Internal ratings-based model introduction to a

retail portfolio and its impact on a bank's

capital requirement

Aleksandr Peussa

University of Helsinki
Faculty of Social Sciences
Economics
Master’s Thesis
December 2019
Tiedekunta - Fakultet - Faculty Laitos - Institution - Department

Faculty of Social Sciences Department of Political Science and Economics

Tekijä - Författare - Author

Aleksandr Peussa
Työn nimi - Arbetets titel - Title

Internal ratings-based model introduction to a retail portfolio and its impact on a bank's capital
requirement
Oppiaine - Läroämne - Subject

Economics
Työn laji/ Ohjaaja - Arbetets art/Handledare - Level/Instructor Aika - Datum - Month and year Sivumäärä - Sidoantal - Number of pages

Master’s Thesis October 2019 26 pages

Tiivistelmä - Referat - Abstract

The bank as a business tries to maximize the profit measured by the return on allocated capital
(ROAC). This measure is calculated by dividing net income on capital allocated, thus the bank should
both maximize net income and minimize capital allocated on the portfolio.

The idea of the research is to check whether the higher return on allocated capital may be achieved if a
bank calculates a capital requirement based on an internal ratings-based (IRB) approach instead of a
standardized approach. The research question hypothesizes that the introduction of the IRB approach
may lower the capital allocated on the retail portfolio, as the bank currently uses a standardized
approach for the capital requirement calculation. In case the hypothesis is valid, the introduction of the
IRB model lowers minimum capital requirements and through it, the higher ROAC is achieved.

The capital requirement under the IRB approach is calculated through the estimation of the probability
of default (PD) and loss given default (LGD) based on the bank's historical data. Probability of default
is estimated using logistic regression and for loss given default estimation, the simple linear formula is
applied due to unavailability of relevant data.

The research proves that the application of the IRB approach significantly lowers bank's capital
requirement and as a result increases its return on allocated capital.
Avainsanat – Nyckelord - Keywords

Credit risk, banking, capital requirement, Basel Accords

Muita tietoja - Övriga uppgifter - Additional information
Contents

1. Introduction .................................................................................................................... 1
2. Literature review ............................................................................................................ 3
3. Basel Accords ................................................................................................................ 5
3.1.Value at Risk .................................................................................................................. 5
3.2.Capital Requirements ..................................................................................................... 6
3.3.Standardized Approach .................................................................................................. 7
3.4.Internal rating based Approach ...................................................................................... 7
4. Credit measures and risk weight formula ...................................................................... 9
4.1.Probability of default ..................................................................................................... 9
4.2.Loss given default ........................................................................................................ 11
4.3.Exposure at default ...................................................................................................... 12
4.4.Risk Weight under A-IRB ........................................................................................... 13
5. Research question ........................................................................................................ 15
5.1.Return on allocated capital........................................................................................... 15
5.2.The research question hypothesis ................................................................................ 15
6. Data sources ................................................................................................................. 17
6.1.Application data ........................................................................................................... 17
6.2.Credit bureau data ........................................................................................................ 18
6.3.Behavioral data ............................................................................................................ 18
6.4.Collateral data .............................................................................................................. 18
7. Optimal model selection .............................................................................................. 20
7.1.PD model selection ...................................................................................................... 20
7.2.LGD model selection ................................................................................................... 22
8. Estimation .................................................................................................................... 23
8.1.Data ............................................................................................................................. 23
8.2.PD model .................................................................................................................... 24
8.3.LGD model ................................................................................................................. 29
8.4.Risk-weighted assets .................................................................................................... 30
9. Results .......................................................................................................................... 31
10. Conclusion .................................................................................................................. 33
11. Bibliography ............................................................................................................... 34
 1. Introduction

Banks and other financial institutions receive cash inflow from various sources. For instance,
there are bank deposits like saving accounts or bonds issued. A bank uses the money obtained
in order to make different investments, which always contain a risk associated with the fact
that obligor could default and not pay the loan due to various reasons. As the government
provides the banks’ customers with deposit insurance and as financial institutions play a big
role in the economic system, the government aims to ensure that financial institution is well
protected against the risks it is exposed to. Hence, a bank is supposed to hold enough
shareholder capital to withstand an unexpected loss in order to avoid insolvency. (Baesens
2016, 5.)

According to the Basel Accord, the financial institution has a choice between two broad
methodologies for calculating their minimum capital requirements for credit risk:
standardized approach and internal ratings-based approach. The approaches differ in terms of
their complicity and level of flexibility related to banks’ ability to use own credit risk
estimates. (Joseph 2013, 289.)

The standardized approach refers to the fact that the banks are required to use measures
defined by the regulator in order to estimate a minimum capital required. The values of
measures depend on the structure of the product and the type of borrower. (Witzany 2017,
112.)

The internal ratings-based (IRB) approach is more complicated. Under the approach, the
banks are allowed to use their quantitative models in order to estimate credit risk measures.
The appeal of the internal ratings-based approach is that it may allow banks to obtain a lower
level of capital requirement compared to the standardized approach. However, banks can use
IRB only after approval from their local regulators - central banks. (Joseph 2013, 291.)

The bank as a business tries to maximize the profit measured by the return on allocated
capital (ROAC). This measure is calculated by dividing net income on capital allocated,
thus, the bank should both maximize net income and minimize capital allocated on the
portfolio.

1
The idea of the research is to check whether the higher return on allocated capital may be
achieved if the bank calculates a capital requirement based on the IRB approach instead of a
standardized approach. The research question hypothesizes that the introduction of the IRB
approach may lower capital allocated on the retail portfolio, as the bank currently uses a
standardized approach for the capital requirement calculation. In case the hypothesis is valid,
the introduction of the IRB model lowers minimum capital requirements and through it, the
higher ROAC is achieved. (Balin 2008, 8.)

The first part of the thesis focuses on the theoretical aspects of the research question. Firstly,
the idea behind Basel Accords and methodologies proposed for minimum capital
requirements calculation are explained. Then the credit risk measures, the data and
econometric methods for these measures estimation and validation are reviewed. Next, paper
shows how credit risk measures are used in the internal rating-based model.

The second part contains empirical analysis, which is based on the theoretical part and the
data from a bank. The bank operates in a private segment and provides secured products like
cars and other asset financings. The bank does not provide mortgages or revolving products,
thus the data contains non-mortgage and non-revolving products for private customers only.
Based on the empirical data from the bank, the purpose of the research question is to prove
whether the introduction of the IRB approach would lower the bank's capital requirement and
through this would increase portfolio profitability measured by ROAC.

