Contents

Motivation .............................................................................................................................................. 3
Literature review ..................................................................................................................................... 4
GME-general aspects ........................................................................................................................... 5
The model ............................................................................................................................................... 6
Data ........................................................................................................................................................ 8
Implementation..................................................................................................................................... 10
GME implementation ........................................................................................................................ 11
GME vs OLS ....................................................................................................................................... 14
Conclusions ........................................................................................................................................... 15
References ............................................................................................................................................ 18
Table 1 Initial set of macroeconomic variables ......................................................................................... 9
Table 2 OLS results ................................................................................................................................ 10
Table 3 GME results ............................................................................................................................... 12
Table 3 Bootstrap experiment ............................................................................................................... 14


Motivation
In this paper we propose to ascertain the macroeconomic determinants of the probabilities
of loan defaults in Romanias case. In our exercise we use the conditional probability of default
methodology (CoPoD), proposed by Miguel A. Segoviano Basurto.
Generally, in the last two decades, the credit cycles were linked to macroeconomics and
asset prices cycles. It was observed that after a financial crisis the level of empirical frequencies
of loan default increased considerably. Hence, the measurment of a bank credit risk should take
into account the changes in macroeconomic conditions.
In late 2008 the global financial crisis hit Romania and the macroeconomic conditions
deteriorated significantly. The banking system was severely affected by this economic downturn
as it faced liquidity problems due to limited access to external sources of funds: withdrawal of its
parent funding and a rapid increase of nonperforming loans that reached a peak in the last quarter
of 2013 (22%).
According to Segoviano (2006) CoPoD (conditional probability of default methodology)
improves the measurement of the impact of macroeconomic developments on loans credit risk
by making a twofold contribution. First, econometrically, the proposed methodology, based on
the J aynes (1957) generalized maximum entropy rule (GME), recovers estimators that in the
setting of finite samples are superior to OLS estimators under the Mean Square Error (MSE)
criterion. Second, economically, on the basis of a hypothesis that is consistent with economic
theory and empirical evidence, a procedure is proposed to select the set of explanatory variables
that have a significant effect on loans credit risk.
As we confront with a data limitation problem: the National Bank of Romania had been
publishing the quarter non-performing loans ratio for the aggregate banking since the end of
2007- we believe that this methodology based on generalized maximum entropy rule is a good
solution from an econometric point of view.
The motivation behind this paper was represented by the importance for the banking
system and for the financial stability to be able to forecast the credit risk challenges . As told by
Moinescu (2013): A general message that can be drawn from the study of past boom-bust credit
cycles, is that these cycles show high resemblances, and therefore lessons can be drawn from an
analysis of their commonalities.
Literaturereview
In the recent literature the probability of default in macroeconomic context has been
assessed using two main approaches: early warning systems and multivariate probit or logit
regression models.
The main model applied in macro credit risk modeling follows the methodology proposed
by Wilson (1997) that use the logistic function in order to assess the relation between credit risk
and macroeconomic variables.
Boss (2002) used this methodogy to capture the probability of default at the portfolio
level under stressed macroeconomic scenarios for Austrian corporate and household sectors:
GDP, private consumption, the unemployment rate and industrial production as well as the ratios
of equipment investment to GDP and exports to GDP are significant assuming a lag of four
quarters.
Fiori, Foglia and Iannotti (2009) investigates the correlation between Italian corporate
default and the macroeconomic risk factors or to other possible unobservable factors arising from
business inter-connections. This analysis considered the contagion across sector while estimating
the effects of macroeconomic conditions to default rates. The most important macroeconomic
determinats for all the sectors, as per this study, are real GDP growth, loan to GDP and the real
exchange rate with US dollar .
A recent study in this direction was performed by Moinescu (2013) using panel data of
nonperforming loans from Central and Eastern European countries. Econometric results
confirm the repute of the economic growth indicators as the key risk factor for the NMS debtors
to face default events. Real GDP growth and the change in output gap were almost equally
important. Moreover, money market interest rate 3M, inflation and exchange rate changes are
also statistically relevant. However, fixed effects were not found statistically relevant. This result
indicates that credit discipline is homogeneous across CEE region.
Logit and probit models use OLS estimation in order to determine credit risk as a
function of macroeconomics variables. As the series with the empirical frequency of loans
defaults contain few observations, the OLS estimators represent an important limitation for this
models as the number of observations barely exceed the number of macroeconomics
determinants.
An alternative to logit models, based on OLS estimation, is represented by conditional
probability of default methodology based on generalized maxim entropy estimation. Segoviano
(2006) using this method investigates the probability of default for small business enterprises as
function of macroeconomics variables. The dataset used in this study is represented by non-
performing loans ratio registered in Mexico and Norway. Besides the macroeconomic
determinats mentioned in the studies presented above Segoviano introduced macroeconomic
variables linked to financial liberalization. According to Kaminsky and Reinhart (1999)
financial liberalization helps to explain the occurrence of banking crises in large samples of
developing and developed countries.
GME-general aspects
The entropy concept was introduced in science more than 150 years ago, but it was only
since the middle of the 20th century that its applications have disseminated to many research
fields. At the root of this movement, there were the works of Shannon (1948) who introduced
the entropy concept in information theory as a measure of uncertainty in a random variable.
After a decade E.T. J aynes lay down the maximum entropy principle according to which
if we maximize the Shannons entropy function and we take into consideration our current state
of knowledge about the possible outcome of a random variable we may select the probability
distribution function of the random variable. In this manner we choose the probability
distribution which leaves us the largest remaining uncertainty , the maximum entropy, consistent
with our constraints, without introducing additional assumptions or biases into our calculations.
More recently Golan and Judge (1996) introduce the estimators obtained through
Generalized Maximum Entropy Rule (GME). ). According to their studies in this area, applying
GME lead us to robust estimators even when we have an ill-conditioned or ill-posed problem;
In the last two decaded GME was used in several areas of economics. One of the first
paper was represented by the study of Golan (1996) that used GME in order to estimate the
market shares of 20 companies in different industries, using only information from four different
concentration indicators to estimate all market shares.
Randall and Campbell (2005) in their study about determinants of poverty rate in
California used GME and OLS. Based on a Monte Carlo experiment they concluded that GME
estimators present the smallest variance.
Recently Ferreira and Dionisio (2012) used GME, as an alternative to OLS in the
estimation of utility function. The study results revealed that GME estimators are more accurate
than OLS estimators.
The model
Starting from Merton model (1957) that assume that the value of the assets of the
borrower at time t follows a geometric Brownian motion:
J S
t

