Introduction
Traditional financial modeling has long struggled with the "micro-to-macro" transition. Classic rule-based Agent-Based Models (ABMs) often rely on oversimplified, homogeneous agents that fail to capture the messy, irrational nuances of human psychology—such as overconfidence, loss aversion, and herding. While Large Language Models (LLMs) have shown promise in simulating human-like reasoning, existing research often overlooks the social dimension: how information spreads through a network and how collective interactions shape market dynamics.
TwinMarket is a novel multi-agent simulation framework that leverages LLMs to model the complex interplay between individual investor behavior and macro-level market phenomena. By utilizing a Belief-Desire-Intention (BDI) cognitive architecture and a dynamic social network, the system successfully replicates real-world financial "stylized facts" and demonstrates how micro-level biases can trigger systemic events like bubbles and crashes.
The TwinMarket Framework
The core of TwinMarket is its BDI (Belief-Desire-Intention) framework, which structures the cognitive process of each agent to mirror human decision-making:
- Belief: Agents perceive the environment through news, social media posts, and historical data, forming a "sentiment score" across five dimensions, including economic fundamentals and self-confidence.
- Desire: Based on their beliefs, agents autonomously generate queries to seek specific information, reflecting the active nature of investor research.
- Intention: Agents commit to actions, which include both financial moves (buy, sell, or hold) and social moves (liking, unliking, or reposting content).
This cognitive layer operates within a dual-layered environment:
- Social Environment: A dynamic graph where connections are formed based on trading similarity (weighted Jaccard similarity) with exponential time decay. It uses a "hot score" mechanism to rank posts, simulating how information cascades through a community and how "opinion leaders" emerge.
- Market Environment: An order-driven trading system based on the China A-share market's call auction rules. This ensures that prices emerge organically from actual agent transactions rather than predefined curves.
Data Foundations and Realism
To ensure behavioral realism, the researchers grounded TwinMarket in extensive real-world datasets:
- Investor Behavior: Transaction records from 639 real users (~12,000 transactions) from the Xueqiu platform were used to initialize agent personas and estimate behavioral biases.
- Social Dynamics: Over 83,000 transactions from the Guba Stock Forum were used to train the recommendation system.
- Market Data: Fundamental stock data for the year 2023 was sourced from CSMAR.
- World Knowledge: A massive database of over 1 million news articles and 5,600 company announcements from Sina, 10jqka, and CNINFO provided the information environment for the agents.
Key Insights and Emergent Phenomena
The simulation revealed several critical insights into the nature of digital financial markets:
The Scaling Law of Simulation
One of the most significant findings is that as the number of active agents increases (up to 1,000 in this study), the model’s predictive error (RMSE/MAE) decreases. This suggests that larger agent populations lead to more robust "price discovery" and more realistic emergent behavior, mirroring the liquidity benefits of real-world markets.
Social Media as a Volatility Multiplier
The study demonstrates that social networks do not merely share information; they polarize it. High-degree "opinion leaders" emerge naturally, and their influence can synchronize agent behavior. This synchronization often leads to "herding," a primary driver of market volatility and the precursor to crashes.
The Necessity of Cognitive Realism
Ablation studies showed that removing the BDI framework—essentially making agents "reactive" rather than "reasoning"—significantly degraded the model's ability to replicate financial realities. Without the BDI structure, the simulation failed to produce characteristic traits like fat-tailed distributions and volatility clustering.
Empirical Results
TwinMarket successfully replicated four major "Stylized Facts" of financial markets, outperforming traditional ABM baselines:
- Fat-tailed Distributions: Extreme market moves occurred more frequently than a normal distribution would predict.
- Leverage Effect: Market volatility increased more significantly following price drops than price gains.
- Volume-Return Relationship: Large price changes correlated strongly with high trading volume.
- Volatility Clustering: Periods of high or low turbulence tended to persist over time.
In a "Rumor Injection" experiment, the authors demonstrated the system's sensitivity to misinformation. By replacing neutral news with negative rumors, the sell-to-buy ratio nearly doubled (from 0.495 to 0.997), triggering a self-reinforcing market downturn that mirrored real-world panic selling.
Limitations and Future Directions
While TwinMarket represents a significant advancement in behavioral finance simulation, the authors acknowledge several limitations. The current model is tailored to the China A-share market’s daily call auction and does not yet account for continuous trading or diverse global market structures. Additionally, the simulation currently operates in a zero-sum environment, which may not capture the full complexity of long-term economic growth.
Future work will focus on investigating "agent trust" more deeply, exploring human-LLM alignment in financial contexts, and scaling the simulation to even larger populations and more diverse asset classes.
Summary updated