GAN-Diffusion Framework Generates Realistic Financial Time-Series

A new generative framework combining GANs and diffusion models produces high-quality synthetic financial time-series data. The approach addresses data scarcity and enables counterfactual market scenario generation. The framework introduces CoMeTS-GAN (Correlated Multivariate Time Series GAN), a conditional GAN that jointly generates mid-price and volume time-series for correlated stocks. This GAN is then integrated into diffusion models to enhance correlation realism. The work is published on arXiv (2605.27113) and targets financial institutions facing data limitations.