Unified Financial Market Simulation Engine: MarS in the Age of Generative Foundation Models

Generative models have been making waves in various fields, and Microsoft Research is at the forefront of this innovation with the creation of the Financial Market Simulation Engine (MarS). By leveraging generative foundation models, such as the large market model (LMM), tailored to the financial industry, researchers have opened up a world of possibilities for empowering financial experts with customizable generative models for diverse scenarios.

The success of generative models, which excel at handling high-quality training data, tokenization of core information, and comprehensive data modeling for implicit reasoning, has paved the way for transformative applications in the financial sector. The vast order data in financial markets offer a rich playground for these models, thanks to their fine granularity, large scale, and structured nature, making them ideal for tokenization and sequential modeling.

Order flow data plays a pivotal role in generative modeling for finance, providing insights into both micro-level pricing behavior and macroscopic market dynamics. Microsoft Research’s LMM captures the nuanced feedback and overarching dynamics by modeling individual orders and order sets over time. With advanced tokenization techniques, these models can simulate complex market behaviors with a high degree of accuracy.

To unlock the full potential of financial data, Microsoft researchers have implemented two scaling strategies for their generative models, based on the Transformer architecture. By analyzing historical trading data, they have optimized the model to generate order flows that align with actual market intricacies, enabling more precise time-series modeling for financial applications.

The culmination of these efforts is the MarS, a flexible financial market simulation engine built on the foundation of the LMM. This customizable generative model is designed to adapt seamlessly to various financial scenarios, outperforming traditional models and offering a new paradigm for applying generative solutions to finance. The integration of generative models with domain-specific data is poised to revolutionize efficiency and insight generation in the financial industry.