Paper: arXiv 2503.20787
Authors: Cheng Wang, Chuwen Wang, Shirong Zeng, Jianguo Liu, Changjun Jiang
Abstract
The high-order complexity of human behaviour is likely the root cause of extreme difficulty in financial market projections. We consider that behavioural simulation can unveil systemic dynamics to support analysis. Simulating diverse human groups must account for the behavioural heterogeneity, especially in finance. To address the fidelity of simulated agents, on the basis of agent-based modeling, we propose a new paradigm of behavioural simulation where each agent is supported and driven by a hierarchical knowledge architecture. This architecture, integrating language and professional models, imitates behavioural processes in specific scenarios. Evaluated on futures markets, our simulator achieves a 13.29% deviation in simulating crisis scenarios whose price increase rate reaches 285.34%. Under normal conditions, our simulator also exhibits lower mean square error in predicting futures price of specific commodities. This technique bridges non-quantitative information with diverse market behaviour, offering a promising platform to simulate investor behaviour and its impact on market dynamics.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper presents a complex agent-based modeling paradigm with a hierarchical knowledge architecture (high math), backed by specific numerical results on futures markets like 13.29% deviation and MSE comparisons (moderate empirical rigor).
Research Flowchart
flowchart TD
A["Research Goal:<br>Simulate financial market<br>systemic projection"] --> B{"Methodology:<br>Agent-Based Modeling"}
B --> C["Key Innovation:<br>Hierarchical Knowledge Architecture<br>Language & Professional Models"]
C --> D["Simulation Engine:<br>Agents + Heterogeneity<br>vs. Futures Market Data"]
D --> E{"Evaluation"}
E -- Crisis Scenario --> F["Findings: Low Deviation<br>13.29% (Price: 285.34%)"]
E -- Normal Conditions --> G["Findings: Lower MSE<br>Predicting Futures Prices"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a hierarchical knowledge architecture called “Model Tower” for generative agents to enable behavioural simulation with domain-specific financial knowledge.
Source sentence
In this article, we propose a hierarchical knowledge architecture called “Model Tower” for generative agents that emulates the knowledge structure of human cognition as displayed in the bottom of Fig. 1.
Assumes 4
- Futures markets are chosen for their relatively straightforward participant structures and influencing factors, making them suitable for modeling under current computational constraints. market
Source sentence
Compared to stock markets and other financial derivative markets, futures markets feature relatively straightforward participant structures and influencing factors. Under current computational constraints, futures markets represent an appropriate research subject.
- Agents can process non‑quantitative information such as news to affect their behaviour. model
Source sentence
The inability to process non-quantitative information further diminishes heterogeneity in ABM simulations of financial markets.
- Each agent is endowed with personality and individual knowledge through the system configuration of its LLM. model
Source sentence
Moreover, each agent is endowed with personality and individual knowledge through the system configuration of its LLM.
- Trading follows defined rules R and matching engine E without considering slippage or market impact beyond the rule‑based matching. execution
Source sentence
... The engine E automatically withdraws all remaining unmatched requests in Omatch.
Data
- WIND database price data market-prices commercial
Source sentence
Price data are retrieved from WIND database.
- Eastmoney news data text-news public
Source sentence
News data in Chinese are from https://www.eastmoney.com/.
Costs
Not applicable
Finds 5
- The simulator achieves a 13.29% deviation when simulating crisis scenarios with price increase rates up to 285.34%. performance
Source sentence
our simulator achieves a 13.29% deviation in simulating crisis scenarios whose price increase rate reaches 285.34%.
- In the Tsingshan Nickel incident simulation, the error between actual return rate (283%) and simulated return rate is 13%, compared to a 99% relative error for the best general temporal model. predictive-accuracy
Source sentence
In terms of prediction accuracy, the related error between actual return rate (283%) and simulated ones is 13%, while the best general temporal model has a relative error of 99% (Fig. 2a).
- For futures contracts SC2501 and TA501, the proposed method consistently outperforms baseline models over a three‑day horizon when news information is incorporated. performance
Source sentence
For SC2501 and TA501, our method consistently outperforms the baseline models over a three-day horizon, indicating that the simulation framework effectively captures the market impact of news.
- The method performs poorly on IH2412, ranking near the lowest among compared approaches, likely due to complex dynamics not captured by limited news information. negative
Source sentence
However, for IH2412, even with the inclusion of news information, the performance of our method is suboptimal, ranking nearly the lowest among the compared approaches.
- Agents with hierarchical knowledge architecture successfully reproduce Glencore Group’s long positions and subsequent market sentiment changes leading to Tsingshan Group’s forced liquidation. other
Source sentence
within three simulations of the Tsingshan Nickel Incident, the agents successfully reproduce Glencore Group’s long positions leading to market sentiment changes and subsequently Tsingshan Group’s substantial forced liquidation.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 32 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.