A0721
Title: Computational efficient time series generator through maximum mean discrepancy
Authors: Chu-Lan Kao - National Yang Ming Chiao Tung University (Taiwan) [presenting]
Yi-Jing Wang - National Yang Ming Chiao Tung University (Taiwan)
Abstract: Recent developments in generative adversarial networks (GANs) provide a way to augment imbalanced data by simulating data similar to the minority class. For instance, the conditional Sig-Wasserstein GAN (SigCWGAN) can generate time series data, such as stock price processes, during rare events. However, such methods are usually time-consuming, limiting their real-world applications. Therefore, an alternative method, Sig-MMD, is proposed, aiming to reduce computational time while preserving performance. The idea is to consider the maximum mean discrepancy (MMD) between path signatures, a distance measure that can be computed at a lower computational cost compared to traditional Wasserstein distances and similar metrics. The simulation studies show that the proposed Sig-MMD indeed yields results similar to those of SigCWGAN, but with significantly shorter computation time. In addition, it is observed that Sig-MMD actually outperforms SigCWGAN in capturing the marginal distribution, especially for more complex models, although SigCWGAN performs better when considering the autocorrelation structure. Applications to real financial data further support the above claims. Further analysis of hyperparameters, such as the signature order, is also provided.
