The Wayback Machine - https://web.archive.org/web/20230320104713/https://medium.

com/@willempretoriu…

Search Medium

Willem Pretorius Follow

Mar 6 · 20 min read

An Introduction to Credit Risk in Banking:

BASEL, IFRS9, Pricing, Statistics, Machine
Learning — PART 1

Hello, and welcome to my blog series! I have always wanted to share my thoughts and
insights on credit risk in the banking industry. As a Junior Writer and Data
Scientist/Quant, I may not have the same level of experience as some of my peers, but
I’m eager to learn and share my perspectives with others. I have a deep passion for
statistics (machine learning), quantitative analytics, regulatory aspects of banking and
application to credit modelling.
In this series, I’ll be exploring various topics related to credit risk & statistics, including
different models and techniques used by banks, current trends and challenges in the
industry, and the impact of regulations and standards such as Basel II, III, IV and IFRS9
(definitions to follow). It has to be mentioned that I will focus a lot on the banking
perspectives in Europe (European Banking Authority — EBA / European Central Bank
— ECB) and Africa (South African Reserve Bank — SARB) as this is where I have
gained my experience. Nonetheless, the techniques and ideas that will be discussed are
insights and challenges throughout the world. Join me as I dive into the complex and
fascinating world of credit risk management in banking.

Recently, Machine Learning (ML) — which is basically statistics on steroids— have

become a debated and hot topic in the credit risk space. The ECB has been actively
exploring the use of ML in the context of banking and finance. In November 2021, the
ECB published a report titled “EBA DISCUSSION PAPER ON MACHINE
LEARNING FOR IRB MODELS” which provided an overview and questionnaire of the
different applications of ML in Internal Rating-Based Models (IRB) — more on this
later. The Bank for International Settlements (BIS) have also been exploring various
topics in ML together with the ECB and the SARB.

These reports highlighted the potential benefits of using ML techniques, such as

improved accuracy in forecasting, early warning signals for financial stability risks, and
more efficient fraud detection. However, it also noted some of the challenges associated
with using ML in central banking, such as data quality issues, model interpretability, and
potential biases in the data. The use of these statistical outputs are now more important
than ever (as is the data feeding into these models).

Overall, the ECB recognizes the potential benefits of using ML in banking and finance,
but also emphasizes the need for careful consideration of the potential risks and
challenges associated with these techniques. I’ll be exploring general statistical & ML
ideas for credit modelling and validation with code & explanation — the advantages and
the pitfalls I see in the industry.

W hat I am aiming to achieve in these posts are to share knowledge with new
data scientists/credit quants entering the workspace, and perhaps to people in
different industries who are interested in these applications. Credit Risk is a fascinating
topic with a mixture of quantitative and qualitative ideas. It is a space where the
convergence of statistics, expert opinions, and business objectives occurs — and it is
rather not focused on in universities (unlike market risk with repetitive analysis of
options and deriving the Black-Scholes formula).

On the quantitative side, credit risk management (on the banking book) involves
statistical modelling, risk assessment, and financial analysis. Financial institutions use
various tools to measure and manage credit risk, such as credit ratings, credit scoring,
and Probability of Default (PD) models. These models provide a quantitative basis for
assessing the likelihood of default and the potential losses associated with credit risk.
These models are specifically used in Acceptance, Pricing, Expected Loss (EL), Stress
Testing and Regulatory Capital (RC) calculations. We will get to each of these in the
coming posts (as well as discussing the model lifecycle of each — development,
validation and maintenance).

On the qualitative side, credit risk management involves factors such as industry trends,
management quality, and macroeconomic conditions. Credit Quants consider these
qualitative factors when assessing the creditworthiness of a borrower or a portfolio of
loans. They also monitor market trends, regulatory changes, and emerging risks to
anticipate potential credit risks.

The combination of quantitative and qualitative analysis makes credit risk management
a challenging and dynamic field. Banks and financial institutions must continuously
update their credit risk models and adapt to changing market conditions to manage their
credit risk effectively.

Overall, credit risk is a fascinating topic with both quantitative and qualitative aspects
that requires continuous analysis and adaptation to changing market conditions. In-
depth knowledge of statistics and risk management are needed as a starter (typically at a
Master’s level). Additionally, it has become much more important to be able to
effectively code these large calculations that were mentioned earlier. Being familiar with
two or more coding languages (such as Python, R, SAS, SQL, or C) can be a great asset
for a young quant from university entering the job market.

In my discussion, I will aim to indicate the route I took and what university and online
courses I would recommend for young quants to get started in these coding languages
and to obtain the necessary statistical basics needed for credit modelling with traditional
and ML models.
Table of Contents
The Basics To Credit Risk in Banking: BASEL

The Credit Loss Distribution & Credit Definitions

The Asymptotic Single Risk Factor (ASFR) Framework

Advanced Internal Rating-Based Framework

Rating Philosophies — Part 2

The Basics To Credit Risk in Banking: IFRS9 — Part 2

Use Cases — Part 2

Important Statistical Concepts & Models For The Credit Model Lifecycle — Part 3

The Basics To Credit Risk in Banking: BASEL

The BASEL Accords are a set of international regulations for the banking industry.
These accords were developed by the BASEL Committee on Banking Supervision
(BCBS), which is made up of central banks and regulatory authorities from around the
world. The importance of the BASEL Accords lies in their ability to promote financial
stability by ensuring that banks have sufficient capital to withstand economic downturns
and financial crises. Major banks around the globe follow these accords, which serve as
the fundamental basis for most credit models in banks.

