LLM Agents for Time-Series: A Survey

ACL ARR 2026 March Submission1652 Authors

Keywords: LLM agents, time-series reasoning, tool use, agent memory, multi-agent systems

Abstract:

Recent advances in large language models (LLMs) have accelerated the development of agentic systems for time-series analysis.
While traditional time-series methods typically make predictions or decisions based on a given set of evidence, many real-world applications require agentic systems that can autonomously plan workflows, reflect on intermediate results, and leverage external tools and memory.
This survey presents a systematic review of LLM-based agents for time-series tasks.
We adopt a problem-driven taxonomy, organizing existing systems into four problem categories: Forecasting \& Reasoning, Data Augmentation \& Synthesis, Anomaly Detection \& Diagnosis, and Decision Support.
For each category, we analyze how agent behaviors are implemented through architectural design, external tool integration, and memory mechanisms.
We further discuss datasets, environments, and evaluation protocols, and outline open challenges and future directions for LLM-based time-series agents.
Overall, our goal is to provide researchers with a structured view of how LLM-based agents can be designed to address different time-series problems.

Paper Type: Long

Research Area: LLM agents

Research Area Keywords: LLM agents, time-series reasoning, tool use, agent memory, multi-agent systems

Contribution Types: Position papers, Surveys

Languages Studied: English

Submission Number: 1652