Should stress test models be country specific?
Why UK banks would fail the 2014 tests based
on Spanish models.
Burkhard Heppe, Open Source Investor Services B.V. (OSIS)
[email protected], 17 December 2014, Version 0.1
1 Abstract
We explore whether results form recent supervisory stress tests across Europe reflect a level
playing field where local supervisors approved broadly consistent models. Did banks take
advantage of different regulatory regimes making stress test results and capital ratios hard
to compare across countries? We find that banks use widely different sensitivities to macroeconomic variables like GDP, house price changes and the unemployment rate in the recently
published EU-wide stress test conducted by the European Banking Authority (EBA) and the
European Central Bank (ECB). Based on the detailed and unprecedented disclosure of EBA
we estimate implied macro sensitivities for a large number of bank, country and exposure
class combinations and find significant differences between countries and banks. We find
that UK or German banks use smaller sensitivity parameters on average compared to the
more conservative parameters used by the beleaguered Italian or Spanish banks. We look at
historical performance data in different countries to check whether the observed differences
may be justified. As a robustness check for the recent stress test, we estimate the credit
impairments of UK and German banks using the more conservative parameters that we estimated for Spain or Italy. We estimate that the four major UK banks would have failed
the EBA stress test (and also the more stringent Bank of England stress test) using Spanish
average sensitivity parameters indicating that the playing field may not have been level.
2 Introduction
The recent financial crisis that started in 2008 has been felt all over Europe, but the relative magnitude of the impact in terms of asset price declines (especially for real estate),
unemployment rates and loss of output has varied widely with southern European countries
suffering more and northern countries like Germany, the UK or Scandinavia suffering less.
The burst of the house building bubble in Spain and Ireland resulted in large increases in
unemployment that persist to this day whereas the recession of 2009 was short lived in Germany achieving a balanced federal budget in 2014 and UK house price indices have recently
hit all time highs. Moreover, credit risk indicators like the nonperforming loan (NPL) ratio
1
have increased several fold in Spain and Ireland since 2009 whereas they have been quite
stable in Germany and the UK. Hence, it appears that credit dynamics are indeed highly
country specific.
Based on the data recently disclosed by EBA for the 2014 EU-wide stress test we estimate
simple stress testing models and their key sensitivities to macro economic variables. We find
that UK or German banks use much smaller sensitivity parameters on average compared to
the more conservative parameters used by Italian or Spanish banks. Using publicly available
aggregate measures of credit risk we then investigate whether the observed model discrepancies between large European economies can be justified. As noted before (Heppe 2014b),
there is an important analogy between macro sensitivities and the asset correlation parameter in the Basel II capital requirements for which Basel prescribes the same parameters
across all countries. Correlation and macro sensitivity parameters are hard to estimate unless long historical time series are available spanning several economic cycles. In practice for
the bottom up approach that was followed by EBA, such data are rarely available resulting
in uncertain model parameters. The annual Pillar 3 reports for many banks include some
information about the length of historical data available to estimate regulatory stress test
models. From looking at a small sample of Pillar 3 reports we can see that banks in many
instances have less than 10 years of data available potentially resulting in models with a high
degree of uncertainty. In light of this, we argue that a complementary and transparent top
down approach with consistent and perhaps more homogenous model assumptions across
countries and banks could be useful to achieve a more level playing field for banks across
Europe.
First, we provide a brief overview about the academic literature on stress testing credit risk in
multiple countries. Second, we summarize our analysis of the EBA EU-wide stress test data
and the estimation of implied macro sensitivities. Third, we compare macro sensitivities
for a number European countries and banks. Fourth, to benchmark the implied models
from the EBA data, we estimate country specific stress test models for the UK and Spain
against longer histories of aggregated credit performance measures. Fourth, we estimate
credit impairments for banks using more homogenous model parameters across countries.
In particular, we calculate the CET1 ratio for some UK and German banks using more
conservative macro sensitivities similar to the ones used in the EBA stress test in Italy and
Spain. While the publicly available loan performance data suggest some differences between
countries, we recommend that banks should also be required to conduct their stress tests
with more homogeneous top down assumptions for a more level playing field in Europe.
