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

Maxime Kawawa-Beaudan; Srijan Sood; Kassiani Papasotiriou; Daniel Borrajo; Manuela Veloso

cs.LGFeb 27, 2026

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

Authors: Maxime Kawawa-Beaudan, Srijan Sood, Kassiani Papasotiriou, Daniel Borrajo, Manuela Veloso

Organizations: J.P. Morgan AI Research, New York, NY, USA

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.

Figures & tables

Model	Input	Assets	Params	Zero-Shot
DeepLOB	Full LOB	5	60K	No
LOBS5	Full LOB	2	6.3M	No
MaRS	Full LOB	500	$\sim$ 1B	No
TradeFM	Event stream	$>$ 9K	524M	Yes

Table 1: Comparison with related microstructure models. TradeFM uniquely combines partial observability, large-scale multi-asset training, and zero-shot geographic generalization.

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.

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.

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.

$\Delta t_{r}$	Kolmogorov-Smirnov			Wasserstein
	ZI	Hawkes	TradeFM	ZI	Hawkes	TradeFM
10	0.376	0.039	0.013	0.0006	0.0004	0.0001
30	0.439	0.075	0.024	0.0015	0.0012	0.0002
60	0.429	0.101	0.035	0.0027	0.0023	0.0004
120	0.429	0.098	0.043	0.0043	0.0043	0.0007

Table 2: Log return distributional fidelity. Mean K-S and Wasserstein distances from real data across return intervals

\Delta t_{r}

(sec), averaged over 9 assets, 3 liquidity tiers, and 9 held-out months. TradeFM achieves 2–3

\times

lower K-S distance than Compound Hawkes.

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.

Quantity	Kolmogorov-Smirnov			Wasserstein
Quantity	ZI	Hawkes	TradeFM	ZI	Hawkes	TradeFM
Spreads	0.400	0.218	0.238	0.375	0.302	0.400
Interarrival Times	0.651	0.515	0.281	0.415	0.626	0.318
Price Depth	0.436	0.281	0.169	0.390	0.348	0.339
Order Book Imbalance	0.237	0.155	0.142	0.200	0.165	0.099
Bid Volume	0.460	0.296	0.386	0.616	0.278	0.130

Table 3: Distributional fidelity of generated order flow. Mean K-S and Wasserstein distances across 9 assets, 3 liquidity tiers, and 9 held-out months. TradeFM achieves lowest distance on most quantities. Per-month breakdowns in Appendix D.3 , Figure 14 .

Figure 5: 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.

Figure 6: 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.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Tick-based vs. relative features. Properties of tick-based vs. relative feature construction for the sample feature $\Delta p_{t}$ , across liquidity profiles. Relative features generalize better across assets than absolute, tick-based features.

Figure 8: Mid-price estimator comparison. Our proposed EW-VWAP provides a more stable and responsive estimate than standard VWAP or EWM, closely tracking the executed fill prices across different time scales and liquidity regimes.

Time-	Time-	Asset	Avg. Daily	Midprice	Action	Side	Order	Vol.
step	stamp		Vol. (shs)	($)			Price ($)	(shs)
$\vdots$
42	09:45:30	AAPL	53,496,022	182.45	ADD	BUY	182.44	500
43	09:45:38	AAPL	53,496,022	182.48	ADD	SELL	182.50	750
44	09:45:52	AAPL	53,496,022	182.50	CANCEL	BUY	182.49	300
$\vdots$

Appendix

Table 4: Example trade sequence. Toy example of trading activity for an imaginary AAPL sequence, demonstrating the multi-feature and heterogeneous nature of our data pre-tokenization.

Model Size	Num. Train Tokens	Batch Size	GPUs	Train Time / Epoch (hrs)	Optimization
125M	2.6B	24	3xA100	17	Accelerate
250M	6.4B	32	4xA10G	29	DeepSpeed ( Rajbhandari et al., 2020 )
500M	10.7B	24	3xA100	128	Accelerate

Appendix

Table 5: Training configuration. Setup details for different model sizes, including token counts, batch sizes, hardware, and training time per epoch.

Panel of Figure 9 — Figure 9: Training dynamics. (a) Learning curve for TradeFM-500M. (b) Out-of-sample perplexity across model sizes and epochs.

Figure 10: Scaling law analysis. Test loss (negative log likelihood) on held-out data one month in advance of the training data cutoff. The black dashed line represents the power law fit to the minimum loss frontier.

Country	Number of Assets	Date-Asset Pairs	Tokens
US	6,885	81,203	857,017,219
China	4,926	68,925	37,408,529
Japan	2,932	37,235	286,476,052

Appendix

Table 6: Geographic held-out datasets. Dataset statistics for US, China, and Japan. All geographies are evaluated on Jan. 2025 data.