S
t

=p

Jt +o

Jw
t

wherep

is the instantaneous asset return, o

is the instantaneous asset volatility andw

t

is a standard Brownian Motion.

If we assume that the initial logarithmic asset valueln(S
t

) =0, then
ln|S
t

] ~N__p

1
2
o

2
] (I t),o

I t_
We may rewrite the standardized logarithmic asset value of this borrower at time T as
s(I) =
ln|S
1

] [p

1
2
o

2
(I t)
o

I t
onJ s(I)~(0,1)
This borrower default occurs if the return on a borrowers assets rate of return, s
1
, falls
below a certain barrier o
t
, the default threshold:
P(
t
=1) =P(s
1
<o
t
)
As the borrowers assets rate of return asset is standard normally distributed, we may
say that the P(s
1
<o
t
) is also standard normally distributed and we may formulte the
probability of default as follows:
PoB = (o
t

) (1)
where(.) is the standard normal cumulative distribution function (cdf)
We assume that in the economy there is a portfolio of borrowers, with an asset rate of
return that follows equation (1), than the probability of default, at aggregate level in the economy
is equivalent with the borrowers proability of default
The empirical frequencies of loan defaults are grouped in a T-dimensional vector: PoD.
Since the observations in the vector of PoD are restricted to be in the interval (0,1) we
make the following transformation:
o =
-1
(PoB)
where(.) is the standard normal cumulative distribution function (cdf)
We want to estimate the default threshold: o as a function of macroeconomic variables:X:
o =X[ +c (2)
where whereo is a T-dimensional vector of noisy observations (transformation of the
PoDs), X is a known (T x K) matrix of macroeconomic and financial series and is a K-
dimensional vector of unknown coefficients that we are interested in estimating.
We express each estimator as discrete variable with 2M< possible outcomes:
[
k
= z
km
p
km
M
m=1
wit p
k
=1 onJ p
k
(0,1)
In the same manner we reformulate the vector of disturbances, c:
We reformulate equation (2) as follows:
o =XZp +Iw (3)
Following J udge and Golan (1996) Generalized Maximum Entropy rule (GME): we
select p andw in order to maximize the entropy function:
E(p,w)=| p
m
k
ln(p)
m
k M
m=1
K
k=1
] | w
]
t
ln(w)
]
t
]
]=1
1
t=1
] (4)
Taking into consideration the following constraints:
o
t
= x
tk
z
m
k
p
m
k
M
m=1
K
k=1
+:
]
t
w
]
t
]
]=1
(5)
1 = p
m
k
M
m=1
, 1 =w
]
t
(6)
]
]=1
As we have a maximum problem, we may apply Lagrange function:
I = _ p
m
k
ln(p)
m
k
M
m=1
K
k=1
_ _w
]
t
ln(w)
]
t
]
]=1
1
t=1
_ +z
t
1
t=1
_o
t
x
tk
z
m
k
p
m
k
M
m=1
K
k=1
:
]
t
w
]
t
]
]=1
_
+ 0
k
_1 p
m
k
M
m=1
_ +
K
k=1
:
]
_1 w
]
t
]
]=1
_ (7)
1
t=1
In order to recover the probability vector w andp we maximize the Lagrangian function
described in equation (7) and we obtain the following entropy solution:
p
m
k
(z
`
) =
expj z
`
t

