Computer Science editorial
Open AccessOA2026
$t_0$: A Time-Series Foundation Model for Forecasting with Context
Open-weights 102M and 256M parameter models that condition forecasts on target history, past covariates, and known-future covariates without task-specific retraining
Lucas Meyer; Claudio Sole; Huikan Xiang; Nicolas Li; Lucas Franceschino; Arnau Quera-Bofarull; Maarten P. Scholl; Joachim Fainberg; Geoffrey Négiar· 2026· DOI 10.48550/arXiv.2609.24559
The core problem
Time-series forecasting has increasingly adopted the foundation-model paradigm, where a single pretrained model is applied zero-shot across diverse tasks. The authors present , a family of open-weights foundation models designed for forecasting with multivariate context. The first two members are with 102M parameters and with 256M parameters. Both condition their forecasts on three information sources: the target history, past covariates, and known-future covariates. Crucially, they operate without task-specific retraining. The models produce probabilistic forecasts through quantile predictions, and their transformer layers alternate attention along time and across variates. Pretraining combines curated public data with synthetic generator families constructed to contain covariate-to-target dependencies. The work targets practical forecasting settings where exogenous drivers are known in advance, such as electricity demand and prices.
Innovation
On GIFT-Eval, reaches an aggregate CRPS of 0.4941, while achieves a CRPS of 0.4738 and a MASE of 0.6865. Both rank third on these metrics and are within 4.0% of the best zero-shot TSFM. On fev-bench, they score 42.2 and 46.7 in skill, with again third and 2.0 points behind the leader. An ablation on shows that known-future covariates raise its skill by 6.3 percentage points across 30 tasks. On the Victoria electricity-demand benchmark, is among the most accurate models when given a context of nearly a year. In an independent Macrocosm evaluation of hourly ERCOT prices over 29 months, both models cut the MAE of the lagged-price baseline by 38%. These results indicate strong zero-shot performance and effective use of contextual information.
Time-series forecasting has increasingly adopted the foundation-model paradigm, where a single pretrained model is applied zero-shot across diverse tasks. The authors present , a family of open-weights foundation models designed for forecasting with multivariate context. The first two members are with 102M parameters and with 256M parameters. Both condition their forecasts on three information sources: the target history, past covariates, and known-future covariates. Crucially, they operate without task-specific retraining. The models produce probabilistic forecasts through quantile predictions, and their transformer layers alternate attention along time and across variates. Pretraining combines curated public data with synthetic generator families constructed to contain covariate-to-target dependencies. The work targets practical forecasting settings where exogenous drivers are known in advance, such as electricity demand and prices.
The architecture is a transformer whose layers alternate two attention modes: attention along the time dimension and attention across variates. This design allows the model to capture both temporal dynamics and cross-series dependencies. Given a target history , past covariates , and known-future covariates , the model estimates a conditional quantile function:
Why it matters
The authors analyze in depth, covering calibration, rollout strategy on long horizons, and robustness to missing data. The 6.3 percentage point gain from known-future covariates across 30 tasks highlights the value of explicitly modeling exogenous drivers. The alternating attention scheme along time and across variates appears effective for multivariate context, while quantile predictions provide a practical probabilistic output. The models' third-place rankings on GIFT-Eval and fev-bench, within 4.0% and 2.0 points of the leaders respectively, suggest that open-weights foundation models can be competitive without task-specific retraining. The 38% MAE reduction on ERCOT prices over 29 months demonstrates real-world impact in energy forecasting. Limitations include the need for known-future covariates and the focus on two model sizes; future work could scale parameters and expand generator families. Overall, advances accessible, context-aware time-series forecasting.
Who should read this
CS practitioners and researchers
Opening member content…