Back to AI Research

AI Research

Timestral combines horizon forecasting and future covariates in four time-series models

Key Takeaways

  • A model-family report describes synthetic pretraining, mixed post-training and reported forecasting benchmark results.
  • [Timestral](https://arxiv.org/abs/2610.37585) introduces four time-series foundation models for probabilistic forecasting, with 5 million, 28 million, 123 million and 209 million parameters.
  • The report describes their design and results on named forecasting benchmarks.
  • ## Separate the horizon from the observed context
  • The first design element is Decoupled Horizon Forecasting.

Timestral introduces four time-series foundation models for probabilistic forecasting, with 5 million, 28 million, 123 million and 209 million parameters. The report describes their design and results on named forecasting benchmarks.

Separate the horizon from the observed context

The first design element is Decoupled Horizon Forecasting. It separates prediction tokens from the causally encoded context, allowing longer forecast horizons to be trained without masking observed context.
The authors also describe a covariate framework that supports known-future covariates within a causal architecture. Such inputs are information available for a future interval, rather than future target values that an evaluator may only know after the event. Keeping that distinction explicit matters when constructing a forecasting test.

Two stages of training

Timestral pretrains exclusively on a diverse synthetic corpus, then post-trains using real and synthetic data with extended context. The report presents this pipeline alongside the forecasting architecture as ingredients of the model family.
The authors report that the 209-million- and 123-million-parameter models take the top two places on every metric in GIFT-Eval, TIME and fev-bench. These are results reported for the paper’s benchmark settings. They do not establish that the same ranking will hold for every organisation’s series, horizon or covariate pattern.

Weights and a workload-specific comparison

The paper says weights for all four models are available on Hugging Face and inference code on GitHub. That makes it possible to investigate different model sizes rather than judging the family only by its largest member.
A practical evaluation should retain a realistic forecast cutoff and identify which covariates would actually have been known at that point. Compare forecast quality across the horizons and uncertainty measures the application needs, together with runtime. The report provides a starting point for that test, rather than a guarantee that a foundation model replaces domain-specific evaluation.

Comments