Sunteți pe pagina 1din 13

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

Factors that influence LGD for retail loans in financial


institutions
Natalia Cordeiro Zaniboni
Alcides Carlos de Ara
ujo

Alessandra de Avila
Montini

Abstract
The Basel regulations require attention to financial risks, in particular, credit risk.
The capital required for credit risk consists of three components, including the loss
given default (LGD). This study aimed to analyze the factors that influence the
characteristics of LGD in the European financial institutions regarding retail loans,
using the logistic regression model. The results suggest that the segment and exposure (EAD) affect the LGD. The model achieved an important result for the risk
management studies, as it correctly classified 92% of the observations of high loss
(LGD 50%).
Keywords: Basel, Loss Given Default, Logistic Regression, Credit Risk
Resumo
A regulac
ao da Basileia requer atencao para os riscos financeiros, particularmente,
o risco de credito. O capital necessario para protecao do risco de credito consiste
de tres componentes, incluindo a perda em funcao da inadimplencia (LGD). O objetivo deste estudo e analisar os fatores que influenciam as caractersticas do LGD
em instituic
oes financeiras europeias considerando emprestimos e utilizando o modelo de regrress
ao logstica. Os resultados sugeriram que o segmento e a exposicao
(EAD) afetam o LGD. O modelo atingiu um resultado importante para a area de
gestao de riscos, tendo classificado corretamente 92% das observacoes de altas perdas (LGD 50%).
Keywords: Basileia, Perda em funcao da inadimplencia, Regressao logstica, Risco
de Credito

Universidade de S
ao Paulo - FEA/USP; contato: natyzaniboni@yahoo.com.br
Universidade de S
ao Paulo - FEA/USP; contato: alcides.carlos@usp.br

Universidade de S
ao Paulo - FEA/USP; contato: amontini@usp.br

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

Introduction

The risk associated with financial transactions is a widely studied subject throughout the financial system. In order to protect these system against the instability of the
risks and prevent financial institutions from being exposed to high risks, an international
committee of banking regulations and supervisory practices was established in 1974, the
Basel Committee (Kapstein, 1994).
The Basel II, known as the International Convergence of Capital Measurement and
Capital Standards: A Revised Framework, first published in 1999 and completed in
2004, has as its main objective to standard the capital at risk from financial institutions,
especially for credit risk (Santos, 2001) and consists of three pillars (BIS, 2004):
1. Minimum Capital Requirement (Pillar 1);
2. Supervisory Review Process (Pillar 2);
3. Market discipline (Pillar 3).
The minimum capital required is the financial reserve required for financial institutions
to cope with their risk exposures. This capital is composed by three risk components:
PD (probability of default), LGD (loss given default) and EAD (exposure at default).
The Basel committee suggests two approachs for calculating this capital: the standard
approach, where the regulator itself provides the value of the risk components, and internal
based approach, where the bank calculates its components based on its credit risk history
and exposure characteristics, supervised by the local regulator.
The Basel II presents the main definitions of the internal approach methodology (part
2, section III). In paragraph 252, the committee says that banks must develop methodologies and calculate their own estimates of PD (probability of default), LGD (loss given
default) and EAD (exposure at default) for retail exposures.
According Silva, Marins, and Neves (2008), the component LGD is the percentage of
losses of a risk exposure at the time of default, which includes principal loss, the loss
arising from opportunity costs and collection and recovery loss.
Given the international requirements, the study aims to analyze characteristics and
the factors that most influence the risk component LGD, considering the information
available in Pilar 3 risk management reports of financial institutions.
The first studies regarding the risk component LGD began in 1995, with Asarnow
and Edwards (1995), and focused on LGD descriptive statistics, its distributions and
factors that influence its variability. Latest research, starting in 2008, focused on LGD
statistical models. This paper approach both subjects, and uses data from several financial
institutions, most in Europe.
A Logistic Regression model was ajusted for the LGD data, dividing the data in two
groups: low LGD and high LGD, creating a binary responde variable. Several authors used
this kind of approach to predict LGD. Belotti and Crook (2008) used logistic regression
and obtained good results. Loterman et al. (2012) and Leow and Mues (2012) used binary
models to predict LGD and found that this kind of model has better results in predicting
LGD.

