23 March 2016 FINANCIAL STABILITY
European Central Bank
Directorate General Communications
Sonnemannstrasse 20, 60314 Frankfurt am Main, Germany
Tel.: +49 69 1344 7455, E-mail: media@ecb.europa.eu
Website: www.ecb.europa.eu
A bank-level early warning model and its uses in macroprudential policy
Read HERE the full Report First Macroprudential Bulletin published
3.1 Purpose of the analytical tool
The bank early warning model (BEWM) can be used to identify both vulnerabilities in individual systemically important banks and vulnerabilities that build up simultaneously across a number of banks at euro area or country level.
Compared to country-level early warning models, the BEWM has the advantage of providing a micro view on the build-up of vulnerabilities, which can be important when there are nonlinearities and bank heterogeneity. This model can therefore provide information about the build-up of systemic risk in the cross-sectional and the time dimension. The decomposition of bank-level distress probabilities into bank-specific, aggregate banking-sector and macro-financial factors can further support the identification of the main drivers of vulnerabilities. This can, in turn, inform macroprudential policy decisions. 3.2 Description of the analytical tool The BEWM is built around a bank-level logit model that uses bank-specific, aggregate banking-sector and macro-financial indicators as predictive variables to warn of future bank distress. The model is based on the method and dataset described in Lang, Peltonen and Sarlin (2015)30, applied to euro area banks. The logit model is estimated via the logistic LASSO (least absolute shrinkage and
3.2 Description of the analytical tool
The BEWM is built around a bank-level logit model that uses bank-specific, aggregate banking-sector and macro-financial indicators as predictive variables to warn of future bank distress. The model is based on the method and dataset described in Lang, Peltonen and Sarlin (2015)30, applied to euro area banks. The logit model is estimated via the logistic LASSO (least absolute shrinkage andselection operator) method31, where the shrinkage parameter is chosen by means of cross-validation, in order to obtain a parsimonious model specification that optimises the model’s out-of-sample forecasting performance.
The optimal signalling threshold for the logit model probabilities is derived by maximising the usefulness of the model for a policymaker who has a loss function that is defined in terms of Type I errors (missed crises) and Type II errors (false alarms), and that also accounts for the unconditional probability of events, as proposed by Sarlin (2013)32. The relative weight assigned to Type I errors is set at 0.9, which roughly corresponds to balanced preferences between Type I and II errors in the loss function framework developed by Alessi and Detken (2011)33, given that the unconditional probability of a bank being in a vulnerable state (experiencing a “pre-distress event”) is around 10% in the sample in question.
The bank distress events that are used to define the
vulnerability indicator of the model comprise
bankruptcies, defaults, liquidations, state-aid cases and
distressed mergers (see Betz et al. (2014)34 for details).
Vulnerable states are defined as the eight quarters prior
to a distress event. The model is estimated for a large
number of euro area banks using data from the first
quarter of 2000 to the last quarter of 2014. An overview
of the distribution of distress events across the various
subcategories can be found in Chart 3. Based on the
dataset and the optimal penalty parameter obtained
through cross-validation, the LASSO method
automatically identifies the variables that best predict
bank distress events over a two-year horizon.
The
optimal forecasting model contains 11 risk drivers: five
bank-specific variables, four banking-sector variables
and two macro-financial variables. In order to account
for real-time publication lags, all of the variables are
lagged by one or two quarters.
The five bank-specific variables identified as optimal
predictors by the LASSO method relate respectively to bank leverage, asset quality,
funding costs, profitability and trading activities. The four banking-sector indicators
identified as optimal predictors relate to banking sector size, the change in the loanto-deposit
ratio and the level of and change in the share of market-based funding.
Finally, the two macro-financial variables identified by the LASSO method relate to developments in residential real estate prices and government bond yields. Table 1 shows a number of in-sample and out-of-sample performance measures, which can be used to assess how well the optimal parsimonious model explains and predicts the data. Starting with the in-sample fit of the model, it can be seen that the parsimonious model seems to explain the data reasonably well. The AUROC, a measure of the global signalling performance of the model independent of policy preferences, is fairly high, at 0.847. In addition, the relative usefulness for a policymaker with a relative preference of 0.9 for Type I errors is around 53%, indicating that the model could offer considerable benefits for a policymaker who is relatively concerned about bank failures. The model only fails to signal less than one third of pre-distress events, while just 14% of calm periods are incorrectly classified as pre-distress events.
3.3 Illustrative results
The BEWM can be used to identify both vulnerabilities in individual systemically important institutions and vulnerabilities that build up simultaneously across anumber of banks at euro area or country level. Aggregate euro area and countrylevel distress probabilities can be calculated as a weighted average of the individual distress probabilities for all the banks (for which data is available) in the country or region. In addition, a decomposition of distress probabilities into contributing factors can support the analysis carried out to guide macroprudential policy. It can help to gauge which factors drive the build-up of vulnerabilities, and can direct attention to relevant areas for further investigation.
Chart 4 shows the aggregate distress probability for the euro area up to the final quarter of 2014, together with a decomposition of the factors contributing to this probability into bank-specific, aggregate banking-sector and macro-financial factors. As can be seen, the model captures the build-up of vulnerabilities prior to the global financial crisis fairly well. The model starts to issue warning signals for the euro area aggregate at the beginning of 2006. While the aggregated vulnerability for the euro area is currently well below the peaks reached before the financial crisis, it increased somewhat in the second and third quarters of 2015, partly driven by developments in Greece (not shown here).
The decomposition of distress probabilities in the first quarter of 2016 into contributing factors suggests that remaining vulnerabilities in the euro area banking sector are mainly linked to bank-specific and countrylevel banking sector factors, while macro-financial factors, such as residential real estate prices and government bond yields, are currently playing a lesser role in most countries. A further breakdown of distress probabilities reveals that the remaining bank-specific vulnerabilities are, in most cases, strongly linked to weak asset quality, highlighting the need for comprehensive action to be taken to deal with non-performing loans. Similar aggregations and decompositions can also be produced at the country level.
Chart 5 shows an example of how the BEWM can be used to identify the build-up of vulnerabilities over time for individual systemically important banks. The coloured area represents the evolution of the forward looking distress probability for a given bank. Past distress events and pre-distress events for the bank are highlighted by the shaded areas. Further information on what is driving the distress probability is provided by the high-level decomposition of the distress probability into bankspecific, aggregate banking-sector and macro-financial factors. Finally, the signalling threshold provides information as to whether the estimated distress probability should raise warning signals that could lead to further investigation into the vulnerabilities identified in respect of this particular bank. Chart 6 shows another decomposition of the latest estimated distress probability into specific driving factors (within the three main groups of distress factors). This decomposition can help to identify the priority areas that a more detailed follow-up analysis should focus on.
page source http://www.ecb.europa.eu/
Directorate General Communications
Sonnemannstrasse 20, 60314 Frankfurt am Main, Germany
Tel.: +49 69 1344 7455, E-mail: media@ecb.europa.eu
Website: www.ecb.europa.eu
