PaperlayerView exact revision on arXiv

Source-linked AI summary

TimeART: Towards Agentic Time Series Reasoning via Tool-Augmentation

Xingjian Wu, Junkai Lu, Zhengyu Li, Xiangfei Qiu, Jilin Hu, Chenjuan Guo, Christian S. Jensen, Bin Yang

arXiv:2601.13653v1cs.LG

TL;DR

Time series analysis often relies on human data scientists, limiting automation. TimeART combines LLM reasoning with 21 analytical tools, TimeToolBench’s 100k expert trajectories, and a four-stage training strategy; an 8B TSRM achieves state-of-the-art performance across multiple TSQA benchmarks.

  • Problem

    Existing time series reasoning models still struggle to replace data scientists, while conventional workflows rely heavily on expert knowledge and lack automation.

  • Method

    TimeART integrates 21 analytical tools with TSRMs, while TimeToolBench provides 100k tool-use trajectories and a four-stage strategy trains models on expert and early self-reflection experience.

  • Results

    An 8B TSRM equipped with TimeART achieves state-of-the-art performance on multiple TSQA benchmarks, with MAE reductions of 9%–68% across all reported settings.

  • Takeaways & Limitations

    TimeART demonstrates a fully agentic approach to time series reasoning through strategic tool-use and competitive performance with much fewer parameters than closed-source models.

Abstract

from arXiv · show

Time series data widely exist in real-world cyber-physical systems. Though analyzing and interpreting them contributes to significant values, e.g, disaster prediction and financial risk control, current workflows mainly rely on human data scientists, which requires significant labor costs and lacks automation. To tackle this, we introduce TimeART, a framework fusing the analytical capability of strong out-of-the-box tools and the reasoning capability of Large Language Models (LLMs), which serves as a fully agentic data scientist for Time Series Question Answering (TSQA). To teach the LLM-based Time Series Reasoning Models (TSRMs) strategic tool-use, we also collect a 100k expert trajectory corpus called TimeToolBench. To enhance TSRMs' generalization capability, we then devise a four-stage training strategy, which boosts TSRMs through learning from their own early experiences and self-reflections. Experimentally, we train an 8B TSRM on TimeToolBench and equip it with the TimeART framework, and it achieves consistent state-of-the-art performance on multiple TSQA tasks, which pioneers a novel approach towards agentic time series reasoning.

1. Introduction

Time series reasoning models struggle with numerical reliability, complex inputs, and strategic tool use. TimeART addresses these limitations through tool augmentation, expert trajectories, and a staged training strategy, with an 8B model achieving state-of-the-art performance on multiple TSQA benchmarks.

  • Limitations of existing TSRMs: TSRMs struggle with numerical hallucination, cognitive defects on long or complex inputs, and limited reliability in forecasting and anomaly detection.Discrete tokenization makes structured numerical processing difficult, while flexible questions and long time series further challenge reasoning.
  • TimeART framework: TimeART integrates 21 analytical tools, enabling TSRMs to invoke statistical methods and lightweight time series foundation models during reasoning.The framework is designed to let models use tools autonomously, robustly, and efficiently.
  • Training resources: TimeToolBench contains 100k high-quality tool-use trajectories for strengthening TSRMs' strategic tool invocation.The corpus contains ReAct-style trajectories from multiple real-world domains and supports fine-tuning open-source LLMs.
  • Training challenges: Conventional tool-use training can cause low generalization through imitation and entropy collapse under sparse outcome-only rewards.Behaviour cloning may encourage copying expert trajectories, while reinforcement learning provides limited process feedback for understanding the environment.
  • Training strategy: The proposed training strategy decomposes agentic reinforcement learning principles into stages that use early exploration experience and self-reflection.It aims to improve perception of the external environment and action space while preserving generalization capability.
  • Empirical result: An 8B TSRM trained on TimeToolBench and equipped with TimeART achieves state-of-the-art performance on multiple TSQA benchmarks.The reported result evaluates the combined framework and training approach across multiple benchmarks.

2. Methodology

TimeART combines 21 analytical tools with ReAct-style reasoning, a 100k tool-use corpus, and a four-stage training strategy to teach TSRMs strategic and generalizable tool use.