2
 2. Literature review

There are no straightforward references to similar researches. The assumption is that due to
compliance and corporate security reasons there is no easy access to unique bank's data in
order to cover similar research questions. However, based on multiple pieces of evidence IRB
approach is widely used in the banking industry. Based on European Banking Authority at
least 102 financial institutions have the IRB approach applied to their portfolios, which cover
64% of EU institutions’ total credit risk-weighted exposures (European Banking Authority
2017, 13.)

This is a strong signal that the IRB approach has a good potential to be applied for a bank's
retail portfolio. However, it is not inevitable that the introduction of the IRB approach would
necessarily decrease the capital requirement. The capital requirement under the IRB approach
depends on portfolio agreements' probability of default (PD) and loss given default (LGD).
Hence, in order to calculate capital requirements under the IRB approach, the PD and LGD
estimations are required.

The PD estimation part is mainly covered in my previous research (Peussa, 2016), which
shows that customers’ probability of default can be estimated by logistic regression and
shows what type of data is required for PD model estimation. However, the method of PD
model estimation in the current research is enriched by multicollinearity analysis through
correlation matrix and economic power considerations by using Kullback-Leibler information
divergence.

The focus of previous research is to show that the statistical model can be used for
underwriting purposes and prove that the credit policy based on the statistical model is more
efficient compared to manual underwriting policy. The previous research is based on the data
from another bank specialized in unsecured products. The generalized result of that research
is that every bank might benefit from using statistical models for underwriting purposes.

The purpose of current research is to check whether the introduction of the IRB approach
would increase the bank's profitability. While this research partly utilizes findings of my
previous research, the purpose and contribution of new research differ remarkably from the
previous one. The PD model estimation is only one part needed for capital requirement
calculation under the IRB approach. This research covers the loss given default (LGD) part

3
also using the collateral data, which is unavailable in the previous research. Hence, the
outcome of current research depends on the portfolio's PD and LGD distributions and the
bank would not necessarily benefit from the PD model estimation in terms of capital
requirement regardless of the contribution level of PD model itself to underwriting methods.

4
 3. Basel Accords

The banks traditionally have a high level of leverage, as it is obvious that banks funded
completely by equity are not feasible. The leverage means that a bank is funded mainly by
deposits, bonds, and other liabilities. Hence, equity capital represents just a small part of
banks’ assets, which is the main reason why banking is the most regulated private business in
the world. (de Servigny 2004, 387.)

3.1. Value at risk

The main guiding principle of the Basel regulation is presented in Figure 3.1. Banks and
other financial institutions regularly suffer from expected and unexpected credit losses. The
expected part of the losses – expected loss (EL) is considered as the cost of doing business
and should be covered not from the capital but from annual revenues, which means that EL is
factored into the products’ pricing. In other words, expected losses are provisioned for, where
provisions are losses in the P&L statement. This explains why expected loss is not part of the
capital requirement calculation and is covered by another banking regulation – IFRS9, which
is out of the research question scope.

The expected loss is calculated according to the formula shown below:

𝐸𝐿 = 𝑃𝐷 × 𝐿𝐺𝐷 × 𝐸𝐴𝐷 (3.1)

where PD is the probability of default, LGD is loss given default and EAD means exposure at
default. The next section focuses on a detailed description of these parameters.

The unexpected loss (UL) is the difference between the value at risk (VaR) and the expected
loss (EL). The bank or financial institution calculates the value at risk in a similar way as an
expected loss but uses downturn estimates for PD, LGD, and EAD, which means that these
measures are stressed in order to reflect economic downturn – adverse economic scenario.
(Baesens 2016, 11.)

5
Figure 3.1 Value-at-Risk model

The unexpected loss is not provisioned through profit and loss statement, therefore it must be
charged against the capital. Banks and financial institutions have to ensure that they have
sufficient capital to cover UL in order to protect the depositors and other creditors.
Otherwise, bank insolvency or failure happens, which means that the bank is not able to
fulfill its obligations anymore and the government should cover depositors’ losses by deposit
insurance. Hence, the goal of the regulation is to establish a procedure estimating the
potential unexpected loss on a regulatory probability level. (Witzany 2017, 15.)

3.2. Capital requirements

The pillar of Basel regulation, which reflects the minimum capital requirement, is called the
capital adequacy ratio (CAR) and is defined as the regulatory capital divided by the risk-
weighted assets (RWA):

𝑅𝑒𝑔𝑢𝑙𝑎𝑡𝑜𝑟𝑦 𝐶𝑎𝑝𝑖𝑡𝑎𝑙
𝐶𝐴𝑅 = ≥ 8% (3.2)
𝑅𝑊𝐴

The regulatory capital is an amount of capital a bank should have according to regulation in
order to protect itself against unexpected loss (UL).

Risk-weighted assets (RWA) is a sum of agreements’ exposure at defaults (EAD) multiplied

by the agreement’s specific risk weights (RW):

6
 𝑛

𝑅𝑊𝐴 = ∑ 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖 (3.3)

𝑖=1

where 𝑛 refers to a number of agreements in the portfolio.

The regulatory capital can be derived from the formula(3.2):

𝑅𝑒𝑔𝑢𝑙𝑎𝑡𝑜𝑟𝑦 𝐶𝑎𝑝𝑖𝑡𝑎𝑙 ≥ 𝑅𝑊𝐴 × 0.08 (3.4)

Hence, the research focuses on RWA estimation, which can be calculated using several
approaches: standardized (SA) and internal rating-based (IRB).

3.3. Standardized approach

Under the standardized approach, the risk weights are set by regulation based on the product
and the type of borrower. For instance, the government and central bank obligations have a
risk weight equal to 0%, while retail exposures have a risk weight equal to 75% (non-
mortgage). The risk weights provided by the regulation do not have a direct connection to the
IRB approach, thus the regulation does not provide information under what credit parameters
or other assumptions these risk weights are estimated.

Although the standardized approach looks simple and transparent, it suffers in terms of
coverage of various types of exposures. The retail exposures are discriminated only in terms
of mortgage or non-mortgage. (Baesens 2016, 10.)

As the research covers non-mortgage and non-revolving retail portfolios, the straightforward
conclusion is that under the standardized approach the risk weight is equal to 75%.

3.4. Internal rating based approach

Under IRB banks are supposed to use their quantitative models to estimate PD (probability of
default), EAD (exposure at default), LGD (loss given default) and other parameters required
for calculating the RWA (risk-weighted asset). These parameters should be in line with the
value-at-risk model, thus, PD and LGD should be stressed against the economic downturn.

