PaperlayerView exact revision on arXiv

Source-linked AI summary

TimeSeriesScientist: A General-Purpose AI Agent for Time Series Analysis

Haokun Zhao, Xiang Zhang, Jiaqi Wei, Yiwei Xu, Yuting He, Siqi Sun, Chenyu You

arXiv:2510.01538v2cs.LG

TL;DR

Time series forecasting often requires substantial manual work to preprocess, validate, and ensemble predictions for heterogeneous short and noisy series. TSci automates this end-to-end workflow with four specialized agents that use tool-augmented LLM reasoning, multimodal diagnostics, validation, adaptive ensembling, and transparent reporting. Across eight benchmarks, it outperforms statistical and LLM-based baselines while producing rigorous, interpretable reports.

  • Problem

    Heterogeneous, short, noisy time series make preprocessing, validation, and ensembling labor-intensive, while existing forecasting systems remain too model-centric for general end-to-end automation.

    Source: §1 · pp. 1–2

  • Method

    TSci orchestrates Curator, Planner, Forecaster, and Reporter agents to diagnose data, select and validate models, adaptively ensemble forecasts, and synthesize transparent reports.

    Source: §1 · p. 2

  • Results

    TSci reduces forecasting error by 10.4% versus statistical baselines and 38.3% versus LLM-driven baselines across eight public benchmarks spanning five domains.

    Source: §1 · p. 2

  • Takeaways & Limitations

    TSci provides a unified, domain-agnostic approach that combines forecasting accuracy with interpretable and auditable workflow reporting.

    Source: §1 · p. 2, §5 · p. 10

Abstract

from arXiv · show

Time series forecasting is central to decision-making in domains as diverse as energy, finance, climate, and public health. In practice, forecasters face thousands of short, noisy series that vary in frequency, quality, and horizon, where the dominant cost lies not in model fitting, but in the labor-intensive preprocessing, validation, and ensembling required to obtain reliable predictions. Prevailing statistical and deep learning models are tailored to specific datasets or domains and generalize poorly. A general, domain-agnostic framework that minimizes human intervention is urgently in demand. In this paper, we introduce TimeSeriesScientist (TSci), the first LLM-driven agentic framework for general time series forecasting. The framework comprises four specialized agents: Curator performs LLM-guided diagnostics augmented by external tools that reason over data statistics to choose targeted preprocessing; Planner narrows the hypothesis space of model choice by leveraging multi-modal diagnostics and self-planning over the input; Forecaster performs model fitting and validation and, based on the results, adaptively selects the best model configuration as well as ensemble strategy to make final predictions; and Reporter synthesizes the whole process into a comprehensive, transparent report. With transparent natural-language rationales and comprehensive reports, TSci transforms the forecasting workflow into a white-box system that is both interpretable and extensible across tasks. Empirical results on eight established benchmarks demonstrate that TSci consistently outperforms both statistical and LLM-based baselines, reducing forecast error by an average of 10.4% and 38.2%, respectively. Moreover, TSci produces a clear and rigorous report that makes the forecasting workflow more transparent and interpretable.

1 Introduction

Time series forecasting remains labor-intensive because heterogeneous, short, noisy series require substantial manual preprocessing, validation, and ensembling. TSci addresses this gap with a four-agent, tool-augmented workflow and reports consistent benchmark gains over statistical and LLM-based baselines.

1 Introduction

  • Forecasting organizations manage short, noisy series with heterogeneous sampling, missing values, and shifting horizons, making preprocessing and evaluation pipelines a dominant practical cost.End-to-end pipelines still require substantial human effort to tailor preprocessing, validation, and ensembling for each collection of series.Source: §1 · p. 1
  • Existing expert, universal, AutoML, and LLM forecasting systems remain model-centric, with limited reasoning about data quality, temporal structure, tool adaptation, and natural-language justification.These limitations motivate treating forecasting as sequential decisions over preparation, model selection, validation, and ensembling.Source: §1 · pp. 1–2
  • TSci automates the full univariate forecasting workflow through Curator, Planner, Forecaster, and Reporter agents that diagnose data, select and validate models, ensemble predictions, and synthesize reports.The framework uses tool-augmented LLM reasoning, multimodal diagnostics, and natural-language rationales to make the pipeline adaptive, interpretable, and extensible.Source: §1 · p. 2
  • TSci reduces forecasting error by 10.4% versus statistical baselines and 38.3% versus LLM-driven baselines across eight public benchmarks spanning five domains.The reported evaluation also finds that each module contributes materially to performance.Source: §1 · p. 2
  • TSci also evaluates generated reports for technical rigor and communication quality, supporting use in settings that demand transparency and auditability.The report evaluation is presented as a distinct contribution alongside predictive benchmarking.Source: §1 · p. 2

