StDDPM: Sequential Financial Synthetic Data Generation using T-distribution based Diffusion Model
Research Square preprint, 2025 · Submitted
Omid Tarkhaneh, Hamideh Mehri, Wanglong Lu, Farzaneh Shoeleh, Vinicius Veloso de Melo
Brief description:
This work studies sequential synthetic financial data generation using diffusion models with a Student t-distribution denoising process. The model targets privacy-preserving financial sequence generation for model development, testing, and analysis when real sensitive transaction data cannot be directly shared.
The method combines Student-t diffusion with an LSTM synthesizer and was evaluated on 1.06 million Czech banking transaction records across 4,500 accounts. The manuscript has been submitted; the linked version is a Research Square preprint, not an accepted journal article. Its statistical privacy evaluations do not constitute a formal differential-privacy guarantee. Author manuscript.