Default Risk Model Using Conditional Generative Adversarial Net by High-Dimensional Financial Time-Series Generation

Generative adversarial nets (GAN) have shown a remarkable success in many fields of applications to produce realistic data artificially, and the works on images dominate most of them. In this article, the authors propose to apply a set of GANs, including GAN, CGAN, WGAN, CWGAN, RGAN and RCGAN, to generate high-dimensional financial time-series data with proposed evaluation metrics. They show that the conditional GAN (CGAN) model demonstrates robust and satisfactory performance in generating high-dimensional time series by various testing sets, including using the data of the COVID-19 period as the test set. The authors aim to develop a default risk model to calculate the jump-to-default risk. They propose to use the CGAN as an alternative to the industry practice based on the Monte Carlo (MC) simulation. Their studies on two real-data portfolios demonstrate the CGAN model, as a cutting-edge nonparametric approach, addresses the drawbacks of the industry standard approach and generates more realistic returns with convincing defaults and smoother tail loss distribution. Overall, the CGAN model demonstrates its ability to generate high-dimensional financial time series and shows a strong performance to model default risk. It proves a promising alternative to the popular MC simulation.