Paper: arXiv 2502.15800

Authors: Thomas Henning, Siddhartha M. Ojha, Ross Spoon, Jiatong Han, Colin F. Camerer

Abstract

This paper explores how Large Language Models (LLMs) behave in a classic experimental finance paradigm widely known for eliciting bubbles and crashes in human participants. We adapt an established trading design, where traders buy and sell a risky asset with a known fundamental value, and introduce several LLM-based agents, both in single-model markets (all traders are instances of the same LLM) and in mixed-model “battle royale” settings (multiple LLMs competing in the same market). Our findings reveal that LLMs generally exhibit a “textbook-rational” approach, pricing the asset near its fundamental value, and show only a muted tendency toward bubble formation. Further analyses indicate that LLM-based agents display less trading strategy variance in contrast to humans. Taken together, these results highlight the risk of relying on LLM-only data to replicate human-driven market phenomena, as key behavioral features, such as large emergent bubbles, were not robustly reproduced. While LLMs clearly possess the capacity for strategic decision-making, their relative consistency and rationality suggest that they do not accurately mimic human market dynamics.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper relies on a well-defined experimental design with clear statistical metrics (MSE, correlation) and visualizations of market behavior, indicating moderate empirical rigor, but its mathematical content is relatively light, focusing on behavioral comparisons rather than dense theoretical derivations.

Research Flowchart

  flowchart TD
  RQ["Research Goal<br>Do LLM agents replicate<br>human market trading<br>and bubble formation?"] --> I["Inputs<br>Experimental Finance Design<br>(Classic Bubble/Crash Paradigm)"]
  
  I --> M["Methodology<br>Agent-Based Modeling<br>LLM Agents as Traders"]
  
  M --> MP["Market Scenarios<br>1. Single-Model Markets<br>2. Mixed-Model Battle Royale"]
  
  MP --> CP["Computational Process<br>Simulated Trading Sessions<br>Asset priced near<br>fundamental value<br>vs. Human trading data"]
  
  CP --> F["Key Findings<br>1. LLMs: 'Textbook-Rational'<br>2. Muted Bubble Formation<br>3. Low Strategy Variance<br>4. Do NOT replicate<br>human market dynamics"]
  
  F --> O["Outcome<br>LLM-only data is insufficient<br>for replicating<br>key human behavioral features"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes to compare Large Language Models with human traders in a classic experimental finance paradigm to assess their ability to replicate human market dynamics.

Source sentence
In this study, we compare Large Language Models (LLMs) with human traders in a classic experimental-finance paradigm where prices are determined endogenously.

Assumes 6

  • The asset has a known constant fundamental value. market
    Source sentence
    We adapt an established trading design, where traders buy and sell a risky asset with a known fundamental value
  • Dividends are drawn i.i.d. from a Bernoulli distribution over {0.4, 1.0} with equal probability. data
    Source sentence
    In each period, a unit of stock pays a stochastic dividend drawn i.i.d. from a Bernoulli distribution over {0.4, 1.0}, with equal probability.
  • Cash holdings earn a constant risk‑free return. market
    Source sentence
    Cash holdings earn a constant risk-free return of…
  • Human participants receive instructions that mirror the LLM prompts. execution
    Source sentence
    The human participants were given instructions that mirror the LLM prompts
Show 2 more
  • Risk‑elicitation tasks are orthogonal to trading and do not affect outcomes. other
    Source sentence
    The task was exogenous: lottery choices did not affect cash balances, inventory constraints, or market clearing.
  • LLM agents are prompted after each round to describe their strategy, enabling chain‑of‑thought reasoning. model
    Source sentence
    After each round, LLM agents were prompted to provide details of the strategy they used to drive their trading behavior.

Data

  • Human participant data · prolific.com synthetic public
    Source sentence
    e participants recruited through prolific.com.
  • LLM market simulation runs synthetic unknown
    Source sentence
    We run homogeneous (mono‑agent) markets with single‑model LLM agents and heterogeneous “Battle Royale” markets with multiple LLM models.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 5

  • Claude‑3.5‑Sonnet and GPT‑4o achieve the lowest MSEs (≈0.5‑0.8) and near‑zero correlation with human price paths, keeping prices close to fundamentals. performance
    Source sentence
    Claude-3.5 Sonnet and GPT-4o yield the lowest MSEs and nearly zero correlation with human price paths, keeping prices close to fundamentals
  • LLM agents generally price the asset near its fundamental value, showing “textbook‑rational” behavior and only muted bubble formation. performance
    Source sentence
    LLM‑agents typically converge on prices near the fundamental value, thereby exhibiting more “textbook-rational” behavior than human subjects.
  • LLM markets display lower portfolio‑value variance than human markets, indicating less strategic heterogeneity. statistical
    Source sentence
    Both also show low portfolio variance, suggesting consistent strategies across instances.
  • All AI models outperform human forecasters; human one‑round forecast error ≈1.67 versus AI errors ranging from 0.00 to 0.14. predictive-accuracy
    Source sentence
    Thus, all AI models outperform human forecasters in terms of forecast accuracy.
  • Adding risk‑elicitation lotteries does not materially change LLM trading outcomes relative to the baseline. negative
    Source sentence
    Adding risk‑elicitation did not materially change outcomes relative to the baseline per model

Checked by others

Supported by

  • Financial Stability Implications of Generative AI: Taming the Animal Spirits (2025)
    On the question: Do AI/LLM agents reduce asset‑price bubbles by pricing assets near their fundamental value?
    Compare the findings

    This paper: LLM agents generally price the asset near its fundamental value, showing “textbook‑rational” behavior and only muted bubble formation.