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

2
2.1

Theoretical Framework
Calculation of LGD

There are a few methodologies for the calculation of LGD. Chalupka and Kopecsni
(2008) defined these as Market, Workout and Implied Market LGD. The Market LGD is
calculated using the prices of marketable securities and loans after the default, the Workout LGD is calculated using the cash flow resulting from the recovery period, discounted
to present value at the time of default and the Implied Market LGD is calculated using
the prices of risky assets, but not yet in default, using a theoretical model of asset pricing.
The most commonly used approach for calculation of LGD is the Workout LGD (Silva
et al., 2008). According to Schuermann (2004), this is the most complex calculation, which
involves some important decisions, such as the definition of the recovery time, the caution
to appropriately include the costs of such recovery and the definition of the discount rate
to calculate the present value of this flow. Workout LGD formula considering all losses
given by Silva et al. (2008) is presented in equation (1).
P ayments Costs
(1)
EAD
Where payments are indicated by the amount recovered after the default, the costs
are the costs of recovery and EAD is the exposure at the time of the default.
LGD = 1

2.2

Factors that influence LGD

Studies on LGD models began in the 90s, with Asarnow and Edwards (1995), who
analyzed the behavior of the rate of loss during 24 years in the U.S. bank Citibank.
They showed that there is a difference of the LGDs between commercial and industrial
operations and structured operations (loans with greater monitoring are more structured
and have greater governance). Operations with securities have average LGD of 12% and
operations without securities have average LGD of 35%. Hurt and Felsovalyi (1998)
developed a study similar to that of Asarnow and Edwards (1995), as they also analyzed
the behavior of the rate of loss of Citibank, but it focuses on Latin America and covers 27
years of database. This study resulted in a calculation of LGD of 32%, and the exposure
amount influence on LGD.
Gupton, Gates, and Carty (2000), in a study conducted by Moodys with 181 bank
loans, indicated that the average of the LGDs of secured loans is 52% and the average
LGD of unsecured loans is 69%. Gupton and Stein (2002) documented Mooodys model
for LGD, the LossCalc. It is a multivariate model developed with 1,800 securities, loans
and preferred shares. They used factors grouped into four categories: type of debt and
seniority, the companys capital structure, industry indicators and macro-economic factors. Schuermann (2004) mentioned the distribution of two peaks of LGD using data from
Moodys financial securities, resulting in an average LGD of 40%, with the presence and
quality of the security, the companys industry, the degree of subordination of the bond
and the economic situation are factors that influence on LGD.
Querci (2005) studied the behavior of 15,827 loans at an Italian medium sized commercial bank and found that the LGD of retail loans has an average of 53.3%, while small
and medium enterprises have an average of 48.4%. He also mentioned the distribution of
two peaks of LGD data, and studied the following explanatory variables: place where the
client lives, segment, type of loan, security, size of the recovery time (workout) and the

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

client code. Only the client code successfully explains the variability of the LGD (this
variable represents some characteristics of the client).
Dermine and Carvalho (2006) developed a study with 371 credit operations of companies from a bank in Portugal, found an average LGD of 29% and showed that variables
such as the loan amount, security and the companys industry influence the LGD. The
distribution of two peaks in this study was also verified.
Belotti and Crook (2008) developed predictive models of LGD with credit card products in Britain. Variables such as length of relationship, income, number of credit cards,
time at address, type of employment, credit or behavioral scores, debit balance and region
of residence are variables that can influence the LGD value of the contract.
Chalupka and Kopecsni (2008) studied losses of companies at default and the average
of the LGD was 52%. It is also possible to observe the distribution with two peaks, and
the exposure, recovery time, type of industry and company time are variables that can
affect the loss of contracts.
Belotti and Crook (2009) developed another study of predictive models of LGD with
credit card products in Britain, adding macroeconomic variables to their previous study
and found that, due to the period selected, it was not possible to prove that these macroeconomic variables influence the LGD just as the previously mentioned variables.
Qi and Yang (2009) studied mortgage loans and showed that LGD can be explained
by the percentage borrowed relative to the total value of the property, also called loanto-value, the depreciation of the property, the amount of the loan, the type of customer,
the purpose of financing, if the property is used for housing and the time passed since the
beginning of the loan.
Silva et al. (2008) studied 9,557 operations registered in the SCR (Credit Information
System of the Central Bank of Brazil) in the modalities: overdraft secured-check, current
account overdraft and working capital. They also observed the distribution of two peaks
of LGD, a mean of 47% and that the outstanding balance, the length of customer relationship, presence of security, customer size, rating, presence of renegotiations and the
companys industry sector affect the value of the LGD. The analysis of the LGD was also
recommended, focusing on the segment of contracts (retail, specific financing and others).
2.3

