💰 quantitative finance

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

The paper introduces M3, a state-event generative foundation model trained on large-scale order-level data that captures the dynamic interaction between order flow and limit-order-book liquidity to enable realistic counterfactual market microstructure simulations for applications like forecasting and stress testing.

Original authors: Yanzhi Zhang, Yu Ma, Yilin Cheng, Jian Li, Yitong Duan

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yanzhi Zhang, Yu Ma, Yilin Cheng, Jian Li, Yitong Duan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Financial markets are vast, complex ecosystems where millions of decisions happen every second, driven by people and algorithms buying and selling assets. At the heart of this system lies a digital ledger known as the limit order book, a constantly shifting record of who wants to buy, who wants to sell, and at what price. This ledger is not just a static list; it is a living environment that changes the moment a new order arrives, and in turn, that changing environment influences the next order that comes in. Because we can only observe the single path history actually took, we cannot see the countless other futures that might have unfolded if a trader had acted differently or if a large order had arrived a second earlier. This limitation makes it difficult to answer critical questions about how markets would react to stress, how a large trade might move prices, or how a strategy would perform under different conditions. To understand these possibilities, researchers need a way to simulate realistic market futures, not just predict the next price tick.

A team of researchers has developed a new tool called M3, a foundation model designed to simulate the intricate dance of liquidity and order flow in electronic markets. Unlike previous attempts that treated the list of orders and the state of the market as separate problems, this model learns to generate entire sequences of future trading events while constantly accounting for how the market's current state shapes those events. The researchers trained M3 on a massive dataset containing roughly 31.9 billion individual order events from the Chinese stock market, covering hundreds of different stocks over nearly two years. By feeding this model a snapshot of the market at a specific moment, it can generate thousands of possible future paths, showing how prices, spreads, and available liquidity might evolve. The model does not just guess the next price; it simulates the actual process of orders arriving, being matched, and updating the market, creating a realistic replay of how the financial world might behave.

The core innovation of M3 is its ability to handle two very different types of information simultaneously: the irregular, chaotic stream of individual orders and the structured, multi-level view of the market's liquidity. In the past, models often struggled to connect these two worlds, treating them as isolated sequences. M3 bridges this gap by translating both the chaotic order events and the structured market snapshot into a unified language of tokens, allowing a powerful neural network to learn the deep, dynamic relationship between them. When the model generates a new order, it does so while considering the current depth of the market, and when it updates the market state, it does so based on the rules of the exchange. This creates a closed loop where the model learns how the market state influences order flow, and how order flow, in turn, reshapes the market. The researchers found that as they increased the size of the model and the amount of data it was trained on, its ability to learn improved in a predictable and consistent way, suggesting that this approach can be scaled up to handle even more complex financial systems.

To verify that the model was truly capturing the essence of the market, the researchers tested it against several well-known patterns that real markets exhibit. They found that the simulated market paths reproduced key statistical features of real trading, such as the tendency for price changes to have extreme outliers and the way volatility tends to cluster together in time. The model also demonstrated a strong ability to predict short-term price movements and future volatility, outperforming older methods that did not account for the full state of the market. Perhaps most importantly, the model proved useful for "what-if" scenarios. When the researchers simulated the effect of a large, one-sided buying or selling pressure, the model responded in a way that matched economic theory, showing prices moving in the expected direction but with a diminishing impact as the size of the order increased. This behavior, known as the square-root law of market impact, is a fundamental characteristic of real markets, and the fact that the model reproduced it without being explicitly programmed to do so suggests it has learned the underlying mechanics of liquidity.

The implications of this work extend beyond simple prediction. By providing a realistic engine for simulating market microstructure, M3 offers a new way to stress-test trading strategies and evaluate the potential risks of large trades before they are executed. The researchers showed that the model could be used to analyze how a trading strategy would perform under different market conditions or to estimate the cost of executing a large order without actually moving the market. This capability is crucial for financial institutions that need to manage risk and optimize their trading activities in an environment where every action has a reaction. While the model is currently trained on data from a specific set of stocks, the researchers believe the approach is scalable and could be adapted to other markets and asset classes. The success of M3 suggests that treating the market as a generative world model, capable of producing multiple realistic futures, is a powerful new direction for understanding and navigating the complexities of modern finance.

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

Problem Statement

Financial markets are complex adaptive systems where dynamics emerge from the interaction of heterogeneous participants, information acquisition, and liquidity conditions. A fundamental limitation in analyzing these systems is that historical market data reveals only a single realized trajectory. Consequently, many critical questions—such as how prices might evolve under different interventions, how a trading strategy performs across regimes, or how a large execution program impacts subsequent order flow—are inherently counterfactual.

Existing financial generative models often fail to address this at the microstructure level. Prior approaches typically model either Limit Order Book (LOB) states or order flow events in isolation, treating the other as auxiliary conditioning. This overlooks the tight, dynamic coupling between order events (which supply, consume, and redistribute liquidity) and LOB states (which determine execution outcomes and price impact). There is a need for a generative model that explicitly captures this coupled state-event dynamics to enable realistic, trajectory-level market simulation for forecasting, stress testing, and impact analysis.

Methodology

The authors propose M3 (Market Microstructure Model), a state-event generative foundation model designed to simulate electronic market microstructure. M3 operates within a unified token-based framework that integrates heterogeneous order events and structured LOB snapshots.