    That paper: AI agents exhibit less herd behavior than human financial professionals, suggesting a potential reduction in asset‑price bubbles.

    Both findings show AI agents (LLM agents in A, AI agents in B) price near fundamentals and exhibit less herd behavior, leading to muted bubbles; B’s result supports A.

  • What Does ChatGPT Make of Historical Stock Returns? Extrapolation and Miscalibration in LLM Stock Return Forecasts (2024)
    On the question: Do LLM forecasts provide more accurate and better calibrated stock return predictions than human forecasters?
    Compare the findings

    This paper: All AI models outperform human forecasters; human one‑round forecast error ≈1.67 versus AI errors ranging from 0.00 to 0.14.

    That paper: LLM forecasts are considerably better calibrated than human forecasts, indicating improved risk assessment.

    Both studies find AI outperforms humans: A reports LLMs are considerably better calibrated; B shows human error 1.67 versus AI errors 0.00‑0.14, confirming superior accuracy and calibration.

Contradicted by

  • Machine Spirits: Speculation and Adaptation of LLM Agents in Asset Markets (2026)
    On the question: Do LLM agents generate speculative price bubbles that deviate from the asset's fundamental value?
    Compare the findings

    This paper: LLM agents generally price the asset near its fundamental value, showing “textbook‑rational” behavior and only muted bubble formation.

    That paper: Some LLMs form price bubbles that overprice the asset up to 14.9 times its fundamental price, and these bubbles are almost always speculative.

    Finding A shows agents price near fundamentals (no bubbles); Finding B reports bubbles up to 14.9× the fundamental, almost always speculative.

  • Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations (2025)
    On the question: Do LLM agents replicate human market trader behavior in asset pricing and bubble formation?
    Compare the findings

    This paper: LLM agents generally price the asset near its fundamental value, showing “textbook‑rational” behavior and only muted bubble formation.

    That paper: Interactions among LLM agents generate realistic market dynamics, including price bubbles and corrections, highlighting both capabilities and potential systemic risks.

    A shows near‑fundamental pricing with muted bubbles; B shows realistic dynamics with bubbles and corrections, indicating a contradiction.

Cited by 4 papers in the archive (4 by independent authors): Shared Models, Selective Trading, and Order Flow (used as a baseline); Machine Spirits: Speculation and Adaptation of LLM Agents in … (discusses its findings)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (83)

