TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets¶
🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
TwinMarket is a multi-agent LLM simulation framework for financial markets that couples a BDI (Belief-Desire-Intention) cognitive architecture with a dynamic social network and an order-driven trading system. It scales to 1000+ heterogeneous agents and reproduces four canonical stylized facts (fat tails, leverage effect, volume–return correlation, volatility clustering) more faithfully than rule-based ABM baselines. The framework is calibrated on real Chinese stock-market data (SSE 50 index, 2023) and demonstrates emergent macro phenomena—bubbles, wealth polarization, herding—from micro-level individual decisions.
Problem¶
Rule-based Agent-Based Models (ABMs) assume homogeneous agents with static behavioral rules, failing to capture irrational factors (loss aversion, overconfidence, herding) emphasized in behavioral economics and the emergent effects of agent-to-agent information propagation. Prior LLM-based market simulators either use fixed prompts (CompeteAI), ignore inter-agent interaction (EconAgent, ASFM), or do not scale beyond a handful of agents.
Method¶
Each agent is an LLM (GPT-4o in experiments) structured by the BDI framework: - Belief: initialized from real user profiles (639 Xueqiu users, 11,965 transactions); updated daily via a sentiment score (average of five self-assessed market dimensions); encoded as a system prompt with demographic and investment-style traits. - Desire: autonomous query generation against a knowledge base of stock data, news, and company announcements. - Intention: final trading decision (buy/sell/hold) submitted to an order-driven matching system, plus social media actions (post/repost/like).
The macro environment consists of two subsystems: 1. Order-driven trading system: clears transactions and updates stock indices. 2. Dynamic social network: modeled as \(G = (V, E)\) where edge weights are weighted Jaccard similarities computed from time-decayed trading intensity \(w(u,i) = \sum_{t \in T(u,i)} e^{-\lambda \cdot \Delta t}\). Each agent's perception field is restricted to neighbors with similarity \(S(u_t, u) > \tau\); posts are ranked by a hot-score \(h(p) = \frac{\log_{10}(u_p - d_p + 1)}{(T_p + 1)}\) with \(T_p = \max\!\left(\frac{t_c - t_p}{\alpha}, \epsilon\right)\), decaying over ~2 weeks.
Validation uses 100 agents over a 5-month simulation (June–November 2023) against real SSE 50 data and two ABM baselines (HPM herding model, BH hyperbolic-discounting model).
Key Contributions¶
- Real-world grounding: profiles synthesized from actual retail investor transaction data; simulated stock data, news, and announcements from live Chinese market feeds.
- BDI cognitive transparency: each agent's reasoning chain is structured and inspectable, enabling micro-to-macro causal tracing.
- Dynamic social network: time-weighted Jaccard similarity graph updates continuously; personalized information aggregation via hot-score–ranked feeds.
- Scalability: demonstrated at 1000+ agents across multiple assets, with explicit group-size scaling experiments.
- Emergent phenomenon reproduction: self-fulfilling speculative bubbles, wealth divergence (rising Gini coefficient), and information cascade–driven market crashes.
Results¶
- Stylized Fact I – Fat tails (Kurtosis): Real 7.26 | TwinMarket 5.24 | ABM-HPM 4.47 | ABM-BH 4.99 — TwinMarket closest to real.
- Stylized Fact II – Leverage effect (Negative Return Autocorrelation): Real 0.14 | TwinMarket 0.11 | ABM-HPM 0.05 | ABM-BH 0.19.
- Stylized Fact III – Volume–Return Correlation: All systems achieve \(p < 0.01\); TwinMarket matches real data.
- Stylized Fact IV – Volatility Clustering (GARCH \(\alpha+\beta\)): Real 0.95 | TwinMarket 0.89 | ABM-HPM 0.82 | ABM-BH 0.72.
- Wealth inequality: Gini coefficient rises monotonically over the simulation; top 10% achieve +6.65% return at 4.02% turnover; bottom 50% average −10.52% return at 7.03% turnover — consistent with real retail-investor overtrading findings.
Limitations¶
- All experiments use GPT-4o; sensitivity to LLM choice, prompt design, and model version is not characterized.
- Validation is limited to 100 agents; scalability experiments with 1000+ agents are not cross-validated against real data.
- The social network is derived from trading similarity, not actual follow/friend graphs; opinion-leader emergence may differ from real platforms.
- Simulation covers a single market (SSE 50, 2023) with a specific regulatory and cultural context; generalizability to other markets is untested.
- Information propagation experiments (rumors, policy shocks) are described but results are truncated in the provided text.
- Computational cost of running 100+ LLM agents per simulated day is not reported.
Relevance to Harnesses / Meta-Harnesses¶
TwinMarket is structurally a simulation meta-harness: it orchestrates the lifecycle of 1000+ LLM agents simultaneously, routing environment observations (market data, social feeds) through a shared information aggregation layer and collecting actions (orders, posts) back into two coordinated subsystems. The BDI scaffold functions as a per-agent cognitive harness that enforces a standardized perception→reasoning→action loop, directly analogous to the observe–think–act pipelines common in agentic AI harnesses. The framework's design—decoupled perception fields, dynamic graph state, hot-score ranking for information gating—illustrates reusable architectural patterns (selective context injection, tiered agent coordination) applicable to non-financial multi-agent harness design. It also demonstrates how calibrating a harness against real-world behavioral data (rather than synthetic benchmarks) can ground emergent multi-agent dynamics in empirically verifiable outcomes.