1. Data Representation and Tokenization

To handle the heterogeneity of market data, M3 employs a Vector-Quantized (VQ) Order Tokenizer:

Input Features: Order events are represented as feature vectors including relative price (anchored to the daily open price to avoid inference-time feedback loops), log-transformed volume, inter-arrival time, order action (add/cancel), and side (bid/ask).
Vector Quantization: Instead of discretizing attributes independently (which leads to a multiplicative explosion of the vocabulary), M3 uses a causal encoder-decoder architecture with a VQ bottleneck. Continuous order features are mapped to a compact, non-uniform discrete codebook ( $K=32,768$ ).
Reconstruction: The tokenizer ensures high-fidelity reconstruction of order events, specifically preserving price trajectories relative to the daily open, which avoids the accumulation of errors seen in mid-price-relative baselines during autoregressive generation.

2. Model Architecture

M3 utilizes an Auto-Regressive Transformer conditioned on the initial market state:

LOB Encoder: The initial LOB snapshot (top 10 levels of bid/ask depth) and auxiliary state variables are processed by a lightweight Transformer encoder. This generates a sequence of continuous "prefix tokens" (state tokens) that summarize the liquidity environment.
Causal Modeling: The model concatenates these LOB state tokens with the sequence of order event tokens. A causal Transformer backbone then predicts the next order token based on the joint history of the initial liquidity state and preceding order events.
Simulation Loop: During inference, generated order tokens are decoded into executable orders and processed by a deterministic Matching Engine. The engine updates the LOB state according to exchange rules, and the resulting trajectory serves as the basis for the next prediction step. This creates a closed-loop simulation where the model learns state-conditioned order flow, and the engine realizes event-induced state transitions.

3. Training and Scaling

M3 is trained on large-scale, real-world order-level data from the CSI 300 and CSI 500 indices (covering ~800 equities from Jan 2024 to Nov 2025, totaling ~31.9 billion order events). The training objective is standard next-token prediction. The authors empirically study the model's scaling behavior, fitting a Chinchilla-style scaling law to characterize how validation loss varies with model size ( $N$ ) and training token budget ( $D$ ).

Key Contributions

Unified State-Event Framework: M3 introduces a foundation model that represents both order events and LOB states within a single token-based framework, explicitly modeling their coupled dynamics rather than treating them as separate sequences.
Scalable Market World Modeling: The authors demonstrate that market modeling follows predictable scaling laws. They fit a Chinchilla-style scaling law ( $L(N, D) = E + \alpha N^{-\beta} + \gamma D^{-\delta}$ ) to the financial domain, showing that loss decreases predictably with increased model size and data volume.
Empirical Validation on Downstream Tasks: M3 is validated across multiple simulation tasks, demonstrating its ability to reproduce market stylized facts, support predictive diagnostics, and enable counterfactual analysis.

Experimental Results

Stylized Facts Reproduction

M3-generated trajectories successfully reproduce key statistical properties of financial markets:

Heavy Tails and Aggregational Gaussianity: The distribution of mid-price log-returns exhibits heavy tails at short horizons and converges toward Gaussianity at longer horizons, matching empirical data more closely than Zero-Intelligence (ZI) or Compound Hawkes Process baselines.
Volatility Clustering: The model reproduces the weak linear autocorrelation of raw returns and the persistent positive autocorrelation of absolute returns.
Metrics: M3 achieves lower Kolmogorov-Smirnov statistics and Wasserstein distances compared to baselines across multiple time horizons (10s to 300s).

Predictive Diagnostics

Mid-Price Prediction: Conditioning on the initial LOB state significantly improves directional accuracy for short-horizon price movements compared to baselines without LOB input or the DeepLOB model. The model effectively distinguishes between "Up," "Flat," and "Down" regimes.
Volatility Prediction: The model can forecast future realized volatility. When prompts are grouped by predicted variance quintiles, the realized volatility of the corresponding future paths increases monotonically, indicating the model captures distinct market volatility states.

Counterfactual and "What-If" Analysis

Directional Stress Testing: When subjected to controlled one-sided order-flow pressure (e.g., amplifying buy-side add events), the simulator produces economically consistent mid-price responses. Buy stress raises prices, and sell stress lowers them, with the impact scaling sublinearly.
Market Impact (Square-Root Law): In TWAP (Time-Weighted Average Price) meta-order simulations, M3 exhibits a concave price impact scaling. The estimated log-log slopes for buy and sell meta-orders are approximately 0.53 and 0.52, respectively, closely matching the theoretical square-root benchmark ($0.5$) with high $R^2$ values ( $>0.8$ ).

Significance and Claims

The paper claims that M3 establishes a scalable foundation-model paradigm for counterfactual market simulation at the microstructure level. By successfully modeling the feedback loop between order flow and liquidity, M3 moves beyond simple point prediction to generate executable, realistic market trajectories.

The authors assert that this approach provides a useful foundation for:

Simulation-based analysis: Generating diverse market scenarios for risk evaluation.
Strategy development: Testing trading strategies across different market regimes and interventions.
Market impact estimation: Quantifying the cost of large trades under varying liquidity conditions.

The work suggests that order-level generative world models can capture the complex, non-linear dynamics of financial markets, offering a promising direction for studying market microstructure beyond the limitations of single-trajectory historical data.