OB-GAN: a novel approach to modeling oscillatory latent dynamics for enhancing realistic time-series generation

Abstract Time series data augmentation remains challenging due to complex temporal dependencies, regime shifts, and oscillatory dynamics that conventional generative models often fail to capture. While GAN-based approaches improve distributional realism, they frequently suffer from mode collapse, gradient instability, and poor long-range temporal coherence. In this study, we propose Oscillatory Bifurcation Generative Adversarial Network (OBGAN), a novel framework that integrates Hopf-inspired bifurcation dynamics and controlled latent oscillations to enhance temporal structure preservation. By embedding oscillatory mechanisms within the generator and incorporating a bifurcation-aware regularization strategy, OBGAN promotes stable training while improving higher-order temporal consistency. We evaluate OBGAN against multiple baselines, including Vanilla GAN, WGAN-GP, VAE-based models, and recurrent GAN variants, across diverse datasets, such as Plant Humidity, Jena Climate, and USDT-USD cryptocurrency time series. Quantitative analysis using FID, MMD, Wasserstein distance, JS divergence, KS statistic, autocorrelation, and reconstruction metrics demonstrates that OBGAN consistently enhances temporal coherence and regime-transition fidelity while maintaining competitive distributional accuracy. Ablation studies further confirm the contribution of oscillatory coupling and multiscale components to stability and performance. This indicates that bifurcation-aware oscillatory modeling provides a dynamic mechanism for improving generative time series augmentation, particularly for datasets characterized by nonlinear dynamics and long-range temporal dependencies. The complete implementation is available at: https://github.com/avokhuese/Oscillatory-BifurcationGAN .