Back to AI Research

AI Research

EvoTS-Agent: A Self-Evolving LLM Agent for Financia... | AI Research

Key Takeaways

  • EvoTS-Agent is an autonomous system designed to identify structural changes in financial time series—such as shifts in volatility or market regimes—without r...
  • 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.
  • 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.
Paper AbstractExpand

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.

EvoTS-Agent is an autonomous system designed to identify structural changes in financial time series—such as shifts in volatility or market regimes—without requiring manual model selection or hyperparameter tuning. By using a validation-guided, self-evolving framework, the agent automates the creation of detection pipelines that adapt to the specific statistical characteristics of different financial datasets.

Automating Financial Change-Point Detection

Financial time series are often noisy and non-stationary, meaning a single detection algorithm rarely performs well across all assets or market conditions. Traditionally, analysts must manually select models, design features, and tune parameters, which limits scalability. EvoTS-Agent addresses this by treating change-point detection as an empirical search problem. It uses a Large Language Model (LLM) to propose, execute, and refine experimental pipelines, ensuring that the detection strategy is tailored to the data at hand rather than relying on a fixed, one-size-fits-all approach.

How the Agent Evolves

The agent operates through a closed-loop process that begins with Exploratory Data Analysis (EDA). This stage characterizes the time series using features like lag-1 autocorrelation and spectral concentration, which helps the LLM select the most suitable candidate models from a pre-built bank.
Once initialized, the agent improves its performance through three evolutionary operators:

  • Revision: The agent refines the current best-performing pipeline based on recent successful experiments.

  • Alternative Strategy: If progress stagnates, the agent shifts to a fundamentally different modeling direction or tests an untried model suggested by the initial EDA.

  • Recombination: In the final stage, the agent synthesizes successful components from multiple high-performing trajectories to create a final, optimized experiment.
    Throughout this process, the agent maintains a "trajectory pool" that records the experimental plan, code, and validation results. This allows the agent to learn from its own history while using an incumbent-preserving rule to ensure that unsuccessful experiments do not degrade the best-validated solution.

Performance and Reliability

According to the researchers, EvoTS-Agent consistently outperforms existing LLM-based agents across four benchmark datasets. A key finding is the system's robustness: it maintained a 100% execution success rate across all backbone LLMs evaluated. By framing the task as an empirical search over executable scientific experiments, the agent provides a transparent, validation-driven workflow that adapts to the specific statistical properties of the input data.

Considerations for Implementation

EvoTS-Agent is designed for inference-time transformation rather than updating the underlying language model itself. The framework relies on validation feedback to guide its evolution, meaning it requires a validation split with reference boundaries to evaluate configurations. While the agent automates the search for an effective detection pipeline, the quality of the final output remains dependent on the initial model bank and the agent's ability to interpret the dataset's meta-features during the EDA stage.

Comments (0)

No comments yet

Be the first to share your thoughts!