Multivariate Inference for Dynamic Systemic Risk Measures

This paper introduces a system perspective on inference for standard dynamic systemic risk measures. In particular, we provide a multivariate GARCH-type framework to analytically quantify confidence and prediction intervals of marginal expected shortfall (MES) and delta conditional value-at-risk (∆CoVaR) type measures in a multivariate system setting. We establish the asymptotic properties for estimators of both types of measures and show how the estimation uncertainty in the multivariate case can be decomposed into dynamic univariate marginal and potentially time-varying dependence components. Our finite sample study shows good performance of our methodology for estimation and prediction risk in cases with constant and dynamic dependence. In an empirical application, we provide new results for the analysis of systemic risk contributions of 50 large US financial institutions in a recent period from the financial crisis to the COVID crisis (2010-2020). Our findings highlight the critical role of comprehensive multivariate forecast intervals in systemic risk assessment, particularly with regard to the interpretation of systemic risk rankings.