There are three main BASEL accords:

1. BASEL I:

This accord was issued in 1988. The accord focused on three pillars. The BCBS
established a set of regulations to ensure that banks maintain sufficient capital to fulfill
their financial commitments and withstand financial and economic pressures. Minimum
Capital Requirements: The first pillar requires banks to maintain a minimum amount of
capital, based on a percentage of their Risk-Weighted Assets (RWAs), to absorb
Unexpected Losses (UL). The higher the risk of a bank’s assets, the more capital it must
hold. This was the main objective of BASEL I. Banks were required to maintain a
minimum Capital Adequacy Ratio (CAR) of 8%, calculated as a percentage of their
RWAs. The CAR is the ratio of a bank’s capital to its RWAs, where RWAs are calculated
by assigning different risk weights to different categories of assets based on their
perceived riskiness. The risk weights under BASEL 1 were divided into five categories,
ranging from zero for government securities to 100% for certain types of loans and other
risky assets. The minimum capital requirement of 8% meant that for every €100 of
RWAs, banks had to hold at least €8 of capital.

Regulatory Capital ≥ 8%× Risk-Weights × Asset

Values
The risk weight categories included:

0% risk weight for fully secured home loans, government bonds and cash.

20% risk weight for claims on OECD banks and certain multilateral development
banks (loans to other banks & public sector companies in these OECD —
Organization for Economic Co-operation and Development countries).

50% risk weight for claims on non-OECD banks, securities firms and certain other
financial institutions (home loans that are not fully secured — not covered 100% by
collateral).

100% risk weight for most corporate loans, mortgages, and claims on non-financial
institutions.

200% risk weight for certain exposures like derivatives, bridge loans, and non-traded
assets.

BASEL 1 also introduced the concept of Tier 1 and Tier 2 capital, with Tier 1 capital
consisting of the highest quality capital it can hold namely equity capital (ordinary
shares) and disclosed reserves, and Tier 2 capital consisting of subordinated debt (a type
of debt that ranks below other debts in the event of bankruptcy or liquidation), loan-loss
reserves, and certain types of hybrid capital. Banks were now monitored on their CAR
(which represents a bank’s ability to meet all of its financial obligations, not just absorb
losses) as well as their Common Equity Tier 1 (CET1) ratios (a measure of a bank’s core
capital compared to its RWAs — which represents a bank’s ability to absorb losses and
continue operating without needing external funding).

CAR = (Tier 1 + Tier 2) / Risk-Weighted Assets

 CET1 = Tier 1 / Risk-Weighted Assets

Supervisory Review: The second pillar requires banks to have a process for assessing their
overall risk profile and determining whether their capital levels are adequate. Regulators
(such as the ECB or SARB) are responsible for supervising this process and ensuring
that banks have sufficient capital to cover their risks. All quantitative models built under
Pillar 1 needs to be reviewed by overseeing supervisors. Market Discipline: The third pillar
aims to promote market discipline by requiring banks to disclose information about
their risk profile, capital adequacy, and risk management practices to the public. The idea
is that market participants will use this information to make more informed investment
decisions and put pressure on banks to manage their risks effectively.

Overall, the BASEL 1 framework laid the foundation for modern RC (not to be mistaken
for economic capital which is the best estimate of required capital that financial institutions
use internally to manage their own risk and to allocate the cost of maintaining RC among
different units within the organization) requirements and helped improve the stability of
the global banking system. However, it was criticized for its simplicity and for not taking
into account the varying degrees of risk within different asset classes (and collateral).
Operational Risk & Market Risk was also not taken into account. As a result, the Basel
Committee continued to refine the framework, leading to the introduction of Basel 2 and
Basel 3.

2. BASEL II:

This accord was issued in 2004 and introduced a more risk-sensitive approach to capital
requirements. BASEL II used a set of formulas to determine the amount of capital banks
needed to hold based on the risks they were taking on (formulas below). The same rule
for 8% of minimum capital requirement hold, however a minimum of 4% of CET1 capital
was now imposed. It also introduced a new framework for assessing credit risk,
operational risk, and market risk. Under BASEL II, banks are required to use internal
models to calculate their capital requirements for credit risk, subject to supervisory
approval. Smaller banks were provided with simpler formulas to compute the minimum
amount of capital they must hold to safeguard against risk, known as the “Standardized
Approach (SA).” Meanwhile, larger banks were offered the option of using the “Advanced
Internal Ratings-Based (AIRB)” approach (with the Vasicek distribution as the
fundamental description of credit losses and the measurement of credit risk) which
enabled them to devise their own models to determine their risk capital. This is where
PD, Loss Given Default (LGD) and Exposure at Default (EAD) calculations/models
became mandatory under this approach. Banks used these models to calculate how much
the it stands to lose when borrowers default (Loss = PD x LGD x EAD). An additional
approach was also introduced as the the “Foundation Internal Ratings-Based (FIRB)”
approach, where the bank uses its own internal ratings to determine risk weights for
different types of assets. However, some of the model parameters are still determined by
the regulators, and it is considered less sophisticated than the AIRB Approach. Once the
PD, LGD, and EAD are known, risk-weight functions provided in the BASEL accord can
be used to calculate the RC. We will discuss these approaches in-depth later.

