Timestral models lead GIFT-Eval, TIME and fev-bench
Original title: Timestral
Four probabilistic forecasters from 5M to 209M parameters ship with Hugging Face weights and GitHub inference code.
Timestral is a family of probabilistic time-series foundation models with 5M, 28M, 123M and 209M parameters. Timestral-209M and Timestral-123M take the top two places on every metric of GIFT-Eval, TIME and fev-bench. The design uses Decoupled Horizon Forecasting, native known-future covariates, synthetic-only pretraining and mixed real-and-synthetic post-training.
- Model sizes are 5M, 28M, 123M and 209M parameters.
- Decoupled Horizon Forecasting separates prediction tokens from context.
- Pretraining uses only a diverse large-scale synthetic corpus.
- Weights for all four models are available on Hugging Face.
From the original article
1 min read at arxiv.orgWe introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric.
By Adrien Petralia · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.