LLM trading agents show measurable pre-failure signatures before market drawdowns

A new study by Weicheng Xue (arXiv:2605.28850) introduces TradeArena, an auditable trading-agent testbed that lets researchers analyze how LLM agents reason and behave under market stress. The paper reveals measurable pre-failure signatures: planning embeddings drift away from normal-state centroids, and fused plan-risk representations separate normal from pre-drawdown states. The author uses 80 rolling failure anchors across 8 LLM trajectories and shows that effective-rank contraction persists across multiple embedding probes (hash, LSA, Transformer, white-box hidden states).

Key experiments show that structured risk feedback can act as an external alignment signal without fine-tuning, though it's not a universal performance booster: true audit feedback improves calibration for some models, while placebo feedback sometimes yields higher short-horizon returns but weaker alignment diagnostics. A 51-stock intraday experiment reveals a correlation blind spot—LLM rationales often justify concentrated exposure to coupled assets that the risk layer repeatedly clips. The takeaway is not profitability but auditable risk feedback and representation trajectories that reveal when LLM financial reasoning is aligning, drifting, or failing.