  • TimeART framework: TimeART organizes agentic reasoning as repeated Thought, Action, and Observation states between a Query and Final Answer.Actions select and parameterize one of 21 analytical tools.
  • TimeART framework: The framework targets numerical and long-series reasoning by providing atomic tools for statistical computation, correlations, anomaly detection, and forecasting.The tools compensate for TSRMs’ shortcomings in directly processing numerical values.
  • TimeToolBench: TimeToolBench contains over 100k expert tool-use trajectories generated from TSQA samples and checked for answer consistency and semantic alignment.Fixed-option answers must match the reference, while open-ended answers must exceed a BERT-Score threshold σ.
  • TimeToolBench: Trajectory quality is further screened by sampling logical chains and retaining trajectories only when all LLM judges approve them.This combines diversity from different judges and evaluated subsequences with strict unanimous filtering.
  • Training strategy: The four-stage strategy first teaches tool capability boundaries, then strategic use, and finally self-reflection explanations jointly with expert actions.Early-experience training models tool transition rules and improper invocations; self-reflection explains why an expert action is preferable to alternatives.
  • Training strategy: Self-reflection training helps TSRMs move beyond imitation toward strategic tool use and more generalizable decision criteria.The model summarizes why an expert action is preferable based on differences between resulting observations, then predicts the explanation and action together.

3. Experiments

TimeART is evaluated on forecasting and reasoning benchmarks, with ablations testing tool use and the proposed four-stage training strategy. The framework achieves strong forecasting and reasoning results, while the ablations support the value of strategic tool use, TimeToolBench, and self-reflection.

  • Experimental Settings: TimeART is evaluated on MTBench and TimeMQA using forecasting errors and task accuracy rates, against closed-source and open-source baselines.Forecasting uses MSE, MAE, and MAPE; reasoning uses accuracy, including 3-way and 5-way settings.
  • Main Results: 9%–68% reduction in MAE is reported across all forecasting settings compared with baselines.The reported reductions cover Stock Price Forecasting and Temperature Forecasting.
  • Main Results: TimeART outperforms all open-source baselines on each reported reasoning task and remains competitive with GPT-4o and Qwen3-max using fewer parameters.The reasoning evaluation covers MTBench and TimeMQA tasks involving temporal correlations, domain texts, understanding, perception, reasoning, and estimation.
  • Ablations on the TimeART framework: Tool use improves both Qwen3 8B and Qwen3-max, with larger gains for Qwen3-max across all four TimeMQA task categories.The ablation compares original and trained Qwen3 8B, Qwen3-max, and their tool-use variants.
  • Ablations on the TimeART framework: A finetuned Qwen3 8B acquires strategic tool-use ability and exceeds closed-source Qwen3-max with TimeART.This result is reported in the framework ablation on four TimeMQA reasoning categories.
  • Ablations on training stages: Stage 1 consistently improves all TSQA tasks, while Stages 3 and 4 make the complete four-stage strategy consistently best across them.Stage 1 teaches tool capability boundaries, and Stages 3 and 4 use self-reflection rather than fixed manually designed preferences.

4. Related Works

Prior work applies LLMs to conventional time-series analysis and TSQA, but largely overlooks LLMs’ numerical weaknesses. TimeART addresses this by combining analytical tools with LLM reasoning for agentic time-series analysis.

  • Earlier studies transfer LLM generalization to forecasting and anomaly detection when large-scale time-series corpora are scarce.
  • Recent TSQA systems target cross-modality and interpretability across domains including finance, weather, and aircraft-engine analysis.
  • TSQA benchmarks such as MTBench, TimeSeriesExam, and TimeSeriesGym evaluate Time Series Reasoning Models.
  • Most TSQA efforts improve LLM reasoning while neglecting numerical weaknesses that limit time-series handling.
  • TimeART combines analytical tools for numerical processing with LLM reasoning to support agentic time-series reasoning.
  • Time-series agents also pursue automation in representation learning, classification, and forecasting AutoML, although conventional methods already solve these tasks well.

5. Conclusion

The paper concludes by presenting TimeART, TimeToolBench, and a four-stage training strategy for agentic time-series reasoning. An 8B TSRM equipped with TimeART achieves state-of-the-art performance on multiple benchmarks.

  • TimeART integrates multiple out-of-the-box analytical tools, spanning statistical functions and time-series foundation models.
  • TimeToolBench provides 100k expert trajectories, while the four-stage strategy trains TSRMs on expert data and their own early experience.
  • An 8B TSRM equipped with TimeART achieves state-of-the-art performance on multiple benchmarks.
  • Future work will address efficient sensing of time series with numerous variables and excessive lengths.

A. Implementation details

The implementation uses LangChain for robust tool calling and supervised fine-tuning for all training stages. Training uses LLaMA-Factory with DeepSpeed ZeRO-3, BF16, FlashAttention-2, and LoRA.

  • TimeART is implemented with LangChain, and APIs for all 21 tools are documented in Appendix D.
  • Because the training strategy does not use reinforcement learning, every stage is organized as supervised fine-tuning.
  • LLaMA-Factory and DeepSpeed ZeRO-3 support the training stages and efficient model training.
  • Fine-tuning uses BF16 precision, FlashAttention-2, and LoRA; the training curve appears in Figure 7.

Training Data

TimeToolBench is organized in ReAct style, making the reasoning structure explicit through ordered Thought, Action, and Observation steps.

