大语言模型驱动的市场交易行为模拟

doi:10.11871/jfdc.issn.2096-742X.2026.01.001

数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (1): 2-13.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.01.001

doi: 10.11871/jfdc.issn.2096-742X.2026.01.001

大语言模型驱动的市场交易行为模拟

王成^1,^2,³(),曾诗容¹,王楚文¹,蒋昌俊^1,^2,^3,^*()

1.同济大学，计算机科学与技术学院，上海 201804
2.同济大学，嵌入式系统与服务计算教育部重点实验室，上海 201804
3.上海人工智能实验室，上海 200232

收稿日期:2025-02-28 出版日期:2026-02-20 发布日期:2026-02-02
通讯作者: 蒋昌俊
作者简介:王成，同济大学计算机科学与技术学院，教授，主要研究方向为智能风控与网络计算。
本文中承担的工作为整体研究框架与实验设计。
WANG Cheng is a professor at the School of Computer Science and Technology, Tongji University. His research interests include AI-powered risk management and network computing.
In this paper, he is mainly responsible for the overall research framework and experiment design.
E-mail: cwang@tongji.edu.cn|蒋昌俊，中国工程院院士、同济大学教授、英国工程技术学会会士、伦敦布鲁内尔大学荣誉教授，主要研究领域为网络计算与金融安全。
本文中承担的工作为整体研究框架设计。
JIANG Changjun is an academician of the Chinese Academy of Engineering and a professor at Tongji University. He is also an IET fellow and an honorary professor at Brunel University London. His research interests include network computing and financial security.
In this paper, he is mainly responsible for designing the overall research framework.
E-mail: cjjiang@tongji.edu.cn
基金资助:
国家自然科学基金(62372328);国家自然科学基金(72342026)

Market Trading Behavior Simulation Driven by Large Language Models

WANG Cheng^1,^2,³(),ZENG Shirong¹,WANG Chuwen¹,JIANG Changjun^1,^2,^3,^*()

1. School of Computer Science and Technology, Tongji University, Shanghai 201804, China
2. The Key Laboratory of Embedded System and Service Computing Ministry of Education, Tongji University, Shanghai 201804, China
3. Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China

Received:2025-02-28 Online:2026-02-20 Published:2026-02-02
Contact: JIANG Changjun

摘要/Abstract

摘要：

【目的】市场中的参与者决策主要体现在市场交易行为中。研究市场交易有助于理解市场的运行机制。本文旨在评估大语言模型（Large Language Models，LLMs）能否复现人类在市场交易中的有限理性行为，探究LLMs在市场交易模拟中的潜力和局限。【方法】行为经济学中的框架效应、禀赋效应和心账理论揭示了认知偏差和情感等因素对人类市场交易行为的影响。基于此，本文设计并实施了四类实验：问卷实验、公平价格实验、兑换券和商品交易实验。通过分析LLMs在模拟实验中的回答及交易决策，评估其模拟人类行为的能力。【结果】实验结果表明，LLMs在问卷实验中的表现与人类高度一致，人设、问题表述和事件描述顺序均会影响其决策。但LLMs在定价决策和商品价值评估方面与人类存在差异，未能复现心账理论和禀赋效应。【结论】LLMs在特定情境下能模拟人类的市场交易决策，但在完全复现人类的有限理性决策方面仍存在局限。

关键词: 市场交易, 大语言模型, 行为经济学, 有限理性, 模拟能力

Abstract:

[Objective] In markets, participants' decisions are reflected through market trading. Understanding market trading is crucial for grasping the operational mechanisms of markets. This paper aims to assess whether Large Language Models (LLMs) can replicate bounded rationality in human market trading decisions and to explore the potential and limitations of LLMs in market trading simulation. [Methods] Behavioral economics theories, such as framing effect, endowment effect, and mental accounting, highlight how cognitive biases, emotional factors, and external information influence human market trading decisions. These theories form the basis for understanding bounded rationality in human decision-making. Building on these classic theories, this paper designs and implements four types of experiments: a questionnaire simulation, a fair price simulation, a voucher exchange simulation, and a product trading simulation. This study evaluates the human-like consistency of LLMs by analyzing the alignment between LLMs' responses and human decisions in the questionnaire experiments and their trading decisions in the exchange experiments. [Results] The experimental results show that LLMs align closely with humans in the questionnaire experiments. Factors such as persona, phrasing of questions, and the sequence of event descriptions significantly influence LLMs' decisions. However, LLMs differ from humans in pricing decisions and value assessments and fail to reproduce key behavioral patterns implied by mental accounting and the endowment effect. [Conclusions] In conclusion, while LLMs can simulate human market trading decisions in specific contexts, they still face limitations in fully replicating bounded rationality in human decision-making.

