Evaluating banking crisis predictions in EU and V4 countries .

Abstract

Relying on a recently published database of financial crises, this paper assesses an early warning model for predicting banking sector distress. The exercise employs discrete choice models and a signaling approach to evaluate the performance of an existing model based on credit-to-GDP change and real house price growth in regard to predominantly post-crisis data for EU and Visegrad Group countries. As such, unbalanced panel data for 27 EU countries, spanning with annual frequency at longest the period of 2003-2017, as well as unbalanced panel data for 4 Visegrad Group countries covering at most the period 2008Q1-2017Q4 with quarterly frequency were analyzed. The results are generally in line with other empirical research featuring the same model and indicate that the model retains most of its predictive capabilities even when currently available data are used. However, the analysis identifies that the indicator of real house price growth may not be as useful of a predictor of banking crises in more recent periods for EU countries, as it might have been before the 2008 financial and economic crisis. Consequently, a simpler univariate early warning indicator approach might be sufficient for banking sector risk monitoring and management in EU and Visegrad Group countries in regard to identifying periods of distress similar to those in 2008.

Every crisis does matter: Comparing the databases of financial crisis events

“…probit (Berg & Pattillo, 1999;Frankel & Rose, 1996;Mulder et al, 2016), logit (Arregui et al, 2013;Davis et al, 2016;Davis & Karim, 2008;Domonkos et al, 2017a;Ostrihon, 2020;Valinskyt ̇e & Rupeika, 2015), or multinomial logit (Caggiano et al, 2016;Ciarlone & Trebeschi, 2005) models with random effects. Given our objective, we opt for a simple bivariate pooled logit model with random effects, as in Boyd et al (2009).…”

“…Discrete choice models belong to the most commonly used approaches to assess the predictive properties of EWSs in the literature (Berg & Pattillo, 1999; Kaminski et al, 1997). In general, studies use either linear probability models (Davis et al, 2016), probit (Berg & Pattillo, 1999; Frankel & Rose, 1996; Mulder et al, 2016), logit (Arregui et al, 2013; Davis et al, 2016; Davis & Karim, 2008; Domonkos et al, 2017a; Ostrihon, 2020; Valinskytė & Rupeika, 2015), or multinomial logit (Caggiano et al, 2016; Ciarlone & Trebeschi, 2005) models with random effects. Given our objective, we opt for a simple bivariate pooled logit model with random effects, as in Boyd et al (2009).…”