High-Quality Synthetic Financial Time-Series using a GAN-Diffusion Framework

Researchers present CoMeTS-GAN, a hybrid generative framework combining GANs and diffusion models to create realistic synthetic financial time-series data that accurately reproduce stock market stylized facts and inter-asset correlations. The approach addresses data scarcity challenges for financial institutions while improving upon existing general-purpose generative architectures.

The development of synthetic financial data generation represents a critical advancement for quantitative finance and risk management. Financial institutions increasingly require realistic market simulations for stress testing, algorithm validation, and regulatory compliance, yet generating data that captures the complex statistical properties of real markets remains technically challenging. This research tackles that problem by integrating two complementary generative approaches—conditional GANs for multi-asset correlation modeling and diffusion models for enhanced quality control.

The framework's innovation lies in its dual-architecture design. CoMeTS-GAN first learns joint distributions of correlated assets, capturing how price movements and trading volumes interact across securities. The GAN's internal critic component then serves as a quality monitor within the diffusion process, ensuring generated synthetic data maintains realistic correlation structures rather than producing statistically valid but economically nonsensical scenarios. This is particularly valuable because many synthetic datasets fail to preserve cross-asset dependencies that are fundamental to portfolio behavior.

For the financial industry, this development has practical implications for machine learning model development, where training data scarcity often limits algorithm sophistication. Synthetic data enables backtesting across more diverse market scenarios without requiring decades of historical records. The approach also supports privacy-preserving data sharing among institutions—synthetic datasets reveal statistical patterns without exposing proprietary trading information.

The research demonstrates superior performance against existing generative models in capturing stylized facts like volatility clustering and fat tails. Future applications may extend to derivative pricing, portfolio optimization, and regulatory scenario analysis. As financial AI systems become more sophisticated, reliable synthetic data generation becomes increasingly critical infrastructure.

• →Hybrid GAN-diffusion framework successfully generates realistic synthetic stock market data while preserving inter-asset correlations.
• →CoMeTS-GAN's critic module acts as a quality control mechanism, ensuring generated time-series maintain learned market structure.
• →Addresses critical financial industry need for privacy-preserving synthetic data in model training and stress testing.
• →Framework captures stylized facts like volatility clustering better than existing general-purpose generative architectures.
• →Lightweight implementation enables practical deployment for risk management and algorithmic backtesting applications.