M3: A State-Event Generative Foundation Model for Market Microstructure Dynamics

Battle Count: 78

Highly relevant to quantitative trading. The paper directly addresses market microstructure simulation at the order level, which is fundamental to algorithmic execution, market impact estimation, liquidity risk assessment, and strategy backtesting. The model reproduces key stylized facts (heavy tails, volatility clustering, square-root impact law), enables counterfactual stress testing, and supports practical applications like TWAP execution analysis. The scaling law findings provide guidance for model development. The LOB-conditioned generation framework is directly applicable to execution optimization and market-making strategies.

High implementation complexity due to multiple interacting components: (1) VQ-VAE tokenizer with 32,768 codebook entries, causal encoder-decoder architecture, and composite loss function; (2) LOB prefix encoder with cross-attention pooling; (3) LLaMA2-style autoregressive Transformer backbone (up to 1.27B parameters); (4) Deterministic matching engine for order replay and LOB state updates; (5) Complex preprocessing pipeline with open-price anchoring and feature normalization. Training requires large-scale order-level data (31.9B events) and significant compute. The closed-loop simulation with feedback into autoregressive context adds further complexity.

Code is available on GitHub (ArthurZhang02/m3-market-microstructure) and model weights on HuggingFace (Arthur210/M3-market-microstructure). The paper provides detailed architecture specifications, hyperparameters, training procedures, and evaluation protocols. However, the raw order-level market data (CSI 300/500 constituents, Jan 2024 - Dec 2025) is proprietary and not publicly released, limiting full end-to-end reproduction. The matching engine rules and preprocessing pipeline are described in detail.