Figure 11: Simulator fill validation. Stylized facts of market simulator fills: (left) fill volumes; (right) lot counts. Correlations between simulated and actual fills are 0.91 for volumes and 0.98 for lot counts.

Figure 12: Feature stationarity over time. Kernel-density estimation of feature distributions from the tokenizer calibration period of Feb. 2024 to one year later in Feb. 2025. Our feature engineering successfully makes these features stationary, enabling generalization to out-of-distribution temporal regimes.

Figure 13: Tokenizer drift over months. Kolmogorov-Smirnov and Wasserstein distances between feature distributions during the tokenizer calibration month and held-out months. Raw mid-price, a non-stationary feature, is included for context. (OPD and IA Time in the legend correspond to Price Depth and Interarrival Time, respectively.)

Figure 14: Distributional fidelity over time. Wasserstein distance and Kolmogorov-Smirnov statistic of feature distributions and emergent market factors from various methods over nine held-out months.

Figure 15: Order-level distributional fidelity. (Top) Simulated order volumes follow the power-law observed in real data. (Bottom) Simulated interarrival times match the exponential distribution of real data.

Figure 16: Multi-step mid-price forecast. Rollout-based forecast for an imaginary asset initially priced at $100. The average trajectory and 50% confidence interval over 256 simulations show the model generates plausible, non-stationary market paths.

Figure 17: Counterfactual stress testing. The model’s generated price path responds realistically to injected anomalous order flow (10x normal frequency), demonstrating its utility for impact analysis.

Quantitative Finance

Quantitative finance uses mathematical models, statistics, and machine learning to value assets, measure risk, and guide trading and portfolio decisions.

See all 191 papers

What Do Foundation Models for Market Microstructure Learn? A Controlled Study with Synthetic Limit Order Books
Oct 14, 2026Fernando AceroLimit Order Book Modeling Time-Series Foundation Models
J.P. Morgan QTR AI Research, United Kingdom
Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models
Apr 18, 2026Yuchen Pan, Soung Chang LiewLLM Evaluation Financial Reasoning in Language Models
Department of Information Engineering, The Chinese University of Hong Kong
Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books
Oct 20, 2026Valentin Mohl, Alexandre Bismuth, Aramis Marti-Shahandeh +6Limit Order Book Modeling State Representation Learning
British Open-ended Learning and Discovery Lab (BOLD), Department of Computer Science, University of Oxford, Oxford, United Kingdom · Harvard University, Cambridge, MA, United States · British Open-ended Learning and Discovery Lab (BOLD), Department of Engineering Science, University of Oxford, Oxford, United Kingdom +3

Oct 14, 2026J/K move · Enter open · S save

What Do Foundation Models for Market Microstructure Learn? A Controlled Study with Synthetic Limit Order Books

Fernando Acero

J.P. Morgan QTR AI Research, United Kingdom

Limit Order Book Modeling Time-Series Foundation Models Quantitative Finance Limit Order Book Forecasting Synthetic Data Pretraining Bayesian Filtering

Open paper AskPDF arXiv

Autoregressive transformers trained on limit order book (LOB) message streams are increasingly proposed as foundation models for market microstructure. Because the generative process of real markets is unknown, these models can be judged only by the realism of their output, leaving open what they actually learn. We make the question answerable by inverting the setup: we train GPT-style models on order flow from simulators whose latent parameters are known, so that the Bayes-optimal predictor, the generating state, and the exact data entropy serve as references. Measured against them, next-token pretraining on order flow implements amortized Bayesian inference over market regimes. The learned predictive matches the exact Bayes optimum to within a millinat per token and approaches the oracle floor even on a nonstationary family whose latent state drifts within an episode, a regime that requires filtering. The latent regime is linearly decodable from frozen activations, well above baselines, so a frozen-model read-out calibrates simulator parameters more accurately than windowed maximum likelihood at a fraction of the cost. In our experiments these capabilities are not produced by scale: capacity and token volume saturate early, and it is the diversity of training regimes, not the token count, that determines generalization to unseen markets. On real NASDAQ data, zero-shot transfer is set by how well the simulator prior covers real flow: an enriched mixture nearly closes the gap to an in-domain baseline, and pretraining beats from-scratch training at every budget across held-out tickers. Known ground truth turns average next-token loss, which is uninformative near the entropy floor, into an exact comparison against the Bayes-optimal and oracle predictors.

Abstract

Quantitative Finance

Explore similar work

What Do Foundation Models for Market Microstructure Learn? A Controlled Study with Synthetic Limit Order Books

Learning to Trade Like an Expert: Cognitive Fine-Tuning for Stable Financial Reasoning in Language Models

Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books