2 Related work

Forecasting research has progressed from classical and deep models to foundation-style approaches, but most systems remain centered on modeling rather than the full workflow. TSci is positioned as a multi-agent, domain-agnostic framework that combines reasoning, tool use, and automation for interpretable forecasting.

2 Related work

  • Classical methods exploit linear trends and seasonalities, while global deep learning models capture nonlinear patterns and long-term dependencies.The literature therefore spans statistical, deep learning, and newer foundation-style approaches.Source: §2 · pp. 2–3
  • TSci uses a collaborative four-agent workflow in which Curator diagnoses data, Planner narrows model configurations, Forecaster produces predictions, and Reporter generates the final report.The framework is presented as an automated analogue of a human scientist’s forecasting process.Source: Figure 2 · p. 3
  • Foundation-style and prompt-based forecasters broaden zero-shot and few-shot capabilities, yet preprocessing, evaluation design, and ensemble synthesis still rely heavily on manual effort.This gap motivates an end-to-end LLM-powered agentic framework spanning the forecasting workflow.Source: §2 · p. 3
  • Multi-agent frameworks coordinate specialized agents through communication and tool use, but existing systems rarely address heterogeneous sampling, irregular multimodal data, and transparent forecast ensembling together.This leaves an opportunity for a domain-agnostic forecasting system emphasizing interpretability and auditability.Source: §2 · p. 3
  • Curator transforms raw data into a cleaned dataset, generates tailored visualizations, and extracts trends, seasonality, and stationarity for downstream analysis.Its workflow is designed to expose temporal structure while supporting interpretability.Source: Figure 3 · p. 4

3 TimeSeriesScientist

TSci is an end-to-end agentic framework that automates univariate forecasting through four specialized agents, from data diagnostics and model selection to ensembling and transparent reporting. Its pipeline combines data-aware preprocessing, validation-driven planning, adaptive forecasting, and workflow documentation.

3 TimeSeriesScientist

  • TSci coordinates Curator, Planner, Forecaster, and Reporter to automate diagnostics, model selection, forecasting, ensembling, and report generation.Each agent has a distinct role within the end-to-end workflow.Source: §3 · p. 4

3.1 Problem Formulation

  • The forecasting objective estimates an H-step horizon from T historical values while minimizing mean average error between predictions and ground truth.Source: §3.1 · p. 4

3.2 Curator

  • Curator diagnoses statistics, missing values, and outliers, applies targeted preprocessing, generates visualizations, and profiles trend, seasonality, and stationarity.Its visualization suite includes overview, decomposition, and autocorrelation plots to support downstream analysis.Source: §3.2 · pp. 4–5

3.3 Planner

  • Planner uses multimodal data patterns to form a candidate model pool, optimize hyperparameters on validation data, and rank tuned models by validation MAPE.The selected models, hyperparameters, and validation metrics become the foundation for ensemble construction.Source: §3.3 · pp. 5–6

3.4 Forecaster

  • Forecaster chooses among single-best, performance-aware averaging, and robust aggregation, fixes the strategy before testing, and evaluates the final forecast on the held-out horizon.This design aims to balance performance and stability while avoiding test-set leakage.Source: §3.4 · p. 6

3.5 Reporter

  • Reporter produces a report containing forecasts with confidence intervals, model and ensemble test metrics, interpretability explanations, visualizations, and full workflow documentation.The report records model rationales, ensemble-weight derivations, confidence, assumptions, limitations, and decisions made throughout the pipeline.Source: §3.5 · pp. 6–7, Figure 4 · p. 7
  • Natural-language rationales at each decision point make model selection, hyperparameter choices, and ensemble construction transparent to users.The framework presents automated optimization as an interpretable and extensible workflow.Source: §3.5 · p. 7

4 Experiment

Across diverse benchmarks, TSci is evaluated for forecasting accuracy, report quality, module contributions, and interpretability. The experiments show broad performance gains, rigorous reports, preprocessing as the largest ablation contributor, and stable ensemble behavior in a seasonal case study.

