Credit Risk
Management
Public Lectures
22-23.04.2024 Masaryk University, Brno
About
Lukas
Prorokowski
• Academic Engagement:
• University of
Luxembourg
• Wroclaw University of
Economics
• Research:
• Early Warning Systems in
Credit Risk
• Real Estate Collateral
Valuation Models
• Model Risk
Quantification
• Professional:
• Nordea Bank
• Banque Internationale a
Luxembourg
• Lloyds Bank
• CITI
• BNY Mellon
• HSBC
• RBS
Credit Risk can be found in:
•Absent or downgraded ratings or risk profiles of
customers, which could also bias the financial ratios
•Some types of customers acceptable or not
acceptable to the bank either due to customer profile
or specific type of activity
•Any limitations on markets or specific geographies
eligible for finance
•Risks concerning short vs. long term lending and any
specific maturity limitations
•Preferred type of finance products, e.g. high-risk loans
•Any specific collateral requirements
•Interest rate or currency mismatch and any related
hedging requirements
•Industry specific sustainability requirements
•Industry specific climate related and environmental
risks
Credit Risk Management
What is Credit Risk management?
• It is a continuous process of:
• Identifying risk
• Analysing risk
• Modelling risk
• Mitigating risk
• Monitoring risk
• Predicting risk
Identify
Analyse
Quantify
Mitigate
Monitor
Credit risk appetite, strategy and credit
risk limits
The credit risk appetite should be
implemented with the support of
appropriate credit risk metrics and
limits. These metrics and limits should
cover key aspects of the credit risk
appetite, as well as client segments,
currency, collateral types and credit risk
mitigation instruments.
Credit Risk Management: KRI
Design
Statement Metric Limit/Thershold
Interconnectedness: appropriate mitigation techniques
are in place to address potential risks stemming from
the uncertainty about interconnectedness of the
shadow banking entities. (14.a)
Each Legal Entity (or relevant Line of
Business) has a clear policy/procedure on
mitigating the risk of uncertainty about
interconnectedness of SBEs
Techniques 100%
documented
Institutions should set an aggregate limit to their
exposures to shadow banking entities relative to their
eligible capital (17)
Aggregate Limit as % of Eligible Capital
x% of the Eligible
Capital
When setting individual limits for shadow banking
entities, as part of their internal assessment process,
the institutions should take into account information
available about the portfolio of the shadow banking
entity, in particular non-performing loans (19c)
% of non-performing loans in the shadow
banking portfolio
0%
https://www.eba.europa.eu/sites/default/documents/files/documents/10180/1310259/f7e7ce6b-7075-
44b5-9547-5534c8c39a37/EBA-GL-2015-
20%20Final%20report%20on%20GL%20on%20Shadow%20Banking%20Entities.pdf
Identify
Analyse
Quantify
Mitigate
Monitor
•Credit Risk Modelling Team
•Credit Risk Data Management Team
•Model Implementation Team
•Model Risk Manager (Credit Risk Unit)
First Line of
Defence
• Model Validation (Independent
Model Validation Unit)
Second Line
of Defence
• Internal Audit
Third Line of
Defence
Credit Risk Management is conducted through the 3 Lines of Defence model. All lines are coordinating tasks to set up
the governance framework. The objective followed in credit risk policies and procedures should be to promote a
proactive approach to monitoring credit quality, identifying deteriorating credit early and managing the overall credit
quality and associated risk profile of the portfolio, including through new credit-granting.
Models in Regulatory Framework for Credit Risk
• Basel II:
• Credit Risk measurement approach:
Standardised
• Use of regulatory
prescribed weights for
exposure classes;
• Over-reliance on
agency ratings;
• Some flexibility on the
choice of risk weight
assignment.
FIRB (Foundation)
• Use of own models for
Probability of Default
(PD) – subject to the
regulatory approval;
• Use of regulatory
prescribed LGD models
(e.g. regulatory floors).
AIRB (advanced)
• Use of own models for:
• PD
• LGD
• EAD
• Credit risk models are
subject to regulatory
review.
