Awesome Agentic Time Series

A curated bibliography of 243 papers on LLMs and agents applied to time series, MIT-licensed, 188★/10 forks, created 2026-06-01 and last pushed 2026-08-10 (read 2026-08-10). It is the companion repository to a survey by Hu, Yang, Dai and eleven co-authors, The Landscape of Agentic Time Series Systems: Architectures, Reliability, and Frontiers, whose PDF ships in the repo.

The spoke’s second bibliography after ml-system-design-case-studies, and the same kind of object: it tells you where a literature is, not what is true in it.

What it covers, by weight

Counted from the README, because the shape of the counts is the finding:

The distribution is the argument

Sixteen papers on evaluating forecasts and agentic decisions, against 133 on building the systems. Adding the reliability section, the whole “does it work and can it be trusted” side of this field is roughly 23 papers out of 243 — under a tenth — and the survey’s own title puts reliability in second position.

That is the same imbalance demo-to-production-gap names, arriving from the literature side rather than the practitioner side, and it carries further than any individual entry. This wiki’s time-series-forecasting page ends by saying it cannot tell whether a pretrained forecaster beats classical alternatives, because its one source benchmarked against seasonal-naive on synthetic data. A 243-paper index confirms the question is live and does not answer it: the papers that would are the thinnest category in the list.

What it adds concretely

A named taxonomy of agent roles for one task family — perception, reasoning, planning and action, memory and knowledge, world-model and data — plus reliability as a separate concern. The survey draws the field as four streams (benchmarks, foundation models, LLM4TS, agents) converging on a closed-loop system.

Evidence that timesfm is not alone. Thirty-one foundation-model entries name Falcon-X, Toto 2.0 (“forecasting enters the scaling era”), Timer-S1 at billion scale, AME-TS, TimeCAP, TS-Memory. The pretrain-once move this spoke recorded in a second modality is now a crowded field rather than one Google model.

Domains outside finance. Weather captioning and multi-scale forecast reporting, IIoT anomaly detection, motion time series, multimodal classification. Time series is not a synonym for markets here, which is the reason this routed to the ML spoke rather than the trading one.

Provenance and what it cannot do

T3. A maintainer-curated paper list, three months old, with a self-declared “200+” that counts to 243, and no inclusion criteria published beyond the section headings. Every claim it makes about a paper is that paper’s title. The survey behind it is unpublished as far as the repository says, and the BibTeX entry has no venue or year.

A bibliography is a lead, not a finding. Nothing here was read; the counts above are of titles. The value is knowing what exists and in what proportion, and the proportion is the part this page stands behind.

Related

time-series-forecasting · timesfm · timesfm-2-5-forecasting-tutorial · demo-to-production-gap · ml-system-design-case-studies · machine-learning

Cross-spoke. The agent architecture vocabulary — perception / planning / memory / world-model, and reliability as its own layer — is ../agentic-tooling-wiki’s subject; that spoke owns agent-memory and agent-guardrails and should pick this up if agents-for-a-task-domain becomes a body there. The finance slice (agentic trading, alpha-factor discovery, FinHEAR) belongs to ../quant-trading-wiki, which already holds financial-time-series-foundation-model and kronos-financial-foundation-model. Both noted, neither duplicated.