4.1 Performance Analysis

  • TSci outperforms LLM-based baselines across eight benchmarks, reducing MAE by an average of 38.2% versus the second-best baseline.Figure 5 visualizes comparisons across datasets and horizons, supporting the reported robustness across heterogeneous domains.Source: §4.1 · p. 7, Table 1 · p. 8, Figure 5 · p. 8
  • TSci generally outperforms statistical methods across ETT-small datasets and horizons, although simple linear models can match or slightly exceed it at short horizons.The aggregate advantage becomes clearer as forecast length increases or patterns depart from near-linear dynamics.Source: §4.1 · p. 7, Figure 6 · p. 9

4.2 Generated Report Evaluation

  • TSci reports outperform all baselines across five rubrics, with win rates above 80% for Analysis Soundness and Model Justification and mostly above 75% for Interpretive Coherence and Actionability Quotient.Structural Clarity shows smaller advantages, but the reports remain consistently structured and professional.Source: §4.2 · pp. 8–9, Table 2 · p. 9

4.3 Model Analysis

  • Removing preprocessing increases MAE by 41.80% on average, the largest degradation among the three modules, especially as prediction horizons increase.The results emphasize cleaning, resampling, and outlier handling for robustness and long-horizon forecasting.Source: §4.3 · p. 9, Figure 7 · p. 10
  • Removing data analysis harms MAE by 28.3% on average, indicating that series profiling supports downstream model choice and horizon-specific settings.Two minute-level cases improve slightly after removal, suggesting limited settings where preprocessing and search may need further tuning.Source: §4.3 · p. 9
  • Removing model optimization produces a 36.2% average MAE drop and marked deterioration on long horizons or high-variance series.The module searches parameters such as horizon chunking and window sizing for selected models.Source: §4.3 · p. 9

4.4 Case Study

  • On an ECL case with H = 96, TSci’s ensemble remains stable and follows seasonal cycles while individual ARIMA and Prophet forecasts accumulate errors.The report links model choices to seasonality and non-stationarity and identifies risks such as irregular spikes.Source: §4.4 · p. 10

5 Conclusions and Future Work

The paper presents TSci as an end-to-end agentic approach to univariate forecasting that combines LLM reasoning with prediction accuracy and report interpretability. It positions the framework as an initial step toward unified, domain-agnostic forecasting while identifying multimodal, external-knowledge, and efficiency extensions.

5 Conclusions and Future Work

  • TSci automates univariate time series forecasting through an end-to-end agentic framework based on LLM reasoning.The conclusion frames the system as a unified approach bridging traditional forecasting and foundation-model capabilities.Source: §5 · p. 10
  • Experiments across diverse benchmarks show consistent gains over state-of-the-art LLM baselines while preserving report interpretability.The reported contributions cover both prediction accuracy and the interpretability of generated reports.Source: §5 · p. 10
  • Future work includes multimodal settings, external knowledge, and efficiency-oriented designs to improve applicability, interpretability, and scalability.These directions are presented as extensions for broader use of the framework.Source: §5 · p. 10

A Data Processing Strategies

This section specifies leakage-safe data processing, validation-driven ensemble choices, and a 21-model forecasting library for heterogeneous univariate series.

A.1 Outlier Detection

  • Local rolling statistics support non-stationary data, while outliers can be detected with IQR, Z-score, or percentile rules.The toolkit estimates statistics on rolling windows; thresholds include α=1.5 for IQR, α∈[3,4] for online Z-scores, and examples such as 1%–99% percentiles.Source: A Data Processing Strategies · p. 15, §A.1 · pp. 15–16

A.2 Outlier Handling

  • Detected outliers and missing values are repaired with context-sensitive clipping, interpolation, filling, local replacement, or causal smoothing.The procedures include segment-aware interpolation, forward/backward fill, local mean or median replacement, semantic-zero filling, and small-window causal smoothing.Source: §A.2 · p. 16

B Ensemble Strategy

  • Ensembling selects a validation-best model, performance-weighted average, or robust median/trimmed mean without using test data.Weights depend on validation scores and preset hyperparameters; median and trimmed mean address substantial disagreement among candidate predictions.Source: B Ensemble Strategy · p. 17

C Model Library

  • The model library contains 21 forecasting models spanning five categories, including statistical, regression, tree-based, neural, and specialized time-series models.The detailed inventory is reported in Table 3.Source: C Model Library · p. 18, Table 3 · p. 18

D.1 Implementations

Experiments use GPT-4o, sliced benchmark data, multiple horizons, and MAE/MAPE evaluation across eight established datasets.