Statistical Models for LGD

Belotti and Crook (2008) developed predictive LGD models for credit cards in Britain,
and compared ordinary least squares regression with tobit regression and decision tree.
OLS models are as good as tobit regression and decision tree.
Bastos (2010) compared fractional regression and nonparametric regression tree using
data on loans to small and medium enterprises from a bank of Portugal, and suggests
that regression tree presented estimates closer to the true values when the workout time
recovery are 12 to 24 months.
Matuszyk et al. (2010) used decision tree to predict LGD for loans from a financial
institution in the UK, covering a period from 1989 to 2004. The decision tree indicates
a two stage modeling process: First LGD was reclassified as binary, where Yi = 0 if
LGD 0 and Yi = 1 if LGD > 0, and a logistic regression model was used, identifying
that exposure amount, history of payment delays, the time in which the client lives at
the same address and if there is another loan applicant influence the binary variable.
From this models response, if the loan is classified as LGD > 0, the value of LGD
was predicted from a linear regression using historical payment delays, the clients score,
exposure amount and time to default as predictive variables.
4

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

Zhang and Thomas (2012) compared a linear regression model with survival analysis
in predicting LGD for unsecured and credit cards loans from a financial institution in the
UK, covering a period from 1987 to 2003. Linear regression obtained better results than
survival analysis for both cases: where LGD was modeled directly and where a mixture
of distributions was used as a two stage model similar to Matuszyk et al. (2010).
Loterman et al. (2012) compared LGD modeling techniques. One stage techniques
such as linear regression, beta regression, robust regression, ridge regression, splines regression, neural networks and regression tree, and two-stage models, that combine more
than one technique, were compared. The study indicated that most of the LGD variability can not be explained, and nonlinear models, such as neural networks, achieve better
results than more traditional linear models. Two-stage models that combine linear and
nonlinear techniques also obtain good accuracy.
Leow and Mues (2012) compared a two-stage model, which creates a logistic regression
model for the probability of loss and a linear regression model to the severity of the loss,
with a one stage linear regression model. The authors found that the two-stage model
has better performance and accuracy in predicting LGD.
3
3.1

Methodology
Data

According to Basel II, institutions must subject to a pillar called Market Discipline
(Pillar 3) in which, under paragraph 809, market participants can access key information on capital, risk exposures, risk management processes and the capital adequacy of
institutions. These institutions usually release an annual risk management report for the
market.
In this paper, we analyzed 12 risk management reports of financial institutions (Australia and New Zealand Banking Group, Barclays, Commonwealth, Credit Suisse, Danske
Bank, Deutsche Bank, National Australian Bank, NIBC, RBS, Santander, SNS Banks
and Westpac) from December/2008 to March/2011.
The selected data refer to retail exposures of homogeneous risk groups. In accordance
with paragraphs 401 and 402 of Basel II, each retail exposure must be allocated in a homogeneous risk group, which should provide risk differentiation, homogeneous exposures and
allow accurate and consistent estimations of loss characteristics. It should also consider
the characteristics of the borrower, the operation and default (BIS, 2004). The data base
is composed of 214 homogeneous risk groups of credit loans from these 8 countries.
PD, LGD, EAD (in thounsands of euros) and Basel segment classification were selected. Basel segment classification is defined by Basel II Accord, dividing loans as small
and medium enterprise loans, residential mortgage loans, qualifying revolving retail exposures (such as credit cards) and other loans.
3.2

Logistic Regression

Since the purpose of this paper is to analyze the factors that influence the classification
of loss given default and, consequently, the credit recovery, the binary logistic regression
analysis is the most appropriate.
We defined the binary variable as Yi = 1 for groups with LGD greater than or equal
to 50%, that is, high loss and Yi = 0 for groups with LGD lower than 50%, that is, low
loss. The model used to explain the probability of Yi = 1 is the binary logistic regression
5

