MarS

MarS simulates a market at the level of individual orders: an event-driven exchange and matching engine (mlib.core), an agent framework on top of it, and a generative Large Market Model trained to produce order flow. The paper frames the combination as a training environment for trading agents, a forecasting tool, and a platform for market-impact analysis, and was published as a poster at ICLR 2025 (arXiv:2409.07486). A bundled report implements eleven of Rama Cont’s stylized facts — heavy tails, volatility clustering, the leverage effect, asymmetry in timescales, and the rest — as the yardstick for whether a simulated tape behaves like a real one.

The model is not available. download.py fetches the Hugging Face repo microsoft/mars-order-model, which returns HTTP 401; the README describes the model as private awaiting final review approval. Everything downstream of it — forecasting, market impact, the stylized-facts report — therefore cannot be run. What runs without weights is the simulation core: market_simulation/examples/run_simulation.py builds an exchange from mlib.core, registers a trade-info state, drives it with a noise agent over a one-hour session, and plots the resulting price trajectory. The README’s OneDrive fallback for the remaining prerequisites is reported broken (issue #15), and the issues asking for the model and the training code (#9, #11, #20) are open and unanswered.

Treat it as research code rather than a maintained platform. main has not moved since 2025-05-26 — later repository activity is dependabot branches, not commits to main — there is no PyPI package, and Docker on Linux with CUDA is the only supported install path. A third-party Hugging Face upload under a similar name is not a Microsoft artifact and is not the model release.