The internal models have to satisfy minimum standards set by the local regulator in terms of
methodology, data quality, length of the observation period, etc. (Witzany 2017, 111.)

For instance, the minimum length of the observation period is 5 years for PD estimation and
7 years for LGD estimation. In addition, the parametric models are preferred to non-
parametric, as the main advantage of parametric models is their interpretability.

7
Now, the total required capital is calculated as a fixed percentage of the estimated RWA
according to the formula (3.4).

IRB benefits banks and financial institutions to hold lower capital requirements as having a
low default portfolio, which means the portfolio of agreements with lower probability of
default compared to the assumption made in the standardized approach. (de Servigny 2004,
400.)

The internal rating-based approach can be divided into two sub approaches – foundation (F-
IRB) and advanced (A-IRB). In the foundation IRB, only PD is calculated internally, while
LGD is defined by Basel Accords or local regulation. In the advanced approach, the bank
estimates all parameters internally and then utilizes these parameters in risk weight
calculation. For retail portfolio the foundation approach is not permitted, hence the research
focuses on the advanced IRB approach (A-IRB). (Baesens 2016, 11.)

8
 4. Credit measures and risk weight formula

This section contains the information regarding credit risk measures needed for the risk
weight estimation under the A-IRB approach. According to the formula (3.3), RWA is a sum
of agreements’ exposure at defaults (EAD) multiplied by the agreements’ specific risk
weights (RW) and reciprocal of the minimum capital ratio. Hence, the risk weight is
estimated on an agreement level; accordingly, all credit measures required for risk weight
estimation are estimated on an agreement level as well.

4.1. Probability of default

A probability of default (PD) describes the likelihood of a default event. The default
definition used in the research is Basel's definition, which is based on payment delinquency
of 90 days or more within 12 months after loan origination. In other words, a defaulter is
defined as an obligor who has days past due (DPD) more than 90 days within the next 12
months after loan origination. (Baesens 2016, 138.)

According to Basel Accords, the bank should stress PD estimate via the concept of a worst-
case default rate given a virtual macroeconomic shock based on a confidence level of 99.9%
and a sensitivity to the microeconomic conditions that are based on the asset correlation. This
logic is included in the risk weight formula, which means that no other PD adjustments are
required. (Joseph 2013, 293.)

The idea of the research is to apply logistic regression to a sample of historical loan
agreements in order to estimate the probability of default for an existing portfolio. The
decision to use the logistic regression is based on a publication of Desai, Crook and
Overstreet (1996), where the authors have proved that logistic regression is a most
appropriate approach compared to other parametric and non-parametric (machine learning)
approaches for binary response variables.

In binary logistic regression, the response variable 𝑦 ∈ {0,1} follows a Bernoulli distribution.
The response variable reflects the default defined by Basel Accords on an agreement level:

1, 𝑖𝑓 𝐷𝑃𝐷 > 90 𝑤𝑖𝑡ℎ𝑖𝑛 12 𝑚𝑜𝑛𝑡ℎ𝑠 𝑎𝑓𝑡𝑒𝑟 𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑡𝑖𝑜𝑛

𝑦= {
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

9
The vector of agreements 𝒚 = (𝑦1 , 𝑦2 , … 𝑦𝑛 ) 𝑇 should be independent, where 𝑛 indicates the
number of agreements in the sample. This assumption holds for the financial institution’s
dataset, as each customer has only one loan agreement and customers’ payment behavior is
expected to be independent on a portfolio level.

The logistic regression has the logit link function, which is an inverse of a standard
cumulative distribution function of the logistic distribution. The logistic regression model has
a linear form for the logit:

𝑃𝐷𝑖
𝑙𝑜𝑔𝑖𝑡(𝑃𝐷𝑖 ) = log ( ) = 𝜷𝑇 𝒙𝒊 (4.1)
1 − 𝑃𝐷𝑖

where 𝑃𝐷𝑖 = 𝑃(𝑦𝑖 = 1) is a probability of default for observation 𝑖, 𝜷 is a regression

coefficient vector and 𝒙𝒊 is a𝑛 (𝑚 + 1) vector containing 𝑚 explanatory variables and a
constant term. (Agresti 2015, 2.)

The possible explanatory variables for the probability of default estimation can be the
duration of living at current address, employment length, customer's age, education, payment
behavior, and other variables gathered from various sources of the data discussed more in
detail in section 6.

The probability of default 𝑃𝐷𝑖 can be derived from equation (4.1) by using the exponential
function:

exp{𝜷𝑇 𝒙𝒊 }
𝑃𝐷𝑖 = (4.2)
1 + exp{𝜷𝑇 𝒙𝒊 }

The standard way to estimate a logistic regression model is a maximum likelihood estimation.
(Agresti 2007, 6.). The likelihood function 𝐿(𝜷) is the probability for the occurrence of a
sample 𝒚 = (𝑦1 , 𝑦2 , … 𝑦𝑛 ) 𝑇 given the Bernoulli probability density:
𝑛

𝐿(𝜷) = ∏ 𝑃𝐷𝑖 𝑦𝑖 (1 − 𝑃𝐷𝑖 )1−𝑦𝑖 (4.3)

𝑖=1

The log-likelihood function 𝑙(𝜷) is equal to:

𝑛

𝑙(𝜷) = log(𝐿(𝜷)) = ∑ 𝑦𝑖 log(𝑃𝐷𝑖 ) + (1 − 𝑦𝑖 ) log(1 − 𝑃𝐷𝑖 ) (4.4)

𝑖=1

Its first derivative with respect to 𝛽𝑗 where 𝑗 ∈ {1, … , 𝑚 + 1} is equal to:

10
 𝑛
𝑑𝑙(𝛽)
= ∑(𝑦𝑖 − 𝑃𝐷𝑖 log(𝑃𝐷𝑖 ))𝑥𝑖𝑗 (4.5)
𝑑𝛽𝑗
𝑖=1

The maximum likelihood estimates for 𝜷 can be found by setting each of the (𝑚 + 1)
equations defined in (4.5) equal to zero and solving for each 𝛽𝑗 .

4.2. Loss given default

The loss given default (LGD) means a ratio of the loss on exposure due to the default of an
obligor to the amount outstanding at default. LGD is calculated as below:

𝐿𝐺𝐷 = 1 − 𝑅𝑅 (4.6)

where RR means Recovery Rate. Recovery rate varies from 0% to 100% and refers to the
recoverable part of the credit asset. The recovery rate is equal to the present value of all cash
flows to the bank (payments from obligor or collateral realization) after the date of default
divided by exposure at default, thus, LGD can be defined as:

𝐸𝐴𝐷 − ∑𝑇𝑡=1(𝐶𝐹𝑡 /(1 + 𝑟𝑡 )𝑡

𝐿𝐺𝐷 = (4.7)
𝐸𝐴𝐷

where 𝐶𝐹𝑡 is the cash flow and 𝑟𝑡 is a discount rate for time 𝑡.