Source

FIRB vs AIRB. Source: Baesens, Bart, Roesch, D, & Scheule, H. (2016). Credit Risk Analytics — Measurement
Techniques, Applications and Examples in SAS. Wiley.

Overall, BASEL II was designed to be more flexible and risk-sensitive than its
predecessor, BASEL I. However, it has also been criticized for being too complex and
allowing banks to use internal models to manipulate their capital ratios. As a result, the
BCBS has continued to refine and update the BASEL framework over time.
3. BASEL III:

This accord was issued in 2010 in response to the global financial crisis of 2008. BASEL
III focused on strengthening bank capital requirements and introducing new liquidity
and leverage ratios to reduce the risk of bank failures. It also placed a greater emphasis on
stress testing.

Some of the key differences between the BASEL accords include the risk weights
assigned to different types of assets, the methods used to calculate capital requirements
(Tier 1 capital increased from 4% to 6%), and the emphasis on liquidity and leverage
ratios. Additionally, each subsequent accord has built on the previous one to provide a
more comprehensive framework for regulating the banking industry.

BASEL II vs BASEL III Requirements. Source: Baesens, Bart, Roesch, D, & Scheule, H. (2016). Credit Risk
Analytics — Measurement Techniques, Applications and Examples in SAS. Wiley.

4. BASEL IV:

This accord further strengthen the regulation, supervision, and risk management
practices of the banking industry. The reforms aim to address issues and weaknesses
identified in the previous BASEL frameworks, particularly with regards to the
calculation of RWAs and the use of internal models.

Some of the key features of BASEL IV include:

1. Output Floor: The introduction of an output floor for RWA, which means that
banks using internal models will be required to hold a minimum level of capital
regardless of the risk weightings produced by their models. Banks capital
requirements will be floored to a certain percentage of the standardized
requirement, i.e. from 50% to 72.5%.

2. Credit Risk: Changes to the calculation of credit risk RWA, including the removal of
certain IRB approaches and the introduction of more granular risk weightings for
exposures to small and medium-sized enterprises (SMEs).
The implementation of BASEL IV is ongoing, with various jurisdictions expected to
adopt the reforms at different times. The reforms are expected to significantly impact
the banking industry and require banks to hold higher levels of capital. This will
continue to improve comparability among capital levels of banks.

The Credit Loss Distribution & Credit Definitions

Source

The Credit Loss Distribution is a statistical representation of the potential losses that a
financial institution may incur due to credit risk. It is used to estimate the EL and UL of a
credit portfolio.

The credit loss distribution is based on the probability distribution of credit losses,
which can be modelled using various techniques, such as the Normal distribution,
Poisson distribution or binomial distribution, depending on the nature of the credit risk.

The EL is the average amount of loss (that are incurred when obligors fail to pay or
default) that is expected to be incurred over a certain time period based on historical data
and probability distributions. It is calculated as the product of the PD, LGD and EAD.
This is the average/mean loss that should be provisioned for. Banks and corporations set
aside provisions for such losses. However, these values interrelate and merge in intricate
manners, leading to unforeseen ULs. The CAR does not account for ELs on bank runs.
Bank runs occur when a large number of depositors withdraw their money from a bank
due to concerns about the bank’s solvency or liquidity. Bank runs can lead to liquidity
shortages and can potentially cause a bank to fail. This suggests that a bank may need to
hold additional capital to mitigate the risk of losses from bank runs.

The UL, on the other hand, is the potential loss that can occur beyond the expected loss.
It is based on extreme scenarios and is calculated as the difference between the EL and
the worst-case loss that can occur in the tail of the distribution. Here the difference is
between the Credit Value at Risk (CVaR) and the ELs = dispersion of ELs (it is the
capital of a bank should cover losses that exceed provisions — and can be regarded as
part of economic capital). EL and UL changes continuously due to macroeconomic
factors and portfolio sizes. This is where the BASEL equations play a vital role in capital
management for banks. According to Basel II, the credit risk capital charge should aim to
cover ULs and account for rare events, specifically at the 99.9th percentile. The RC is
used to cover the UL. The sum of the provisions and the RC should equal the 99.9% loss.
If banks do not insure themselves, they may be vulnerable to losses beyond this
percentile.

By analyzing the credit loss distribution, banks can determine the capital required to
cover potential losses from credit risk, as well as the overall risk of the credit portfolio. It
is an important tool for risk management and regulatory compliance, particularly under
Basel II, III and IFRS 9.

Exposure at Default (EAD)

This refers to the calculation of a bank’s potential exposure to a counterparty in the event
of the counterparty’s default. It is measured in currency and is estimated for a period of
one year or until maturity, whichever comes first. The EAD for loan commitments (the
expected outstanding balance if the facility defaults, which is equal to the expected
utilization plus a percentage of the unused commitments including unpaid fees and
accrued interest) is based on BASEL Guidelines and measures the portion of the facility
that is likely to be drawn in the event of a default. BASEL II requires banks to provide an
estimate of the exposure amount for each transaction in their internal systems. The
purpose of these estimates is to fully capture the risks associated with an underlying
exposure.

