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

cs.AIAug 18, 2026

EvoTS-Agent: A Self-Evolving LLM Agent for Financial Time Series Change Point Detection

Authors: Lei Jiang, Ye Wei, Xinyu Xi, Jordan Langham-Lopez, Yifan Bao, Raad Khraishi, Yihao Ang, Anthony K. H. Tung,

Organizations: Alan Turing Institute · University of Oxford · National University of Singapore · NatWest AI Research ·

Abstract

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: Revision exploits the current best solution, Alternative Strategy explores fundamentally different modeling directions when progress stagnates, and 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.

Figures & tables

Dataset	Model	F1 $\uparrow$				Hausdorff Distance $\downarrow$				Precision $\uparrow$				Recall $\uparrow$				Success Rate (%) $\uparrow$
Dataset	Model	GPT-4o	GPT-5.4	Sonnet-4.6	Sonnet-5	GPT-4o	GPT-5.4	Sonnet-4.6	Sonnet-5	GPT-4o	GPT-5.4	Sonnet-4.6	Sonnet-5	GPT-4o	GPT-5.4	Sonnet-4.6	Sonnet-5	GPT-4o	GPT-5.4	Sonnet-4.6	Sonnet-5
OU-based Dataset	TS-Agent	$0.728^{\pm 0.02}$	$\mathbf{0.696}^{\pm 0.04}$	$0.633^{\pm 0.08}$	$0.680^{\pm 0.05}$	$\mathbf{47.67}^{\pm 0.58}$	$\mathbf{48.94}^{\pm 1.99}$	$53.89^{\pm 3.52}$	$\mathbf{46.44}^{\pm 1.54}$	$0.708^{\pm 0.01}$	$\mathbf{0.666}^{\pm 0.05}$	$0.570^{\pm 0.12}$	$0.676^{\pm 0.03}$	$\mathbf{0.889}^{\pm 0.00}$	$\mathbf{0.889}^{\pm 0.00}$	$\mathbf{0.889}^{\pm 0.00}$	$0.815^{\pm 0.13}$	100	100	100	100
	DS-Agent	$0.631^{\pm 0.28}$	$0.671^{\pm 0.11}$	$0.502^{\pm 0.12}$	$0.583^{\pm 0.09}$	$57.83^{\pm 17.03}$	$64.67^{\pm 20.30}$	$58.00^{\pm 12.60}$	$64.28^{\pm 8.77}$	$0.620^{\pm 0.32}$	$0.658^{\pm 0.14}$	$0.431^{\pm 0.11}$	$0.671^{\pm 0.01}$	$0.741^{\pm 0.26}$	$0.815^{\pm 0.08}$	$0.833^{\pm 0.00}$	$0.648^{\pm 0.20}$	100	100	100	100
	ResearchAgent	$0.473^{\pm 0.31}$	N/A	$\mathbf{0.794}^{\pm 0.00}$	$\mathbf{0.794}^{\pm 0.00}$	$80.22^{\pm 24.81}$	N/A	$48.17^{\pm 0.00}$	$47.83^{\pm 0.00}$	$0.484^{\pm 0.33}$	N/A	$\mathbf{0.806}^{\pm 0.00}$	$\mathbf{0.806}^{\pm 0.00}$	$0.574^{\pm 0.37}$	N/A	$\mathbf{0.889}^{\pm 0.00}$	$\mathbf{0.889}^{\pm 0.00}$	100	0	33.3	33.3
	EvoTS-Agent	$\mathbf{0.767}^{\pm 0.05}$	$0.626^{\pm 0.18}$	$0.671^{\pm 0.11}$	$0.717^{\pm 0.13}$	$47.89^{\pm 0.48}$	$57.56^{\pm 13.88}$	$\mathbf{47.78}^{\pm 4.10}$	$47.33^{\pm 1.17}$	$\mathbf{0.769}^{\pm 0.06}$	$0.552^{\pm 0.24}$	$0.625^{\pm 0.16}$	$0.731^{\pm 0.13}$	$\mathbf{0.889}^{\pm 0.00}$	$\mathbf{0.889}^{\pm 0.06}$	$0.852^{\pm 0.03}$	$0.796^{\pm 0.16}$	100	100	100	100
Mean-variance Shift	TS-Agent	$0.355^{\pm 0.18}$	$0.568^{\pm 0.13}$	$0.608^{\pm 0.04}$	$0.789^{\pm 0.13}$	$128.11^{\pm 5.91}$	$79.11^{\pm 10.78}$	$74.22^{\pm 9.33}$	$33.83^{\pm 30.31}$	$0.537^{\pm 0.38}$	$0.634^{\pm 0.13}$	$0.620^{\pm 0.08}$	$0.833^{\pm 0.14}$	$0.315^{\pm 0.07}$	$0.556^{\pm 0.10}$	$0.611^{\pm 0.00}$	$0.759^{\pm 0.13}$	100	100	100	100

Table 1. Time series change point detection performance on four benchmark datasets. Each metric is averaged over three runs. The best result for each LLM (per column) across different agents is bolded.

Figure 3. Visualization of raw input data. One representative sample from each dataset: (a) OU process, (b) Mean-variance Shift, (c) ADIA, and (d) Bee-Dance.

Figure 4. Average agent ranks across four LLM backbones for each dataset–metric pair. Lower average ranks, plotted farther outward, indicate better performance.

Figure 5. A case study of our EvoTS-Agent on the Bee Dance dataset. The F1 scores shown are evaluated on the validation set.

