Early-Signal Detector

model · first seen 4 Oct 2026

time series foundation model

Also: TSFM

A large pretrained model for forecasting across diverse time series datasets.

Time series foundation models are large models pretrained on many heterogeneous time series so they can forecast zero-shot on new data, following the pattern of language models. The sources show a mature research area rather than a new idea: papers on scaling laws across 21 checkpoints, retrieval-based correction of frozen models, efficient Mamba-3 based architectures, per-instance fine-tuning, sampling-frequency robustness tests of MOMENT, Chronos-2, Moirai-MoE and TimesFM, and probing of Datadog's Toto. New entrants include Mistral's Timestral release and small open-source energy forecasting repos. Arxiv dominates the volume and weekly counts are flat, which fits steady output in an existing field.

Reliable zero-shot forecasting would let teams in energy, retail, sensing and observability skip per-dataset model training, and entry by a general-purpose lab like Mistral suggests the category is being treated as a standard model type.

Analyst note (established), 17 Nov 2026.

Mentions per week

01319 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 56 on 14 Nov 2026 · 47 mentions in 28 days · 47 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration -0.20

Earliest sources

Latest sources

Mentioned together with

low-rank adaptationjoint-embedding predictive architectureprobabilistic forecastingretrieval-augmented forecastingworld modelscaling lawdistribution shiftin-context learning for time series anomaly detectioncounterfactual supervisiontime series anomaly detection