GroundTrader: Multi-Agent Transition Supervision for LLM-Based Financial Trading

Large language model (LLM)-based financial agents use market information to generate trading actions together with textual reasoning. However, the resulting executable position transition may conflict with its supporting analysis, while the effect of intervening on that transition becomes observable only after the market evolves. We propose GroundTrader, a reliability-oriented multi-agent supervisory framework for LLM-based financial trading. GroundTrader coordinates market-state, coherence, and record analysts to assess market context, review action–reasoning consistency, and use previously settled interventions to regulate subsequent deferral proposals. GroundTrader supervises selected transitions through bounded one-period deferrals and counterfactually evaluates each proposed deferral after the subsequent market return becomes available. The resulting settlements update a recoverable budget that regulates later intervention strength. We evaluate GroundTrader on nine financial assets using fixed upstream action–reasoning trajectories. Experimental results show that GroundTrader supervision improves annualized return and Sharpe ratio and reduces maximum drawdown relative to direct execution of the fixed upstream trajectories across all nine evaluated assets. Ablation results further show that contextual record reflection, action–reasoning coherence review, and adaptive budget control all contribute to the supervisory process.