While downturn PD is extrapolated from bank-reported PD via mapping function, which is

part of the risk-weight formula, in case of LGD, the bank is asked to provide downturn LGD
by own mapping function or based on their internal assessment of LGD during adverse
conditions. (Bank for International Settlements 2005, 7.)

The U.S. Department of the Treasury, Federal Reserve System, and Federal Deposit
Insurance Corporation (2006) proposed a linear relationship between the downturn LGD
(DLGD) and the expected LGD. The formula implies a floor of 8% and a cap of 100%:
(Baesens 2016, 460.)

𝐷𝐿𝐺𝐷 = 0.08 + 0.92 × 𝐿𝐺𝐷 (4.8)

In order to estimate expected LGD on the agreement level, the bank utilizes marginal LGD
modeling. This means that LGD is estimated based on the data from defaulted obligors only
and do not consider those cases in which no default happened.

Unlike the probability of default (PD) which follows Bernoulli distribution, LGD is a variable
which primarily follows a Beta distribution. (Yang, Tkachenko, 2012)

11
The beta distribution for the LGD has two parameters 𝛼 and 𝛽 and has the form:

1
𝑓(𝐿𝐺𝐷) = 𝐿𝐺𝐷𝛼−1 (1 − 𝐿𝐺𝐷)𝛽−1 (4.9)
𝐵(𝛼, 𝛽)

where 𝐵(𝛼, 𝛽) is the beta function.

The standard model for LGD estimation is the beta regression model, which is closely related
to the beta distribution. Both parameters of the beta distribution are transformed into a
location (mean) parameter 𝜇 and a shape parameter 𝛿 such that parameter 𝛼 = 𝜇𝛿 and
parameter 𝛽 = (1 − 𝛼)𝛿. (Baesens 2016, 296.)

A regression model is then applied to the location and the shape parameter transformations:

exp{𝛽𝜇𝑇 𝑋}
logit(𝜇) = 𝛽𝜇𝑇 𝑋 ⟺ 𝜇 = (4.10)
1 + exp{𝛽𝜇𝑇 𝑋}

log(𝛿) = 𝛽𝛿𝑇 𝑋 ⟺ 𝛿 = exp{𝛽𝛿𝑇 𝑋} (4.11)

where 𝛽𝜇 and 𝛽𝛿 are regression coefficient vectors and 𝑋 is an 𝑛 × (𝑚 + 1) model matrix

containing 𝑛 observations of 𝑚 explanatory variables and a constant term.

Both coefficient vectors 𝛽𝜇 and 𝛽𝛿 are estimated by maximum likelihood, similar to the one
described in Section 4.1. (Section with capital “S”)

The possible explanatory variables for LGD estimation can be collateral type, the size of
down payment, age at default, and other variables gathered from various sources of the data
discussed more in detail in the next section.

4.3. Exposure at default

Exposure at default (EAD) represents the expected level of usage of the facility utilization
when default occurs. Theoretically, as maturity increases, risk increases, so the probability of
credit loss increases as well. As the research concentrates on a retail portfolio, which does not
include revolving products, the exposure at default is equal to exposure drawn amount and no
further modelling is required. In other words, EAD is defined as the nominal outstanding
balance on an agreement level.

12
 4.4. Risk weight under A-IRB

According to Basel II, the formula for the agreement specific risk-weight (RW) calculation
under advanced IRB for retail exposures is given below:

1 𝑅
𝑅𝑊 = 12.5 × (𝐿𝐺𝐷 × 𝑁 ( × 𝐺(𝑃𝐷) + √1−𝑅 × 𝐺(0.999)) − 𝐿𝐺𝐷 × 𝑃𝐷) (4.12)
√1−𝑅

where 12.5 refers to reciprocal of the minimum capital ratio of 8%, 𝑁 refers to cumulative
normal distribution and 𝐺 to the inverse of cumulative normal distribution. Accordingly,
𝐺(0.999) is an inverse of cumulative normal distribution variable for 99.9% confidence
interval, which means that the institution is expected to suffer losses that exceed regulatory
capital on average once in a thousand years. (Bank for International Settlements 2005, 11.)

The correlation factor 𝑅 depends on PD and asset class. The correlation factor for non-
revolving and non-mortgage exposures uses the next formula:

1 − exp{−35 × 𝑃𝐷} 1 − exp{−35 × 𝑃𝐷}

𝑅 = 0.03 × + 0.16 × (1 − ) (4.13)
1 − exp{−35} 1 − exp{−35}

This formula gives a correlation between 3% and 16% depending on the probability of
default value.

In order to illustrate the dependency of risk weight from PD and LGD under the IRB
approach for a retail portfolio, the next figure is presented based on a formula (4.12).

Figure 4.1 Combinations of PD and LGD values satisfying 75% risk weight

13
This figure shows what PD and LGD levels are required to have a 75% risk weight on the
portfolio level. For instance, if PD is equal to 4% (0.04) then LGD should be equal
approximately to 55% (0.55) in order to reach risk weight equal to 75%.

The risk weight equal to 75% is selected because it is a level defined by a standardized
approach. In other words, the curve shows what combinations of PD and LGD levels within
the portfolio are expected by Basel accords under the standardized approach.

If the bank's average PD and LGD combination lies below the line, the introduction of the
IRB approach would decrease capital requirement. Otherwise, the introduction of the IRB
approach would increase capital requirement compared to a standardized approach. Hence,
the IRB approach is not necessarily beneficial for a bank but the outcome depends on the
portfolio's PD and LGD levels.

14
 5. Research question

The scope of the research question is a non-revolving non-mortgage portfolio and a private
customer segment. The scope is defined by the customer segment and the product space the
bank operates with. The idea of the research is to check whether the higher return on the
allocated capital may be achieved if the bank calculates its capital requirement based on the
IRB approach instead of the standardized approach. The research question hypothesizes that
the introduction of the IRB approach may lower the capital allocated on the retail portfolio, as
the bank currently uses the standardized approach for the capital requirement calculation.