x
tk
z
m
k 1
t=1
[
jcxp| z
t

x
tk
z
m
k 1
t=1
][
M
m=1
(8)
w
]
t
(z
`
) =
expjz
`
t

:
]
t
[
_cxp jz
`
t

:
]
t
[_
]
]=1
(9)
Now we may state the estimator under GME as:
[
`
=Zp (10)
Data
As a proxy for the empirical frequencies of default we selected the nonperforming loans
ratio, at aggregate level registered by the Romanian credit institutions in the period 2008
Quarter1 -2013 Quarter 4.
The nonperforming loan ratio is defined as the report between nonperforming loans and
the total gross value of loans. A nonperforming loan is a credit with at least 90 days past due or
with a debtor declared insolvent.
The initial set of macroeconomic explanatory variables have been selected based on the
theoretical and empirical literature about crisis periods and from other studies about
macroeconomic determinants of credit risk that have been available for Romania and we believe
that could be relevant for our analysis.
As presented in the previous section of literature review section the most relevant
macroeconomic indicators for credit risk are the economic growth variables: GDP growth rate,
consumption, index of equity prices.
Foreign exchange rate we intuit that has an important impact on Romanian
nonperforming loans because one of the main issue for Romanian borrowers was represented by
credits in foreign currency corroborated with a depreciation of the domestic currency.
From a similar category we included the ratio between the M2 monetary aggregate and
foreign exchange reserves that captures the extent to which the liabilities of the banking system
are backed by international reserves. In the event of a currency crisis, individuals may rush to
convert their domestic currency deposits into foreign currency, so this ratio seems to capture the
ability of the central bank to meet those demands (Calvo and Mendoza, 1996).
A summary of the variables that were analyzed, the code that we used to identify them
and their source is in Table 1
Table 1 Initial set of macroeconomic variables
CODE Variable
GDP Real GDP
REER Real Foreign Exchange rate ;
LTIR Long Term Interest rate-EMU convergence criterion bond yeld
GVDovGDP Ratio of Government Debt to GDP
CONovGDP Ratio of Consumption to GDP
BET Share Price Index- Bucharest Stock Exchange Index
CRE Real aggregate credit to private sector
CREovGDP Ratio of real aggregate credit to GDP
FDI Foreign direct investments
FDOovGDP Ratio of foreign direct investments to GDP
M2ovFXR Ratio of M2 monetary aggregate to foreign exchange reserves
MMIR Money market interest rate
UNEM
Unemployment rate
The purpose of thiss paper is analyze how the fluctuation in the selected macroeconomic
variables have an effect on credit risk. We computed this fluctuation wuth respect to previous
period
For all the series selected and presented in the table above we performed unit root tests:
Augmented Dickey-Fuller and KPSS- Kwiatkowski, Phillips, Schmidt and Shin. The variables
expressed as growth under both test have no unit roots.
The series for the explanatory variables contains data from the period 2005 Q1- 2013 Q3.
We used the dataset publised by EUROSTAT and the National Bank of Romania and the
indicators expressed as percent of GDP and Ratio of M2 monetary aggregate to foreign exchange
reserves we computed by us.
Implementation
In order to select the set of macroeconomic determinants relevant for assessing the
nonperfoming loans ratio for Romanian credit institutions, we run multiple multivariate OLS
regression and we explored different combination of variables as number and different
combinations for the lag. We selected 3 models based on R-squared criteria and Akaike criteria
and we choose models that have an economic signification. For the regression results of the
models selected please refer to Annex 1.
Table 2 OLSresults
Model 1 contains the results with two explanatory variables: growth rate of the ratio of
vernment debt over GDP and the growth rate of Bucharest Stock Exchange index BET. These
variables are lagged 10 quarters and 8 quarters, respectively. As can be seen the parameters of
this variables are highly significant with p-value less than 1%. Taking into consideration that we