Key words: market trading, Large Language Models (LLMs), behavioral economics, bounded rationality, simulation capability

王成,曾诗容,王楚文,蒋昌俊. 大语言模型驱动的市场交易行为模拟[J]. 数据与计算发展前沿, 2026, 8(1): 2-13.

WANG Cheng,ZENG Shirong,WANG Chuwen,JIANG Changjun. Market Trading Behavior Simulation Driven by Large Language Models[J]. Frontiers of Data and Computing, 2026, 8(1): 2-13, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.01.001.

图/表 7

图1

图2

表1

不同LLMs在框架效应问卷实验上的匹配率，整体匹配率下的最佳结果加粗表示"

LLMs设置		$M$ /%	$M c 1$ /%	$M c 2$ /%	$M c 3$ /%	$M c 4$ /%
有人设	glm-4-9b	72.73	100	66.67	100	50
	internlm3-8b	54.55	50	66.67	50	50
	Llama-3-8b	72.73	100	66.67	50	75
	Qwen2.5-7B	63.64	50	66.67	50	75
	Qwen2.5-14B	54.55	50	66.67	100	25
无人设	glm-4-9b	54.55	100	33.33	100	25
	internlm3-8b	72.73	50	100	100	50
	Llama-3-8b	72.73	50	100	100	50
	Qwen2.5-7B	54.55	50	33.33	50	75
	Qwen2.5-14B	63.64	50	66.67	100	50