5.1. Return on allocated capital (ROAC)

The bank maximizes the profit measured by the return on allocated capital (ROAC). This
measure is equal to net income divided by regular capital allocated:

𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒
𝑅𝑂𝐴𝐶 = (5.1)
𝑟𝑒𝑔𝑢𝑙𝑎𝑟 𝑐𝑎𝑝𝑖𝑡𝑎𝑙 𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝑑

By inserting formulas (3.3) and (3.4) into (5.1), the return on allocated capital can be
expressed as:

𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒
𝑅𝑂𝐴𝐶 = (5.2)
0.08 × ∑𝑛𝑖=1 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖

where risk weight (RW) is applied to agreement’s exposure at default and either estimated on
an agreement level under A-IRB or defined by regulation under the standardized approach.

The bank currently uses the standardized approach, where the risk weight for retail exposures
is equal to 75%, thus the bank calculates return on allocated capital for a retail portfolio as:

𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒
𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑 = (5.3)
0.08 × ∑𝑛𝑖=1 0.75 × 𝐸𝐴𝐷𝑖

5.2. The research question hypothesis

The hypothesis of the research question is that ROAC under A-IRB approach 𝑅𝑂𝐴𝐶𝐼𝑅𝐵 ,
where agreement based risk weights (RW) are estimated internally by the formula (4.12), is
higher compared to ROAC under bank’s current (standardized approach) 𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑 ,
where the risk weight for retail exposures is equal to 75%.
15
Mathematically, the expected result of the research question can be expressed as:

𝑅𝑂𝐴𝐶𝐼𝑅𝐵 > 𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑

𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒 𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒

𝑛 >
0.08 × ∑𝑖=1 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖 0.08 × ∑𝑛𝑖=1 0.75 × 𝐸𝐴𝐷𝑖

which means that

𝑛 𝑛

∑ 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖 < ∑ 0.75 × 𝐸𝐴𝐷𝑖

𝑖=1 𝑖=1

Hence, the research question aims to estimate agreement based risk weights under A-IRB for
the bank's retail portfolio and to test the hypothesis. If the research proves the hypothesis
validity, the bank has incentive to initiate the transition from a standardized approach to A-
IRB in order to increase its portfolio's profitability and attractiveness from the investors'
perspective.

16
 6. Data sources

The financial intermediary has to collect all data, which is relevant for risk management. The
data available for customers’ credit assessment is of three types: those derived from the
application form, those available from a credit bureau search, and those describing the
transaction history of the borrower. (Peussa 2016, 5.)

However, these types mentioned above cover only the data valid for the probability of default
(PD) estimation. In order to estimate the loss given default (LGD), the collateral data is
required.

6.1. Application data

Application data is information provided by the potential customer. The application form
could contain such characteristics as salary, occupation, number of children, other loans and
so on. When designing an application form, a financial institution faces a trade-off between
the simplicity of the form and the quantity of the information. (Peussa 2016, 5.)

A detailed application form, with a great variety of variables, is attractive from the risk
management perspective. Clearly, the more variables available a lender has, the better the
model can be constructed. The representative application form contains the next variables:
age, gender, education, employment type and length, accommodation type, income, loan
obligations, etc.

However, a long or too detailed application form decreases the probability of its completion,
which cuts sales. Hence, there is a pressure on the lender to make the form as simple as
possible. Furthermore, some questions are not permitted for legal reasons. For instance, the
U.S. Equal Credit Opportunity Acts of 1975 and 1976 made it illegal to discriminate in the
granting of credit on the grounds of race, color, religion, national origin, sex, marital status,
or age. (Thomas 2002, 124.)

The application data is a significant instrument from the risk modeling perspective but it
contains problems as well. First, the data should be carefully looked through and validated.
The most frequently arising problems with the application data are frauds and errors. An
applicant could unintentionally violate information, putting the wrong income, for example.

17
This problem could be partly solved by revealing and eliminating impossible or inconsistent
answers. (Peussa 2016, 6.)

The frauds are a more serious problem. Here, an applicant intentionally violates application
data in order to receive a positive loan decision or better offer. Moreover, if the intermediary
has economic incentives for sending good customers, they might be tempted to advise the
applicant on suitable answers. Therefore, from a modeling perspective, the usage of only
application data could lead to major problems in the end. This is the reason why credit bureau
data is a significant part of IRB modeling.

6.2. Credit bureau data

A credit bureau or credit reference agency is an organization which collects data from various
sources and provides consumer credit information on individual consumers. The data
provided by the credit bureau includes its estimate of customers' probability of default, socio-
demographics, applicant's borrowing, and bill-paying habits. The purpose of the credit bureau
is to reduce the impact of asymmetric information between borrowers and lenders. (Peussa
2016, 6.)

From the IRB modeling perspective, the position of a credit agency is very important. It has
two advantages compared to a bank: the credit bureau validates the data using municipality
information and its customer base, which contains millions of applications and historical
records and is much larger than the bank's portfolio. Hence, a lender protects himself against
violated information, which could happen if the bank uses application data only.

6.3. Behavioral data

Behavioral data is a history of existing customers' transactions and cash flows. In other
words, the data is a conglomeration of customers' payment characteristics and habits. The
most common ones are minimum, maximum or average balance, total value or regularity of
both debit and credit transactions, payment defaults and other delinquency indicators. (Peussa
2016, 10.)

6.4. Collateral data

The collateral data includes information about customers’ paid down payment and the
financed asset's value. The customers' paid down payment increases the level of loans'
securitization and via this decreases the bank's collateral risk.

18
The collateral value of the asset financed is not equal to the purchase price of the asset. The
reason behind this is that collateral value should cover unfavorable scenarios like poor
maintenance of the asset or negative changes in the market demand for the asset. In other
words, it should express the expected value of an asset if customer defaults and the bank
realizes the asset. The collateral value estimate generally is based on variables like asset type,
mileage, and age. The bank provides either its collateral value estimate or orders it from the
third parties like asset evaluation companies.

19
 7. Optimal model selection

Input parameters of risk weight formula - PD and LGD estimates are based on internal
modeling, which is not covered by regulation, which means the better PD and LGD models
perform, the more precise risk weight estimate is achieved. This section describes the
methodology utilized for constructing an optimal model. The optimal model means the best
model from a performance perspective with respect to the data available.

7.1. PD model selection

The response variable 𝑃𝐷 ∈ {0,1} follows a Bernoulli distribution, which means that a
receiver operating characteristic curve (ROC curve) may be used for the model’s
performance estimation.

