🏆 NeurIPS 2025 & ICLR Financial AI Best Paper Award

TwinMarket:

A Scalable Behavioral and Social Simulation for Financial Markets

A novel multi-agent framework that leverages Large Language Models to simulate socio-economic systems, modeling individual investor behaviors, social interactions, and emergent market phenomena

🤖
1000+
LLM Agents
📊
50
SSE Stocks
📈
4
Stylized Facts
🧠
BDI
Framework
👥
Real
User Data
⚡
Scalable
Architecture
📄 Read Paper View Code 📝 Cite This Work

Overview

TwinMarket Overview
🎯

Core Problem

Traditional ABMs struggle to capture the diversity of human behavior, particularly irrational factors in behavioral economics

⚡

Our Solution

LLM agents that account for cognitive biases, emotional fluctuations, and non-rational influences in market simulations

💡 TwinMarket examines how individual behaviors, through interactions and feedback mechanisms, give rise to collective dynamics and emergent phenomena in financial markets. Unlike traditional rule-based ABMs, our LLM-based approach captures the full spectrum of human behavior complexity.

🎯 Through experiments in a simulated stock market, we demonstrate how individual actions trigger group behaviors, leading to emergent outcomes such as financial bubbles, market recessions, and other collective socio-economic patterns that arise from the complex interplay between individual decision-making and market dynamics.

Key Features

🎯

Real-World Alignment

Grounded in established behavioral theories and calibrated with real-world data, ensuring realistic human behavior modeling

🔄

Dynamic Interaction Modeling

Captures diverse human behaviors and their interactions, particularly in the context of information propagation and social influence

📈

Scalable Market Simulations

Supports large-scale simulations, allowing researchers to analyze the impact of group size and interaction complexity on market behavior

Framework Components

👤 Micro-Level Simulation: Individual Behaviors

🧠

BDI Framework

  • 💡 Belief: Agent's understanding of the market
  • 🎯 Desire: Agent's objectives or preferences
  • ⚡ Intention: Concrete actions the agent executes
📊

Behavioral Biases

💪 Overconfidence 📉 Loss Aversion 🐑 Herding 🎲 Risk Preferences

Influencing trading decisions and contributing to market heterogeneity

🌐 Macro-Level Simulation: Social Interactions

🌐

Social Network Construction

  • 🔗 Dynamic Connections: Network evolves based on trading patterns
  • 📋 Behavior Similarity: Agents connect through similar strategies
  • ⏱️ Time Decay: Recent interactions weighted more heavily
📡

Information Propagation

  • 📦 Information Aggregation: Collecting data from social connections
  • 👑 Opinion Leaders: Influential agents shape market sentiment
  • 🔄 Echo Chambers: Formation of polarized belief groups

Data Sources

Data Sources
👤

Real User Profiles

From Xueqiu social media platform

100K+ users
💱

Transaction Details

Historical trading data from Xueqiu

1M+ trades
📈

Stock Data

SSE 50 index from CSMAR database

50 stocks
📰

News & Announcements

From Sina Finance, 10jqka, and CNINFO

Daily updates

Experimental Results

Information Propagation

Opinion Leader Emergence

Influential nodes emerge and shape network opinions through cascading influence

Information Polarization

Information spreads differently across groups, creating echo chambers and polarized beliefs

Opinion Leaders

Simulations reveal the emergence of opinion leaders who exert significant influence on the network, shaping market sentiment and trading behaviors

Information Polarization

Different information signals lead to the formation of distinct groups with divergent beliefs, creating self-reinforcing feedback loops

Behavioral Polarization Under Rumors

Behavioral Polarization

Belief Divergence

Rumors lead to a divergence in user beliefs and the formation of distinct echo chambers

Trading Volatility

Rumors make users more likely to sell, leading to significant increase in sell/buy ratio

Market Turbulence

Rumors cause the market to suffer sharp declines

Market Dynamics - Stylized Facts

Market Dynamics
📊

Fat-tailed Returns

Extreme price movements occur more frequently than normal distribution

📈

Volatility Clustering

High volatility periods are followed by high volatility periods

⚖️

Leverage Effect

Negative returns correlate with increased future volatility

💹

Volume-Return Relationship

Trading volume positively correlates with price changes

Emergent Group Behaviors: The framework reveals self-fulfilling prophecies where collective expectations drive market trends, and information cascades where traders rely on perceived consensus rather than fundamental analysis.

Scalability

Scalability Analysis

TwinMarket is designed to scale to large populations, with simulations involving up to 1,000 agents. The framework maintains realistic market dynamics even at larger scales, providing a robust platform for studying complex socio-economic systems.

How to Cite

If you use TwinMarket in your research, please cite the following paper:

@misc{yang2025twinmarketscalablebehavioralsocialsimulation,
      title={TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets},
      author={Yuzhe Yang and Yifei Zhang and Minghao Wu and Kaidi Zhang and
              Yunmiao Zhang and Honghai Yu and Yan Hu and Benyou Wang},
      year={2025},
      eprint={2502.01506},
      archivePrefix={arXiv},
      primaryClass={cs.CE},
      url={https://arxiv.org/abs/2502.01506},
}
View on arXiv