MODEL COMPLEXITY
Input
Assumptions
Data
Scenarios
Expert Judgment
Process
Methodology
Implementation
Calculation Engine
Aggregation
Output
Quantitative Estimates
Forecasts
Input Processes Output
Identify
Analyse
Quantify
Mitigate
Monitor
EBA Guidelines on credit risk mitigation for institutions applying
the IRB approach with own estimates of LGDs
The collateral is only utilised by a bank in a case of bankruptcy of the obligor. Figure below shows that a bank assesses the
value of the collateralised asset and applies a haircut that reduces the asset value. This would mean that a collateralised
asset valued at EUR 100 would be worth less upon the application of a haircut (e.g. EUR 80 if the haircut is 20%). The
challenge faced by banks is to find an appropriate value for a given type of collateral that should be sensitive to the
changing macro conditions and balanced between being conservative enough to deliver a minimum level of credit risk
protection and being least detrimental to the borrowers.
Loan →→→ Borrower
Proceeds ($$$) →→→ Credit Institution ←←← Collateral Asset
Haircut Collateral Value Default
Liquidation of Collateral
Seminar
Case
Studies
Data Quality
Dimension Explanation
Completeness Values must be present in the attributes that require them. Is defined as the availability
of the required information. Completeness checks are carried out to detect missing
information.
Timeliness Data values are up to date.
Validity Data are founded on an adequate and rigorous classification system.
Availability / Accessibility Data are made available to the relevant parties.
Consistency A given set of data can be matched across different data sources of the institution.
Accuracy Data is substantively error-free. Is interpreted as the absence of mistakes and exact
correspondence of the reported values with the underlying concept for each data point.
Accuracy is ensured by a set of validation rules that have to be respected by the
reported data.
Uniqueness Aggregate data are free from any duplication from filters or other transformations of the
source data.
Traceability The history, processing and location of the data under consideration can be easily
traced.
Representativeness Historical data are representative of current portfolio.
Data Quality – Spotlight on Completeness
• Completeness is defined as the data property which covers the way how the
data is identified, defined and present against the following aspects:
SCOPE
• Scope of the historical data equal
to the model scope and
corresponding data requirements
• All primary keys are present in the
dataset
• All relevant ID fields are in the
dataset
• All relevant date and timestamp
columns are in the dataset
• All potentially relevant risk drivers
are present in the dataset
• Variables functional forms are in
line with requirements
DATA SOURCES
• Complete set of systems and data
sources used
• Presence of legacy systems,
historical systems migration or
changed processes
• Data reconciliation between
different sources providing the
same data based on the functional
specification
• Performed at primary keys level
with the focus on completeness of
the base dataset (where all
relevant facilities and snapshots
should be present)
HISTORICAL WINDOW
• The length of the underlying
historical observation period used
shall be at least five years for at
least one source. If the available
observation spans a longer period
for any source, and these data are
relevant, this longer period shall be
used
Data Quality – Spotlight on Completeness
• Completeness is defined as the data property which covers the way how the
data is identified, defined and present against the following aspects:
MISSING VALUES
• No missing values are allowed for
primary inputs and mandatory
fields
• Variables with share of
uninformative missing values above
5% should not be used in the
model development unless well
justified
• Added value of external variables
has to be assessed on a case-bycase
basis, as share of missing
values can be significant
MISSING SNAPSHOTS
• Number of snapshots between the
first snapshot date and the last
snapshot date for a facility must be
equal to number of unique records
for that facility
• Difference between the initial and
ultimate snapshot must be checked
against documented data request
OTHER
• Plot number of obligors / facilities,
total portfolio outstanding, default
rate, loss
Data Quality – Spotlight on Timeliness
• Timeliness refers to whether the information around the different data aspects
(e.g. clients, credit agreements, facilities, etc.) is up to date and available at the
time of the model’s predictions:
TIMELINESS
• Most recent historical data (given
the predefined performance
period) is available
• Credit bureau data is no older than
12 months
• Financial information is no older
than 12 months
• Internal behavioral information is
no older than 1 month
CreditRiskData
Credit Bureau
Financial
Statement
Ongoing Portfolio
Monitoring
Data Quality – Spotlight on Validity
• Validity refers to which extent the data is founded on an adequate and rigorous
classification system:
Unreasonable Values
Erroneous Entries
Wrong Data Feed
Validity
• Field observed data type match with the expected
one as defined in the source system, (e.g. variables
such as age of obligor or year are expected to be
an integer)
• Values observed in the variables are in line with
internal policies
• If the observed value is not in line with basic logic,
it has to be checked with Business
• If the observed value falls outside predefined
range / domain, it has to be checked with Business
• Some fields (especially primary keys such as obligor
or facility identifiers) often have a fixed length.