  • TimeToolBench raw data follows a ReAct-style format with a clear reasoning structure.
  • Each trajectory follows the established Thought-Action-Observation order.

A ReAct-style Example

The example organizes time-series reasoning as repeated thought, tool-action, and observation steps before a final answer. It also presents the same trajectory in a multi-turn conversation format for training.

  • Tool-use sequence: The model separates trend and seasonality detection into sequential tool calls.It first invokes trend_classifier, then seasonality_detector, using each observation to guide the next step.
  • Training constraint: The training format is intended to preserve context logic but is not suitable for supervised fine-tuning in a question-answering paradigm.This limitation is stated during discussion of the multi-turn organization.
  • Conversation format: The same reasoning sequence is represented as alternating human, GPT, function, and observation messages.The format places human and observation messages in odd positions, while GPT and function messages occupy even positions.
  • Outcome: The example ends with a final answer after the seasonality tool reports an estimated period of 24.The final answer placeholder follows the trend and seasonality observations.

Benchmarks

The paper evaluates time-series reasoning across multimodal and multitask benchmarks covering diverse domains and analytical tasks. It selects benchmark subsets to maximize task diversity while avoiding overlap with TimeToolBench.

  • MTBench: MTBench evaluates reasoning across paired time-series and textual data in finance and weather domains.Its tasks include forecasting, technical indicator prediction, semantic trend analysis, and news-driven question answering.
  • Time-MQA: Time-MQA combines time-series forecasting, imputation, anomaly detection, classification, and open-ended reasoning in one question-answering framework.Its TSQA dataset contains approximately 200,000 question-answer pairs across 12 domains.
  • Evaluation design: The evaluation uses most MTBench tasks and a categorized subset of TimeMQA.The subset is chosen to avoid overlap with TimeToolBench and cover as many task types as possible.
  • Application scope: Time-series analysis spans fields including economy, transportation, health, weather, and energy, with forecasting among its critical tasks.The related-work discussion situates the benchmarks within this broader application scope.
  • Tool support: TimeART adopts time-series foundation models as out-of-the-box tools for zero-shot forecasting, anomaly detection, and classification.These tools are used to provide numerical capabilities within the framework.

D. Details of Tools

TimeART provides tools for metadata retrieval, statistical analysis, pattern detection, cross-channel relationships, anomaly detection, and forecasting. Its training materials teach stepwise tool selection and ReAct-style reasoning with explicit intermediate states.

  • Basic and statistical tools: The tool suite retrieves metadata, datapoint values, summary statistics, returns, autocorrelation, rolling statistics, quantiles, and volatility.These operations support inspection of sequence structure, values, ranges, and local variability.
  • Pattern analysis: Pattern-analysis tools classify trends, detect seasonality, identify change points, profile noise, and test stationarity.Trend analysis supports global or window-based segments, while seasonality returns an estimated period and strength.
  • Cross-channel analysis: Relationship tools measure channel correlation, cross-correlation, dynamic time-warping distance, shape similarity, and Granger causality.They characterize alignment, similarity, association, or statistical predictiveness between channels.
  • Advanced tools: The anomaly_detection and forecaster tools use DADA and LightGTS for zero-shot multivariate analysis.DADA detects anomalies from reconstruction error, while LightGTS returns predicted values for all channels.
  • Self-reflection: The training prompts represent situations with expert actions, alternative actions, and expected outcomes for self-reflection.The reflection process compares possible actions, justifies the expert action, and considers constraints or consequences.
  • Reasoning behavior: The reasoner is trained to reconstruct intermediate steps, use tools without relying on them completely, and produce structured ReAct trajectories.At least one tool call is required in one training instruction, with additional calls encouraged.

F. Case Study

The case studies demonstrate tool-augmented reasoning for forecasting, trend analysis, pattern recognition, volatility evaluation, and moving averages. They also show the model revising an overly coarse tool output through segmented analysis and direct inspection.

  • Moving average: The 3-point moving averages are approximately 4874.43, 4875.79, 4878.00, and 4879.64.The sequence of averages is interpreted as an upward trend.
  • Forecasting: The forecasting example compares the last observed day with the first predicted day using summary statistics.The reported means are 19.3875 for the last observed day and 22.871 for the first predicted day, producing an increasing answer.
  • Trend prediction: A trend-prediction example uses windowed analysis after a global flat label conflicts with the visible decline.The windowed result identifies a down segment followed by stabilization, supporting a flat or very gradual decline forecast.
  • Pattern recognition: Pattern recognition selects the statement that the data mostly decreases with some increase.The reasoning supplements global classification with visual inspection and change-point analysis.
  • Volatility evaluation: Rolling volatility is approximately 0.0049 across all segments with window size 5.The case study interprets this consistent value as low volatility and minimal variability.
Loading 2601.13653v1…