Constrained synthetic financial data

Synthetic data is valuable when real financial records are sensitive, but it has to meet two standards at once. It must be statistically faithful, reproducing marginal distributions and relationships between columns. And it must be valid: tabular financial data carries deterministic rules between fields, and rows that violate them break downstream use.

Generative models are good at the first standard and do not guarantee the second, because they learn distributions and not rules.

1. Conditional generation. A PyTorch conditional deep learning model learns the joint structure of the table.
2. Constraint enforcement. A post-hoc enforcer applies the table's deterministic constraints to generated rows, so validity is guaranteed by construction instead of hoped for.
3. Validation. Aggregate statistics of the synthetic data are validated against the source data.

Moved from research into production use at Fidelity, with throughput and cost measured alongside statistical fidelity.

• An aggregate score summarizes many statistics. It can hide weaker fidelity in a specific column or rare segment, which is why the individual statistics matter too.
• Post-hoc enforcement guarantees validity but can slightly shift distributions near constraint boundaries.