3 Literature review
Several studies have shown that credit performance indicators vary with the economic cycle
with higher defaults and losses being correlated with rising unemployment or falling GDP
or residential real estate prices (e.g. Hirtle et al. 2014 for the US and Henry Kok 2013 for
Europe). Most research on credit stress testing focuses on exposures in a single country and
central banks regularly publish country specific credit risk analysis and macro stress test
results in their financial stability reports (Čihák 2012). Cross-country comparisons of stress
2
test models are rare. Published stress test models designed to fit individual countries vary
in their econometric specification and in terms of the choice of macro variables used, which
makes a cross-country comparison impossible without access to the underlying data.
Jakubik Schmieder (2008) compare credit stress testing models in Germany and the Czech
Republic and investigate both the corporate sector and the household sector. They find
that credit risk for corporates can be modeled based on similar macroeconomic variables for
both countries, despite fundamental differences in the default rate time series patterns. In
contrast, Simons Rowles (2009) find that a top down stress test model for Dutch corporate
credit risk does not work well in Austria. From a practitioners view, Van Gestel (2014) claims
good global performance of simple econometric models for corporate defaults based on GDP
growth as a sole explanatory variable, however, without providing specific evidence. Beck
et al. (2014) study the nonperforming loan (NPL) ratio across 75 countries over one decade
using panel models with country fixed effects and pooled macro sensitivities. In other words,
while the authors allow for the baseline NPL ratio to vary by country (i.e. there is a country
specific intercept in each regression), the panel models estimates one sensitivity parameter
for GDP and one parameter for the lending rate applicable to all countries in addition to
other macro risk factors. The authors do not state whether non-pooled country-specific
models would have been statistically preferable.
Focusing on rating scores for corporate borrowers in 32 European and 3 non-European jurisdictions, Altman et al. (2014) shows that a simple score model based on four financial ratios
performs reasonably well across all jurisdictions. This is surprising as most practitioners
and researchers would approach credit risk modeling on a country-by-country basis claiming
improved accuracy over a global model. Altman and co-workers do make the point that the
model fit and predictive performance can be improved by adding country specific risk factors,
but looking at their results the solid performance of a simple score model is more striking
than the incremental performance gains from additional country-specific model complexity.
The complexity of bank risk management and regulation has been critized by Haldane (2013)
who suggested that overly complex regulations could be gamed by insiders making banks
less robust. Bank risk managers may be overwhelmed by the tens of thousand of model
parameters in use at large international banks with significant model risk and risk of overfitting. The US FED stress test program has recently been criticized by Dowd (2014) who
argued for less complexity but also for less prescriptive stress test models. If all banks have
to manage to the risk assessment of one regulator or one set of models this would fuel the
build up of systemic risk. This aspect is also mentioned in the recent publication by the
Bank of England (2014) for the UK variant stress test.
4 Empirical stress test models
To set the scene and to justify our choice of models, we note that the well-known Basel
Vasicek model underying the calculation of risk-weighted assets can be interpreted as simple
model for the PD PIT as follows.
3
(
P DitP IT
Φ−1 (P DiT T C )
√
=Φ
−
1 − ρi
√
ρi
ft
1 − ρi
)
Here, Φ denotes the cumulative distribution function of the standard normal distribution
with inverse Φ−1 , t is the time index measured in years and i is the index for an individual
borrower or a homogeneous cohort of borrowers. A homogeneous cohort could consist of all
borrower with the same through-the-cycle P DT T C , corporate borrowers in the same industry
segment or in the same country or those sharing all these characteristics together (e.g. a BBBrated UK utility). Kupiec (2009) noted that this model shares some important characteristics
of historical default data, but quantitatively does not fit corporate bond data well. He writes
the model as a linear panel regression model in the probit-transformed dependent variable
yit = Φ−1 (P DitP IT )
yit = ai + bi ft + ϵit .
