Self-Improving Traders: Meta-RL for LLMs in Evolving Market Regimes

TL;DR: What if LLM traders could learn from their mistakes in real time? We'll combine LiveTradeBench with Yang’s reinforcement learning framework to let agents continuously refine their strategies.

Research Question: Can meta-RL enable LLM agents to autonomously adapt their portfolio allocation strategies during live trading?

Hypothesis: Meta-RL agents will outperform static LLMs by 30% in non-stationary markets (e.g., during regime shifts like bull-to-bear transitions).

Experiment Plan: - Setup: Integrate Yang’s QR-DQN with LiveTradeBench agents, using portfolio Sharpe ratio as reward.

• Data: Live trading data + simulated regime shifts (e.g., interest rate hikes).
• Metrics: Cumulative returns, adaptation speed (epochs to recover from drawdowns).
• Expected Outcome: RL-augmented agents will dynamically hedge risks during black swan events.

References: ['Yang, X. (2025). Feature-Generating Networks for Latent Feature Discovery in Deep RL Trading Strategies. Advances in Economics.', 'Chen, D., et al. (2024). Efficient Sequential Decision Making with LLMs. EMNLP.']