Content area
Simulation has become an essential tool across various scientific disciplines for exploring complex and dynamic phenomenon. One area that has attracted significant research interest is the modeling of financial markets. The macroscopic modeling of financial time series, first established over a century ago with the development of Brownian motion, is typically formulated as an ad hoc stochastic process. However, these explicit mathematical formulations often struggle to replicate various empirical regularities of financial time series. Empirical studies have shown that the seemingly random nature of financial time series share several non-trivial statistical properties that are common across assets, markets, and time periods. Researchers have linked the emergent properties and the highly nonlinear, interconnected dynamics exhibited by markets to properties of complex adaptive systems, which have been successfully modeled using bottom-up approaches.
The aim of this work was to develop a high-fidelity market simulation platform for interactive strategy evaluation and all-purpose scenario generation. Toward this aim, we made three key contributions: (1) the development of a framework for validating and benchmarking models, (2) a baseline implementation of a scalable market environment, and (3) a deep generative model for bottom-up market simulation. We revisited several previously identified empirical regularities of financial markets and found that many still accurately describe modern markets. These results informed the design of a generalized measure to evaluate how well a synthetic time series captures actual market dynamics—serving as a criterion for simulation fidelity. Next, we outlined our approach to generating financial time series from first principles. Using agent-based modeling, we reproduced many statistical properties of markets by explicitly incorporating known structural elements (market microstructure) while making minimal assumptions elsewhere—highlighting the importance of integrating central trading mechanisms to produce macroscopic phenomenon. Finally, after establishing a solid foundation for the design and analysis of simulated market environments, we addressed the need for realistic agent behavior in our market environment using data-driven methods. Inspired by the recent success of generative large language models (LLMs), we developed a single agent in the form of a generative pre-trained transformer (GPT) to replace the population of agents used in prior work. This new agent acted as a responsive order generation engine within an interactive discrete event market simulator. Our results show that, despite being trained on the most granular data available (microstructure messages), our model reproduces several empirical regularities and data distributions at the macro scale. Collectively, we have developed a robust modeling and evaluation framework that establishes new benchmarks and represents a significant step toward achieving high-fidelity interactive market simulation.
