Systemic risk in empirical interbank networks: disentangling balance sheets and network effects

We study how the level of systemic risk empirically observed in an interbank network differs from the expected risk implied by the balance sheets composition of the participating banks, abstracting from the specific configuration of bilateral exposures. To this end, using different contagion models (i.e. interbank asset valuation functions), we measure observed systemic risk on e-MID network data from 2005 to 2012 and compare it with the expected risk generated by a null model: an ensemble of interbank networks whose links represent possible market clearing configurations. This counterfactual is obtained through a maximum-entropy approach that preserves banks' balance sheet variables and diversification patterns, while randomizing the configuration of exposures. We find that the aggregate levels of observed and expected systemic risks are generally consistent but diverge significantly during turbulent periods–specifically following the default of Lehman Brothers (2009) and the VLTRO implementation by the ECB (2012). At the individual level, however, banks often appear more or less risky than implied by their balance sheet alone, due to their position in the network. Our results show, on the one hand, that balance sheet information embedded in a suitable maximum-entropy framework provides reliable estimates of aggregate systemic risk; on the other hand, they highlight the importance of detailed network information for accurate stress testing of individual banks, particularly after systemic events.