Example: Let’s say a bank extends a €10000 credit line to a customer. The customer uses
€5000 of the credit line and pays off €2000, leaving a current outstanding balance of
€3000. The bank estimates that if the customer were to default at this point, they would
likely draw down the entire remaining balance of €3000. Therefore, the EAD in this
example would be €3000. This represents the amount that the bank would be exposed to
if the customer were to default.

Typically, EAD is a linear formula and is calculated per loan product. There are different
ways to estimate EAD (for example through conversion factors) and we will discuss these
later.

Loss Given Default (LGD)

The EBA defines LGD as the proportion of an exposure that is not expected to be
recovered in the event of default. It is calculated as the difference between the EAD and
the amount recovered through collateral or other means, expressed as a percentage of the
EAD (loss is defined as the difference between the observed exposure at default and the
sum of all the discounted cashflows where loss rate = observed loss as a percentage of
observed EAD).

Example: Suppose a borrower takes out a €10000 loan from a bank. If the borrower
defaults on the loan, the bank will try to recover the amount owed by selling any
collateral (e.g., property) put up against the loan. Suppose the collateral is sold for
€8000. Then the LGD would be:

LGD = (Total amount of loan — Amount recovered) / Total amount of loan LGD =
(€10000 — €8000) / €10000 LGD = 0.2 or 20%

This means that in the event of default, the bank would lose 20% of the total amount of
the loan, or $2000 in this case.

Usually after default occurs a client can either cure (defaulter paying off all outstanding
debt), Restructuring (loan characteristics and terms are changed and modified),
Liquidation/Foreclosure (banks repossess the collateral).

Typically LGD is calculated per loan product and involves statistical estimation using
either regression (Linear, Logistic, Survival Models, etc) or historical averages. As per
BASEL II guidelines, it is recommended that banks and other financial institutions
calculate the Downturn LGD (loss given default during a downturn in the business cycle)
for regulatory purposes. This helps to reflect the potential losses that may occur during
an economic downturn. We will look into different definitions of LGD, Loss Rates
(implied historical loss rates and workout loss rates) and statistical estimation later.

PD & The Definition of Default

Lending money to both retail and non-retail (e.g. corporate) customers is a primary
function of banks. To ensure responsible lending practices, banks must have effective
systems in place to determine who is eligible for credit. Credit scoring is a critical risk
assessment technique used to analyze and quantify the credit risk of a potential
borrower. The objective of credit scoring is to measure the probability that a borrower
will repay their debt (binary outcome of default or non-default — additionally cure
models can also be built separately which is the probability that the client will cure or
payoff the loan which is also binary). The result of this process is a score that reflects the
creditworthiness of the borrower (Logit models are very popular in scorecard
development). To ensure there are consistencies across banks on what is regarded as a
default, regulation has set tight measures on this.

The EBA defines default as the occurrence of one or more of the following events:

1. When a bank considers that the obligor is unlikely to pay its credit obligations in full,
without recourse by the bank to actions such as realizing security (i.e., the obligor is
“unlikely to pay” criterion).

2. When the obligor is past due more than 90 days on any material credit obligation to
the bank (i.e., the “90 days past due” criterion).

3. When the bank has reason to believe that the obligor has entered into bankruptcy or
other financial reorganization (i.e., the “bankruptcy” criterion).

The EBA requires banks to use the above criteria to identify and report defaulted
exposures in their portfolios, and to ensure that they have adequate policies, procedures,
and systems in place to accurately identify and report such exposures. We obtain
(through External ratings — provided by rating agencies, but some regulators prohibit
their usage, for example, S&P, Fitch or Moody’s ratings for corporations or
governments), estimate or interpolate (through the use of Scorecards, Logit or Probit
models) the PD of a particular rating/grade/pool as the long-run average of the one-year
default rates.

There are many interesting discussions around PDs such as Stressed and Unstressed
PDs; Through-the-cycle (TTC) and Point-in-Time (PIT); Estimation (either through
traditional statistics or ML). We will look at these topics in-depth later when we discuss
the model lifecycle.
The Asymptotic Single Risk Factor (ASFR) Framework
The Asymptotic Single Risk Factor (ASRF) framework is a credit risk model used to
calculate RC requirements for credit risk under the BASEL II framework. It assumes
that the portfolio’s risk can be represented by a single systematic factor, known as the
“factor model” and uses this factor to estimate the portfolio’s risk. When describing
default events, it is commonly assumed that a single factor, which is the state of the
world’s economy, is tied to all loan values and obligor’s default probabilities in a simple
manner (by the correlation between). This approach allows for the calculation of the RC
charge for credit risk to be made using a single-factor model, which is much simpler than
trying to model each individual exposure in a portfolio. The ASRF approach is
particularly useful for banks with large and diverse portfolios, as it allows them to
estimate RC requirements with reasonable accuracy while minimizing the computational
burden.

The ASFR formulas involve calculating the EL and UL of a portfolio of exposures, and
then applying a capital charge based on the UL. The formulas take into account the
correlation between obligor defaults and the systematic factor (usually GDP), as well as
the variability of the factor over time. The ASFR approach is considered a simplified
version of the more AIRB approach.