The parameter ai is called a fixed effect and ft a random effect. As the random effect is the
same for all cohorts (no index i), different cohorts are correlated with each other causing
systematic undiversifiable risk to the lender. The panel model can now be estimated with
different assumption for the model coefficients. A frequently used pooled model with fixed
effects would leave the ai cohort-specific, but would estimate the same parameter b for all
cohorts. In the Basel framework, the asset correlation ρ and hence b is prescribed to be the
same for each retail asset class (e.g. ρ = 0.15 for residential mortgages and ρ = 0.04 for
qualifying revolving retail exposures) and does not vary by bank or country. In other words,
a Spanish bank has to hold the same amount of capital against their residential mortgages
as a German or UK banks, if the loans share the same P DT T C and loss-given-default (LGD)
parameter. The asset correlation for corporate loans is prescribed as a function of the P DT T C ,
but does not vary by country or industrial sector. These assumptions in Basel were made
for simplicity and the empirical support for the assumed asset correlation values and the
PD-dependency of corporate asset correlations is generally weak.
By comparison, top down stress testing models are often formulated as linear panel models
or time series models for the P DitP IT possibly after a suitable transformation like the probit
transformation above. The key difference to the Basel model is two-fold. First, analysts
often use the time lagged dependent variable in the regression for an improved fit of the
data. Second, rather than using a latent random effect ft , analysts use one or more observable
macroeconomic variables like the annual growth in GDP or the unemployment rate to capture
the systematic risk resulting from the macro environment. The macro variables can appear
contemporaneously at time t or with a time lag t-1, t-2, … , reflecting the expectation that
changes in the macroeconomic environment take time to be reflected in the loan performance
statistics. In statistical terms, such a stress test model (also called satellite model) is specified
as an auto-regressive distributed lag model. For one macro variable using no and one time
lag only, the model can be written as
yit = β0i + β1i yt−1 + β2i xt + β3i xt−1 + ϵit ,
4
where β0i are constant intercepts, β1i are the auto-regression coefficients, xt is an observable
macro economic variable, and ϵit is the random error term. The model coefficients β2i and
β3i drive the sensitivity of the performance measure to the macro variable xt and we will
simply refer to these parameters as the macro sensitivities. Comparing the Basel model with
the satellite model makes it clear that the satellite model approximates the latent random
effect of the Basel model with observable macro variables and the lagged dependent variable.
In the following we also consider a simplified version of the satellite model which is more similar to the Basel model and drop the auto-regressive term in the dependent and independent
variables
yit = β0i + β1i xt + ϵit .
While an auto-regressive specification often fits the historical data better, especially if
monthly or quarterly data are available, for a three year forecast based on annual data, the
auto-regressive model has the undesirable feature of slowing down the response to changes
in the macro environment. We restrict ourselves to a small number of variables for which explicit macro scenarios have been published: the annual growth in GDP, the unemployment,
and house prices and interest rates which are all common choices in credit modelling. In
principle there are a large number of macro variables that could be used and some will be
more predictive for the loan performance measure than others. A number of statistical variable selection methods have been applied to stress testing modeös of US mortgages (Heppe
2014a). For exposures secured by real estate for which banks report an average loan to value
(LTV) ratio at year end 2013, we also calculate a stressed LTV based on updated indexed
house price values.
The models above can be estimated with frequentist methods using any commonly used
statistical software. If data are scarce and time series short as is often the case in credit risk
modeling, we prefer to estimate the models using a Bayesian approach to directly capture the
uncertainty in the model parameters. Researchers from the ECB also advocate the benefits
of the Bayesian approach (Henry Kok 2013 p20): “A Bayesian approach to modeling is
particularly useful for modeling banks’ risk parameters to account for the inherent model
uncertainty related to the fact that for many risk parameter variables, the data quality
is often imperfect and the historical time-series are typically rather short.” For detailed
examples of Bayesian model estimates using US and European credit data we refer to Heppe
(2014a and 2014b).
