In the financial markets, Central Clearing Counterparties (CCP) serve a crucial role as a guarantor between large amounts of trade, thereby assuming significant risks. To ensure the resilience and stability of CCPs more complex methods for stress testing are required to better understand the potential losses a CCP could accrue during extreme but plausible market scenarios. This thesis presents a method utilizing a hybrid approach of Variational Autoencoder (VAE) and Wasserstein Generative Adversarial Networks with a Gradient Penalty (WGAN-GP) for constructing a scenario generator capable of generating new stock and options price changes for stress testing purposes.
This thesis work can be divided into four different parts. Firstly, implied volatilities were obtained from the market prices of options, these implied volatilities were then utilized to construct an implied volatility surface (IVS) for each trading day. Secondly, the changes in the IVS between days were used to train a Variational Autoencoder (VAE) to reconstruct these changes with minimal reconstruction error while maintaining a meaningful lower-dimensional feature representation. The third step was to train a WGAN-GP model to generate synthetic latent feature observations inseparable from the original data from a distributional perspective. Finally, the output from the WGAN-GP model was decoded to get new synthetic scenarios of option and stock price changes. A subset of the most stressed scenarios was then selected to represent extreme but plausible market scenarios.
Comparing the original data with the VAE’s reconstruction indicates that the data characteristics were well retained, and inspecting the synthetic latent space observations generated by the WGAN-GP reveals features similar to the historical data. Applying both historical and synthetic scenarios on portfolios led to comparable losses in general. But, for the worst losses of extreme but plausible scenarios a systematic underestimation was observed, indicating that the hybrid approach creates a less extreme reality the historical data shows. Nevertheless, when generating increasing amounts of synthetic scenarios increasing losses were observed, suggesting that the presented method could identify new potential risks and thereby reduce the potential exposure of a CCP.