SABCEMM-A Simulator for Agent-Based Computational Economic Market Models

Financial markets exhibit complex, nonlinear behaviors that challenge classical theories like EMH. Stylized Facts such as fat-tailed distributions, volatility clustering, and income inequality have been documented across markets globally. Agent-Based Models (ABMs) emerged as a powerful approach to simulate these phenomena by modeling heterogeneous agents with diverse strategies and interactions. Notable models include Lux-Marchesi, Levy-Levy-Solomon, and Cont-Bouchaud, each capturing different stylized facts. Despite progress, existing simulation tools are limited in scale, flexibility, and reproducibility, hindering comprehensive analysis of market dynamics and systemic risks. The need for a unified, high-performance simulation platform is urgent to advance both theoretical understanding and practical applications.

Current ABCEM implementations are often fragmented, with models coded in various languages and lacking standard interfaces. Large-scale simulations are computationally demanding, and the impact of random number generators on performance and results is underexplored. Reproducing published results is difficult due to incomplete documentation and diverse computational environments. These issues impede systematic studies of finite-size effects, model comparison, and sensitivity analysis. Developing a scalable, flexible, and reproducible simulation environment is essential to overcome these bottlenecks and enable rigorous, large-scale investigations into market phenomena.

This work introduces SABCEMM, a modular, object-oriented C++ simulator supporting millions of agents. Key innovations include: 1) XML-based configuration for flexible model assembly; 2) integration of multiple PRNGs with systematic performance evaluation; 3) abstract model classes simplifying extension and comparison; 4) validation across multiple models demonstrating robustness and efficiency. The system enables large-scale, reproducible simulations, addressing the limitations of existing tools. Its design facilitates rapid prototyping, parameter tuning, and comprehensive statistical analysis, significantly advancing the state-of-the-art in ABCEM modeling.

• �� Develop an abstract model class defining core components: agents, market mechanisms, environment.
• �� Implement modular agent types (e.g., fundamentalists, chartists) with customizable decision rules.
• �� Design market mechanisms based on excess demand and price adjustment equations, including rational and irrational models.
• �� Integrate multiple PRNGs (e.g., Mersenne Twister, PCG) for performance testing.
• �� Use XML files for model configuration, enabling quick assembly and comparison.
• �� Validate models by reproducing known stylized facts and benchmarking simulation times.
• �� Conduct large-scale experiments with up to 2 million agents, analyzing output stability and efficiency impacts of PRNGs.

Simulations employed Levy-Levy-Solomon, Cross, and Harras models, with varying agent numbers up to 2 million. Performance metrics included runtime, scalability, and output statistical properties. Different PRNGs were tested to evaluate their influence on speed and result fidelity. The models' ability to reproduce stylized facts was assessed through distribution analysis, autocorrelation, and volatility measures. Sensitivity analyses explored parameter impacts, while reproducibility was verified via multiple runs under identical configurations. Hardware setups ranged from standard laptops to high-performance servers, ensuring broad applicability.

The simulator achieved over two million agents with simulation times reduced by 30% using optimized PRNGs like PCG. Random number generator choice significantly affected runtime, with some generators doubling efficiency. Large-scale outputs consistently reproduced stylized facts, including fat tails (kurtosis > 3) and volatility clustering (autocorrelation decay). Results were highly reproducible across runs, confirming system robustness. Comparative analysis demonstrated the platform's flexibility in model testing and parameter sensitivity, establishing a new benchmark for ABCEM simulation tools.

The platform supports academic research in market microstructure, systemic risk, and policy impact analysis. Financial institutions can use it for stress testing, scenario analysis, and risk management. Regulators may employ the simulator to evaluate macroprudential policies' effects. Its scalability allows exploration of rare events and large-system behaviors, providing insights into market stability and crisis prevention. Integration with real data and macroeconomic variables will further enhance its practical utility.

Despite high scalability, the simulator requires substantial computational resources, limiting accessibility for some users. The current models focus mainly on micro-level interactions, with macroeconomic factors needing further development. Numerical stability may be challenged in extreme parameter regimes. The discretization approach introduces approximation errors, which could affect results in highly sensitive scenarios. Future work should address these issues to improve robustness and realism.