used only two explanatory variables the goodness of fit is quite high.
In model 2 we augument the number of explanatory variables to four variables. The most
significant variables with a p-value below 1% are real GDP growth rate and the growth rate of
Bucharest Stock Exchange index BET, lagged 11 quarters and 9 quarters. From the three new
introduced variables the real GDP growth rate and ratio of FDI (foreign direct investments) over
GDP are highly significant. Ajusted R-squared and Akaike criterion show that goodness of fit
improves.
In model 3 we include the growth rate of FDI lagged 5 quarters. The parameter of this
macroeconomic indicator is highly significant with a very small p-value. Another new variable :
Foreign exchange rate RON/EUR lagged 7 quarters is significant at 3.30% significance level.
The Government Debt expressed as percent of GDP lagged 11 quarters is significant, with a very
low p-vlaue, similar to the significance level obtained in model 1 for this variable. For model 3
we may observe that the goodness of fit improves with an Adjusted R-squared of 0.84.
GME implementation
As we have a small sample of observations for the nonperforming loans ratio, the OLS
estimators may be sensitive to small changes in the data, thus making the measurement of the
impact of macroeconomic variables on loans credit risk imprecise.
As mentioned in the previous sections the general maximum entropy rule enable us to
obtain estimators that show greater robustness than OLS under the finite sample properties.
Based on the models specified above we will generate GME estimators as follows:
For each model we will estimate the vector of coefficients[
0LS
using:
o =X[ +c
In order to simulate[
0LS
parameters we used Bootstrap with 10,000 trials ;
As we have said earlier in the model specification section we try to express each
estimator as discrete variable with 2M< possible outcomes
According to Judge and Golan (1996) , the best results are obtained when we
increase the number of possible outcomes: M , from 3 to 5. Based on this result
we choose M=5.
With the distributions of [
0LS
coefficients obtained using Bootstrapping, we
calculated the standard errors, , for each coefficient, and used these standard
errors to define the bounds of Z, using a three-sigma rule
1
:
z
k1
= S,z
k5
=3,z
k3
= ,z
k2
=
z
k1
+z
k3
2
and z
k4
=
z
k5
+z
k3
2
We used the same approach for the vector of disturbancesW
We defined the vector Z and W, we know X , the macroeconomics explanatory
variables anda, vector of observations, transformation of the Proability of default
(nonperforming loans ratio)
We apply the Generalized Maximum Entropy rule in order to recover proabability
vectorsp andw and then we obtain the GME estimators:
[
`
= Zp
The GME estimators for the three models detailed in the previous section are detailed in
the table below:
Table 3 GME results
The adverse macroeconmics events are reflected with a lag in the nonperforming loans
ratio. The explanation of this situation is quite intuitively, as it takes time, once the financial
1
Chebychevs inequality may be used as a conservative means of specifying bounds. For any random variable, x,
such that E(x) =0 and V ar(x) =2, the inequality provides Pr [| x| <v] v2 for arbitrary v >0. An example is the
familiar 3 rule that excludes at most one-ninth of the mass for v =3.
problems begin until the borrower is in a default situation. Another factor that may increase the
lag is the fact that banks may delay declaring a loan as a nonperforming one due to capitalization
requirments.
According to our results from the first model the governement debt, lagged four
quarters is a very strong indicator: a significant growth of this variable can signal 11 quarters
ahead an increase in nonperforming loans ratio. The other variable: Bucharest Stock Exchange
index lagged 8 quarters has an adeverse effect on credit risk: a decrease of this index may signal
an increase in nonperforming loans ratio. This result is in accordance with the results obtained by
Demirgc-Kunt and Detragiache (1998) that find variables that contribute to systemic banking