This approach is based on the asset value factor model and has its origin in Merton’s
structural approach (1974). Models based on asset value propose that the probability of a
firm’s default or survival is contingent on the value of its assets at a specific risk
measurement horizon, typically at the end of that horizon. If the value of its assets falls
below a critical threshold, its default point, the firm defaults, otherwise it survives. The
origins of asset value models can be traced back to Merton’s influential paper published
in 1974. Vasicek (2002) adapted Merton’s single asset model which can be used to model
a credit portfolio, by creating a function that transforms unconditional default
probabilities into default probabilities tied to a single systematic risk factor. The AIRB
approach employs an analytical approximation of CVaR using the Vasicek distribution.

The derivation for the logic follows as:

 Source: BIS & Author Adjusted

Banks assess each outstanding loan on an individual basis to determine the associated
risk, including the likelihood of default by the obligor. For retail loans such as mortgages,
factors like income, age, residence, loan nature, and macroeconomic indicators like house
prices and unemployment rates they all play a role in determining the risk of a loan.
However, since loan defaults are correlated with one another, the BASEL capital
requirements mandate that capital must be calculated for a bucket of loans with similar
characteristics. To determine the input parameters for capital requirements (PD, LGD,
and EAD), they must represent the entire bucket. The PD for the entire bucket is the
average of all individual PDs. The methodology described above in the image can be
extended to determine the default fraction distribution of a credit portfolio, which is a
bucket of loans. However, this requires an important assumption that the credit
portfolio is infinitely granular, meaning that it contains an infinite number of loans and
no individual loan represents a significant portion of the total portfolio exposure. In such
a portfolio, there is no idiosyncratic risk since all idiosyncratic risk is diversified. Now
consider the following.
 Source: BIS & Author Adjusted

The final equation can be interpreted as the default fraction in the infinitely granular
portfolio, conditional on y. The BASEL framework (and IFRS9 TTC PDs to PIT PDs
which we will discuss later) builds upon these equations where credit losses and
measurement are described using the Vasicek one-factor distribution. This model
assumes the presence of a single risk factor, typically measured by GDP. Additionally, all
loans are linked to this single risk factor through a single correlation value. Longer-term
loans are considered riskier than short-term loans, and the correlation value varies with
the PD. However, the LGD is not correlated with the PD.

Advanced Internal Rating-Based Framework

Source: BIS & Author Adjusted

Source: BIS & Author Adjusted

 Source: BIS & Author Adjusted

For a defaulted exposure/past due exposure, that is where PD=1 or 100% in the above
RWA formula, the capital requirement (K) is equal to the greater of zero and the
difference between its LGD and the bank’s best estimate of expected loss (BEEL). The
RWA formula remains the same.

I have focused a lot on the theoretical aspects here and I would like the reader to note
that to build or validate credit models effectively, one can’t necessarily skip these basics
as the banking credit model frameworks are built around these definitions & formulas.
Of course, assuming the bank follows the IRB approach — for SA approach it is much
more straight forward in the sense that the regulators provide the formulas to use
directly. In the case of AIRB — banks estimate PDs, EADs, LGDs through the use of
various statistical (and ML models).

The series will focus a lot on different modelling and validation techniques through
quantitative coding perspectives but I will refer back to this post for the general basics. In
the following posts we will look at rating philosophies, IFRS9, use cases of AIRB models
in a bank (stress testing, etc), statistical modelling (Logistic Regression & ML models)
and statistical validation techniques (discriminatory power, calibration, stability —
where there will be a lot more coding involved).
References
The Internal Ratings-Based Approach (2001), Basel Committee on Banking
Supervision.

Duffie, Darrell and Kenneth J. Singleton (2003). Credit Risk: Pricing, Measurement, and
Management. Princeton University Press.

Lando, David (2004). Credit Risk Modeling: Theory and Applications. Princeton University
Press.

High-level summary of Basel III reforms (2017), Basel Committee on Banking

Supervision.

EBA DISCUSSION PAPER ON MACHINE LEARNING FOR IRB MODELS (2021),

European Banking Authority.

Conclusion
Thank you for reading! I hope you found this article interesting and helpful (please feel
free to share with me). I am always trying to improve my writing and knowledge sharing.
If you know of any tools or tips please let me know in the comments.

For now, I have no newsletter or active mail list so best is to add me on LinkedIn. Here I’ll
let you know if I have any new posts. Otherwise, best is to give me a follow on Medium
directly.

If you enjoy reading articles like this and want unlimited access to my articles and all
those supplied by Medium, consider signing up using my referral link below.

https://medium.com/@willempretorius/subscribe

You can also buy me a coffee if you’d like :).

https://www.buymeacoffee.com/wlpretorius

Disclaimer: The views expressed in this article are my own and do not represent a strict outlook or the
view of any corporation.
 Machine Learning Credit Risk Management Regulation Data Science Quant

About Help Terms Privacy

Get the Medium app

The Wayback Machine - https://web.archive.org/web/20230313205426/https://medium.com/@willempretoriu…

Search Medium

Willem Pretorius Follow

Mar 11 · 19 min read

An Introduction to Credit Risk in Banking:

BASEL, IFRS9, Pricing, Statistics, Machine
Learning — PART 2

Welcome back! In my previous post (PART 1) I introduced the basic terminology &
building blocks to credit risk management in banking. In this part, I will continue
building upon these basics. I will discuss Rating Philosophies in banking, IFRS9,
important regulatory frameworks in the EU, Use Cases & different models for credit risk
and finally a list of important statistical concepts that is needed to build & validate credit
models. In later posts, the focus will shift towards statistics/machine learning and
implementation of this list with examples in Python.
Table of Contents
International Financial Reporting Standard 9 (IFRS 9)

