A financial market simulation engine powered by a generative foundation model for realistic, interactive, and controllable order generation.
MarS is a financial market simulation engine powered by a generative foundation model called the Large Market Model (LMM). It generates realistic, interactive, and controllable financial market orders to simulate market dynamics, enabling applications like forecasting, impact analysis, and strategy testing without real-world financial risk. The engine models order-level events to produce emergent market behavior.
Quantitative researchers, financial engineers, and data scientists working in algorithmic trading, market microstructure analysis, and risk management who need to simulate and analyze market behavior in a controlled environment.
Developers choose MarS for its novel approach of using a generative foundation model to simulate order-level market dynamics, providing a more realistic and controllable alternative to traditional price prediction models. Its ability to validate simulations against established market stylized facts and enable counterfactual analysis for trading strategies offers unique value for financial research and application development.
MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Generates individual market orders that interact to produce emergent prices and trends, rigorously validated against 11 stylized facts like volatility clustering and heavy tails.
Includes ready-to-use tools for forecasting, market impact analysis, and interactive Streamlit demos, enabling visual exploration and counterfactual testing without real-world risk.
Built on Ray Serve for deployable batch inference, designed to handle large-scale simulations with parallel processing across multiple GPUs.
Based on a peer-reviewed paper accepted to ICLR 2025, demonstrating verification of scaling laws and rigorous methodology for financial domain modeling.
The core Hugging Face model is private awaiting review, crippling full functionality of demos and examples until released—a major limitation admitted in the README.
Requires Docker setup, CUDA, and significant GPU resources (e.g., 128 GPUs for research simulations), making it inaccessible for teams without high-performance computing access.
The README explicitly states that production deployment needs optimizations like replacing Ray with vLLM and model compression, indicating it's currently a research prototype, not a polished tool.