5 Macro sensitivities calculated from the 2014 EUwide stress test
We have published an interactive tool to analyze the EBA EU-wide stress test results free
of charge on www.os-is.com. Users can explore the EBA data online with a wealth of
information on credit risk. We estimate the point-in-time probability of default (PD PIT)
assuming that banks adhere to the methodology of calculating impairments as prescribed by
5
EBA (EBA 2014b, OSIS manual 2014 on www.os-is.com). We obtain six PD PIT estimates
for each year 2014, 2015, and 2016 in each the baseline and adverse scenario. We then
regress the PD PIT with the macroeconomic scenarios applied by the banks in the stress
test to arrive at simple models for the PD PIT in dependence of a small number of macro
factors. While many results look reasonable, we do not know whether a bank adhered to
the published methodology from EBA or to what extent the estimates were impacted by
regulatory overwrite. We know that 23 banks had permission by EBA not to use the static
balance sheet assumption. In those 23 cases, our estimates are likely to be inaccurate and
we exclude then im the calculation of country averages.
Figure 1 shows the estimated probit-transformed PD PIT for a number of banks lending to
residential mortgage borrowers in the UK (EBA exposure class Retail - Secured by on real
property).
Figure 1: Estimated probit PD PIT of UK mortgages versus indexed LTV.
The simple univariate probit model can be displayed as a trend line when plotting the probit
PD PIT against the macro variable whereby the macro sensitivity equals the slope of the
trend line. Big circles indicate that the p-value of the regression is less than 0.2 whereas small
data points indicate lack of statistical significance. The trend lines appear to all have similar
slopes and all models are significant. Deutsche Bank is not a high street lender in the UK,
their PD PIT are much higher than those of the other banks shown which is probably due to
a higher portion of non-conforming mortgages. The estimated PD PIT for Handelsbanken
are small. HSBC and Barclays are not shown as their extremely small impairment rates give
rise to negative PD PITs (cf. Figure 16 in Annex 1). In this graph, we include Lloyds despite
the banks’ exemption from the static balance sheet assumptions.
Figure 2 adds some banks with their residential mortgage exposure in Spain. It is obvious
6
Figure 2: Estimated probit PD PIT of UK and Spanish mortgages versus indexed LTV.
that the macro sensitivity for the Spanish exposures are significantly higher. The difference
exists for Spanish banks in the UK (e.g. Santander) and UK banks in Spain (e.g. Barclays).
ING Spain being the exception with PD PIT even lower than in the UK (the two trend lines
at the bottom of the graph). Perhaps surprisingly, the starting point PD PIT in the baseline
2014 scenario are quite similar across banks in both countries seen in the left part of the
graph whereas the higher stresses in the adverse scenario are located on the right side of
the graph corresponding to higher LTVs. If the baseline PD PIT were close to the throughthe-cycle PD PIT then the banks in Spain and the UK would be required to hold similar
amounts of regulatory capital against the respective mortgage pools given the same asset
correlation prescribed by Basel (and assuming similar LGD). Annex 1 provides more charts
for comparing the risk weights and impairments between selected UK and Spanish banks.
In contrast, the stress test results point to much higher defaults in an adverse scenario in
Spain compared to the UK.
The model differences persist if we use the annual change in unemployment rather than
the indexed LTV (Figure 3) or if we consider non-real estate secured retail exposures (EBA
exposure class Other Retail, Figure 4). A similar behaviour can be found for corporate
exposures. For example, the exposure weighted average macro sensitivity for corporate
loans versus the change in unemployment is 3.73% in Spain, 1.97% in Italy, but only 0.68%
in the UK and 0.52% in Germany. For Other Retail exposures, the sensitivities to the change
in unemployment are 2.63% (Spain), 3.23% (Italy), 0.15% (UK), and 0.80% (DE).
7
Figure 3: Estimated probit PD PIT of UK and Spanish mortgages versus annual change in
unemployment rate.
Figure 4: Estimated probit PD PIT of UK and Spanish Other Retail loans versus annual
change in unemployment rate.