sector fragility may be in place (signalling) for two years (on average), before problems become
manifest
In model 2 we introduce 3 new macroeconomic indicators: real GDP growth rate, lagged
11 quarters and FDI growth rate expressed as percent of GDP, lagged 8 quarters. Both variables
have an adverse effect on probability of defaults in economy: an increase of the growth rate of
this variables may signal a decrease in nonperforming loans ratio. The sign and the lag of real
GDP growth rate is in line with the empirical studies performed by Borio and Lowe (1994). A
growth in the indicator report between M2 monetary aggregate and foreign international
rezerves, lagged 3 quarters could signal a growth in nonperforming loans ratio. Ratio of M2
over reserves is a proxy that indicates the extent to which the liabilities of the banking system
are backed by international reserves according to Calvo and Mendoza (1996). In the event of a
crisis individuals may rush to convert their deposits in domestic currency into foreign currency,
So we expect that after an event like the one described above, in short term, less than one year
the probability of default to increase.
In model 3 we fiind a new macroeconomic determinant of credit risk: variable the
foreign exchange rate RON/EUR. The lag and the sign of this variable are in accordance with the
studies of Kaminsky and Reinhart (2000). Based on the situation of the volume of loans in
foreign currency we expected that the depreciation of foreign exchange rate to have an impact
on worsening impact on nonperforming loans ratio.
GME vs OLS
Golan, J udge and Miller (1997) claim that in finite sample settings, the GME
esstimators exhibit reduced mean squared error (MSE) due to their properties minimum
variance..
For this reason in order to quantify the small sample properties of GME estimator we
performed the following experiment:
We used Model 1 specification: the matrix of explanatory variables, X and the
vector of observationsa to compute the vector of coefficients ols .
We used the matrix X , the vector of observations a and assuming the values of
the OLS estimators in order to obtain the vector of residuals:
R =o X[
0LS
We used Bootstrap with 10,000 random trials. Each trial represent a random value
of X andR.
With this element we computed simulated values for the vector of observation a
and we were able to recover the OLS and GME estimators.
Based on the distribution of OLS and GME estimators obtained we computed
Mean Square Error criterion:
HSE|[
`
] =E|([
`
[)]
We present in the table below the results of our experiment:
1 1 1
Variance ols 0.0041 0.1688 0.7462
Variance gme 0.0036 0.1028 0.4492
Variance (difference) 12% 39% 40%
MSE (difference) 133% 120% 84%
Table 4 Bootstrap experiment
Based on the results of this experiment we mai state that GME estimators show greater
robustness than OLS estimators in finite sample settings under the Mean Square Error criterion.
Conclusions
We propose the Conditional Probability of Default Methodology for modelling the
probabilities of loan defaults as functions of identifiable macroeconomic and financial variables.
This methodology makes a twofold contribution. From the econometric point of view,
produces estimators that, in the setting of finite samples, are superior to OLS estimators under
the mean square error criterion.
From an economic point of view, based on theoretical arguments and empirical evidence,
the methodology presented involves a procedure to select a set of explanatory variables that
have a significant effect on loans credit risk.
Some aspects of our macroeconomic credit risk model would certainly require further
elaboration: our model estimates at aggregate level the probability of default and thus we may
obtain less accurate estimates than those obtained with industry-specific default rates.
Since our probability of defaults are explained by lagged values of relevant explanatory
variables, it is possible to obtain ex-ante measures of probabilities of loan defaults given a set of
realised or simulated (in the case of stress testing) values of macroeconomic explanatory
variables.