The ROC curve is a graphical plot that illustrates the performance of a binary classifier
model, thus it is mainly used for Bernoulli distributed response variables. The ROC curve is
created by plotting the true positive rate (TPR) against the false positive rate (FPR) at various
threshold points. The threshold point means the value of the model's estimate at which the
agreement is regarded as positive. In terms of the research question model, positive means
defaulted obligor (PD = 1), while negative means non-defaulted obligor by Basel definition.

The TPR defines how many correct positive results occur among the sample's positive
results. FPR, on the other hand, defines how many incorrect positive results occur among the
sample's negative results. For instance, the green point in figure 6.1 shows that at this specific
threshold, the model predicts correctly 80% of positive ones and 27% is a share of false-
positive ones from the sample's negative ones. The red curve in figure 6.1 represents a ROC
curve for a model estimated. The diagonal line on the figure shows the “random guessing”
model which is expected to have no separating power at all that turns into the equal TPR and
FPR rates. Intuitively, the bigger distance is between diagonal and the ROC curve, the better
is a model’s predictive power.

20
Figure 7.1 ROC curve

Based on the ROC curve, the Gini coefficient may be calculated, which is used as a model
performance indicator. The Gini coefficient is defined as the ratio of the area between the
diagonal and ROC curve to the area of the triangle formed by the diagonal and axes.

If the ROC curve is approximated on each interval as a line between consecutive points, then
the Gini coefficient can be approximated with trapezoids as:
𝑛

𝐺𝑖𝑛𝑖 = 1 − ∑(𝐹𝑃𝑅𝑘 − 𝐹𝑃𝑅𝑘−1 )(𝑇𝑃𝑅𝑘 + 𝑇𝑃𝑅𝑘−1 ) (7.1)

𝑘=1

where 𝐹𝑃𝑅𝑘 and 𝑇𝑃𝑅𝑘 are indexed in increasing order such that 𝐹𝑃𝑅𝑘 > 𝐹𝑃𝑅𝑘−1 and 𝑇𝑃𝑅𝑘
> 𝑇𝑃𝑅𝑘−1, so that:

 𝐹𝑃𝑅𝑘 is the false positive rate for k = 0,…,n with 𝐹𝑃𝑅0 = 0 and 𝐹𝑃𝑅𝑛 = 1
 𝑇𝑃𝑅𝑘 is the true positive rate for k = 0,…,n with 𝑇𝑃𝑅0 = 0 and 𝑇𝑃𝑅𝑛 = 1

In other words, the ROC curve consists of a set of points, where every point represents the
false and true positive rates of the sample picked according to a cut-off presented by this
particular point. The cut-off means an estimate of the probability of default provided by the
model.

21
The value of the Gini coefficient may fluctuate between zero and one, where one means the
perfect classification model, and zero means "random guessing" model. According to the
bank’s internal governance, it is forbidden to use models with the Gini coefficient lower than
10% (0.1), which gives a minimum requirement for PD modeling.

The Gini coefficient for a set of candidate models is used and the model with the highest Gini
coefficient is selected.

7.2. LGD model selection

In contrast to PD, LGD is not Bernoulli distributed, thus another performance measurement
metric is used. Given a set of candidate models, the Bayesian information criterion (BIC)
supports the selection of the most appropriate model, which has a balance between the
goodness of the model, measured by the maximum value of the likelihood function, and
simplicity, measured by several explanatory variables in the model. (Peussa 2016, 20.)

It is defined as

𝐵𝐼𝐶 = −2 ln(𝐿̂) + 𝑘ln(𝑛) (7.2)

where 𝐿̂ is the maximum value of the likelihood function for the model, 𝑘 is the number of
parameters in the model and 𝑛 is the number of observations.

The most appropriate model from the selection criterion perspective minimizes the BIC
value.

22
 8. Estimation

This section contains models estimated for the probability of default (PD) and loss given
default (LGD). Based on these models, the agreements' specific risk weights and through this
the risk-weighted assets are calculated, which are required for capital requirement calculation
under the IRB approach.

8.1. Data

The research is based on the bank's internal data, stored in the bank's data warehouse (DWH).
Every type of data presented above is stored in the own SQL database, thus the aim is to
prepare a final data set that includes all the data required for the research question purposes.
Therefore, an SQL query is created, which combines and links all the data required in one
sample. Due to the bank's operational challenges related to the data, the research is based on
the constrained sample in terms of date range limitation and absence of part of LGD related
data like final loss data.

The scope of the research question is a non-revolving non-mortgage portfolio and private
customer segment. Hence, the data set includes household-level data from consumer secured
products like car financing. Mortgages and revolving financial products like credit cards are
not included in the data set as they are out of the scope of the research question. The data is
based on a period from the beginning of July 2017 till March 2018. The initial dataset
includes 2093 observations and 52 variables, from which

 48 variables are potentially explanatory for PD model

 1 variable is a response variable for the PD model, thus it reflects the default event
defined by Basel Accords
 1 variable, which reflects customer’s paid down payment
 1 variable, which reflects the asset’s collateral value
 1 variable, which reflects an agreement's exposure at default (nominal balance)

Potential explanatory variables are based on the data from the application, the credit bureau
and the bank's internal data reflecting customers' past behavior. However, due to compliance
and business reasons, all potential explanatory variables are not raw data but are already
transformed according to the bank's internal policy standards into a rule-based format. Hence,

23
all potential explanatory variables in the data set are binary. Due to corporate security
reasons, the exact nature of potential explanatory variables is not revealed. That is why labels
of these variables are recoded into “A_” form.

The data regarding the collateral value of the asset is provided by a third-party – asset
evaluation company.

8.2. PD model

Based on the bank's internal recommendations the data set used for PD model estimation is
split by train (2017 H2) and test (2018 Q1) samples according to a timestamp. The training
sample contains 1588 observations and the test sample contains 505 observations. The
training sample is used for model development and the test sample is used for model
validation. The timestamp-based split is explained by the need to both ensure that there is no
overfitting issue and that model is stable in time.

For the optimal model selection, a stepwise methodology is utilized based on backward
elimination. In other words, the method involves starting with a model performance check for
the model including all candidate variables and testing the model performance again after
deletion of the least important variable. The importance of the variable is measured by both
economic power (information value) and statistical significance (p-value).

