MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model

Source

Status And Credibility

MarS was first posted to arXiv on 2024-09-04 and the inspected arXiv metadata reports version 2, updated 2025-03-13. The arXiv comment states ICLR 2025. It is referenced by the Microsoft TimeCraft repository as a financial application source.

Core Claim

MarS is a financial market simulation engine powered by a Large Market Model trained on order-level financial data. It generates realistic and controllable market trajectories, supports injected orders and market matching rules, and is evaluated as a forecasting, anomaly-detection, what-if, and reinforcement-learning environment.

Key Contributions

  • Frames order-level market simulation as a generative foundation-model task.
  • Trains Large Market Models over trading orders and order-batch representations.
  • Combines generative order models with a simulated clearing house.
  • Evaluates realistic, interactive, and controllable simulations and downstream financial tasks.

Evidence And Results

The paper reports scaling behavior for order sequence modeling, stylized-fact realism, interactive market impact simulations, forecasting, anomaly detection, what-if analysis, and reinforcement-learning agent training. The source is one of the more world-model-adjacent generation papers in this batch because generated futures are used for action evaluation and agent training.

Limitations

  • The domain is financial order books; results should not be generalized to generic multivariate time series without testing.
  • Market realism depends on matching rules, data coverage, and assumptions about participant behavior.
  • The paper demonstrates useful simulation tasks but does not solve causal market-impact identification in the general case.
  • Treat the reinforcement-learning environment as a domain-specific simulated transition model under LMM and clearing-house assumptions, not as evidence of a true causal market transition kernel without external market-impact validation.
  • Public evaluation may be hard to reproduce if data access or market microstructure details are restricted.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Action-conditioned world modelspartially closesSimulates market trajectories under injected orders and supports what-if analysis and RL-agent training.Needs clearer causal validation, multi-agent policy interaction, and reproducibility details.
Event streams and trajectoriespartially closesModels order-level event streams and order batches rather than only regular numeric observations.Interface is finance-specific and not a generic event-stream benchmark.
BenchmarksadjacentEvaluates stylized facts, forecasting, detection, market impact, and RL utility.Needs public benchmark harness and leakage controls.