Time Series Generation

Time series generation studies how to create realistic sequences that reproduce trends, periodicity, cross-variable dependencies, irregular observation patterns, and rare events. Common approaches include diffusion and flow-matching models, GANs and VAEs, autoregressive token models, and continuous-time methods such as Gaussian processes and neural differential equations; spectral and wavelet representations can help preserve fine and long-range structure. Generated data can augment forecasting and classification, support privacy-sensitive sharing, and enable controlled simulations in fields such as medicine, finance, energy, and neuroscience. A central challenge is balancing distributional realism and controllability with faithful temporal dynamics, extremes, and missingness, while evaluation increasingly considers downstream utility and privacy as well as statistical similarity.