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

model. The binary logistic regression relates a binary response variable with explanatory
variables, which can be categorical, continuous or discrete.
According to Hosmer and Lemeshow (2000) the ideal function to model binary cases
is the logit function. The model estimates the probability of one of the binary responses,
in this case we estimated the probability of high loss, which is presented in the equation
2.
e0 +1 x1 +2 x2 ++k xk
(2)
1 + e0 +1 x1 +2 x2 ++k xk
Where B0 , B1 , . . ., k are the model parameters related to the k explanatory variables,
(x) is the estimated probability of high-loss and xk is the k-th explanatory variable.
(x) =

4
4.1

Results
Descriptive Statistics

Table 1 shows the distribution of homogeneous risk groups for the countries. The
database consists of 214 groups of contracts arising from 8 countries. The groups were
developed to be homogeneous with respect to the risk of retail exposures and were obtained
from 12 financial institutions.
Table 1: Distribution of groups for the countries
Country

Number of Groups

Australia
Scotland
Denmark
Spain
Germany
Netherlands
Switzerland
England
Total

83
31
30
21
18
14
9
8
214

38.79%
14.49%
14.02%
9.81%
8.41%
6.54%
4.21%
3.74%
100.00%

Tables 2, 3 and 4 present the descriptive analysis of the LGD variables (%), EAD
(million dollars) and PD (%) for each country used.
Note that England is the country with the highest mean value of LGD, with the
highest median of PD and highest minimum, mean value, median, maximum and standard
deviation of exposure (EAD), indicating the highest risk among the countries analyzed.
In addition, it presents the highest standard deviation of the LGD component. In the case
of the PD component, England has the lowest coefficient of variation, the highest median,
the highest minimum value and the lowest maximum value, indicating concentration and
little discrimination of risk.
In the case of component LGD (loss given default), Australia and Scotland have the
highest means after England (approximately 50%). These countries also have the highest
means of PD, indicating that the percentage of expected loss of these countries, represented by the PD multiplied by the LGD (BIS, 2004) are higher than the others.
Netherlands is the country that indicates the lowest average of loss given default (LGD)
and the lowest average of exposure at default (EAD), thus being one of the countries with
the lowest risk compared with the others.

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance


Table 2: Descriptive Analysis of LGD (%) for the countries
Country
Australia
Scotland
Denmark
Spain
Germany
Netherlands
Switzerland
England

Min. (%)

Average (%)

Median (%)

Max. (%)

S.D.(%)

C.V.

20.0%
1.9%
12.0%
13.0%
9.3%
8.2%
16.7%
14.9%

53.2%
50.8%
26.4%
39.4%
32.9%
13.2%
37.6%
60.6%

62.0%
70.6%
24.5%
39.3%
38.5%
9.8%
35.7%
67.4%

97.9%
100.0%
44.0%
79.5%
53.3%
36.7%
60.0%
97.1%

27.5%
29.1%
10.7%
24.3%
13.2%
7.7%
14.8%
30.5%

52%
57%
41%
62%
40%
59%
39%
50%

Table 3: Descriptive Analysis of Average PD (%) for the countries


Country
Australia
Scotland
Denmark
Spain
Germany
Netherlands
Switzerland
England

Min. (%)

Average (%)

Median (%)

Max. (%)

S.D.(%)

C.V.

0.1%
0.0%
0.0%
0.0%
0.0%
0.0%
0.1%
2.3%

9.3%
9.9%
8.7%
6.2%
3.9%
5.4%
3.5%
5.1%

1.5%
1.6%
0.3%
0.8%
0.7%
1.2%
0.5%
4.8%

54.5%
58.6%
62.8%
35.5%
18.3%
32.3%
10.0%
8.5%

17.9%
19.3%
19.0%
11.3%
6.5%
9.2%
4.9%
2.6%

191%
195%
220%
183%
165%
170%
138%
50%

Table 5 shows the number of homogeneous risk groups by Basel Segment for each
country, and note that the Netherlands only contains contracts of Real Estate Loans,
England, Scotland and Australia have an approximately equal distribution of their contracts to all segments and Denmark, Spain, Germany, Netherlands and Switzerland do
not have contracts classified as Small and Medium Enterprises.
The measures of LGD, PD and EAD of England were obtained through the exploratory
analysis of data provided in Table 6, which presents the 8 homogeneous risk groups of
Barclays Bank, the only bank that generated a report of the operations in England. Note
that the database contains a high exposure to real estate loans, and these exposures
indicate the lowest risk, verified by the lowest mean of PD and LGD of this group.
Figure 1 shows the distribution of LGD. Note the existence of two peaks, as noted in
the literature, for example Schuermann (2004).
Table 7 indicates that revolving credit agreements have higher average loss given default (LGD) and mortgage contracts have lower average loss given default (LGD).
Table 4: Descriptive Analysis of EAD (US$ million) for the countries
Country

Min.

