TradeFM: A Generative Foundation Model for Trade-flow and Market Microstructure

Abstract: Foundation models have transformed domains from language to genomics by learning general-purpose representations from large-scale, heterogeneous data. We introduce TradeFM, a 524M-parameter generative Transformer that brings this paradigm to market microstructure, learning directly from billions of trade events across >9K equities. To enable cross-asset generalization, we develop scale-invariant features and a universal tokenization scheme that map the heterogeneous, multi-modal event stream of order flow into a unified discrete sequence -- eliminating asset-specific calibration. Integrated with a deterministic market simulator, TradeFM-generated rollouts reproduce key stylized facts of financial returns, including heavy tails, volatility clustering, and absence of return autocorrelation. Quantitatively, TradeFM achieves 2-3x lower distributional error than Compound Hawkes baselines and generalizes zero-shot to geographically out-of-distribution APAC markets with moderate perplexity degradation. Together, these results suggest that scale-invariant trade representations capture transferable structure in market microstructure, opening a path toward synthetic data generation, stress testing, and learning-based trading agents.

• The paper introduces TradeFM, a generative model that leverages scale-invariant features and a universal tokenization scheme to generalize across thousands of equities.
• It employs explicit trade feature engineering—including price, volume, and interarrival times—to simulate closed-loop market dynamics and reproduce key market stylized facts.
• Empirical results show TradeFM achieves 2–3x lower distributional errors compared to baselines, enhancing synthetic data generation, stress testing, and agent-based trading simulations.

TradeFM: A Generative Foundation Model for Trade-flow and Market Microstructure

Overview and Motivation

TradeFM introduces a generative foundation model paradigm for financial markets, specifically targeting the modeling of trade-flow and market microstructure dynamics. Unlike prior work reliant on full limit order book (LOB) input or asset-specific models, TradeFM leverages partial observations—event streams available to most real-world market participants—and demonstrates that foundation models can generalize across thousands of equities using scale-invariant features and a universal tokenization scheme. The approach is grounded in the need for synthetic data generation, stress testing, and learning-based trading agent development, motivated by the persistent statistical properties (stylized facts) universally observed across financial markets.

Core Trade Feature Engineering

Fundamental trade features in financial markets exhibit heterogeneous distributions:

• Price features: Leptokurtic, heavy-tailed (Laplace-like), corresponding to more frequent extreme returns than a Gaussian model predicts.
• Volume features: Heavy-tailed, empirically following a power-law, across liquidity tiers.
• Interarrival times: Exponential distribution, capturing the asynchronous nature of trade events.

  Figure 1: Trade feature distributions. Canonical distributions for core trade features, conditioned on liquidity. Price features are leptokurtic (Laplace); volume follows a heavy-tailed power-law; and interarrival time is exponential.

To enable generalization across assets and markets with vastly differing scales and liquidity regimes, TradeFM employs explicit scale-invariant representations: log-transformed volumes, normalized price depths (basis points vs. mid-price), and interarrival times, combined with a robust EW-VWAP mid-price estimator to overcome missing LOB context. These engineered features exhibit stationarity over time and assets, supporting OOD generalization and cross-market transfer.

Universal Tokenization Methodology

TradeFM's tokenization aggregates multi-feature trade events into a univariate discrete sequence suitable for Transformers. Bin calibration uses quantile-based binning for price features (fine resolution at the center) and logarithmic bins for volume and time (capturing wide dynamic range), balanced to mitigate capacity wastage on rare tokens.

Figure 2: Calibrated bin edges. Price features (top) use quantile-based binning for high resolution near the mean; volume and time (bottom) use logarithmic bins to capture their wide dynamic range.

Each trade event is encoded as a mixed-base composite integer, combining action, side, price, volume, and time indices into a single token. Contextual features (liquidity, price change relative to open, participant/market-level indicator) are handled separately to enable controllable and conditional generation.

TradeFM Architecture and Closed-Loop Simulation

The model is a 524M-parameter decoder-only Transformer, employing GQA and RoPE enhancements for long-range dependency modeling and inference efficiency. A tabular embedding layer combines feature projections for multi-feature token input.

