Source-linked AI summary
EvoTS-Agent: A Self-Evolving LLM Agent for Financial Time Series Change Point Detection
Lei Jiang, Ye Wei, Xinyu Xi, Jordan Langham-Lopez, Yifan Bao, Raad Khraishi, Yihao Ang, Anthony K. H. Tung, Lukasz Szpruch, Hao Ni
TL;DR
Financial change-point detection is difficult because financial series vary across assets and regimes, while conventional pipelines require extensive manual choices. EvoTS-Agent uses validation-guided trajectory evolution to select and refine detection pipelines, achieving strong performance across four benchmark datasets and 100% execution success across backbone LLMs.
Problem
Financial change-point detection lacks a consistently effective algorithm across noisy, non-stationary assets and market regimes, making pipeline selection dependent on interdependent manual decisions.
Method
EvoTS-Agent combines curated exploratory analysis, diverse detector selection, executable experimentation, and validation-guided evolution of experiment trajectories.
Results
Across four benchmark datasets, EvoTS-Agent consistently achieves strong change-point detection performance across multiple backbone LLMs while maintaining a 100% execution success rate.
Takeaways & Limitations
The results support adapting and evolving detection pipelines to each dataset’s statistical characteristics rather than relying on a fixed workflow.
Takeaways & Limitations
EvoTS-Agent depends on labeled validation data to retain experimental trajectories during evolution, which may require expert annotation or historical ground-truth events.
Abstract
from arXiv · showhide
Financial time series exhibit non-stationary and heterogeneous statistical properties, making change-point detection challenging because no single unsupervised algorithm performs consistently across assets and market regimes. Conventional workflows consequently depend heavily on expert-driven model selection, feature design, and hyperparameter tuning, limiting their scalability and adaptability. We propose EvoTS-Agent, a validation-guided self-evolving LLM agent for autonomous financial time-series change-point detection. EvoTS-Agent first performs curated exploratory data analysis to characterize dataset properties and initialize candidate detection models. It then evolves executable experiment trajectories through three complementary operators: \textit{Revision} exploits the current best solution, \textit{Alternative Strategy} explores fundamentally different modeling directions when progress stagnates, and \textit{Recombination} synthesizes complementary evidence from high-performing trajectories. Validation feedback guides trajectory evolution throughout the search, enabling the agent to adapt its detection pipeline to the statistical characteristics of each dataset while preserving reliable optimization. Experiments across four benchmark datasets demonstrate that EvoTS-Agent consistently outperforms existing LLM-based agents while maintaining a 100\% execution success rate across all evaluated backbone LLMs.
1 Introduction
Financial change-point detection is difficult because financial time series are heterogeneous, non-stationary, and regime-dependent, while effective pipelines require interdependent expert decisions. EvoTS-Agent addresses this challenge with curated EDA, diverse candidate detectors, executable experimentation, and validation-guided self-evolution of experiment trajectories.
- Motivation: Financial time series are noisy, non-stationary, heavy-tailed, and variable across assets, sampling frequencies, and market regimes, so no single detector performs consistently well.
- Motivation: Effective detection pipelines require interdependent choices about algorithms, representations, preprocessing, hyperparameters, thresholds, and post-processing rules.
- Motivation: Expert-driven experimentation and parameter tuning are difficult to scale across assets and datasets and may adapt inefficiently when market statistics change.
- EvoTS-Agent: EvoTS-Agent combines curated exploratory data analysis, model selection, executable experimentation, and validation-guided trajectory evolution in a closed-loop framework.Its EDA characterizes dataset-specific properties and initializes diverse candidate change-point detectors.
- EvoTS-Agent: Revision refines the incumbent, Alternative Strategy explores different modeling directions after stagnation, and Recombination synthesizes evidence from high-performing trajectories.An incumbent-preserving selection rule prevents unsuccessful experiments from degrading the best validated solution; self-evolution transforms experiment trajectories at inference time rather than updating the language model.
2 Related Work
Related work spans assumption-dependent change-point detection methods and increasingly capable LLM-based data-science agents. EvoTS-Agent is positioned relative to these approaches through validated executable experiment trajectories that record modeling decisions, feedback, performance, and lineage.
- Change-Point Detection: Change-point detection methods differ in their assumptions about what remains stable within a regime, shaping which structural changes and application scenarios they can address.Classical statistical methods assume constant distributional parameters within segments.
- Change-Point Detection: Deep learning methods seek regime changes in learned latent representations, while spectral methods detect changes in frequency-domain behavior.KL-CPD is cited as an example of representation learning for subtle structural changes.
- LLM-Based Agents: LLM-based agents automate data-science workflows through planning, code generation, tool use, and execution feedback.ReAct interleaves reasoning and acting, while DS-Agent and ResearchAgent extend the paradigm through solution adaptation and iterative research planning.
- LLM-Based Agents: SE-Agent evolves agent trajectories through revision, recombination, and refinement, primarily for software-engineering problem solving.EvoTS-Agent instead maintains validated executable experiment trajectories spanning plans, model and transformation choices, implementation changes, execution feedback, validation performance, and lineage.
3 Preliminaries and Problem Setup
EvoTS-Agent formulates financial change-point detection as validation-guided empirical optimization, where an LLM agent iteratively proposes, executes, and evaluates pipelines to approximate reference structural-change boundaries. The framework represents experimentation through explicit states, actions, executable trajectories, and lineage.
- Problem formulation: The objective is to identify structural-change locations that approximate the reference boundary set B in a financial time series.N denotes sequence length and d the number of observed variables.
- Problem formulation: Validation feedback guides model selection and trajectory evolution, while reference boundaries are available only on validation data and test annotations remain hidden.Although detectors are unsupervised, test-set metrics are withheld throughout optimization.