表1

表2

表3

表4

图3

参考文献 26

[1]	SHILLER R J. Market volatility and investor behavior[J]. The American Economic Review, 1990, 80(2): 58-62.
[2]	BROCK W A, HOMMES C H. A rational route to randomness[J]. Econometrica: Journal of the Econometric Society, 1997, 65(5): 1059-1095.
[3]	BERGMANN B R. Micro-to-macro simulation: a pri-mer with a labor market example[J]. Journal of Economic Perspectives, 1990, 4(1): 99-116.
[4]	FARMER J D, AXTELL R L. Agent-Based modeling in economics and finance: Past, present, and future[R]. Institute for New Economic Thinking at the Oxford Martin School, University of Oxford, 2022.
[5]	WACHI A. Failure-scenario maker for rule-based age-nt using multi-agent adversarial reinforcement learning and its application to autonomous driving[C]// International Joint Conference on Artificial Intelligen-ce, 2019: 6006-6012.
[6]	RAMAN N, LEIDNER J L. Financial market data simulation using deep intelligence agents[C]// Advances in Practical Applications of Survivable Age-nts and Multi-Agent Systems, 2019: 200-211.
[7]	MIN B, ROSS H, SULEM E, et al. Recent advances in natural language processing via large pre-trained language models: A survey[J]. ACM Computing Surveys, 2023, 56(2): 1-40.
[8]	GAO S, WEN Y, ZHU M, et al. Simulating financial market via large language model based agents[J]. arXiv preprint arXiv:2406.19966, 2024.
[9]	THALER R H. Behavioral economics: Past, present, and future[J]. American economic review, 2016, 106(7): 1577-1600.
[10]	XIE C, CHEN C, JIA F, et al. Can Large Language Model Agents Simulate Human Trust Behavior?[C]// Advances in neural information processing systems, 2025, 37: 15674-15729.
[11]	LIU R, GENG J, PETERSON J C, et al. Large language models assume people are more rational than we really are[J]. arXiv preprint arXiv: 2406.17055, 2024.
[12]	TVERSKY A, KAHNEMAN D. The framing of decisions and the psychology of choice[J]. Science, 1981, 211(4481): 453-458.
[13]	THALER R. Toward a positive theory of consumer choice[J]. Journal of economic behavior & organization, 1980, 1(1): 39-60.
[14]	THALER R. Mental accounting and consumer cho-ice[J]. Marketing science, 1985, 4(3): 199-214.
[15]	KAHNEMAN D, KNETSCH J L, THALER R H. Experimental tests of the endowment effect and the Coase theorem[J]. Journal of political Economy, 1990, 98(6): 1325-1348.
[16]	HORTON J J. Large language models as simulated economic agents: What can we learn from homo silicus?[R]. National Bureau of Economic Research, 2023.
[17]	CHEN Y, LIU T X, SHAN Y, et al. The emergence of economic rationality of GPT[J]. Proceedings of the National Academy of Sciences, 2023, 120(51): e231-6205120.
[18]	GEERLING W, MATEER G D, WOOTEN J, et al. ChatGPT has aced the test of understanding in college economics: Now what?[J]. The American Economist, 2023, 68(2): 233-245.
[19]	GLM T, ZENG A, XU B, et al. Chatglm: A family of large language models from glm-130b to glm-4 all to-ols[J]. arXiv preprint arXiv:2406.12793, 2024.
[20]	DUBEY A, JAUHRI A, PANDEY A, et al. The llama 3 herd of models[J]. arXiv preprint arXiv:2407.21-783, 2024.
[21]	YANG A, YANG B, ZHANG B, et al. Qwen2.5 technical report[J]. arXiv preprint arXiv: 2412.15115, 2024.
[22]	CAI Z, CAO M, CHEN H, et al. Internlm2 technical report[J]. arXiv preprint arXiv: 2403.17297, 2024.
[23]	WANG P, LI L, CHEN L, et al. Large Language Models are not Fair Evaluators[C]// Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024, 1: 9440-9450.
[24]	WEI J, WANG X, SCHUURMANS D, et al. Chain-of-thought prompting elicits reasoning in large language models[C]// Advances in neural information processing systems, 2022, 35: 24824-24837.
[25]	WANG C, WANG C, ZHANG W, et al. Next-Generation Simulation Illuminates Scientific Problems of Organised Complexity[J]. arXiv preprint arXiv:2401. 09851, 2024.
[26]	WANG C, WANG C, ZENG S, et al. Advanced Digital Simulation for Financial Market Dynamics: A Case of Commodity Futures[EB/OL]. TechRxiv, 2025-02-19 [2025-02-28]. http://dx.doi.org/10.36227/techrxiv. 174000482.22151732/v1.

LLMs	交易轮次	实际交易量	预期交易量	实际市场价格/$	预期市场价格/$
glm-4-9b	1	11	11	4.25^-	3.75⁺
	2	13	13	4.75^-	4.75^-
	3	7	6	4.75^-	4.75
internlm3-8b	1	12	11	4.75	4.75⁺
	2	13	13	4.75⁺	4.75⁺
	3	10	10	4.75⁺	4.75⁺
Llama-3-8b	1	10	9	4.25⁺	4.25
	2	10	11	5.75⁺	4.75
	3	11	12	3.75⁺	3.75
Qwen2.5-7B	1	11	11	4.75	4.75^-
	2	13	14	4.75⁺	4.75⁺
	3	9	9	4.75⁺	5.25
Qwen2.5-14B	1	10	9	4.75	4.75⁺
	2	10	10	4.25^-	4.25^-
	3	11	11	3.75⁺	3.75⁺

大语言模型驱动的市场交易行为模拟

Market Trading Behavior Simulation Driven by Large Language Models

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 7

参考文献 26

相关文章 10

编辑推荐

Metrics

本文评价

LLMs设置			整体匹配率/%	原则1	原则2	原则3	原则4
有人设	原始顺序	glm-4-9b	25	—	X		X
		internlm3-8b	75				X
		Llama-3-8b	75				X
		Qwen2.5-7B	50		X		X
		Qwen2.5-14B	100
	调整顺序	glm-4-9b	50	—			X
		internlm3-8b	75				X
		Llama-3-8b	50			X	X
		Qwen2.5-7B	50		X		X
		Qwen2.5-14B	50			X	X
无人设	原始顺序	glm-4-9b	25	X	X		X
		internlm3-8b	50	—			X
		Llama-3-8b	50			X	X
		Qwen2.5-7B	25		X	X	X
		Qwen2.5-14B	75		X
	调整顺序	glm-4-9b	50	X			X
		internlm3-8b	75				X
		Llama-3-8b	50			X	X
		Qwen2.5-7B	25		X	X	X
		Qwen2.5-14B	75				X