Validation occurs via closed-loop simulation, integrating a deterministic market simulator that executes TradeFM's predicted trades, updates LOB and mid-price estimates, and feeds back market state for subsequent model steps. This architecture is designed to evaluate realism via stylized fact reproduction and enables interaction for stress testing and agent training.

Figure 3: Closed-loop simulation architecture. TradeFM predicts a trade, the Market Simulator executes it, and the updated market state is fed back to the model.

Stylized Fact Reproduction and Empirical Evaluation

TradeFM's generative rollouts—conditioned on real market history and executed in simulation—reproduce key stylized facts:

• Near-zero return autocorrelation: Consistent with efficient market hypothesis.
• Volatility clustering: Slow decay of absolute return autocorrelation, indicates long-memory effects.
• Heavy-tails and aggregational Gaussianity: High kurtosis for short intervals, converging to normality over longer horizons.

Quantitatively, TradeFM achieves 2–3x lower Kolmogorov-Smirnov and Wasserstein distances to real log return distributions compared to Compound Hawkes and ZI baselines, demonstrating superior emergent behavior fidelity.

Figure 4: TradeFM model validation. Simulated returns exhibit: (Left) near-zero autocorrelation, (Middle) slowly decaying autocorrelation of absolute returns (volatility clustering), and (Right) heavy tails and aggregational Gaussianity.

Distributional Fidelity and Controllability

Order-level analyses reveal that TradeFM-generated volumes and interarrival times match power-law and exponential distributions in real data. Across multiple microstructure metrics (spread, depth, imbalance, volumes), TradeFM exhibits consistently lower distributional error than baselines, except for bid-ask spreads, where Hawkes models perform marginally better (due to explicit modeling of arrival dynamics).

TradeFM’s output is controllable via conditioning tokens for context (market-level vs. participant-level, liquidity bin), with feature variances modulating in expected directions. This controllability allows for tailored generation under scenario-driven stress tests or agent-based simulations.

Figure 5: Controllability experiments. Standard deviation of generated volumes and interarrival times, conditioned on liquidity ($i_l$) and observation level ($I_{MP}$). The model produces distinct order flow, demonstrating controllable generation.

Temporal and Geographic Generalization

Scale-invariant trade encoding yields stationarity of core tokenized features over time, enabling the model to maintain fidelity in distribution drift regimes. Zero-shot evaluation on APAC markets (China, Japan) yields moderate perplexity degradation and substantial overlap with US distributions, despite stark microstructural differences (batch auctions, price limits, wider spreads), substantiating universal principle learning.

Figure 6: Zero-shot geographic generalization. Perplexity distributions for TradeFM (trained on US equities) evaluated on held-out US, China, and Japan data (January 2025) demonstrate cross-geography robustness of scale-invariant features.

Implications and Future Directions

TradeFM's demonstration of scale-invariance, generalization across assets and geographies, and closed-loop realism in market simulation has several practical and theoretical implications:

• Synthetic data generation: Supports privacy-preserving analysis, backtesting, and augmentation of sparse datasets.
• Impact analysis and stress testing: Enables regulator/researcher-driven counterfactuals, response to anomalous order flow, and systemic risk assessment.
• Multi-agent simulation and RL training: Provides a high-fidelity background market for reinforcement learning agents, adaptable to different market regimes and participants.
• Model scaling and universality: The scaling law analysis shows compute-optimal performance and transferability, positioning TradeFM as a domain-adapted foundation model for financial microstructure.

TradeFM's architecture, tokenization pipeline, and empirical validation suggest that universal market structure can be captured from partial observations by large-scale generative models, and opens avenues for further integration of explicit microstructure theory, hybrid modeling, and domain-adapted scaling.

Conclusion

TradeFM advances the frontier of foundation models in finance by providing a generative, scale-invariant, and controllable approach to trade-flow and microstructure modeling. Its empirical results validate the synthesis of universal latent structure in markets, robust generalization, and practical utility for simulation, synthetic data, and agent training. Further exploration of hybrid approaches, domain-specific scaling, and downstream tasks in trading and risk management is warranted for continued progress in AI-augmented financial systems.