Self-Evolving LLM Agent Detects Financial Time Series Change Points

Financial time series data is notoriously complex, characterized by non-stationary and heterogeneous statistical properties. This makes the task of detecting "change points"—moments where the underlying statistical characteristics of the series shift—particularly challenging. Traditional methods often rely heavily on human experts for model selection, feature engineering, and hyperparameter tuning, which limits their scalability and adaptability across different assets and market conditions. To address these limitations, researchers have introduced EvoTS-Agent, a novel self-evolving LLM agent. This agent is specifically designed for autonomous change-point detection in financial time series. EvoTS-Agent begins by performing exploratory data analysis to understand the dataset's properties and then initializes a set of candidate detection models. The core innovation lies in its evolutionary process, guided by validation feedback. The agent evolves executable experiment trajectories using three operators: "Revision" to refine current best solutions, "Alternative Strategy" to explore new modeling directions when progress stalls, and "Recombination" to synthesize insights from high-performing trajectories. This adaptive approach allows EvoTS-Agent to tailor its detection pipeline to the unique statistical characteristics of each dataset, consistently outperforming existing LLM-based agents while maintaining a perfect execution success rate across various LLMs.