Average

Median

Max.

S.D.

C.V.

Australia
Scotland
Denmark
Spain
Germany
Netherlands
Switzerland
England

2
57
16
460
80
179
19,842

15,324
7,792
4,868
20,404
6,142
3,358
14,689
79,02

2,597
3,695
1,678
3,846
2,175
1,754
2,093
34,562

182,386
68,79
23,151
70,269
31,047
8,869
74,205
232,485

34,036
14,096
6,718
27,053
8,745
3,294
26,119
89,275

222%
181%
138%
133%
142%
98%
178%
113%

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance


Table 5: Number of Homogeneous Risk Groups by Basel Segment for each country
Country

Real Estate Loans

Others

Revolving Credit

Small and Medium Enterprises

24
9
10
7
6
14
4
2
76

24
7
10
7
6
0
4
2
60

23
9
10
7
6
0
1
2
58

12
6
0
0
0
0
0
2
20

Australia
Scotland
Denmark
Spain
Germany
Netherlands
Switzerland
England
Total

Table 6: Homogeneous risk groups


Bank

Date

Barclays
Barclays
Barclays
Barclays
Barclays
Barclays
Barclays
Barclays

Dec/2009
Dec/2009
Dec/2009
Dec/2009
Dec/2010
Dec/2010
Dec/2010
Dec/2010

PD (%)

LGD (%)

EAD (US$ million)

Segment

6.87%
2.42%
3.88%
8.10%
8.49%
2.31%
2.78%
5.75%

61.8%
16.2%
85.1%
73.1%
57.2%
14.9%
79.6%
97.1%

21,489
210,682
46,69
22,433
20,784
232,486
57,753
19,842

Small and Medium Enterprises


Real Estate Loans
Revolving Credit
Others
Small and Medium Enterprises
Real Estate Loans
Revolving Credit
Others

Figure 1: Distribution of LGD

4.2

Logistic Regression

By dividing the response variable, loss given default, into two groups using the cutoff
in the LGD of 50%, it is possible to notice a greater amount of risk groups in the class of
high loss (136 observations), and note also that the average PD is similar for these two
groups. The average EAD is higher in groups with low loss. These data are presented in
Table 8.
Table 9 shows the number of homogeneous groups classified as high and low loss in
each Basel segment classification. Most homogeneous risk groups classified as real estate
loans obtained loss lower than 50%, indicating lower risk in this component. Note also
that most homogeneous risk groups classified as revolving credit obtained loss greater
than or equal to 50%, indicating the highest risk related to this component of all segment
classifications available.
To represent the categorized variable (segment) in a binary logistic regression, we
8

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance


Table 7: Average LGD for each Basel segment
Segment

Average LGD (%)

Real Estate Loans


Revolving Credit
Others
Small and Medium Enterprises
Total

17.0%
65.7%
54.6%
41.7%
43.0%

Table 8: Number of groups, average PD and exposure for the binary categories of LGD
Class of LGD

PD (%)

N. of Observations

LGD < 50%


LGD 50%

EAD (US$ million)

Average

Standard Deviation

Average

Standard Deviation

7.4%
8.8%

15.2%
17.0%

19,283
4,967

38,753
10,134

136
78

Table 9: Number of groups of each Basel segment for the binary categories of LGD

Class of LGD

N. of Observations

Real Estate Loans

Others

Revolving Credit

SME1

LGD < 50%


LGD 50%

136
78

75
1

29
31

17
41

15
5

Small and Medium Enterprizes

created three dummy variables (indicators of 0 and 1 in case the observation presents the
field observed) so that the small and medium enterprises segment is the reference class
and receive the value 0 in the three dummies created , as shown in Table 10.
Table 10: Dummy variables for the Basel segment classification
Field

