PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM Agents for Time-Series: A Survey

Yilong Chen, Xiao Qin, Chenghao Liu, Liang Wu, Noelle I. Samia, Kaize Ding

arXiv:2608.26226v1cs.AI

TL;DR

LLM-based agents for time-series problems vary across task settings, motivating a problem-driven survey of their architectures, tools, and memory designs. The survey organizes systems into four problem categories and synthesizes recurring design patterns, datasets, environments, and reported performance comparisons.

  • Problem

    Design choices for LLM-based agents vary substantially across time-series problems, while a task-oriented understanding of suitable designs remains needed.

  • Method

    The survey uses a problem-driven taxonomy covering forecasting and reasoning, augmentation and synthesis, anomaly detection and diagnosis, and decision support, examining architecture, tool use, and memory within each setting.

  • Results

    The synthesis shows that agent designs are shaped by task requirements, including evidence integration, semantic preservation, adaptation, simulation, validation, and historical retrieval.

  • Takeaways & Limitations

    A task-oriented view provides guidance for selecting and understanding LLM-agent designs for particular time-series problems.

  • Takeaways & Limitations

    Rapid advances may leave some developments uncaptured, heterogeneous evaluations limit unified comparison, and recurring design patterns lack controlled experimental validation.

Abstract

from arXiv · show

LLM-based agents are increasingly being developed for time-series problems, but their design choices vary substantially across task settings. This survey adopts a problem-driven taxonomy that organizes these systems by the time-series problems they address rather than by isolated technical components. We group existing systems into four categories: forecasting and reasoning, augmentation and synthesis, anomaly detection and diagnosis, and decision support. Within each category, we examine how task requirements shape agent architecture, tool use, and memory design. We further summarize representative datasets and environments, and compare reported model performance under shared or closely related settings. Overall, this survey offers a task-oriented guide to designing LLM-based agents for time-series problems and identifies open gaps for future work.

1 Introduction

Time-series analysis spans forecasting, augmentation, anomaly detection, and decision support, motivating LLM agents that can reason, use tools, and maintain memory across evolving workflows. The survey therefore organizes prior systems by task and examines how architecture, tools, and memory should be designed for each setting.

  • Time-series analysis supports applications in finance, transportation, and climate science through tasks including forecasting, augmentation, anomaly detection, and decision support.
  • LLM agents extend prompt-only applications by planning actions, calling tools, and maintaining memory across multiple steps.
  • Evolving observations, changing conditions, and heterogeneous evidence make agentic workflows particularly suitable for many time-series applications.
  • Recent leaderboards suggest agentic systems can be competitive with state-of-the-art methods by coordinating foundation models, retrieval, validation, and domain tools.
  • The survey addresses gaps in existing taxonomies by grouping systems by target time-series problems rather than isolated agent capabilities.
  • The survey covers background, design dimensions, the task taxonomy, practical resources, and future research directions.

2 Background and Foundations

The background frames time-series agents as observe–act–update systems that combine reasoning, planning, tool use, and memory for tasks requiring evidence gathering, feedback integration, or adaptive action selection.

  • Time-series data are represented as observed or tokenized temporal indices with vectors that cover channels, covariates, or spatially distributed sensors.
  • For irregular, noisy, or partially observed data, systems summarize historical context using a representation constructed from observations up to the current time.
  • Time-series modeling transforms observations into forecasts, anomaly scores, learned representations, or explanatory summaries.
  • LLM agents are framed as observe–act–update systems for time-series tasks requiring evidence gathering, feedback integration, or adaptive action selection.
  • The agentic distinction combines reasoning, planning, tool use, and memory, separating these systems from prompt-only LLM applications.

3 Fundamental Design Dimensions

The survey analyzes time-series agents through architecture, tools, and memory, distinguishing single-agent systems from cooperative, competitive, and mixed multi-agent designs. Tools provide data, computation, and feedback, while memory preserves evidence and historical patterns across evolving workflows.

  • The survey uses architecture, tools, and memory as three core design dimensions for time-series agentic systems.
  • Architecture: Multi-agent systems are categorized as cooperative, competitive, or mixed according to how agents divide roles, compare outputs, or coordinate.
  • Architecture: Single-agent systems use one agent to plan reasoning, call tools, maintain memory, and produce outputs or actions from current observations.
  • Tools: Tools connect agents to temporal data, specialized computation, and verifiable feedback unavailable reliably from parametric knowledge alone.
  • Tools: Tool families include database, search and retrieval, data processing, statistical and machine-learning, simulator, solver, and optimizer interfaces.
  • Memory: Memory maintains coherence across reasoning steps through evidence logs and retrievable libraries of historical cases or recurring patterns.

4 Taxonomy

The survey defines LLM agents through multi-step decision-making and organizes time-series systems by four target problems: forecasting and reasoning, augmentation and synthesis, anomaly detection and diagnosis, and decision support. Across these settings, architecture, tools, and memory are shaped by task-specific evidence, validation, adaptation, and control requirements.