- Agent-environment formulation: The experimental state contains dataset profiles, candidate models, executable scripts, validation observations, trajectory pools, and the current incumbent.Available actions include exploration, model selection, planning, script modification, execution, validation, revision, alternative-strategy generation, and recombination.
- Evaluation: Executed experiments receive validation scores, with boundary-aware validation F1 as the primary change-point metric and Hausdorff distance retained diagnostically.Incumbent selection is determined by the primary scalar score q_k.
- Experiment trajectories: Each executable trajectory records its plan, implemented model and script, validation score, execution evidence, code difference, selection decision, and stagnation status.Parent trajectories establish explicit lineage, making the trajectory pool a directed experimental search graph rather than an unstructured conversational history.
4 EvoTS-Agent
EvoTS-Agent uses leakage-safe exploratory analysis and LLM-guided model selection to initialize executable changepoint-detection trajectories, then evolves them through validation-guided Revision, Alternative Strategy, and Recombination operators. Incumbent updates require valid validation scores that strictly improve over the current solution.
- Warm-up: Warm-up experiments implement and revise each selected model, generating 2 × K validated trajectories stored in experiment trajectory memory.The highest-validation-score trajectory becomes the incumbent for optimization.
- EDA and model selection: Leakage-safe EDA summarizes temporal, spectral, change-sensitive, and dataset properties while excluding ground-truth change points and test metrics.Features include lag-1 autocorrelation, trend strength, nonstationarity, spectral concentration, periodicity, local mean and variance discrepancies, sequence length, dimensionality, and missing-value ratio.
- EDA and model selection: The LLM selector receives meta-features, an unlabeled visualization, task metadata, and candidate descriptions, then chooses K primary models and up to two alternatives.Primary models initialize independent warm-up trajectories, while alternatives can be introduced when validation performance stagnates.
- Trajectory evolution: Revision modifies representation, detector configuration, hyperparameters, or post-processing using the incumbent and recent successful trajectories, while preserving the experimental objective.If improvement falls below a predefined threshold or no valid score is produced, the revision is marked stagnant and Alternative Strategy explores orthogonal or untried directions.
- Trajectory evolution: During the final optimization iteration, Recombination synthesizes complementary components from multiple high-performing trajectories, while each experiment is accepted only when its valid validation score strictly improves the incumbent.Small positive improvements are accepted but can still trigger Alternative Strategy when they do not exceed the meaningful-improvement threshold.
5 Experiment
EvoTS-Agent is evaluated on four complementary change-point benchmarks against representative LLM-based agents and consistently delivers strong detection performance with reliable execution. Results and ablations show that adaptive detector selection and trajectory evolution improve accuracy, localization, and search effectiveness.
- Datasets: The evaluation covers four benchmarks spanning parameterized regime changes, compound distributional shifts, and noisy real-world behavioral signals.The datasets include Piecewise Ornstein Uhlenbeck, Mean-Variance Shift, ADIA, and Bee Dance, supporting assessment of robustness and generalization.
- Evaluation Setup: EvoTS-Agent is compared with TS-Agent, DS-Agent, and ResearchAgent using GPT-4o, GPT-5.4, Claude Sonnet-4.6, and Sonnet-5.Detection quality uses F1, precision, recall, and Hausdorff distance, with F1 computed using boundary-aware matching within a ±10-sample tolerance window.
- Overall Results: 100% execution success rate is maintained across all backbone models, while EvoTS-Agent achieves the strongest or highly competitive performance across all four benchmarks.ResearchAgent’s success rates fall to 33.3% for both Sonnet-4.6 and Sonnet-5 on the OU benchmark, whereas EvoTS-Agent completes every run successfully.
- Benchmark Results: 0.833 F1 is achieved by EvoTS-Agent under GPT-5.4 on Mean-Variance Shift, improving over 0.568 for TS-Agent and 0.562 for DS-Agent while also lowering Hausdorff distance.On ADIA under GPT-4o, Hausdorff distance decreases to 14.8 from 212.0 for TS-Agent and 130.6 for DS-Agent, while F1 rises to 0.500 from 0.235 and 0.461.
- Ablation and Analysis: 0.635 F1 is obtained on Bee Dance with GPT-5.4, compared with 0.578 without Alternative Strategy and 0.575 without Recombination.Manually configured detectors vary substantially, while evolutionary selection and refinement modify representation, detector configuration, hyperparameters, and post-processing; Recombination produces a validation F1 of 0.8631 in the final iteration.
6 Conclusion & Limitation
EvoTS-Agent uses validation-guided trajectory evolution for financial time-series change-point detection, achieving strong performance across four benchmark datasets and multiple backbone LLMs with 100% execution success. Its main limitation is dependence on labeled validation data, motivating label-efficient or label-free optimization.
- Conclusion: EvoTS-Agent performs validation-guided trajectory evolution for financial time-series change-point detection rather than repeatedly refining one solution or following a fixed workflow.It maintains executable experiment trajectories and adapts its search strategy.
- Conclusion: 100% execution success rate was maintained across multiple backbone LLMs while EvoTS-Agent achieved strong detection performance on four benchmark datasets.The gains arise from both diverse detection algorithms and adaptive selection and evolution of pipelines suited to each dataset’s statistical characteristics.
- Limitation: EvoTS-Agent depends on labeled validation data to retain experimental trajectories during evolution, despite using unsupervised change-point detectors.Obtaining validation annotations may require expert annotation or historical ground-truth events.
- Limitation: Developing reliable label-efficient or label-free optimization strategies is an important direction for future research.This need follows from the practical difficulty of obtaining validation labels.