8
6 Analyzing nonperforming loan ratios across countries
In this section and the remainder of this paper, we wish to assess whether the different
model sensitivities apparently used by banks in Spain and in the UK for retail assets can
be justified with historical data. There are very few public data sets that measure the
credit performance of bank loans and span at least one full economic cycle across many
countries. One such data set is the ratio of nonperforming loans to all loans reported by the
Worldbank. However, contrary to Basel, which has a standard definition of default, there
is no standardized definition of what constitutes a nonperforming loan across countries or
banks. The advantage of the Wordbank data is that many countries are covered for the
period 2000 to 2013 though the data are not broken down by asset class and thus combine
retail and wholesale exposures. Figure 5 and 6 show the scatterplots for the probit NPL
ratio versus the change in unemployment and GDP, respectively, for a number of countries
(we download the data for NPLs and the macro variables via API from quandl.com).
We note that no consistent picture emerges. Whereas Germany, Spain and Italy show similar sensitivities to the change in unemployment such dependency is not significant for the
UK. Regarding GDP, Spain, Ireland and the US show a relatively large sensitivity, whereas
the UK is significantly smaller and Germany is not significant. France does not show a significant sensitivity to either the change in GDP or unemployment. The example of France
demonstrates that these short time series have to be interpreted with caution as the statistical error is large as indicated by the shaded areas. It is not plausible that a sharp rise
in unemployment in France would not result in higher NPL ratios for French banks. The
insignificant results reflect the relative stability of the last 13 years in France. Beck et al.
(2014) consider a similar but broader NPL data set for 75 countries and find that GDP is
most explanatory, but other variables like the lending interest rate are also significant. Note
that the authors estimate a pooled model with fixed effects and do not estimate country specific sensitivities. We refrain from introducing more macro variables here, but simply repeat
our warning that 13 annual data points provide little information to justify the estimation of
country specific models. However, the data certainly suggest that pooling across countries
may not be statistically preferable and further research with longer time series is required.
7 Analyzing historical data of UK and Spanish retail
loans
In this section, we look at longer historical data for UK and Spanish retail loans taken from
recent financial stability reports. The published time series are also highly aggregated and
only offer a rough proxy for the default rate. On the positive side, these data offer a longer
view spanning several decades.
In Spain, we look at the nonperforming loan ratio for secured and unsecured household loans
as reported in the Spanish credit register and published in the recent financial stability report
9
Figure 5: Nonperforming loan ratio versus change in unemployment.
Figure 6: Nonperforming loan ratio versus change in GDP.
10
of the Bank of Spain (2014). Data for Spanish and UK house price indices (HPI) are taken
from the Bank for International Settlement. For ease of comparison we only report results
where data are aggregated to an annual frequency. Figure 7 shows the fit and forecast of a
Bayesian auto-regressive model of order one versus the historical change in Spanish house
prices (i.e. no lag for the HPI variable).
The forecasts for 2014-16 are shaded grey and we take the projected HPI changes from
the adverse scenario of the 2014 EU-wide stress test for Spain. The Bayesian fan chart
results from a Markov Chain Monte Carlo simulation (using the rjags package in R) and
expresses the forecast uncertainty including the uncertainty in the model parameters. The
usefulness of Bayesian fan charts for stress testing has recently been highlighted by Franta
et al. (2014). The prior distributions for all parameters are assumed uninformed. The
posterior distribution for the sensitivity to the annual change in the Spanish house price
index is shown in Figure 8.
The mean of the posterior macro sensitivity agrees well with the frequentist OLS estimate.
Even with 30 years of data, the parameter is not sharply defined. Using less than 10 years of
data, which banks often encounter in practice, results in even wider uncertainty distributions
(for more Bayesian examples see Heppe 2014b).
In the UK, we use the percentage of mortgage loans in arrears for more than six months
as published in the financial stability report of the Bank of England and in the recent
publication of stress test results (Bank of England 2014). The macro stress shown is the
UK variant of the 2014 stress test. Figure 9 and 10 show the model fit and forecast and the
posterior sensitivity to UK HPI.
While the time series in the UK and Spain are quite different, the difference in the HPI
sensitivities between the two countries is minor in stark contrast to the models used by the
banks in the EU-wide stress test (cf. Figure 2).
Similarly, for unsecured retail loans Figure 11 and 12 show the fit and forecast of the NPL
ratio of unsecured household loans in Spain against the annual change in the unemployment
rate whereas Figure 13 and 14 show write offs of unsecured retail loans in the UK against
changes in unemployment. Again, the means of the posterior distributions of the sensitivity
parameters differ only modestly especially in light of the displayed uncertainty.
