Modeling systemic risk in the banking industry via a network of extremal dependencies - An approach based on Generalized Pareto regression and the LASSO

Abstract

In this master thesis, we investigate the interdependency among 26 global banks, using the daily closings. We especially focus on extreme negative results and, therefore, use a certain threshold to only retain exceedances from the original data. The aim is to find, for each bank from the sample, the most significant subset of covariates explaining its daily returns thanks to past returns. To do so, the proposed methodology is based on Generalized Pareto Regressions and the Least Absolute Shrinkage and Selection Operator (LASSO). If the threshold is selected sufficiently high, we can consider our analysis to be part of the Extreme Value Theory and we can apply Peaks-Over-Thresholds methods. Consequently, it is possible to model the exceedances by a Generalized Pareto Distribution with scale and shape parameters. The selection of significant covariates is performed by the penalizing technique LASSO. Only significant regression coefficients are strong enough to not be shrunk exactly to zero by the penalty. While both of the time-varying parameters are functions of the covariates, the shape parameter will be focused on drawing conclusions. We are truly convinced that the result of this work could be used to create an additional systemic risk measure. This should especially target banks exhibiting asset levels above $1000B. and banks originating from Europe, as these banks are more exposed and interconnected.