Generative Adversarial Networks: A Novel Approach to Predictive Modeling in Finance

Abstract

This thesis examines the application of Generative Adversarial Networks within finance. GANs were initially introduced in Goodfellow, Pouget-Abadie, et al. (2014)’s paper for generating image, audio, and video samples. Since then, research in this area has continued, leading to improved models beyond the traditional framework. For this study, a Wasserstein GAN with a Gradient Penalty (Gulrajani et al., 2017) was employed to investigate its ability to generate synthetic time series data. Initially, the model was trained on data drawn from three independent and identically distributed (i.i.d.) sources, then on dependent data generated from synthetic AR(2) processes, and finally on real return data from a stock. The results were promising as WGAN-GP was able to generate data that was close to the true data in most cases. However, there were some notable differences compared to the financial data, mainly related to the model’s inability to capture the kurtosis and fat-tails of the true distribution.