8 Stress testing UK banks with Spanish models
While we saw some weak evidence in the Worldbank NPL data that macro sensitivities may
be significantly smaller in the UK compared to Spain, we do not find such evidence in our
Bayesian analysis of Spanish and UK retail loans based on longer time series. The average
macro sensitivities of banks in the UK and Germany are much lower than those used by
banks in Spain and Italy. To gauge the importance of the macro sensitivities in the stress
test results, we reclaculate the PD PITs for UK and German banks using the higher average
sensitivities from Spain and Italy and vice versa. Surprisingly the effect of using more or
less aggressive macro sensitivities seems to work mainly in one direction. While the CET
11
Figure 7: NPL ratio of secured loans to Spanish households.
Figure 8: Posterior sensitivity to Spanish house price changes in autoregressive model of
order one.
12
Figure 9: Arrears ratio of residential mortgage loans in the UK.
Figure 10: Posterior sensitivity to UK house prices in autoregressive model of order one.
13
Figure 11: NPL ratio of unsecured retail loans in Spain.
Figure 12: Posterior sensitivity to changes in unemployment rate (Spain.
14
Figure 13: Write offs for unsecured retail loans in the UK.
Figure 14: Posterior sensitivity to changes in unemployment rate (UK).
15
1 ratios of Spanish and Italian banks increase the effect is modest and none of the failed
banks in Spain and Italy would have passed the EU-wide stress test on the basis of the more
benign model parameters.
In contrast, the use of the more conservative Spanish sensitivities would have a dramatic
impact on those big four UK banks which were subject to the EBA stress test and disclosure.
Barclays and HSBC who comfortably passed the 2014 EBA and Bank of England tests
both follow the static balance sheet assumption and both would fail the stress test by a
significant margin. Even though we cannot calculate the PD PITs for Lloyds and RBS due
to their exemption from the static balance sheet assumption it is reasonable to infer that
those two would have failed the stress tests as well if the more conservative Spanish model
parameter were imposed on them. We also recalculated the CET1 ratios for German banks
using average Italian macro sensitivities and again find some material capital shortfalls with
a further three banks failing the stress test in the adverse scenario. Detailed results of the
capital calculation can be made available upon request.
9 Conclusions
The answer to our initial question is negative whether we find a level playing field between
banks in different European countries. Further research is required into the apparent inconsistency of assuming a homogeneous asset correlations in the calculation of risk weighted
assets while allowing banks to use widely differing macro parameters in the stress testing
models.
What are the possible explanations for the apparent inconsistencies of having widely varying
model parameters allowed in the stress test? First, while we think that our conclusions
are robust, we cannot be sure that our calculated PD PITs are exactly those used by the
banks as banks may have deviated from the published methodology of EBA. Second, as we
have seen in the NPL data from the Worldbank in section 6, banks using different macro
variables could come to different conclusions as the historical macroeconomic data do not
always show the expected co-movements that underlie the specific stress scenarios imposed
by EBA/ECB or the Bank of England in the adverse scenario. Third, statistical sampling
errors are large and macro parameters are notoriously hard to determine with data of length
found in practice and all econometric model parameters will be unstable over long enough
time horizons.
We recommend the Bayesian approach for model estimation as it includes the uncertainty of
the parameters directly and we have shown the importance of such uncertainty for forecasting
credit risk and stress testing. The Bayesian fan charts presented above clearly demonstrate
the large forecast uncertainty which includes the uncertainty about the model parameters.
While all stress tests to date seem to focus on point estimates (the 50% central white line
in the forecast fan), we note that stress test predictions are not certain even if the macro
scenario is assumed to be known with certainty. Bank stakeholders requiring a credibility
of more than 50% will find material capital shortfalls at many institutions. The Bayesian
framework has the additional advantage of allowing the use of informed priors to include
16
external information about model parameters in a statistically coherent manner.
