acceptodds
Under review as a conference paper at ICLR 2027

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

16 views

Abstract

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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, multi-agent systems, trading biases, behavioral finance, market simulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dSrFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper measures six trading behaviors in a simulated market with 100 Qwen2.5-7B-Instruct agents. It compares information channels and trader personas with a budget-constrained zero-intelligence population. The experiments examine which behaviors emerge and use the ablations to investigate their sources.

Critical strengths
  1. I find the analysis of individual trading behavior meaningful for understanding LLM agent populations. Studying how roles and information channels change behavior goes beyond aggregate price realism and provides useful evidence for the design and evaluation of agent systems.
  2. The paired seed analyses, multiple-testing correction, and complete condition-by-event results support the empirical comparisons. The paper also reports absent overconfidence and reversed anchoring, and examines the large drift generated by the LLM population. These negative results help define the scope of the findings.
Critical weaknesses
  1. I would like a deeper account of how shared decision policies produce population-level synchronization. Replacing eight personas with one neutral prompt raises herding from 0.060 to 0.167, an informative result about population design. However, this intervention changes both the instructions and strategy heterogeneity. It leaves open how much synchronization reflects a shared trading policy and how much supports the proposed attribution to model training or alignment.
  2. The loss-aversion interpretation would also benefit from a clearer account of endogenous market feedback. The drift-matched LAR comparison of 1.49 versus 1.13 is a useful control. Equal realized drift can still involve different price paths and opportunities to sell, while LAR measures holding times among completed trades. I would like to understand how these effects contribute to the remaining gap before attributing it to a loss-averse preference.
Questions for authors
  1. Beyond matching realized drift, what evidence separates the excess holding time of losing trades from the different price paths generated by the two populations?
  2. What evidence distinguishes model-intrinsic herding from synchronization induced by the shared neutral trading prompt?
Reviewer_yueAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LLMs tested in isolation exhibit typical human-like biases. It remains unknown that such biases survive, compound, or diminish once many agents interact in a market. This paper systematically studies this question, and show that some of the bias is carried by the context layers and personas, and some by the model itself.

Critical strengths
  1. The paper examines whether typical human biases emerge without explicit prompting.
  2. The paper quantifies the total effect of each information channel on these biases.
  3. The paper assesses how persona prompts and the model itself contribute to these biases.
Critical weaknesses
  1. The study is limited by a single model, fixed temperature, and untested prompt robustness.
  2. The test experiment is just a single asset with market-maker execution.
Questions for authors
  1. Is it possible to show some results for evaluating a wider range of models to assess whether the findings generalize across model families and sizes?
  2. Test paraphrased and slightly modified prompts to assess sensitivity to wording.
  3. How to extend evaluation to more realistic and diverse market scenarios, including multiple assets, longer horizons, and varied news conditions?
Reviewer_DvFAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores whether behavioral trading biases emerge in a simulated market of only 100 LLM agents and how information channels and trading personas influence them. The authors find that four bias patterns and effects of information and persona on agent trading behavior.

Critical strengths
  • This paper addresses a relevant question of agents' trading bias.
  • The ablation study and the zero-intelligence baseline provide a useful comparison across behavioral metrics.
Critical weaknesses
  • Limited technical contributions. This paper focuses mainly on prompting and a standard market simulator. Its technical contributions over prior work on LLM bias and LLM trading agents remain largely unclear.
  • Unclear broader significance. Results from one model in a simplified market provide limited evidence of general behavioral guidance for designing reliable trading agents.
Questions for authors
  • What are the specific technical contributions over prior work?
  • How do the authors distinguish behavioral biases from rational responses to shared information, inventory constraints, or market dynamics?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.