Real Estate Segment

Others Segment

Revolving Segment

Real Estate Loans


Others
Revolving Credit
Small and Medium Enterprises

1
0
0
0

0
1
0
0

0
0
1
0

The indicators of exposure (EAD) and segment (real estate, others, revolving and
SMEs) using a descriptive level of 10%, are determinants in the probability of high loss
of the group. The indicator of average PD does not discriminate the loss very well (descriptive level higher than 0.10) and was excluded from the regression model. The values
of the parameters and descriptive levels are presented in Table 11.
Table 11: Parameters and descriptive level of the logistic regression
Descriptive

Parameter Estimator

Descriptive Level

Intercept
Average PD
EAD (US$ million)
Real Estate Segment
Others Segment
Revolving Segment
SME Segment

-1.142600000
0.470400000
-0.000000475
-3.101800000
1.352900000
2.055800000
0.000000000

0.0309
0.6757
0.0922
0.0062
0.0207
0.0005
.

By excluding the variable Average PD, we noted that the variables EAD and Segments
are still significant for the model. It is possible to notice, through the parameter of the
9

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

model, that the real estate segment is the least risky with regard to the high probability
of loss, followed by loans to small and medium enterprises (SMEs), others and, finally,
revolving credit. It is also evident that the greater the exposure (EAD) the lower the
probability of high loss. These results are shown in Table 12.
Table 12: Parameters of the logistic regression without the PD variable
Descriptive

Parameter Estimator

Descriptive Level

Intercept
EAD (US$ million)
Real Estate Segment
Others Segment
Revolving Segment
SME Segment

-1.096200000
-0.000000476
-3.110800000
1.345400000
2.046300000
0.000000000

0.0338
0.0911
0.0060
0.0213
0.0006
.

Table 13 shows the odds ratio for the variable segment. The ODDS indicate that the
real estate segment is the least risky with regard to the probability of loss, followed by
loans to small and medium enterprises, others and, finally, revolving credit. Revolving
credit agreements present approximately seven times more chances of high-loss than the
contracts of small and medium enterprises and other credit contracts present approximately four times more chances than contracts of small and medium enterprises.
Table 13: ODDS Ratio of the Variable Segment
Descriptive

ODDS Ratio

Real Estate Segment vs. SME


Others Segment vs. SME
Revolving Segment vs. SME

0.045
3.840
7.739

As previously mentioned, the logistic regression shows that groups with higher exposure are less likely to indicate high loss. This is due to the fact that the least risky
segment, the real estate credit, indicates exposure values higher than the others. These
two variables are not correlated as a whole (correlation of -0.20) but it is possible to notice
that the average of the exposure to real estate credit is greater than the others. These
results are shown in Table 14.
Table 14: Exposure to real estate credit and other segments (US$ million)
Descriptive

Average EAD (US$ million)

Real Estate Segment


Others Segments

29,334
4,133

The cutoff point that maximizes the percentage of hits of the model is 50%, and with
this cutoff point the model makes the correct prediction for 79% of the observations,
and out of the groups that presented low loss, the model correctly classified 71% of the
observations, and out of the groups that presented high loss, the model correctly classified
92% of the observations. The matrix that indicates the classifications resulting from the
cutoff point 50% is shown in Table 15.
In accordance with paragraph 444 of Basel II, internal models of loss should have a
key role in risk management. Models built solely for purposes of Basel are not acceptable.
This means that financial institutions that were analyzed in the study use their models for
management, and this could be a possible variable to influence loss, because as institutions
10

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance


Table 15: Matrix of assertiveness of the model
Estimate

Performed
Low Loss
High Loss
Total

Low Loss

High Loss

97
6
103

39
72
111

Total
136
78
214

use models in credit recovery, for instance, they could make this recovery more efficient
and achieve a lower loss.
5

Conclusions and recommendations

The study found that the segment of the contract (real estate, revolving credit, others
or small and medium enterprises) and the size of the exposure at the time of default (EAD)
are variables that influence the probability of high loss given default (LGD), confirming
results from Asarnow and Edwards (1995); Hurt and Felsovalyi (1998); Gupton et al.
(2000); Schuermann (2004); Dermine and Carvalho (2006); Belotti and Crook (2008);
Chalupka and Kopecsni (2008); Silva et al. (2008), especially regarding to the loan segment
classification and its collateral. The variable probability of default (PD), for this database,
did not affect the LGD. The logistic model correctly classified 92% of the observations of
high loss given default (LGD 50%) and 79% of total observations. For further studies,
we recommend a study on the influence of the use of models in loss management, as
they are mandatory for the application of Basel. As institutions better differentiate their
customers in credit recovery, for instance, they can make this recovery more efficient and
achieve a lower loss.

