Research map / Image generation / Autoregressive and latent generation
MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the Large Market Model (LMM), an order-level generative foundation model, and the MarS engine built on it for controllable, interactive financial market simulation with downstream forecasting, detection, impact analysis and RL applications.
Problem
Financial market simulation at order-level resolution that is realistic, controllable and interactive, which existing statistical and agent-based simulators lack.
Approach
Treats order generation as conditional autoregressive modeling: an order-level causal transformer tokenizes each order with its LOB context, ensembled with an order-batch model that encodes minute-level order flow as RGB order images tokenized by VQGAN; a simulated clearing house matches generated and user-injected orders in a feedback loop.
Key results
- Scaling curves show order-model and order-batch model performance improves with data (up to 32B and 10B tokens) and model size (up to 1.02B and 3B params).
- Control with replay curve raises trajectory correlation from 0.23 to 0.47; adding interaction reduces it to 0.33.
- The RL agent improves from about -6 BP to 2-6 BP price advantage over the best TWAP baseline during training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add replay-curve control signals | MarS · Replay-curve control | correlation: 0.23 -> 0.47 | Helped |
| add interaction while replay-curve control is enabled (ablation) | MarS · Replay-curve control | correlation: 0.47 -> 0.33 | Hurt |
| include LOB information in order-model tokenization (ablation) | Order Model · Order-model training | training performance | Helped |
| increase data size and model size | LMM · Order Model and Order-Batch Model scaling | performance | Helped |
| condition the ensemble model on order-batch data | Ensemble Model · Order generation | loss | Helped |
Limitations
- Current implementation uses only a fraction of available order-level data due to resource constraints.
- Manipulation detection evidence is qualitative and requires combining multiple metrics for robust conclusions.
- Forecasting and RL results are presented on limited settings with the DeepLOB comparison in appendix.
Experimental setup
Datasets: Chinese stock market order-level data (top 500 liquid stocks, 2017-2023) CSRC market manipulation cases
Benchmarks: stylized facts of market data trend prediction vs DeepLOB square-root law validation CSRC manipulation detection
Metrics: validation loss trend prediction accuracy price-change correlation distribution similarity price advantage in bp
Baselines: DeepLOB TWAP agent
Builds on: LLaMA2 VQGAN DeepLOB Latent Diffusion Models
Closest papers by meaning
- Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.87)
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.87)
- LOB-Bench: Benchmarking Generative AI for Finance - an Application to Limit Order Book Data (ICML 2025 · similarity 0.87)
- FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.86)
- R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization (NeurIPS 2025 · similarity 0.86)
- Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading (ICLR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.