The information value (IV) measures the economic power of the variable and is based on
Kullback-Leibler information divergence:

𝑃(𝑖)
𝐼𝑉 = 𝐷𝐾𝐿 (𝑃‖𝑄) + 𝐷𝐾𝐿 (𝑄‖𝑃) = ∑[𝑃(𝑖) − 𝑄(𝑖)] × log ( ) (8.1)
𝑄(𝑖)
𝐼

where 𝑃 and 𝑄 are the variable’s probability distributions for defaulted and non-defaulted
agreements. In other words, 𝑃 tells how defaulted agreements are distributed between the
variable’s categories and 𝑄 tells how non-defaulted agreements are distributed between the
same variable's categories. The higher the information value variable receives, the higher is
the difference between defaulted and non-defaulted agreements' distributions, the higher is
the economic power of the variable. Based on the bank's internal recommendations, the
variables with economic power below 0.02 should not be included in the model.

24
Figure 8.1 Economic power of candidate variables measured by information value

The majority of the variables are excluded from the candidate list due to their irrelevancy,
economic power weakness or compliance reasons.

For instance, the U.S. Equal Credit Opportunity Acts of 1975 and 1976 made it illegal to
discriminate in the granting of credit on the grounds of race, color, religion, national origin,
sex, marital status, or age. (Thomas 2002, 124.)

After first exclusions based on economic power weakness, compliance considerations like
discrimination risk and business reasons, the candidate list includes 12 potential explanatory
variables.

25
Figure 8.2 Correlation matrix of candidate variables

These candidates are checked by a correlation matrix in order to verify that there is no risk of
multicollinearity.

Based on Figure 8.1 the variable A_28 has exceptionally strong economic power. This is
explained by the nature of the business rule behind the variable, which reveals a high credit
risk level associated with the population for which this business rule is relevant.
Furthermore, based on the finding the bank’s credit policy is updated. According to the credit
policy update, the population for which business rule under variable A_28 is relevant cannot
be approved anymore. Hence, this population is out of the scope of the research question as
the capital requirement is required only for the approved population.

After the exclusion of the population related to variable A_28, the dataset includes 11
potential explanatory variables and 1745 observations from which 1331 observations belong
to the training sample and 414 observations belong to the test sample.

For PD model estimation, the backward elimination principle is used supported by economic
power (IV) considerations and correlation matrix. The idea is to ensure that explanatory
variables remaining in the model would provide the highest possible economic power
keeping the correlation between explanatory variables as low as possible. As the purpose is

26
to use a training sample for model estimation, the model estimation is performed on the
dataset including 1331 observations.

After several iterations, the final model is produced consisting of six explanatory variables
and an intercept. Every explanatory variable is significant at least at the 10% significance
level. The nature of variables in the final model cannot be revealed due to corporate security
reasons, but what is noteworthy is that variables, based on information stated by the
customer, are not part of the final model due to weak significance. This can be explained by
intentional or unintentional information violation by the customer. (Peussa 2016, 30.)

Table 8.1 Coefficient estimates for the probability of default model

Variable Estimate Standard error Significance level

Intercept -2.4135 0.1511 <0.01

A_46 -1.8227 0.5936 <0.01
A_1 -14848 0.7261 <0.05
A_13 1.0416 0.3351 <0.01
A_44 0.6002 0.2339 <0.05
A_45 -1.0390 0.4094 <0.05
A_3 -0.8550 0.4390 <0.1

The final model includes variables based on credit-bureau data and the bank's internal data.
All variables are checked from a business perspective in order to verify that no illogical
behavior exists. Each variable has acceptable economic power and a coefficient sign indicates
what the variable's effect on credit risk is. The positive sign means that the variable has a
positive correlation with credit risk, and a negative sign means that the variable has a
negative correlation with credit risk.

27
The Gini coefficient derived from the ROC curve measures the performance of the final
model based on the training sample.

Figure 8.3 ROC curve of the final PD-model (train sample)

The Gini coefficient of the final model based on the training sample is 38.44%, which
indicates that the performance of the model is on an acceptable level.

In order to verify that the overfitting issue does not appear and the model is stable in time, the
Gini coefficient is derived for the test sample as well:

Figure 8.4 ROC curve of the final PD-model (test sample)

28
The Gini coefficient of the final model based on the test sample is 38.12%, which indicates
that the performance of the model remains on an acceptable level and the model is stable in
time.

8.3. LGD model

Due to sufficient data unavailability for LGD estimation, the simplified LGD model is used
for risk weight estimation based on the linear relationship with securitization level (SL).
Securitization level is defined as an underlying asset's collateral value plus a customer’s paid
down payment divided by the exposure amount:

𝑐𝑜𝑙𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑣𝑎𝑙𝑢𝑒 + 𝑑𝑜𝑤𝑛 𝑝𝑎𝑦𝑚𝑒𝑛𝑡

𝑆𝐿 = 𝑚𝑖𝑛 ( , 1) (8.2)
𝐸𝐴𝐷

Fully secured exposure (securitization level equal to 100%) means that an underlying asset's
collateral value is equal to or higher than the customer's exposure amount. Fully unsecured
exposure has securitization level equal to 0%. Based on the bank’s historical data for
mortgages and unsecured credit cards aligned with Financial Supervisory Authority (FSA),
the downturn LGD estimate for fully secured exposure is 10% and for fully unsecured
exposure is 44%. This means that based on the bank's historical data gathered from a period
of time, which reflects the economic downturn, the bank experienced a loss of 44% from the
exposure at default for defaulted customers in case of credit card products and 10% in case of
mortgage products.

Based on these estimates, the linear equation below is proposed for downturn LGD
estimation:

𝐿𝐺𝐷 = 0.44 − 0.34 × 𝑆𝐿 (8.3)

where 𝐿𝐺𝐷 is downturn loss given default estimate and 𝑆𝐿 is agreement’s securitization
level.

The equation is derived by using DT LGD estimates mentioned above and assuming the
linear relationship between the level of securitization (SL) and loss given default (LGD).
Generally, the equation is calculated using two points: (0, 0.44) and (1, 0.10) where the first
point refers to fully unsecured exposure and the second point refers to fully secured exposure.

29
 8.4. Risk-weighted assets

The main component for capital requirement calculation under the IRB model is an
agreement specific risk weight. Having PD and LGD models in place, agreement specific risk
weight is calculated based on a formula (4.12):

1 𝑅
𝑅𝑊𝑖 = 12.5 × (𝐿𝐺𝐷𝑖 × 𝑁 ( × 𝐺(𝑃𝐷𝑖 ) + √1−𝑅𝑖 × 𝐺(0.999)) − 𝐿𝐺𝐷𝑖 × 𝑃𝐷𝑖 )
√1−𝑅𝑖 𝑖

where 𝑅𝑊𝑖 is agreement specific risk weight, 𝐿𝐺𝐷𝑖 is agreement specific loss given default
estimated by formula (8.2), 𝑃𝐷𝑖 is agreement specific probability of default estimated by PD
model provided in Section 8.2 for agreement 𝑖.

