Titre : Machine learning data augmentation for the sensitivity analysis of extremes

Abstract : Recent interest in studying extreme events stems from their increasing frequency. Yet, identifying the main factors impacting their magnitude remains challenging. The Shapley values emerge as a tool from global sensitivity analysis to tackle this challenge. However, the scarcity of data poses significant challenges to overcome. Therefore, we take advantage of data augmentation procedures based on machine learning methods. More precisely, our approach involves advanced modeling to capture observations exceeding a high threshold. Vine copulas are then employed to explore interdependencies among variables, enabling the generation of simulated values crucial for data augmentation. Given the generated sample, we apply Shapley values to establish variable importance. We apply our methodology in the context of analyzing extreme operational losses at UniCredit Bank.