|
|
In recent years, considerable effort has been invested into research on analyzing the build-up of vulnerabilities for banks and timely predicting possible bank distress events that could impact the soundness of the European banking system. The motivation behind this lies with the recent global financial crisis of 2007-2009 and the subsequent sovereign debt crisis in Europe, which brought a large number of European banks to the brink of collapse and prompted government interventions that led to unprecedented bailout costs. An important contribution to the enormous impact of the financial crises came from the high interconnectivity of the European banking system, which functions as a potential propagation mechanism of bank-specific vulnerabilities, allowing for risks and contagion to flow through the system.
Much of the empirical literature proposes early-warning models based on conventional statistical modeling methods, such as multivariate logit/probit models that deliver distress predictions for individual banks. However, these models do not account for possible bank interdependencies. This paper addresses this short-coming by providing a general-purpose framework that enables combining potential contagion effects with bank distress models. In particular, the paper proposes two-step estimation where the bank failure model of Betz et al. (2014) is complemented with different contagion variables that account for possible vulnerability transmissions among banks. In such way, the vulnerability of one bank depends also on the vulnerability of its neighbours. The benchmark early-warning model to which authors relate covers mainly vulnerabilities coming from bank-specific, sector-level and macro-financial imbalances in order to predict bank distress events. Given that bank defaults are rare in Europe, the definition of distress events includes state interventions and mergers in distress besides the usual bankruptcies, liquidations and defaults.
For comparison purposes, the paper uses two types of contagion mechanisms built either on the location of banks' incorporation (country-level contagion) or an estimated banking network. The estimation of linkages among European banks is based on the extreme negative co-movements between individual bank returns, such that authors capture dependencies beyond what can be expected in normal times and in excess of what can be explained by economic fundamentals. For assessing the out-of-sample performance of different early-warning models authors use signal evaluation concepts for classification problems while accounting for the policymaker's preferences between missing distress signals and giving false alarms.
The main results of the paper confirm that models including estimated dependencies among banks outperform the benchmark model, where no vulnerability transmission is assumed. In out-of-sample evaluations, the results of the network-based contagion outperform those of simpler contagion benchmarks, such as geographically neighbouring banks. The improvement comes from better performance both in terms of missing less crises and giving fewer false signals. When the contagion variables are built using the location of banks' incorporation, there is almost no change in the results compared to the benchmark case. Authors’results are robust to a wide range of variation in model specification, such as different policymaker's preferences and forecast horizons.