Annex 1
Figure 1 Model 1-regression output
Figure 2 Model 2- regression output
Figure 3 Model 3- regression output
References
Borio, C., Lowe, P. (2002), Asset Prices, Financial and Monetary Stability: Exploring the
Nexus,
Boss, M. (2002), A Macroeconomic Credit Risk Model for Stress Testing the Austrian Credit
Portfolio,Financial Stability Report 4, Oesterreichische Nationalbank., p. 64-131.
Caprio, G., Klingebiel, D., Bank Insolvency: Cross-Country Experiences, Washington DC,
World Bank.
Caprio, G., Klingebiel, D., Bank Insolvency: Cross-Country Experiences, Washington DC,
World Bank.
Dooley, M. (1997), A Model of Crises in Emerging Markets, NBER, Working Paper 6300.
Golan, A., Judge, G., Miller, D. (1997), Maximum Entropy Econometrics, Series in Financial
Economics and Quantitative Analysis, Wiley, p. 85-106.
Golan, A., Judge, G., Miller, D. (1997), Maximum Entropy Econometrics, Series in Financial
Economics and Quantitative Analysis, Wiley, p. 85-106.
Goodhart, C., Hofmann, B., Segoviano, M. (2004), Bank Regulation and Macroeconomic
Fluctuations,
Oxford Review of Economic Policy, Vol. 20, No. 4, p. 591-615.
Horowitz, J. L. (2001), The Bootstrap, Handbook of Econometrics, Vol. 5, J.J. Heckman and
E.E.
Kaminsky, G., Lizondo, S., Reinhart, C. (1998), Leading Indicators of Currency Crises, IMF
Staff Papers, 5, 1, 1-48.
Kaminsky, G., Reinhart, C. (1996), Banking and Balance-of-Payments Crises: Models and
Evidence, Board of Governors of the Federal Reserve, Washington, DC, Working Paper.
Kaminsky, G., Reinhart, C. (1999), The Twin Crisis: The Causes of Banking and Balance-of-
Payments Problems, American Economic Review, 89, 473-500.
Kaminsky, G., Schmukler, S. (2003), Short-Run Pain, Long-Run Gain: The Effects of
Financial Liberalization, NBER Working Paper 9787.
Kiyotaki, N., Moore, J. (1997), Credit Cycles, Journal of Political Economy, 105, 211-248.
Segoviano M. (2004), A System for Measuring Credit Risk When Only Limited Data is
Available, in Credit Risk International, London, October 2004.
Wilson, T. (1997a), Portfolio Credit Risk I, in Risk, Vol. 10, No. 9.
Wilson, T. (1997b), Portfolio Credit Risk II, in Risk, Vol. 10, No. 10.