LLM trading agents trust everything equally. This is why they lose money.
Li, Gonsalves, Li, Yoon, and Wang (arXiv:2603.22567) identify the vulnerability as “uniform trust” — treating all information sources (news articles, financial reports, social sentiment) as equally reliable without filtering for consistency or grounding. Human traders don't do this. They weight sources by track record, cross-check claims, and discount information that conflicts with reproducible signals. LLMs skip all of this and act on whatever they were last told.
TrustTrade replaces uniform trust with selective consensus. Multiple independent LLM agents process the same information separately. Their outputs are compared for semantic and numerical agreement. Consistent signals get amplified; divergent or weakly grounded signals get discounted. Deterministic temporal signals — technical indicators with known mathematical properties — serve as anchors that prevent the consensus from drifting into collectively hallucinated territory.
The results across backtesting in high-noise markets (2024 Q1, 2026 Q1): the system moves from extreme risk-return profiles toward a moderate, human-aligned profile. Less hallucination-driven volatility. More stable returns. Not better predictions — better filtering of predictions that shouldn't have been trusted in the first place.
The deeper insight is that LLM hallucination in trading is not a bug in the model but a mismatch in the decision architecture. A single LLM generates plausible-sounding analysis regardless of whether the underlying information supports it. The fluency of the output carries false authority — the same mechanism that makes LLMs convincing makes them dangerous when the stakes involve money. Consensus among independent agents doesn't eliminate hallucination but does reduce its impact: a hallucination that only one agent produces is filtered, while a genuine signal that multiple agents independently identify is reinforced.
The reflective memory mechanism — adapting risk preferences during testing without retraining — addresses a separate problem: markets change faster than models can be retrained. A model calibrated on a calm market panics in a volatile one. The memory lets the system adjust its behavior based on recent experience, without the cost and lag of full retraining.