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

Financial time series fluctuate in ways that are both non-stationary and heterogeneous, meaning no single unsupervised algorithm reliably detects change points across different assets and market conditions. Manual workflows that depend on expert judgment for model selection, feature engineering, and tuning do not scale effectively. The arXiv preprint 2608.17933v1 presents EvoTS-Agent, a validation-guided self-evolving LLM agent that automates this detection process. It begins with a structured exploratory data analysis step to understand dataset characteristics and seed initial candidate models. The agent then iteratively refines executable experiment trajectories using three operators: Revision, which exploits the current best solution; Alternative Strategy, which explores fundamentally different approaches when progress stalls; and Recombination, which blends prior successful experiments. This design aims to diminish reliance on human intervention while adapting to diverse financial scenarios.