Rating Philosophies

Important Regulatory Frameworks in the EU

Use Cases & Different Models for Credit Risk

Additional Concepts for the Future

Statistical Concepts for Modelling & Validation

International Financial Reporting Standard 9 (IFRS 9)

IFRS9, a set of accounting standards, was issued by the International Accounting
Standards Board (IASB) in July 2014, replacing the previous International Accounting
Standards (IAS) 39 which was implemented in 2001. The development of IFRS9 began
in 2008 after the global financial crisis, which highlighted the weaknesses of IAS39
(credit losses were underestimated on the provisioning level required by banks). The
financial crisis of 2007–2009 prompted a response not only in terms of capital, but also
from an accounting perspective. IASB therefore undertook a major project to improve
accounting standards for financial instruments, with the aim of creating a more
transparent and accurate forward-looking standard. Effective from January 1 2018, IFRS9
includes several key changes to the classification and measurement of financial assets and
liabilities, and introduces a new expected credit loss (ECL) model for recognizing
impairment losses on financial assets. The ECL model requires companies to recognize
expected credit losses based on forward-looking information, allowing for initial/earlier
recognition of credit losses and a more accurate assessment of the financial health of
companies. Overall, IFRS9 has been widely adopted and represents a significant
improvement to accounting standards for financial instruments.

IFRS9 mandates the use of models to calculate 12-month and lifetime ECLs for financial
assets. These ECL models allow banks to track financial assets from their initial
recognition to their final maturity. The IFRS9 model recognizes three different stages that
needs to be modelled/calculated. The need for these three stages in IFRS9 is to enable a
more accurate recognition and measurement of credit losses on financial assets. By
tracking the financial assets through their different stages (and their interest revenue),
banks can recognize expected credit losses in a more timely and accurate manner,
providing more transparency and reliability in financial reporting. This enhances the
usefulness of financial information for investors and other stakeholders, and ultimately
contributes to the stability of financial markets.

Stage 1:
In Stage 1, assets that have not experienced a Significant Increase in Credit Risk (SICR
— more on this later) since initial recognition or have low credit risk at the reporting
date are recognized. For these assets, a one-year ECL is recognized/calculated. The one-
year ECL represents the expected loss resulting from default events that could occur
within one year after the reporting date. Here, interest revenue is calculated based on the
gross carrying amount (i.e., without deduction for credit allowance).

Stage 2:

In Stage 2, assets that have experienced a SICR since initial recognition but do not have
objective evidence of impairment are classified. For these instruments, a lifetime ECL is
recognized, while interest revenue is still calculated based on the gross carrying amount.
Lifetime ECL accounts for all possible default events that could occur over the expected
life of the financial instrument.

Stage 3:

In Stage 3, assets that have objective evidence of impairment at the reporting date are
classified. For these assets, a lifetime ECL is recognized, and interest revenue is
calculated based on the net carrying amount (i.e., after deduction for credit allowance).

Definition of SICR & Default:

SICR plays a vital role in the ability to distinguish between the different stages. There is a
presumption that the credit risk has increased significantly, when contractual payments
are more than 30 days past due or late. In practice, the probability of default (PD) is the
most commonly used indicator to assess a credit risk increase. A residual lifetime PD
should be used, which means that the same remaining period is considered for both the
PD at origination and the reporting date. As a practical shortcut, one-year PDs can be
used if changes in one-year PDs are a reasonable approximation to changes in the
lifetime PDs.

This implies another important concept: the definition of default under IFRS9. IFRS9
does not offer a precise definition of default. However, an organization must use a
definition that aligns with its internal credit risk management objectives (BASEL
definitions mentioned in PART 1). There is an assumption, which states that a default is
usually considered when the client is 90 days past due, i.e. 90 days past due is usually the
default trigger used. Additional indicators are also important to take into consideration,
such as internal credit rating downgrades and significant change in the value of the
collateral.

ECL:

Source: Tiziano Bellini, IFRS 9 and CECL Credit Risk Modelling and Validation: A Practical Guide with
Examples Worked in R and SAS (2019).

A few important aspects of the above equation needs to be taken into account when
calculating ECLs. The IFRS 9 standard specifies that the estimation of expected credit
losses for a financial instrument should consider the following factors:

A fair and probability-weighted value of potential losses (discussed below), assessed

by analyzing a variety of possible outcomes (i.e. unbiased & forward-looking). In the
previous post, we emphasized the fact that the BASEL PDs (one-year PDs) are
conservative. Many banks prefer to use a unified framework for both capital
requirements and accounting purposes. A stress-testing approach is sometimes
followed and serves as a support system for one-year Through-The-Cycle (TTC) PD
estimates to produce Point-In-Time (PIT) forward-looking lifetime estimates. I will
provide a definition below, but it is important to note that in BASEL we use TTC
 PDs and in IFRS9 we use PIT PDs (reasons to follow). The same can be said for
LGDs & EADs since BASEL uses Downturn conservative scenarios, it is important
to note that when we calculate ECL that the LGDs and EADs should be PIT,
unbiased and forward-looking. There are straight forward (in most cases linear)
conversion formulas to make these values unbiased and forward-looking which we
will look at in a later post (for PDs it is slightly more complicated, but I will provide a
method with code on how to convert TTC BASEL PDs to PIT IFRS9 PDs in a
following article).

