The signature-based GAN model for time series generation

The time series generation is a challenging problem, as existing generative adversarial networks (GANs) usually cannot capture the temporal dynamics well. Besides, the training of GAN models is computationally expensive and not stable. In this talk, I will present our recent work on a novel GAN framework for time series generation, which uses the principled and universal (log)-signature feature of time series to extract the temporal dependence of time series and design a more compact generator and discriminator. Numerical results show that our method improves the stability and of training computational efficiency while capturing the temporal dynamics of the observed time series.