The EU-wide stress test provides an unprecedented level of disclosure with up to 12,000 data
points per banks. While the EBA stress test has addressed many of the shortcomings of its
predecessors, many practitioners still consider it more a ritual dance than a source of deep
insight into the riskiness of the balance sheet. Here, we aim to demonstrate the value of this
rich data set for a better understanding of credit risk even if further research is required to
overcome the apparent inconsistencies in the models used by banks.
This version of the paper was released shortly after the UK stress test results were announced
by the Bank of England (2014). For the avoidance of doubt, we do not claim that our analysis shows that any particular UK (or German) bank should have failed the EBA or Bank
of England stress tests. We do not claim that the models of UK banks are inadequate. We
highlight that the apparent inconsistencies between different countries require further investigation. The purpose of stress tests is to complement the existing bank capital framework
so one should not expect the same models or conclusions from both approaches. However,
bottom up stress test procedures are still under development and the experience with the
widely divergent calculations of risk weighted assets points towards a need for simple top
down measures to complement the bank’s own calculations which have historically been intransparent to outsiders. The recent publication of stress test data provide unprecendented
insight into the banks balance sheet, but the more information becomes available the more
questions will be raised by outsiders. Providing answers that are broadly consistent across
Europe will remain a formidable challenge for the years to come.
10 References
• Altman, Iwanicz-Drozdowska, Laitinen, Suvas 2014, Distressed Firm and Bankruptcy
prediction in an international context: a review and empirical analysis of Altman’s
Z-Score Model.
• Banco de Espana 2014, Financial stability report, November.
• Bank of England 2014, Stress testing the UK banking system: 2014 results.
• Beck, Jakubik, Piloiu 2013, Non-performing loans: What matters in addition to the
economic cycle? ECB Working Paper No. 1515.
• Čihák, Muñoz, Teh Sharifuddin, Tintchev 2012, Financial Stability Reports: What
Are They Good For? IMF working paper.
• Dowd 2014, Math Gone Mad. Regulatory Risk Modeling by the Federal Reserve, Cato
Institute, No. 754.
• EBA 2014a, Results of 2014 EU-wide stress test
17
• EBA 2014b, Methodological note EU-wide Stress Test 2014, Version 2.0
• Franta, Barunik, Horvath, Smidkova 2014, Are Bayesian Fan Charts Useful? The
Effect of Zero Lower Bound and Evaluation of Financial Stability Stress Tests, Czech
National Bank.
• Haldane 2013, Containing Discretion in Bank Regulation, speech available on
www.bankofengland.co.uk
• Henry, Kok (Editors) 2013, A macro stress testing framework for assessing systemic
risks in the banking sector. ECB Working Papers No. 152.
• Heppe 2014a, Top down stress testing of US mortgages, www.os-is.com.
• Heppe 2014b, Top down robustness check of the 2014 EBA Stress Test, www.os-is.com.
• Hirtle, Kovner, Vickery, Bhanot 2014, The Capital and Loss Assessment under Stress
Scenarios (CLASS) Model, Federal Reserve Bank of New York Staff Reports.
• Jakubík, Schmieder 2008, Stress Testing Credit Risk: Comparison of the Czech Republic and Germany.
• Kupiec 2009, How Well Does the Vasicek-Basel AIRB Model Fit the Data? Evidence
from a Long Time Series of Corporate Credit Ratings Data.
• Nkusu 2011, Nonperforming Loans and Macrofinancial Vulnerabilities in Advanced
Economies, IMF Working Paper 11/161.
• Simons, Rowles 2009, Macroeconomic Default Modeling and Stress Testing. De Nederlandsche Bank.
• Van Gestel 2014, Credit stress testing, presentation.
11 Annex 1: Screenshots from OSIS LoanCracker regarding UK and Spanish banks
18
Figure 15: Risk weights of UK and Spanish residential mortgages (Source: EBA).
Figure 16: Impairment rates of UK and Spanish residential mortgages in adverse scenario
(Source: EBA).
19
Figure 17: Coverage ratios of default UK and Spanish residential mortgages in adverse
scenario (Source: EBA).
Figure 18: Calculated PD PIT of UK and Spanish residential mortgages in adverse scenario
(Source: EBA, OSIS).
20