The impact of time on the value of money (which is why we have a discount factor
and a full prepayment option).

Access to reasonable and supportable information, without incurring significant

costs or effort, about past events, present conditions, and future economic forecasts
at the reporting date (which is why the macroeconomic forecasts are considered
important).

Additionally, IASB (2014) requires a scenario-weighting scheme (that is, probability-

weighted). This means that the final ECL is obtained as a weighted average of ECLs
estimated under a baseline and alternative scenarios (usually two). The baseline
weightage is usually the highest, and typically there are two other scenarios — one for an
economic upturn and another for an economic downturn. This emphasises the
importance of macroeconomic factors and forecasting in ECL modelling.

Rating Philosophies: TTC PDs vs PIT PDs

One-year PDs that are not significantly affected by the credit cycle are known as TTC
credit measures. These measures provide more stability over time. The default
probabilities generated by a TTC model are largely consistent over time (time-
invariant). On the other hand, PIT credit risk measures use all relevant information
available on a specific date to estimate the expected likelihood of default. Unlike TTC
models, PIT models take into account macroeconomic risk factors, making them more variable.

TTC (long-run) PDs should be used for capital requirement calculations. If the PD used
is not TTC here, it will vary with the economic cycle, and the capital requirement will
also vary (not a situation we always want). In TTC PDs grade migrations would therefore
be uncorrelated to the changes due to the economic cycle. One would see the effect of
such changes ex post in how the realized default rates of each rating varies over time. On
the contrary, a complete PIT model clarifies that variations in default rates occur due to
changes in the economy, ratings and conditional PD of multiple obligors. Another
interpretation is that the same rating category corresponds to varying levels of
creditworthiness depending on the state of the economy.

Fully adopting PIT models for determining capital requirements could amplify the
impact of economic cycles during both upturn and downturn periods, which could have
negative consequences for financial stability (Adrian and Shin, 2008). However, relying
solely on TTC models is unlikely to satisfy a bank’s business needs for assessing the risk
of a transaction or monitoring the creditworthiness of a borrower, as the economic cycle
is a crucial risk factor that needs to be considered. Therefore, in regulatory capital
calculations we are interested in TTC PDs but in IFRS9 forward-looking ECL
calculations we are interested in PIT PDs.

Source. Z-Score method for conversion.

A PIT model can directly be used for conversion of PIT PDs to TTC PD or vice versa.
Usually TTC to PIT PDs conversions are done through the Vasicek distribution and
framework introduced in PART 1. PIT PDs can then be easily adjusted to forward-
looking scenarios by calculating the cumulative PDs from the marginal yearly PIT PDs.
Marginal PDs represent the PD for a specific time period, typically one year, for a
borrower who has not yet defaulted up to that point. Marginal PDs are calculated using
conditional probabilities and can be used to estimate the risk of default for a specific
borrower over a given time period. The conditional PD represents the probability of a
borrower defaulting in for example the second year, given that they did not default in the
first year. This calculation requires two pieces of information: the probability of
surviving without defaulting in the first year (P0), and the unconditional PD in the
second year (P1).

For example:
If P0 is 0.5 and P1 is 0.1, the conditional PD (i.e., the PD given survival) would be 0.1
divided by 0.5, or 20%.

Cumulative PDs, on the other hand, represent the PD for a borrower over a specific time
horizon, such as the life of a loan. Cumulative PDs are calculated by adding up the
marginal PDs for each time period up to the end of the time horizon. Cumulative PDs
are used to estimate the overall risk of default for a portfolio of loans, and can be used to
calculate the expected loss and capital requirements for the portfolio.

For example:
Suppose the one-year or 12-month PD is 2% which means that the survival rate is 98%.
Suppose also the 2nd and 3rd year conditional PDs are 4% and 5%, respectively.

Then the 1st year cumulative survival rate (CSR) is same as first year survival rate (SR).
The 2nd year CSR = 1st year CSR * SR of 2nd year = 98% * 96% = 94%. The 3rd year CSR
= 2nd year CSR * SR of 3rd year = 94% * 95% = 89%. The Lifetime PD = 1–89% = 11%.

This brings forth another important distinction that needs to be kept in mind: stressed
vs unstressed PDs. Stressed PD: this PD depends on the risk attributes of borrower but is
not highly affected by macroeconomic factors as adverse economic conditions are already
factored into it (TTC). Unstressed PD: this PD depends on both current macroeconomic
and risk attributes of borrower (PIT). It fluctuates depending on the economic
conditions.
 BASEL vs IFRS9. Adjusted from Source.

TTC PDs to PIT PDs:

I will have a separate post on this calculation with examples and code.