Model	F1 $\uparrow$	Hausdorff Distance $\downarrow$	Precision $\uparrow$	Recall $\uparrow$
PELT	$0.539$	26.50	$0.393$	$0.875$
BottomUp	$0.565$	$26.67$	$0.408$	$0.931$
Window	$0.523$	$36.50$	$0.519$	$0.750$
ChangeForest-RF	$0.275$	$31.28$	$0.160$	1.000
ChangeForest-KNN	$0.310$	$28.89$	$0.187$	$0.944$
BOCD	$0.501$	24.33	$0.351$	$0.903$

Table 2. Direct results of models in Model Bank on the Bee Dance dataset.

Model	F1 $\uparrow$	Hausdorff Distance $\downarrow$	Precision $\uparrow$	Recall $\uparrow$
EvoTS-Agent	$\mathbf{0.635}^{\pm 0.04}$	$37.67^{\pm 11.05}$	$\mathbf{0.660}^{\pm 0.16}$	$\mathbf{0.671}^{\pm 0.13}$
EvoTS-Agent w/o Alternative	$0.578^{\pm 0.06}$	$29.67^{\pm 4.77}$	$0.587^{\pm 0.12}$	$0.639^{\pm 0.18}$
EvoTS-Agent w/o Recombination	$0.575^{\pm 0.12}$	$\mathbf{28.61}^{\pm 3.08}$	$0.616^{\pm 0.10}$	$0.602^{\pm 0.13}$

Table 3. Ablation study on the Bee-Dance dataset using GPT-5.4. Results are reported as mean

\pm

standard deviation over three runs. The best result in each column is bolded.

Change-Point Detection

Change-point detection identifies when the statistical behavior or underlying regime of a sequence changes.

See all 83 papers

Aug 20, 2026

skchange: Fast and Flexible Algorithms for Changepoint Detection

Skchange is an open-source Python library for detecting structural changes in time series. It implements modern change detection algorithms within a unified and extensible framework. The algorithms are modular and composable, and they include changepoint search methods based on both cost minimisation and statistical tests. Key features include the detection of anomalous segments in addition to changepoints; theoretically well-founded fast and approximate search methods; theoretically well-founded algorithms for high-dimensional data, covering settings where either few or many features change simultaneously; utilities for automatic and data-driven penalty calibration, which balances false alarms against missed detections; and a large collection of built-in costs and statistical tests. The design follows established scikit-learn conventions to streamline both user and contributor experience, and Numba is used extensively to achieve high computational performance. Source code and documentation are available at https://github.com/NorskRegnesentral/skchange.

Martin Tveten, Johannes Voll Kolstø, Per August Jarval Moen

Department of Statistics and Machine Learning, Norwegian Computing Center · Department of Mathematics, University of Oslo

Change-Point Detection Time-Series Anomaly Detection

Ask PDF

Sep 23, 2026

TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent

Time series agents answer analytical questions by calling external tools, and which tools they carry is decided by people before the agent runs. However, we identify two failures in this setup. Human-Agent Tool Misalignment: a library of 21 expert-curated tools helps on some tasks and hurts on others, dropping anomaly accuracy under every backbone we test. Silent Harm: one round of generic self-revision changes 147 answers and breaks 56 of them, while the final score moves by less than a point. Both follow from the same gap: whether a tool helps is decided question by question at runtime, while tools are supplied in advance and judged by a single average. To address this, we propose TimeEvo, which clusters an agent's diagnosed failures into capability gaps, plans a measurement for each, synthesizes evidence-only tools that fill them, and admits the candidate library only through a paired admission gate. Experiments on ten time series QA tasks and three backbones show that TimeEvo, starting from an empty library, improves accuracy on every task and every backbone, and that a library grown on a cheap model still gains when it is installed into stronger ones. Code is available at https://github.com/Muyiiiii/TimeEvo.

Jie Yang, Yan Zheng, Jiarui Sun +8

University of Illinois at Chicago · Visa Research · University of Florida +2

Time Series QA LLM Agent Self-Improvement LLM Agents Self-Evolving Agents Data Analysis Agents

Ask PDF

Oct 1, 2026

Have an LLM Write Your Anomaly Detector: Autonomous Discovery of Compact, Interpretable Detectors for Time Series

Time-series anomaly detection trades off predictive accuracy, computational efficiency, and interpretability. We use a large language model not as the detector but as the author of one: an autonomous research loop in which the model repeatedly edits a single short NumPy program under a leakage-free objective, keeping the best-scoring detector it finds. The loop discovers two compact detectors, one for univariate and one for multivariate series, that describe short windows by their local spectral features and compare them with the training-region distribution through a covariance-aware distance. On the TSB-AD benchmark these detectors lead the field across metrics, ahead of the strongest classical, deep, and foundation-model baselines including Time-RCD, yet they train no network and use no GPU, and the multivariate detector is faster than every similarly performing baseline. LLM-driven program search is thus a practical route to accurate, efficient, and transparent detectors.

David Berghaus

Lamarr Institute Germany · Fraunhofer IAIS Germany

Interpretable Anomaly Detection LLM-Based Program Synthesis Unsupervised Anomaly Detection Time-Series Anomaly Detection Anomaly Detection

Ask PDF

Abstract

Change-Point Detection

Explore similar work

skchange: Fast and Flexible Algorithms for Changepoint Detection

TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent

Have an LLM Write Your Anomaly Detector: Autonomous Discovery of Compact, Interpretable Detectors for Time Series