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

A research team has proposed a new generative framework that combines conditional GANs with diffusion models to produce synthetic financial time-series data that more faithfully reproduces real market behavior, according to a paper submitted in 2026 [1]. The framework, detailed in a paper posted to arXiv on 26 May 2026, introduces a component called CoMeTS-GAN, a Conditional Generative Adversarial Network designed to jointly generate mid-price and volume time-series for correlated stocks [2]. Financial institutions have increasingly turned to synthetic data to address data scarcity and to construct counterfactual market scenarios, but existing general-purpose architectures have struggled to reproduce all the statistical properties of financial time series, known as stylized facts [2]. The authors argue that their approach addresses these limitations by integrating two classes of generative methods [2]. Generative AI, the broader field encompassing this work, uses models that learn underlying patterns from training data to produce new data in response to input [3]. The prevalence of such tools has grown sharply since the AI boom of the 2020s, driven by advances in deep neural networks and large language models based on the transformer architecture [3]. Within finance, machine learning techniques have already been applied to tasks such as credit scoring and decision-making [4]. Machine learning itself relies on statistical algorithms that learn from data and generalize to unseen examples, with deep learning enabling neural networks to surpass many earlier approaches [5]. In the proposed framework, the GAN’s Critic network serves as a quality evaluation module that guides the diffusion process, enforcing learned correlation structures in the generated time-series [2]. The authors describe the resulting system as a lightweight and responsive solution for realistic stock market simulation that explicitly models inter-asset correlation structures [2]. Experimental validation against leading generative architectures showed that the framework more effectively captures the stylized facts of stock markets and models inter-asset correlations [2]. The paper was submitted through arXivLabs, a platform that allows collaborators to develop and share new features on the arXiv website [1].