  • Agent definition: LLM agents are systems in which an LLM changes a multi-step time-series workflow through decisions such as tool selection, memory updates, hypothesis revision, or stage coordination.The survey excludes one-shot prompting, static encoding or explanation, and domain-general agents lacking time-series-specific constraints or evaluation.
  • Taxonomy: The problem-driven taxonomy contains four categories: Forecasting & Reasoning, Augmentation & Synthesis, Anomaly Detection & Diagnosis, and Decision Support.The taxonomy groups systems by the problems they address while analyzing architecture, tool use, and memory within each setting.
  • Forecasting & Reasoning: Forecasting and reasoning require evidence beyond fixed context windows, using pattern libraries, evidence logs, and quantitative tools while managing error accumulation in long workflows.Forecasting commonly compares competing hypotheses, whereas reasoning more often uses refinement loops to reduce long-chain errors.
  • Augmentation & Synthesis: Augmentation and synthesis address semantic drift through construction and verification pipelines, with different constraints for numeric perturbations, text alignment, and domain-rule compliance.Numeric augmentation emphasizes semantic preservation; text-guided synthesis relies on structured queries and executable generation settings, while domain-attribute construction filters for attribute consistency.
  • Anomaly Detection & Diagnosis: Anomaly detection and diagnosis use heterogeneous temporal evidence, tool-supported monitoring, sequential pipelines, and memory updates for auditability, retrieval, and deployment-time adaptation.Streaming monitoring requires continual adaptation while controlling contamination so transient anomalies are not absorbed as the new normal.
  • Decision Support: Decision-support designs differ by target: trading favors cooperative planner–executor systems for multimodal evidence, whereas traffic and grid control emphasize simulator-grounded execution, validation, and feasibility checks.Trading systems also use memory for historical patterns and reusable analytical strategies, while infrastructure control evaluates actions in closed-loop environments.

5 Resources

The survey organizes practical resources for implementing and evaluating time-series LLM agents into datasets and repositories, interactive environments, and toolkits.

  • Datasets and Repositories: Resources are organized into four categories with representative examples summarized in Table 5.
  • Datasets and Repositories: Representative forecasting datasets include Electricity, METR-LA, and ETT.
  • Interactive Environments: Interactive environments support closed-loop experiments with explicit state, action, and reward signals.
  • Interactive Environments: Grid2Op, SUMO, and SocioDojo provide representative environments for power systems, traffic control, and trading, respectively.
  • Toolkits: Toolkits provide reusable pipelines, APIs, and baselines for reproducible development.

6 Future Work

The survey identifies open directions involving numerical and domain understanding, causal temporal reasoning, online adaptation, and continual improvement for time-series agents.

  • Numerical Understanding and Domain Knowledge: LLMs remain limited in understanding numerical signals and specialized domain knowledge, even when agents use external tools.
  • Causal and Counterfactual Temporal Reasoning: Time-series agents often rely on correlations, while causal discovery remains assumption-sensitive.
  • Causal and Counterfactual Temporal Reasoning: Diagnosis and decision support require evidence-grounded reasoning about interventions, delayed effects, and counterfactual outcomes.
  • Online Adaptation and Continual Improvement: Agents in anomaly detection and decision support often need post-deployment improvement, but most systems remain static or update memory only.

Limitations

The survey presents a broad overview of LLM-based time-series agents but notes limitations in coverage, quantitative comparability, and empirical validation of design guidance.

  • Scope and coverage: Rapid advances in LLM-based agents mean that some recent developments and emerging directions may not be fully captured.
  • Lack of quantitative comparison: Heterogeneous time-series tasks and evaluation settings make unified and fair quantitative comparison difficult.
  • Design guidance: The survey’s design insights derive from recurring literature patterns rather than controlled experimental validation.
  • Design guidance: Because the guidance is not controlled experimentally, it may not generalize to all practical settings.

A Survey Scope and Related Surveys

The survey defines its scope through a literature search and inclusion rule, then distinguishes its problem-driven taxonomy from related survey taxonomies.

  • Survey Scope: The appendix describes paper selection and differences from related surveys on LLMs for time-series analysis.
  • Survey Scope: The search covered major scholarly databases and paired agent concepts with time-series tasks and application domains.
  • Survey Scope: Included papers had an LLM controlling at least one decision.
  • Related Surveys: Related taxonomies organize methods by LLM adaptation, reasoning structure, agent capability, or time-series problem family.
  • Related Surveys: The problem-driven taxonomy links task settings to architecture, tools, and memory, complementing topology-driven surveys.
  • Related Surveys: The survey covers 47 representative systems and reports per-system annotations in Table 3.

B Representative Prompt Examples by Task Family

The appendix presents abstracted prompt patterns for four time-series agent families, emphasizing task-specific inputs, evidence use, validation, constraints, and structured outputs.

