Research map / Multi-agent systems and games / Human-AI and goal-conditioned
TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TwinMarket, a BDI-driven multi-agent LLM framework that simulates investor behavior and social media in a China A-share style stock market to reproduce stylized facts and study emergent bubbles and rumor-driven crashes.
Problem
Rule-based agent-based models oversimplify human decision-making (homogeneous agents, static rules) and fail to capture behavioral biases and micro-to-macro emergence in financial markets.
Approach
LLM agents structured by Belief-Desire-Intention (BDI) with personas calibrated on real investor transactions, interacting through a dynamic similarity-based social network and a hot-score post recommendation feed within an order-driven market simulation.
Key results
- TwinMarket achieves RMSE 0.02, MAE 0.02, correlation 0.77 on daily normalized index prices, outperforming the w/o BDI and w/o heterogeneity ablations.
- Simulated kurtosis of 5.24 with leverage effect 0.11 and GARCH alpha+beta about 0.89, closer to real data (7.26, 0.14, 0.95) than ABM-HPM and ABM-BH.
- Rumor exposure raises the sell-to-buy ratio from 0.495 to 0.997 and causes a market downturn, while rumor-exposed agents assign valuations 27.5% lower.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TwinMarket instead of ABM-HPM or ABM-BH | stylized-fact market simulation · Financial market stylized facts | kurtosis: 4.47 / 4.99 -> 5.24 | Mixed |
| Remove the BDI module (ablation) | TwinMarket · Daily normalized index price series | correlation: 0.77 -> 0.34 | Hurt |
| Remove agent heterogeneity (ablation) | TwinMarket · Daily normalized index price series | correlation: 0.77 -> -0.61 | Hurt |
| Disable social interaction (ablation) | TwinMarket · Daily normalized index price series | correlation: 0.77 -> -0.50 | Hurt |
| Disable dynamic belief updates (ablation) | TwinMarket · Daily normalized index price series | RMSE: 0.0158 -> 0.0342 | Hurt |
| Remove proactive information-seeking (ablation) | TwinMarket · Daily normalized index price series | RMSE: 0.0158 -> 0.0532 | Hurt |
| Increase LLM temperature from 0.3 to 1.3 (ablation) | TwinMarket agents · Agent trading behavior | average volume: 4818.31 -> 6071.78 | Helped |
| Replace neutral news with negative rumors | TwinMarket agents · Rumor-exposed agents | asset valuation: 27.5% lower | Hurt |
| Expose agents to negative rumors | TwinMarket market · Rumor-exposed versus control market | sell-to-buy ratio: 0.495 to 0.997 | Hurt |
| Vary active trading participation from 10% to 80% (ablation) | TwinMarket · Market prediction | RMSE | Helped |
Surprising findings
- Removing agent heterogeneity yields a negative correlation (-0.61) with real index prices, worse than removing BDI.
- Removing social interaction flips the correlation to -0.50 with the real series.
Limitations
- Limited to the China A-share market with a single daily call auction and a zero-sum environment.
- Relies on synthetic user profiles and simplified market mechanics.
- Results rest on proprietary-style simulation with variable GARCH parameters across runs.
Experimental setup
Datasets: Xueqiu user profiles and transactions (639 users, 11,965 transactions) Guba stock forum transactions (83,246) CSMAR stock data 2023 Sina and 10jqka news articles 2023 CNINFO company announcements 2023
Benchmarks: China A-share industry index price series (2023 simulation period)
Metrics: RMSE MAE correlation kurtosis leverage effect GARCH alpha and beta Gini coefficient volume-return correlation
Baselines: ABM-HPM ABM-BH CompeteAI EconAgent ASFM
Builds on: Belief-Desire-Intention EconAgent CompeteAI agent-based models
Closest papers by meaning
- MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework (NeurIPS 2025 · similarity 0.87)
- MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model (ICLR 2025 · similarity 0.87)
- CompeteAI: Understanding the Competition Dynamics of Large Language Model-based Agents (ICML 2024 · similarity 0.86)
- FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (NeurIPS 2024 · similarity 0.86)
- Can Large Language Model Agents Simulate Human Trust Behavior? (NeurIPS 2024 · similarity 0.86)
- Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading (ICLR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.