Synthetic financial data from diffusion models
Diffusion models are among the largest topics in the AI journals, and synthetic data generation is growing with them. Yet only 6 of 119 recent synthetic-data papers carry a finance keyword. Could synthetic transaction and loan data let banks share data for research without exposing customers, and how close is it to the real thing?
What prompted this
Since October 2025, "diffusion models" appears on 361 papers (357 in AI journals), and 119 papers carry a synthetic-data keyword (114 in AI journals); only 6 of them carry a finance, stock, bank, trading, credit or fraud keyword. "Differential privacy" appears on 72 papers.
Questions
- Do models trained on synthetic loan or transaction data perform on real data as well as models trained on the real thing?
- Do synthetic datasets keep the rare events that matter most, such as defaults, fraud and crashes?
- How well do they protect privacy against attacks that try to find real customers in the synthetic data?
Data that could help
Public financial datasets (LendingClub loans, credit card transactions, stock returns) to train and test the generators, and standard privacy-attack benchmarks.
Next step
Train a tabular diffusion model on LendingClub loans and compare default models trained on the synthetic data with models trained on the real data.