模型	成本	市场价格	朋友			陌生人
模型	成本	市场价格	0/%	5/%	10/%	0/%	5/%	10/%
glm-4-9b	0	5	55	45	0	0	100	0
	0	10	0	100	0	0	1	99
	5	5	0	100	0	0	100	0
	5	10	0	100	0	0	1	99
	10	5	15	85	0	0	95	5
	10	10	0	100	0	0	0	100
internlm3-8b	0	5	36	63	0	0	100	0
	0	10	5	90	5	0	37	63
	5	5	0	100	0	0	100	0
	5	10	0	92	8	0	10	90
	10	5	0	79	21	0	97	3
	10	10	0	5	95	0	1	99
Llama-3-8b	0	5	12	85	1	0	99	1
	0	10	0	97	3	0	3	97
	5	5	0	100	0	0	100	0
	5	10	0	98	2	0	5	94
	10	5	0	95	4	0	97	3
	10	10	0	72	28	0	5	94
Qwen2.5-7B	0	5	29	71	0	0	100	0
	0	10	35	47	18	0	12	88
	5	5	4	96	0	0	100	0
	5	10	14	75	12	0	18	82
	10	5	8	72	20	0	94	6
	10	10	12	4	84	0	0	100
Qwen2.5-14B	0	5	88	10	0	0	100	0
	0	10	5	95	0	0	5	95
	5	5	1	99	0	0	100	0
	5	10	0	98	2	0	1	99
	10	5	1	90	8	0	84	16
	10	10	0	19	81	0	0	100

[1]	武智晖,黄绍晗,张逸飞,齐家兴,肖智文,曾畅,栾钟治. 基于检索增强的日志问答系统[J]. 数据与计算发展前沿, 2026, 8(1): 64-76.
[2]	吴建华,刘振宇,曾睿,王文瑄,易勇,王世轶. 基于LoRA微调大模型在网络安全等级保护测评报告质量检测中的应用研究[J]. 数据与计算发展前沿, 2025, 7(6): 111-123.
[3]	董伟,廖佳纯,姚思诚,陈海粟,阚苏南. 大语言模型赋能的面向微数据脱敏的敏感信息识别分析方法[J]. 数据与计算发展前沿, 2025, 7(6): 35-43.
[4]	刘典玉,刘青凯,肖雨阳,王杰. CAE-Bench: 面向结构力学仿真的大语言模型评估基准[J]. 数据与计算发展前沿, 2025, 7(4): 155-168.
[5]	屈志勇,王晓光,周纯葆,史源香,乔嘉伟. 面向国产超算系统的大模型训练优化方法[J]. 数据与计算发展前沿, 2025, 7(2): 120-129.
[6]	王子健, 李凯, 曹荣强, 周纯葆. 基于检查点的大模型弹性训练方法研究[J]. 数据与计算发展前沿, 2025, 7(1): 135-151.
[7]	韦一金, 陈彦清, 王秀东, 樊景超. 基于大语言模型的《中国小麦品种志》信息提取[J]. 数据与计算发展前沿, 2025, 7(1): 175-185.
[8]	马秋平, 张琪, 赵晓凡. 图表问答研究综述[J]. 数据与计算发展前沿, 2025, 7(1): 19-37.
[9]	裴炳森,李欣,蒋章涛,刘明帅. 基于大语言模型的司法文本摘要生成与评价技术研究[J]. 数据与计算发展前沿, 2024, 6(6): 62-73.
[10]	韦一金, 樊景超. 基于ChatGLM2-6B的农业政策问答系统[J]. 数据与计算发展前沿, 2024, 6(4): 116-127.