFT-Eval corpus with four synthetic generator families (kernel composition, causal graphs, covariate effects, mixing) under a curriculum; t0-alpha trained in 28 hours on four H100 GPUs. On GIFT-Eval, t0-alpha scores 0.4941 CRPS (11th of 17 zero-shot models) while t0-beta reaches 0.4738 CRPS and 0.6865 MASE, third, within 4.0% of the leader; on fev-bench, skill scores are 42.2 and 46.7. Passing known-future covariates lifts t0-alpha's skill by 6.3 points across 30 tasks; past covariates add 2.7. The report further analyzes calibration, tail extrapolation, missing-data robustness, context efficiency, and rollout, but these deep dives concern only t0-alpha; an independent Macrocosm ERCOT evaluation shows both models cutting the lagged-price baseline MAE by 38%. Main limitations: undocumented t0-beta training, no controlled scaling study, mid-table t0-alpha accuracy, and self-run benchmark rankings.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- t0-beta ranks third among released zero-shot TSFMs on GIFT-Eval (CRPS 0.4738, MASE 0.6865) and fev-bench (skill 46.7), within 4.0% and 2.0 points of the leaders respectively.
- Supplying known-future covariates to the same t0-alpha checkpoint raises fev-bench skill by 6.3 percentage points across 30 tasks; past covariates add 2.7 points across 12.
- t0-alpha's pooled quantile coverage stays within 0.014 of nominal at all eleven queried levels on GIFT-Eval, and post-hoc IQF exponential tails restore tail calibration without retraining.
- Alternating time/variate attention with leakage-safe asymmetric masking lets a single open-weights model exploit covariates, with gains concentrating where covariates plausibly drive the target.
Strengths
Unusually complete reproducibility for t0-alpha
Weights on Hugging Face, the tfc-t0 package, a GitHub repo, and a notebook reproducing GIFT-Eval numbers (Sections 1, 7); Section 4.2 discloses corpora mix with draw probabilities (Table 1), curriculum schedule, 2D packing, optimizer settings, and the 28h/4xH100 budget.
Leakage-aware benchmarking
Section 5.6.1 identifies the 2014 Victoria demand data inside the GIFT-Eval pretraining corpus (australian_electricity_demand, elecdemand) and reruns the skforecast backtest on 2024 data; every model's error rises 2-2.6x, quantifying the leakage effect rather than hiding it.
Clean covariate-path isolation
Deployment-relevant analyses rarely reported
Independent third-party evaluation
Weaknesses
Analysis/claim mismatch
t0-beta undocumented
No significance for rank margins
Small self-selected evaluation subsets
Self-run rankings with discretionary exclusions
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Peer Review: t0 — A Time-Series Foundation Model for Forecasting with Context
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the t0 time-series foundation model?
t0 is a family of open-weights forecasting models from The Forecasting Company: t0-alpha (102M parameters) and t0-beta (256M). Both produce probabilistic quantile forecasts for arbitrary series, optionally conditioned on past measurements and known-future covariates such as planned promotions or weather forecasts, with no per-task retraining.
How accurate is t0 compared to other forecasting models?
Per the paper's benchmarks: on GIFT-Eval, t0-alpha scores 0.4941 CRPS (11th of 17 zero-shot models) and t0-beta 0.4738 (third, within 4.0% of the leader TimesFM-3.0 at 0.4557). On fev-bench they score 42.2 and 46.7 skill, third place for t0-beta. t0-alpha alone is mid-table on univariate accuracy.
Do covariates actually improve t0's forecasts?
On fev-bench, feeding known-future covariates to the same t0-alpha checkpoint raised skill from 36.7 to 43.0 (a 6.3-point gain) across 30 tasks; past covariates added 2.7 points across 12 tasks. Gains are task-dependent: 11 of the 30 known-future tasks got worse, and the effect collapses when covariates are misaligned in time (Figure 7).
Where can I get t0's weights and code?
The paper lists Hugging Face weights for both models (theforecastingcompany/t0-alpha and /t0-beta), a GitHub repository (theforecastingcompany/tfc-t0), and a pip package tfc-t0, including a notebook that reproduces the GIFT-Eval results for t0-alpha.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are image-only in the extracted text; win-rate grids (Fig 4-5), calibration curves (Fig 9-11), and rollout curves (Fig 14) were assessed from caption numbers and body text only.
- t0-beta's training configuration is absent from the text; statements about it rest entirely on the authors' claims.
- Rankings reflect the leaderboard snapshot at writing time (Sept 2026) and may have shifted.
- Text appears complete through references and Appendix A (Table 13), but any content after it would be invisible to this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.