11

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

References
Asarnow, E., & Edwards, D. (1995). Measuring loss on defaulted bank loans: A 24-year
study. Journal of Commercial Lending, 77 , 1123.
Bastos, J. (2010). Forecasting bank loans loss-given-default. Journal of Banking and
Finance, 34 , 25102517.
Belotti, T., & Crook, J. (2008). Modelling and estimating loss given default for credit
cards (Tech. Rep.). Edinburgh: University of Edinburgh Business School, Credit
Research Centre.
Belotti, T., & Crook, J. (2009). Loss given default models for uk retail credit cards
(Tech. Rep.). Edinburgh: University of Edinburgh Business School, Credit Research
Centre.
BIS. (2004). International convergence of capital measurement and capital standards - a
revised framework (Tech. Rep.). BASEL COMMITTEE ON BANKING SUPERVISION - BCBS. Retrieved 17 jun. 2011, from http://www.bis.org
Chalupka, R., & Kopecsni, J. (2008). Modelling bank loan lgd of corporate and sme
segments: A case study. Charles University Prague, Faculty of Social Sciences,
59 (4), 360382.
Dermine, J., & Carvalho, C. (2006). Bank loan-loss provisioning, methodology and
application. Journal of Banking and Finance, 30 , 12191243.
Gupton, G., Gates, D., & Carty, L. (2000). Bank loan loss given default (Tech. Rep.).
New York: Moodys Investors Service. Global Credit Research.
Gupton, G., & Stein, R. (2002). Losscalc: Model for predicting loss given default (Tech.
Rep.). New York: Moodys Investor Service, Global Credit Research.
Hosmer, D. W., & Lemeshow, S. (2000). Applied logistic regression (2nd ed.). New York:
John Wiley & Sons.
Hurt, L., & Felsovalyi, A. (1998). Measuring loss on latin american defaulted bank loans, a
27-year study of 27 countries. The Journal of Lending and Credit Risk Management,
80 , 4146.
Kapstein, E. (1994). Supervising international banks: Origins and implications of the
basel accord (Vol. 185; Tech. Rep.). Princeton: Essays in International Finance.
Leow, M., & Mues, C. (2012). Predicting loss given default (lgd) for residential mortgage
loans: A two-stage model and empirical evidence for uk bank data. International
Journal of Forecasting, 28 (1), 183195.
Loterman et al., G. (2012). Benchmarking regression algorithms for loss given default
modeling. International Journal of Forecasting, 28 , 161170.
Matuszyk, A., Thomas, L., & Mues, C. (2010). Modelling lgd for unsecured personal
loans: Decision tree approach. Journal of the Operational Research Society, 61 ,
393398.
Qi, M., & Yang, X. (2009). Loss given default of high loan-to-value residential mortgages.
Journal of Banking and Finance, 33 (5), 788799.
Querci, F. (2005). Loss given default on a medium-sized italian banks loans: an empirical
exercise (Tech. Rep.). Milan: European Financial Management Association.
Santos, J. (2001). Bank capital regulation in contemporary banking theory: A review of
the literature. Financial Markets, Institutions and Instruments, 10 (2), 4184.
Schuermann, T. (2004). What do we know about loss given default? d. shimko, credit
risk: Models and management (2nd ed.). Londres: London UK Risk Book.
Silva, A., Marins, J., & Neves, M. (2008). Loss given default: um estudo sobre perdas em

12

Sixth Brazilian Conference on Statistical Modelling in Insurance and Finance

operacoes prefixadas no mercado brasileiro. In Anais... (pp. 115). Rio de Janeiro:


XXXII EnANPAD.
Zhang, J., & Thomas, L. C. (2012). Comparisons of linear regression and survival analysis
using single and mixture distributions approaches in modelling LGD. International
Journal of Forecasting, 28 , 204215.

13

S-ar putea să vă placă și