  1. Price bubbles with discounting: A web-based classroom experiment (2009) — link
  2. Algorithm aversion: People erroneously avoid algorithms after seeing them err (2015) — link
  3. A Turing test of whether AI chatbots are behaviorally similar to humans (2024) — link
  4. Artificial intelligence and its implications for income distribution and market structure (2020) — link
  5. Multi-agent reinforcement learning in sequential social dilemmas (2017)
  6. Large language models in finance: Analyzing sentiment and predicting market trends (2023) — link
  7. Machine behavior (2019) — link
  8. Bubbles, crashes, and endogenous expectations in experimental spot asset markets (1988) — link
and 75 more
  1. What lies beneath: The linguistic traces of deception in online dating profiles (2012) — link
  2. Chain of thought prompting elicits reasoning in large language models (2022) — link
  3. Can large language models trade? Assessing AI behavior in financial markets (2022) — link
  4. Can Large Language Models beat wall street? Evaluating GPT-4’s impact on financial decision-making with MarketSenseAI (2024) — link
  5. FinRobot: AI Agent for Equity Research and Valuation with Large Language Models — → 2411.08804
  6. Algorithmic Collusion by Large Language Models (2024) — link
  7. Strategic Collusion of LLM Agents: Market Division in Multi-Commodity Competitions — → 2410.00031
  8. GPT-4 Technical Report (2023) — link
  9. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances (2022) — link
  10. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2023) — link
  11. Iterative Refinement with Self-Feedback (2023) — link
  12. Language Models are Few-Shot Learners (2020) — link
  13. Playing Games with GPT: What Can We Learn About a Large Language Model from Canonical Strategic Games? (2024) — link
  14. Playing Repeated Games with Large Language Models (2023) — link
  15. Overconfidence and Bubbles in Experimental Asset Markets (2016) — link
  16. Measuring Massive Multitask Language Understanding (2021) — link
  17. Training Verifiers to Solve Math Word Problems (2021) — link
  18. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (2023) — link
  19. Price Bubbles, Gender, and Expectations in Experimental Asset Markets (2017) — link
  20. Can LLMs Mimic Human-Like Mental Accounting and Behavioral Biases? (2024) — link
  21. Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus? (2023) — link
  22. Do Large Language Models Reason Like Us? (2024) — link
  23. Algorithm of Call Auction Price (2010) — link
  24. Dynamically optimized sequential experimentation (DOSE) for estimating economic preference parameters (2024)
  25. oTree—An open-source platform for laboratory, online, and field experiments (2016) — link
  26. Price bubbles in asset market experiments with a flat fundamental value (2005)
  27. Irrational exuberance and neural crash warning signals during endogenous experimental market bubbles (2014) — link
  28. Gemini 1.5: Unlocking multimodal understanding across millions of tokens (2024) — link
  29. Mistral Large (24.11) (2024)
  30. Grok-2 Beta Release (2024)
  31. GPT-4o System Card (2024)
  32. Introducing Claude 3.5 Sonnet (2024)
  33. OpenAI API (2025) — link
  34. Latent dirichlet allocation (2003)
  35. Software Framework for Topic Modelling with Large Corpora (2010)
  36. Natural language processing with Python: analyzing text with the natural language toolkit (2009)
  37. Thumbs up? Sentiment classification using machine learning techniques (2002)
  38. Decision-making behavior evaluation framework for LLMs under uncertain context (2024)
  39. Comparative analysis of LLM-based market prediction and human expertise with sentiment analysis and machine learning integration (2024)
  40. K-Level Reasoning: Establishing Higher Order Beliefs in Large Language Models for Strategic Reasoning (2024) — link
  41. LLM Knows Geometry Better Than Algebra: Numerical Understanding of LLM-based Agents in a Trading Arena — → 2502.17967
  42. TradingAgents: Multi-agents LLM Financial Trading Framework — → 2412.20138
  43. A Reflective LLM\textendash Based Agent to Guide Zero\textendash Shot Cryptocurrency Trading — → 2407.09546
  44. FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design — → 2311.13743
  45. Pretrained LLM Adapted with LoRA as a Decision Transformer for Offline RL in Quantitative Trading — → 2411.17900
  46. LLM-Based Routing in Mixture of Experts: A Novel Framework for Trading — → 2501.09636
  47. TradExpert: Revolutionizing Trading with Mixture of Expert LLMs — → 2411.00782
  48. Stock Price Prediction with LLM-Guided Market Movement Signals and Transformer Model (2025)
  49. LLM-Augmented Linear Transformer–CNN for Enhanced Stock Price Prediction (2025)
  50. HedgeAgents: A Balanced-Aware Multi-Agent Financial Trading System — → 2502.13165
  51. FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (2024)
  52. A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist — → 2402.18485
  53. Temporal Data Meets LLM–Explainable Financial Time Series Forecasting — → 2306.11025
  54. Large Language Model Driven Technical Analysis: Enhancing Predictive Accuracy and Interpretability in Stock Trading (2025)
  55. Can LLM-based Financial Investing Strategies Outperform the Market in Long Run? — → 2505.07078
  56. FinRL-DeepSeek: LLM-Infused Risk-Sensitive Reinforcement Learning for Trading Agents — → 2502.07393
  57. When AI Meets Finance (StockAgent): Large Language Model-Based Stock Trading in Simulated Real-World Environments — → 2407.18957
  58. Can Large Language Models Effectively Process and Execute Financial Trading Instructions? (2024) — link
  59. When AI Meets Finance (StockAgent): Large Language Model-Based Stock Trading in Simulated Real-World Environments — → 2407.18957
  60. Enhancing Automated Trading with Sentiment Analysis: Leveraging Large Language Models for Stock Market Predictions (2025)
  61. What Teaches Robots to Walk, Teaches Them to Trade too–Regime Adaptive Execution using Informed Data and LLMs — → 2406.15508
  62. Stock Price Trend Prediction Using Emotion Analysis of Financial Headlines with Distilled LLM Model (2024)
  63. Financial LLM for Stock Price Analysis and Investment Recommendation (2025)
  64. A Multi-Agent System Based on LLM for Trading Financial Assets (2024)
  65. Predictive Power of LLMs in Financial Markets — → 2411.16569
  66. Bubbles, Crashes , and Endogenous Expectations in Experimental Spot Asset Markets (1988) — link
  67. Empirical Asset Pricing with Large Language Model Agents (2024) — link
  68. Revolutionizing Finance with LLMs: An Overview of Applications and Insights (2024) — link
  69. Large Language Models and Financial Market Sentiment (2023) — link
  70. Sentiment Trading with Large Language Models — → 2412.19245
  71. Instruct-FinGPT: Financial Sentiment Analysis by Instruction Tuning of Large Language Models — → 2306.12659
  72. Mental accounting and consumer choice (1985)
  73. Prospect theory: An analysis of decision under risk (2013)
  74. Can Generative AI agents behave like humans? Evidence from laboratory market experiments — link
  75. Can AI Replace Human Subjects? A Large-Scale Replication of Psychological Experiments with LLMs — link

20 of 83 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.