The paper addresses the instability and poor risk control of LLM-based trading agents, attributing these issues to a 'uniform trust' bias where all information sources are treated as equally reliable. Inspired by human decision-making, which relies on selective filtering, cross-validation, and memory, the authors propose TrustTrade, a multi-agent framework that aggregates signals from multiple LLM agents and dynamically weights them based on cross-agent consistency. TrustTrade incorporates deterministic temporal signals (e.g., trend, momentum, volatility) as stable anchors and a reflective memory bank that adapts risk preferences at test time without retraining.

Through controlled experiments with human annotators and backtests on 2024 Q1 and 2026 Q1 data, the authors show that TrustTrade significantly improves decision stability and aligns LLM behavior with human risk preferences. The framework reduces maximum drawdown and shifts the risk-return trade-off toward a mid-risk/mid-return regime, outperforming single-LLM agents and rule-based baselines. The results highlight that behavioral alignment, rather than increased model capacity, is key to robust LLM-based trading.

The study also provides detailed analysis of human trading behavior, revealing that humans prioritize temporally grounded signals and exhibit high decision convergence across information stages. TrustTrade's selective consensus and temporal grounding mimic these heuristics, leading to more stable and risk-aware trading. Limitations include reliance on LLM-generated reports, potential suppression of minority-correct signals, and evaluation on a limited set of stocks. Future work aims to extend to diversified portfolios and incorporate stronger anti-leakage protocols.