Therefore, it is useful to check whether potential
duplicates are not a consequence of identifiers
truncation.
CASE
STUDY
Data Quality - Spotlight on Consistency
• Consistency: Evaluate whether:
• The same data has been used in different model use domains such as capital calculations,
business and regulatory reporting or provisions.
• The data collected from different source systems is consistent (the same primary key can
be observed in different data sources used)
• The related input/output fields retrieved from different sources display consistent values
(e.g. a default flag =1)
• The data is consistent (e.g. in terms of number of defaults, clients, exposures)
• CASE STUDY: Goodness of Fit (Excel)
The subscript “c” is the degrees of freedom (n-1). “O” is your observed value and E is your expected value. It’s
very rare that you’ll want to actually use this formula to find a critical chi-square value by hand. The summation
symbol means that you’ll have to perform a calculation for every single data item in your data set.
CASE
STUDY
Credit concentration risk is the
risk of losses arising as a result
of concentrations of exposures
due to imperfect diversification.
This imperfect diversification
can arise from the small size of
a portfolio or a large number of
exposures to specific obligors
(single name concentration) or
from imperfect diversification
with respect to economic
sectors or geographical regions.
Data Quality – Spotlight on Accuracy
• Accuracy: the inputs (risk drivers) and outputs (risk parameters):
• Do not contain errors as descriptive statistics and/or distributions are consistent
throughout the modelling period selected and display values expected for the variable;
• Do not contain large number of unexpected outliers and the existent ones are justified by
the business and appropriately documented;
• Do not contain concentration bias (risk concentration):
𝐻𝐻𝐼 =
∑ 𝑤!
"
(∑ 𝑤!)"
Where wi denotes the EAD-measured credit concentration risk per defined category i. The following
measures of the concentration risk are taken
11.1% < HHI ≤ 24.9%
24.9% < HHI ≤ 34.5%
34.5% < HHI ≤ 47.8%
47.8% < HHI ≤ 77.9%
77.9% < HHI ≤ 100%
CASE
STUDY
Backtesting
Model Performance
Tests
Focus on Financial Collateral
Backtesting of
Credit Risk ModelsBacktesting Process Validation Focus
Calibration Assessment of the deviation of the internal model’s
estimates from the realised observations.
Discrimination Assessment of the extent to which bad realised
observations are assigned low internal model’s
estimates and the good realised observations are
assigned high estimates.
Stability Assessment of the changes to the population over
time that affect the appropriateness of the internal
model.
KAPPA COEFFICIENT
You want to apply the Kappa statistic to determine the
agreement between the two time series of returns. One series
represents the underlying equity in your portfolio of credit
collateral, and the other time series is for the risk proxy that is
used as an aggregate risk indicator for equity-type collateral (a
relative index).
At this point, returns are calculated on the closing prices for both
the equity (collateral) and the index (risk proxy used in the
model).
You assume that for the Index to be an adequate (sensitive) risk
proxy, it must move in unison with the underlying collateral. For
example, if an equity price decreases, then the index price should
decrease as well on the same day. This would prove that the
proxy remains sensitive and accurate, and no unexpected
behaviour is recorded.
CASE
STUDY
Data Quality – Spotlight on Accuracy
• Kappa Coefficient
The Kappa coefficient is defined by the difference between the observed agreement between the two series of calculated returns and the
probability of chance of agreement. Embarking on the formula proposed by Galton (1892), the Kappa coefficient is computed as follows:
Where “𝑃!” is the relative observed agreement among returns between the equity (collateral) and its proxy (index) and “𝑃"#$%"&” is the
hypothetical probability of chance agreement obtained from the observed data to calculate the probabilities of each return being assigned to
the same category (increase/decrease – expressed as a daily change in closing prices). At this point, the Kappa coefficient measures the
agreement in the returns. However, it does not measure the degree of the variability of the returns.
A Traffic Lights Approach indicating the level of agreement between the two time series is implemented to interpret the results:
CASE
STUDY