A GAN approach to extreme financial events: Proposal and applications

Modeling rare events lies at the heart of financial risk management. However, the scarcity of these tail occurrences makes traditional machine learning methods, which demand plentiful data, ill-suited for capturing extreme market outcomes. This thesis investigates integrating Generative Adversarial Networks (GANs) with Extreme Value Theory (EVT) through an adaptation of the Generalized Pareto GAN (GPGAN) framework for financial applications. Building upon the existing precipitation modeling approach, this research presents the first systematic financial adaptation using a triple representation strategy across Standard GAN, DCGAN, and LSTM architectures. Three datasets enable comprehensive evaluation: synthetic multivariate Student-t distributions, ARMA-GARCH standardized residuals, and raw sector returns. The framework processes financial data as vectors, images, and temporal sequences respectively, evaluated across multiple extremeness thresholds. Results reveal significant limitations in GPGAN's transferability to finance and highlight practical constraints of applying adversarial frameworks to extreme value generation, while establishing foundations for future developments in deep learning and financial extreme value modeling.