Synthetic Data and Generative Models for Markets

Synthetic market data is generated rather than observed: simulated return paths, order books, trade sequences, or alternative-data records. It can expand scarce training samples, support stress testing, and let researchers share data without exposing raw records. But a generator that creates attractive charts is not necessarily useful. For trading, it must preserve the distributional features that determine decisions, costs, and risk—including dependencies that are rare precisely when they matter most.

Generative modeling belongs downstream of data hygiene and machine learning for trading. It cannot repair look-ahead leakage, survivorship bias, or an economic hypothesis with no signal.

Match the generator to the task

Task	Useful synthetic object	Critical property
Portfolio stress	multivariate return paths	tails and changing correlation
Execution research	order events / LOB states	spread, queue, and impact dynamics
Rare-event classification	labeled feature rows	conditional class structure
Privacy-controlled collaboration	tabular customer records	disclosure risk and utility
Policy training	market environment	response to the agent’s actions

Classical block bootstrap, GARCH, copulas, and regime-switching models remain strong baselines. GANs, variational autoencoders, normalizing flows, and diffusion models can model richer joint distributions, but their flexibility increases validation burden. An unconditional daily-return generator cannot answer an intraday market-making question; a model trained on passive data cannot faithfully simulate a strategy that changes order flow.

Stylized facts are acceptance criteria

Before comparing strategies, test the synthetic sample against held-out real data. Market realism is multivariate and temporal.

checks = {
    'mean': returns.mean(),
    'volatility': returns.std(),
    'skew': returns.skew(),
    'kurtosis': returns.kurtosis(),
    'acf_abs_return_lag1': returns.abs().autocorr(1),
    'tail_1pct': returns.quantile(0.01),
}

Extend this with volatility clustering, leverage effect, cross-asset correlation conditional on volatility, drawdown length, autocorrelation of order signs, spread distribution, and intraday seasonality. Report confidence intervals across generated seeds. A generator can match marginal returns while erasing joint crashes; it can match average spread while failing to produce liquidity withdrawal during volatility spikes.

Use train-on-synthetic, test-on-real and the reverse as task-oriented checks. If a risk classifier trained on synthetic examples fails on held-out real examples, the generator has missed decision-relevant structure. Conversely, a discriminator’s inability to tell samples apart is insufficient: it may focus on easy dimensions while ignoring rare tail dynamics.

Conditional generation and regimes

Market distributions are conditional on volatility, session, macro state, asset type, and liquidity. Supply those variables to the generator or generate them jointly. A conditional model can create scenarios such as “elevated volatility, widening credit spreads, and a 3% equity gap,” which is more useful than random paths.

Never condition on variables unavailable at simulation decision time. For instance, a synthetic earnings-surprise feature must respect the release timestamp and the historic coverage of its source. Alternative data is particularly vulnerable: a generative model trained on today’s cleaned panel can reproduce revisions or entities unavailable in the past.

The closed-loop problem

Most historical data was generated by other participants, not your policy. A learned market simulator may be accurate for passive replay yet fail when an agent places larger or more aggressive orders. This is distribution shift under intervention. For execution, combine synthetic experiments with conservative impact models, capacity limits, and shadow-mode tests. Do not certify a reinforcement-learning strategy solely because it performs in its own learned simulator.

A useful hierarchy is: use synthetic data for pretraining and hypothesis stress; use real untouched data for model selection; use paper or small-capital deployment for final evidence. Live trading deployment and monitoring remains the last gate.

Leakage, privacy, and governance

Synthetic does not automatically mean private. Generators can memorize rare records, and membership-inference attacks can reveal whether a record was in training. Evaluate nearest-neighbor similarity, holdout disclosure tests, and privacy budgets where relevant. Keep training data provenance, generator code and seed, conditioning variables, and validation results in a model card.

For market research, version generated datasets like any derived feature. Label them synthetic, prevent them from mixing with evaluation data, and make seed variation part of robustness reporting. The goal is decision support, not an artificial enlargement of a backtest until it looks statistically significant.

Key takeaways

• Define the decision task before choosing a generator or a realism metric.
• Validate tails, temporal dependence, and conditional cross-asset behavior—not just charts and moments.
• Test utility on held-out real data and distrust closed-loop execution conclusions.
• Version synthetic artifacts and assess privacy, leakage, and seed sensitivity explicitly.