The Estimation of Value-at-Risk and Expected Shortfall Based on Deep Generative Models

Value-at-Risk and Expected Shortfall are the most popular risk measures in the world. Financial institutes strongly benefit from the more accurate estimation of both measurers,due to substitutional cost savings and enhanced resilience to handle extreme scenarios. For a long time, simple approaches were used to estimate Value- at-Risk and Expected Shortfall. Their performance was acceptable but left room for improvement. In recent years, with the rapid development of Deep Learning, a specific type of models, called Deep Generative Models, were introduced for this problem. While there have already been some attempts to use Deep Generative Mod- els for Value-at-Risk and Expected Shortfall estimation, in this thesis the completely unexplored approach of the usage of Denoising Diffusion Probabilistic Models, the subtype of Deep Generative Models, is researched. Results show promising perfor- mance of Denoising Diffusion Probabilistic Models in the tasks of Value-at-Risk and Expected Shortfall estimation, as well as potential directions for research and further improvements.