[arXiv]score: 0.29

Timestral Time Series Foundation Models Achieve SOTA on Forecasting Benchmarks

October 28, 2026

Timestral introduces four probabilistic forecasting models up to 209M parameters using Decoupled Horizon Forecasting to separate prediction tokens from causal context. The models set new state-of-the-art results on GIFT-Eval, TIME, and fev-bench through a two-stage pretraining pipeline on large-scale synthetic and real data.

HOW THIS AFFECTS YOU

●

builderYou can integrate these specialized foundation models for improved probabilistic forecasting in production systems.

●

researcherThe Decoupled Horizon Forecasting method provides a new way to handle long-horizon predictions without context masking.

read original ↗arxiv.org

DAILY DIGEST

you don't check 9 sources — we do. one email every morning, read in 2 min. free. unsubscribe anytime. privacy

← back to feed