D.1 Implementations

  • OpenAI GPT-4o (OpenAI et al., 2024) is the default backbone model for the experiments.The experiments use slices rather than entire datasets because of a limited budget.Source: §D.1 · p. 18
  • 25 dataset slices use input length T=512 and prediction horizons H∈{96, 192, 336, 720}.Results are averaged across the 25 slices.Source: §D.1 · p. 18

D.2 Dataset Details

  • Eight established benchmarks evaluate univariate forecasting on four ETT datasets, Weather, Electricity, Exchange, and ILI.The datasets are drawn from Wu et al. (2023b).Source: §D.2 · p. 18

E.1 LLM Guided Data Visualizations

TSci generates multimodal visual diagnostics that expose raw behavior, rolling variation, decomposition components, and temporal dependencies before model selection.

E.1 LLM Guided Data Visualizations

  • The visualization pipeline combines overview, decomposition, and autocorrelation analyses to characterize each input series.These diagnostics support preprocessing, analysis, and forecasting-model selection.Source: §E.1 · p. 19
  • Overview plots display raw values plus rolling mean and standard deviation for inspecting global structure and variability.The raw series uses temporal indexing on the x-axis, while rolling statistics summarize local behavior.Source: §E.1 · p. 19, Figure 8 · p. 19
  • Additive decomposition separates the observed series into trend, seasonal, and residual components for structural diagnosis.The residual component helps assess decomposition adequacy and identify anomalies or structural breaks.Source: §E.1 · p. 19, Figure 9 · p. 20
  • ACF and PACF plots reveal lag dependence and guide differencing, autoregressive-order, and ARIMA model decisions.ACF measures lagged linear relationships overall, whereas PACF controls for intermediate lags.Source: §E.1 · pp. 19–20

E.2 Technical Implementation Details

The implementation automates reproducible diagnostic-plot generation, storage, logging, and integration with later analysis and reporting stages.

E.2 Technical Implementation Details

  • Plots use standardized Matplotlib and seaborn styling, configurable rolling windows, additive seasonal decomposition, and 40-lag ACF/PACF analysis.Outputs are saved as high-resolution 300-DPI PDFs with tight bounding boxes.Source: §E.2 · pp. 20–21
  • The visualization process dynamically adapts plot parameters through LLM-driven configuration while preserving reproducible presentation settings.The implementation specifies consistent colors, line widths, grids, and subplot layouts.Source: §E.2 · pp. 20–21

E.3 Output and Integration

  • Generated PDFs use configurable paths, automatic directory creation, and comprehensive logging for audit and debugging.The files are stored for subsequent analysis and reporting phases.Source: §E.3 · p. 21
  • Preprocessing visualizations provide the basis for understanding series characteristics before model selection and forecasting.Their role is explicitly connected to data-driven decisions in later workflow stages.Source: §E.3 · p. 21

F Full Experiment Results

Across eight datasets, TSci achieves superior performance across most datasets and forecasting horizons against LLM-based and statistical baselines. The evaluation includes full tabulated results, best-performance counts, and distribution visualizations across datasets and horizons.

F Full Experiment Results

  • TSci achieves superior performance across the majority of datasets and forecasting horizons against five LLM-based baselines.The full results are reported in Tables 5 and 6, with the 1st Count row recording test cases where each model performs best.Source: F Full Experiment Results · p. 21
  • Figures 12 and 13 show the distributions of MAE and MAPE across datasets and forecasting horizons.Source: F Full Experiment Results · p. 22
  • Tables 5 and 6 mark the best and second-best results using red and blue, respectively, with lower values indicating better performance.Source: Table 5 · p. 22
  • Figure 11 compares TSci with three statistical baselines across eight datasets.Source: F Full Experiment Results · p. 21, Figure 11 · p. 22

G.1 Analysis Summary

The analysis identifies upward movement, recurring but imperfectly regular seasonality, and non-stationarity in the ECL series, while highlighting structural breaks and data-quality concerns as forecasting challenges. The visualization section presents diagnostic and forecast views for comparing methods, horizons, and ensemble behavior.

