Generative models aim to replicate realistic outcomes across various contexts, from text generation to visual effects. While much progress has been made in creating real-world simulators, the application of generative models to virtual environments like financial markets remains relatively untapped.
In a new paper MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model, a Microsoft Research Asia research team introduces MarS, a financial market simulation engine powered by a Large Market Model (LMM). This system addresses the unique demands of modeling the market impact of orders while enabling highly realistic, controllable simulations.
The core of MarS is the generative foundation model (LMM), which is trained on historical order-level financial data. During real-time simulations, the LMM generates dynamic sequences of orders based on user inputs—ranging from interactive orders to broader scenario descriptions—and integrates this with current or recent market data. These generated orders are then processed in a simulated clearing house, resulting in detailed, real-time market trajectories. The flexibility of MarS allows it to support a range of applications, including forecasting, risk detection, analysis, and agent training.
The research team highlights four main use cases for MarS:
- Forecasting Tool: MarS predicts future market trajectories by generating subsequent orders based on recent activity and the limit order book (LOB). This capability allows for more accurate forecasting by analyzing multiple simulated outcomes.
- Risk Detection System: By simulating a variety of future market paths, MarS can reveal hidden risks that may not be obvious from current data. For instance, a sudden decrease in trajectory variance could signal an upcoming major event, allowing for early risk detection and improved management.
- Analysis Platform: MarS offers a highly realistic simulation environment for “what if” scenarios. It can assess the market impact of large trades by comparing traditional market impact formulas with simulated results, providing deeper insights and potential refinements to existing models.
- Agent Training Environment: The realistic, responsive nature of MarS makes it ideal for training reinforcement learning agents. This is demonstrated in an order execution scenario, showing how MarS can be used to develop and fine-tune trading strategies without exposing users to real-world financial risks.
The research team outlines their key contributions:
- They take the first step in creating a generative foundation model tailored to financial markets, confirming the scaling law of the Large Market Model and highlighting its immense potential for domain-specific applications.
- They design a realistic simulation engine that addresses two critical requirements: generating target scenarios and modeling the impact of orders on the market. This unlocks the practical potential of LMM for real-world financial applications.
- They demonstrate four downstream use cases for MarS, showcasing its vast potential for transforming financial industry practices.
In summary, MarS is the first simulation engine to fully leverage fundamental financial market elements in conjunction with specialized AI technology. By providing a unified interface for a variety of financial tasks, MarS has the potential to drive paradigm shifts across the industry, opening new possibilities for forecasting, risk management, and strategy development.
The paper MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model is on arXiv.
Author: Hecate He | Editor: Chain Zhang