TimesFM-3: Native Multivariate Zero-Shot Forecasting Model
ymatias · x · 2026-09-01
Google has released TimesFM-3, a time-series foundation model that ranks first across major public benchmarks like Gift-Eval, FEV-Bench, and Time.
Key Features:
- Native Multivariate Support: Jointly predicts multiple co-evolving time series to capture cross-series dependencies without task-specific fine-tuning.
- Flexible Covariates: Natively incorporates past covariates (e.g., foot traffic) and dynamic future covariates (e.g., weather forecasts) using a clever "lookahead" token strategy.
- Alternating Attention Architecture: Built on a decoder-only transformer, it uses causal temporal and full variate attention in a 2D grid to blend temporal patterns with cross-series relationships.
- Non-Autoregressive Decode: Generates the entire forecasting horizon and 9 probabilistic quantiles in a single forward pass using Contiguous Patch Masking, eliminating iterative latency and error compounding.
More from Research
- FBI method trains for Pass@K with one sample per update at 12.3% of the cost — EmtiyazKhan · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- 5.8M-param Speech Event Detector adds laughs, breaths and背景 sounds to Whisper transcripts — Ayush Kumar Bar · 2026-10-16
- Huawei's SUPRA-Tex turns one image into relightable 3D materials with WYSIWYG fidelity — Chen Yang · 2026-10-16