Emergent Trading Biases in LLM Agent Swarms: Behavioral-Finance Benchmarks and Information-Channel Ablation

Large Language Model (LLM) agents are being widely adopted to simulate stock market behavior, yet beyond aggregate price dynamics we have limited mechanistic understanding of how individual agents trade when interacting in a multi-agent market: do typical behavioral economic biases show up, and if so, what causes them? We answer both questions with a swarm of 100 LLM agents (Qwen2.5-7B-Instruct) interacting in a simulated market with realistic microstructure. With no bias explicitly mentioned in any prompt, agents are given heterogeneous trading personalities and connected over seeded Barabási–Albert social networks. We measure loss aversion, the disposition effect, herding, anchoring, overconfidence, and panic against a budget-constrained zero-intelligence (ZI) baseline, and test whether the biases documented in human traders emerge. A five-condition ablation (full, no_social, no_personal, no_news, no_persona) across seven stress events and a no-event control (20 seeds, 800 runs) measures how much each information channel influences each bias, with all comparisons seed-clustered and FDR-corrected. Four of the six biases emerge in the expected human direction: herding at 15.4× the ZI floor, the disposition effect—on a single-asset adaptation of Odean's PGR–PLR construction—within the range of human traders (0.034 vs. 0.050, a directional comparison rather than a level match), loss aversion higher than the floor (market-drift controlled, 1.49 vs. 1.13), and panic elevated above its own no-event floor in 18 of 20 seeds. Conversely, anchoring arises with a reversed sign, and overconfidence is absent in trade sizing. Some causes are more predictable, such as loss aversion eroding when the personal position layer is ablated (1.94 to 1.38), as prospect theory anticipates. Others are more difficult to predict: removing persona heterogeneity nearly triples herding, and panic, while amplified by news, is diminished by social context. No ablation leads to perfectly rational agents, and a residue persists under every prompt we test. Understanding each information channel's influence, as well as that residue, tells us how bias can be controlled when deploying multi-agent systems.