Dynamic Human-in-the-Loop Live Trading: LLM Agents with Real-Time Expert Feedback

Research Question: Does integrating real-time human feedback into LLM-based trading agents enhance their adaptability, risk management, and live trading performance compared to fully autonomous LLM agents?

Hypothesis: LLM agents augmented with live, in-the-loop expert feedback will outperform both purely automated and purely human traders on key metrics such as Sharpe ratio, drawdown, and adaptation to market shocks.

Experiment Plan: - Extend LiveTradeBench to support a human-in-the-loop mode: experts can approve, veto, or modify LLM-generated trade suggestions in real time.

• Compare performance of (a) LLM-only, (b) human-only, and (c) hybrid teams over identical market periods.
• Use reinforcement learning from human feedback (RLHF) to adapt LLMs based on expert interventions, with reward shaping for alignment.
• Analyze decision logs to see which types of interventions drive the biggest improvements and whether LLMs learn to anticipate or internalize expert corrections over time.

References: ['Pierre H. Richemond et al. (2024). Offline Regularised Reinforcement Learning for Large Language Models Alignment. arXiv.org.', 'Alex Havrilla et al. (2024). Teaching Large Language Models to Reason with Reinforcement Learning. arXiv.org.', 'Prashant Mehta et al. (2025). AI-Driven Psychological Profiling and Risk Management in Margin and Options Trading Using Large Language Models. InTech.']