Quant GANs propose a novel data-driven approach for generating realistic financial time series using generative adversarial networks (GANs) combined with temporal convolutional networks (TCNs). The paper provides the first rigorous mathematical definition of TCNs, which are used to model long-range dependencies such as volatility clustering. The generator is designed as a Stochastic Volatility Neural Network (SVNN), consisting of volatility and drift TCNs and an innovation network, ensuring that the generated process allows transition to its risk-neutral distribution and can be constrained to conditionally normal log returns. Theoretical results prove that all moments of the SVNN process exist, addressing heavy-tail properties via Lambert W transformations.

Numerical experiments on the S&P 500 index from 2009 to 2018 demonstrate that Quant GANs generate synthetic paths that closely match empirical distributional properties for both small and large lags. The generated data captures key stylized facts including volatility clustering, leverage effects, and serial autocorrelations in absolute returns, outperforming a GARCH(1,1) model. The work bridges pure data-driven and model-driven approaches, offering a flexible and powerful tool for financial time series generation with applications in risk management, trading strategy robustification, and data augmentation.