G.1 Analysis Summary

  • The ECL series exhibits a moderate-to-strong upward trend, recurring but irregular seasonal patterns, and apparent non-stationarity.The trend is especially pronounced in the latter half, while the exact seasonal periodicity remains unclear.Source: §G.1 · p. 23
  • Non-stationarity may require differencing, detrending, or other transformations before applying models that assume stationarity.The analysis links changing mean and variance in rolling statistics to this implication.Source: §G.1 · p. 23

G.2 Visualization

  • Figure 12 organizes slice-level MAE by dataset and horizon, while Figure 13 presents the corresponding MAPE distributions on log10 scales where lower is better.Source: Figure 12 · p. 24, Figure 13 · p. 25

G.1 Analysis Summary

  • Potential structural breaks or regime shifts, especially around periods of rapid change, create an additional forecasting challenge.The report recommends addressing these changes when preparing the series for accurate forecasting.Source: §G.1 · p. 24
  • The analysis does not identify explicit data-quality issues, but missing values and outliers should still be addressed before modeling.Source: §G.1 · p. 24

G.2 Visualization

G.3 Comprehensive Report

This section presents the comprehensive report as a structured synthesis of dataset characteristics, model performance, limitations, recommendations, and ensemble selection. It also describes how forecasts, visualizations, and model tools support an explained ensemble strategy whose JSON decision produces the final forecast.

G.3 Comprehensive Report

  • The dataset showed a strong upward trend, weak seasonality, and non-stationarity, while potential structural breaks or regime shifts were noted near the series midpoint.These properties complicate modeling and motivate differencing or other transformations.Source: §G.3 · p. 26
  • The ensemble used weighted averaging of Exponential Smoothing, ARIMA, and Random Forest with weights of 35%, 40%, and 25%, respectively.Its reported errors were MSE 209,950.78, MAE 393.83, and MAPE 11.20%.Source: §G.3 · p. 26
  • Exponential Smoothing achieved the lowest reported errors, with MSE 192,879.83, MAE 377.15, and MAPE 10.73%, outperforming ARIMA and Random Forest.ARIMA reported MSE 205,582.59, MAE 390.10, and MAPE 11.09%; Random Forest reported MSE 242,977.49, MAE 423.16, and MAPE 12.02%.Source: §G.3 · p. 26
  • The report recommends further preprocessing, model refinement and expansion, and reevaluation of ensemble weights based on individual model strengths.Suggested model extensions include SARIMA and advanced machine learning techniques.Source: §G.3 · p. 26

H Prompts

  • The prompts operationalize a transparent forecasting workflow by requiring structured assessments of data quality, missing values, outliers, trends, seasonality, stationarity, visualization insights, and forecasting readiness.Preprocessing prompts select handling strategies and optional resampling, detrending, or feature engineering, while analysis prompts request structured JSON or Python-dict outputs.Source: H Prompts · pp. 27–28
  • Model-selection prompts choose three suitable candidates, propose tailored hyperparameter search spaces, justify choices, and consider model diversity for ensemble robustness.Separate analyst and ensemble-agent prompts support downstream model configuration and integration strategies for accurate and reliable forecasts.Source: H Prompts · pp. 32–33
  • The report reviews individual model forecasts and visualizations before selecting an integration strategy such as the best model, weighted average, trimmed mean, median, or custom weights.Available model tools include ARMA, LSTM, and RandomForest predictors.Source: H Prompts · p. 33
  • If weights are used, the report specifies each model's weight and justifies the choice using model diversity, agreement, and historical performance.The decision also assesses ensemble confidence and records risks or caveats.Source: H Prompts · p. 33
  • The ensemble decision is returned as raw structured JSON containing the integration strategy, applicable weights or selected model, reasoning, and confidence.This structured output is used as the final forecast for the time-series slice.Source: H Prompts · pp. 33–34

Related papers

Loading 2510.01538v2…