  • B.1 Forecasting and Reasoning: Forecasting and reasoning agents use historical observations, horizons or questions, retrieved context, hypotheses, and tool outputs to produce grounded predictions or answers.Their instructions identify temporal patterns, compare evidence, use tools for quantitative claims, and state uncertainty and failure modes.
  • B.1 Forecasting and Reasoning: These agents return an evidence summary, reasoning, prediction, confidence, and limitations, reflecting recurring needs in representative systems.The example is associated with TimeSeriesScientist, TimeXL, TS-Agent, and TS-Reasoner.
  • B.2 Augmentation and Synthesis: Their inputs include seed series, desired labels or scenarios, required temporal properties, and domain constraints, while outputs record construction and validation details.The structured result includes construction mode, constructed item, control attributes, preserved properties, changed properties, and validation notes.
  • B.2 Augmentation and Synthesis: Augmentation and synthesis agents preserve task-relevant temporal semantics, align control attributes, respect domain constraints, and distinguish preserved from changed properties.The workflow supports target augmentation, text-guided synthesis, and domain-attribute synthesis followed by downstream validation.
  • B.3 Anomaly Detection and Diagnosis: Anomaly detection and diagnosis agents combine detector evidence, temporal context, auxiliary logs, and historical cases to classify windows or explain faults.Instructions compare expected behavior, separate evidence sources, check contradictions, identify abnormal intervals or causes, and avoid overclaiming.
  • B.3 Anomaly Detection and Diagnosis: Their structured outputs contain a decision, abnormal interval, supporting evidence, root-cause hypothesis, and confidence, reflecting several surveyed monitoring systems.The example names SAGE, ARGOS, AgentFM, and LLM-TSFD.
  • B.4 Decision Support: Decision-support agents recommend actions from current states, recent history, external evidence, allowed action spaces, constraints, and validation feedback.They compare feasible actions, reject unsafe or dominated choices, state risks, and identify what to monitor next.
  • B.4 Decision Support: Their outputs summarize the state, recommend an action, justify it, check constraints, and specify the next monitoring target across trading, traffic, and grid settings.Inputs may include prices, flows, loads, events, violations, retrieved cases, news, indicators, and forecasts.

D Failure Modes and Design Patterns

The survey consolidates documented failure modes with related design choices and implementation checks, while its resource summary organizes benchmarks and environments by practical characteristics.

  • D Failure Modes and Design Patterns: Table 2 links each documented failure mode to related work, design choices, and implementation checks as a practical inspection checklist.The table is not presented as a causal comparison.
  • D Failure Modes and Design Patterns: Table 5 organizes resources by category, problem type, interactivity, temporal scale, number of series, and latest release when available.The resources include benchmarks, datasets, environments, and toolkits.

F Evaluation Evidence and Metrics

The survey compares evaluation evidence across time-series tasks while emphasizing metric-specific interpretation, shared task settings, and dependence on experimental conditions.

  • Comparable Quantitative Evidence: Comparable evidence is grouped only when task, dataset, and metric match, but scores may still depend on horizons, splits, adjustment rules, costs, or simulator configurations.The table covers forecasting, reasoning, anomaly detection, trading, and traffic control.
  • Error Metrics: Forecasting and regression tasks commonly report MAE, MSE, and RMSE, with lower values indicating better forecasts and scale-free metrics used across differing magnitudes.The formulas evaluate point-wise prediction errors over n evaluated points.
  • Classification and Detection Metrics: Reasoning and anomaly-detection tasks commonly use accuracy, precision, recall, and F1, while score-based detectors may additionally report AU-ROC or AUPR.Anomaly-detection F1 comparisons require the same point-wise, point-adjusted, or event-based rule.
  • Generation and Annotation Metrics: Augmentation and synthesis are often evaluated indirectly through downstream MAE, MSE, accuracy, or F1, alongside distributional and annotation-quality measures.Additional measures include DTW, MMD, autocorrelation, spectral statistics, nearest-neighbor analyses, expert agreement, and label accuracy.
  • Decision and Control Metrics: Trading evaluation combines cumulative return and Sharpe ratio with maximum drawdown, volatility, and turnover to capture return, risk, and trading-cost sensitivity.StockBench stresses multi-month sequential evaluation using daily market signals.
  • Decision and Control Metrics: Traffic-control agents are usually evaluated in closed-loop simulators, where average travel time measures mean vehicle duration and lower values indicate more efficient flow.Related measures include waiting time, queue length, throughput, cumulative reward, and constraint violations.
  • Surveyed Systems: Table 3 groups representative agentic methods by problem type and records publication information, architecture, tools, and memory, while Table 4 counts these choices within families.Table 4's tool and memory categories are non-exclusive across the 47 summarized systems.
  • Surveyed Systems: Figure 4 provides a supplementary visual summary of recurring design patterns across the surveyed task families rather than a universal performance comparison.Its panels cover forecasting, reasoning, augmentation, synthesis, anomaly detection, diagnosis, trading, and traffic and grid control.
Loading 2608.26226v1…