The agreement specific risk weights are calculated for the total dataset excluding variable
A_28 related population, meaning that risk weight formula is applied to 1745 observations.

Now, based on a formula (3.3), the sample’s risk-weighted assets may be calculated under
the IRB approach:

1745

∑ 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖 = 0.27 × 𝐸𝐴𝐷

𝑖=1

where 𝑅𝑊𝑖 refers to agreement specific risk weight based on formula (4.12), 𝐸𝐴𝐷𝑖 indicates
agreement’s specific exposure at default, 1745 shows the number of agreements in the dataset
and 𝐸𝐴𝐷 is a sum of all agreements’ exposures at default in the dataset.

30
 9. Results

The hypothesis of the research question is that ROAC under A-IRB approach 𝑅𝑂𝐴𝐶𝐼𝑅𝐵 ,
where agreement based risk weights (RW) are estimated internally by the formula (4.12), is
higher compared to ROAC under bank’s current (standardized approach) 𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑 ,
where the risk weight for retail exposures is equal to 75%.

So, the expected result of the research question can be expressed as:

1745 1745

∑ 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖 < ∑ 0.75 × 𝐸𝐴𝐷𝑖 (9.1)

𝑖=1 𝑖=1

The second part of inequality may be expressed as:

1745

∑ 0.75 × 𝐸𝐴𝐷𝑖 = 0.75 × 𝐸𝐴𝐷 (9.2)

𝑖=1

where 𝐸𝐴𝐷 is a sum of all exposures at default within a dataset.

Based on agreement specific risk weights, which are estimated by applying the final PD and
LGD model presented in Sections 8.1 and 8.2, the risk-weighted assets under the IRB
approach are equal to:

1745

∑ 𝑅𝑊𝑖 × 𝐸𝐴𝐷𝑖 = 0.27 × 𝐸𝐴𝐷 (9.3)

𝑖=1

According to ROAC definition based on formula (5.2) and utilizing estimation results from
(9.2) and (9.3) standardized and A-IRB based returns on allocated capital may be presented
as:

𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒
𝑅𝑂𝐴𝐶𝐼𝑅𝐵 =
0.08 × 0.75 × 𝐸𝐴𝐷

𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒
𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑 =
0.08 × 0.27 × 𝐸𝐴𝐷

31
In terms of ROAC comparison, the estimated result means that return on allocated capital
under A-IRB is 177.8% higher compared to that return on allocated capital under
standardized model:

𝑅𝑂𝐴𝐶𝐼𝑅𝐵 − 𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑
× 100% = 177.8%
𝑅𝑂𝐴𝐶𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑

Hence, based on estimation results, the research proves the hypothesis validity and supports
the portfolio transition from a standardized approach to A-IRB in order to increase the
portfolio's profitability in terms of return on allocated capital (ROAC).

Figure 9.1 Combinations of PD and LGD values satisfying 75% and 27% risk weights

Figure 9.1 tends to illustrate the difference in PD and LGD combinations for portfolio having
the same risk weight under the IRB approach as in a standardized approach (red line)
compared to the bank’s portfolio (green line). It is clear that the combinations of PD and
LGD for the bank’s portfolio are much lower than expected by Basel Accords under the
standardized approach. Hence, taking into account the bank's PD and LGD on a portfolio
level, it is logical to expect that it would require smaller capital requirements compared to
banks, which tends to operate under the standardized approach.

32
 10. Conclusion

The idea of the research is to check whether the higher return on allocated capital may be
achieved if the bank calculates a capital requirement based on the IRB approach instead of a
standardized approach. The research question hypothesized that the introduction of the IRB
approach may lower the capital allocated to the retail portfolio.

IRB approach uses financial institution's own credit risk related estimates rather than whole
market expectations presented by standardized approach under Basel III. If the bank's
customer base has lower credit risk compared to the market expectations, there is a perfect
sense to the bank to adopt the IRB approach.

The capital requirement under the IRB approach is calculated through the estimation of the
probability of default (PD) and loss given default (LGD) based on the bank's historical data.
The historical data for the probability of default calculation includes application, credit
bureau, and behavioral data. It is noteworthy to mention that behavior and credit bureau data
explain best the probability of default, while application data is proven to be irrelevant based
on research findings. The probability of default is estimated using logistic regression, which
is recommendation-based on the bank's compliance policy. The historical data for loss given
default estimation is unavailable due to technical gaps in the bank's data warehouse; hence,
the simple linear formula is applied based on historical downturn LGD high-level aggregates.

The research covers the bank's retail portfolio and provides evidence that if a bank adopts the
A-IRB approach, the expected profitability increase of the retail portfolio is 177.8%. The
estimate presented in this research is not final, as the research implies some assumptions to
LGD estimation, which decrease the accuracy of the prediction made. However, such a
remarkable increase in profitability proves the validity of the hypothesis even if some margin
of conservatism is applied.

The research proves that the application of the IRB approach significantly lowers a bank's
capital requirement and as a result increases a return on allocated capital.

33
 11. Bibliography

1. Agresti A. (2015). “Foundations of linear and generalized models.” Wiley

2. Baesens B., Rösch D., Scheule H. (2016). “Credit risk analytics.” Wiley
3. Balin B. J. (2008). “Basel I, Basel II, and emerging markets: A nontechnical analysis.”
4. Bank for International Settlements (2005). “An Explanatory note on the Basel II IRB
Risk Weight Functions.” Bank for International Settlements
5. Desai V.S., Crook J.N., Overstreet G.A. (1996). “A comparison of neural networks and
linear scoring models in the credit environment.” European journal of operational
research, 95, 24-37
6. European Banking Authority (2017). "EBA Report on IRB modeling practices".
European Banking Authority
7. Joseph C. (2013). “Advanced credit risk analysis and management.” Wiley
8. Lewis E.M. (1992). "An Introduction to Credit Scoring." Athena Press
9. Peussa A. (2016). “Credit risk scorecard estimation by logistic regression.” Pro gradu.
Department of mathematics and statistics. University of Helsinki
10. De Servigny A., Renault O. (2004). “Measuring and Managing credit risk.” McGraw-
Hill
11. Thomas L.C. (2002). "Credit scoring and Its Applications." SIAM Monographs on
mathematical modeling and computation
12. Witzany J. (2017). “Credit Risk Management.” Springer
13. Yang B.H., Tkachenko M. (2012). “Modelling of EAD and LGD: Empirical
Approaches and Technical Implementation.” Journal of Credit Risk, 8

34