Market Regime Sensitivity: Evaluating LLM Agent Adaptability Across Structural Shifts

Research Question: How do LLM trading agents detect and adapt to major market regime shifts in real time, and can explicit regime-awareness architectures improve their resilience and returns?

Hypothesis: Current LLM trading agents lack dedicated mechanisms for regime detection, making them slow to adapt and prone to large losses during transitions; adding explicit regime-detection modules or training on regime-labeled data can close this gap.

Experiment Plan: - Annotate historical and live trading periods in LiveTradeBench with known regime shifts (e.g., volatility spikes, macro events).

• Evaluate baseline LLM agent performance pre- and post-shift for metrics like return, drawdown, and adaptation lag.
• Develop and integrate explicit regime-detection submodules (e.g., change-point detection, economic indicator classifiers) into LLM agent prompts or architectures.
• Retrain or prompt LLMs to reason about regime context and adjust portfolio/risk parameters dynamically.
• Compare adaptation speed and trading outcomes with and without regime-awareness.

References: ['S. Pareek & Sujit K. Ghosh (2025). Semiparametric Dynamic Copula Models for Portfolio Optimization.', 'C. Tudor & Robert Sova (2024). Enhancing Trading Decision in Financial Markets: An Algorithmic Trading Framework With Continual Mean-Variance Optimization, Window Presetting, and Controlled Early-Stopping. IEEE Access.', 'Qianqian Xie et al. (2024). FinBen: A Holistic Financial Benchmark for Large Language Models. Neural Information Processing Systems.']