Important Regulatory Frameworks in the EU

For a list of all the important regulations one can view:
https://www.eba.europa.eu/regulation-and-policy/credit-risk

The regulations are distinguished between Discussions/Recommendations, Guidelines

and Regulatory Technical Standards (RTSs). These are mostly regulations surrounding
IRB and capital management. For the IFRS9 accounting framework, the ECB published
the following:
https://www.eba.europa.eu/regulation-and-policy/accounting-and-auditing/guidelines-
on-accounting-for-expected-credit

For a complete list of Capital Requirements Regulation (CRR) which reflects BASEL
rules on capital measurement and capital standards for the EU:
https://www.eba.europa.eu/regulation-and-policy/single-rulebook/interactive-single-
rulebook/504

In my previous post I focused on the Risk-Weighted Asset (RWA) calculation related to

exposures for corporates, institutions, central governments and central banks. If you are
interested in Retail exposures and others, you can have a look at Article 153 and onwards.
The formula structure relatively stays the same but, the correlation coefficients changes
slightly.

Use Cases & Different Models for Credit Risk

IRB (Internal Ratings-Based) models are commonly used by banks in various aspects.
Usually IRB risk parameters & PDs, LGDs and EADs are used in various models in a
bank. Chapter 7 on the ECB guide to internal models outlines these uses. One can find
the regulatory articles in this chapter also. These use cases are beneficial for banks to not
build independent models but to rather use the output from IRB models and transform
them for the bank’s specific needs. The specific model might be:

Credit Approval or Acceptance Models: Banks may use the internal ratings and default
& loss estimates in the approval, restructuring and renewal of credit facilities. These are
statistical models used by banks to evaluate and assess the creditworthiness of potential
borrowers. Many Machine Learning (ML) experimentation in this use case is currently
ongoing in banks around the world (specifically non-linear ML classification methods, as
they are able to pick up patterns that traditional models struggle with). From data
collection to the final approval model.

Pricing of Transactions: Risk assessments and internal ratings can be used to determine
transaction pricing. The ECB recommends to use risk-adjusted performance indicators
such as Risk-Adjusted Return Capital (RAROC) or adjusted IRB parameters to ensure
accurate pricing estimation.

Early Warning Systems (EWS): The bank’s EWS can consider the dynamics of PDs or
ratings downgrades, as well as other indicators related to risk measures such as expected
loss, loan-to-value, or overdrafts. The use of ML is becoming prominent in this case.
Non-linear classification methods are being heavily experimented here.

Collection and Recovery Policies and Processes: Banks may implement risk management
procedures that are activated before the exposure defaults, such as early collection calls.
These procedures can be based on various indicators, including internal ratings or risk
drivers. Banks can also establish rules, strategies, or procedures for the recovery process,
which consider factors such as their LGD/Expected Loss Best Estimate (ELBE) values
and set-aside provisions.
Credit Risk Adjustments: The alignment of collective provisioning for both performing
exposures and defaulted assets (or the proportion of defaulted assets) to IRB parameters
may be done. This is where banks can use IRB risk parameter outputs (and PDs, LGDs,
EADs) as input to an IFRS9 model, but they have to ensure to convert these parameters
to point-in-time, unbiased and forward-looking. The output of the IFRSS9 model can
be directly used as an input to a Stress Test model (stress test for future provisions)
which in this case is another use case of IRB rating and parameters. Banks can use the
results of these stress tests (and RWA stress tests) to identify areas of their portfolio that
are most vulnerable to economic shocks and adjust their risk management strategies
accordingly. Banks can use the output of stress testing to further adjust their future
regulatory or economic capital.

Additional Concepts for the Future:

I will not go into depth in the below concepts, as I am still developing my own knowledge
and ideas around these topics. However, I will provide a quick overview and will refer
back to this section in future posts.

Margin of Conservatism (MoC)

This refers to the use of a cautious adjustment to address any deficiencies in the
quantitative assessment of risk parameters. The adjustment is intended to be on the side
of caution, to ensure that the risk is not underestimated, and adequate measures are
taken to mitigate the risk. This is the level of prudence or caution exercised in assessing
the risk associated with a credit exposure.

A conservative approach involves using cautious assumptions and parameters in credit

risk models, resulting in higher expected credit losses and more conservative
provisioning levels. Conversely, a less conservative approach entails the use of optimistic
assumptions and parameters, which may lead to lower expected credit losses and less
conservative provisioning levels. Banks are required to address the identified deficiencies
in data or methods via appropriate adjustments and MoCs. General estimation error is
very important in this regard.

Climate Change
This topic is very actively researched currently. The impact of climate change is felt
worldwide and requires a shift towards sustainable growth to prevent permanent,
catastrophic effects. Recognizing that this shift may affect economic stability, financial
regulators require banks to identify, quantify and mitigate the financial risks associated
with climate change. These risks can be classified as physical risks, resulting directly from
climate change such as extreme weather events, and transition risks associated with the
shift towards a sustainable, low-carbon economy, such as changes in consumer
preferences and raw material costs. Both financial institutions and regulatory authorities
are exploring methods to quantify these risks.

The ECB requires banks to do climate stress tests, and this is only the beginning of the
climate topic in banking.

Expert & Human Judgement

Human judgment refers to the process of incorporating subjective opinions, experience,
and expertise of credit data scientists and risk managers into quantitative models (think
of risk factor selection when building models and deciding which factors are the most
important — sometimes not only statistical aspects are taken into account but also
expert opinions). Since models alone may not capture all relevant information,
anticipate new trends or changes in the credit environment, human judgment plays an
important role in adjusting or complementing model outputs to reflect unique
circumstances or idiosyncratic risks that are not captured by the models.

The use of human judgment can take various forms depending on the stage of the credit
risk assessment/modelling process. For example, it may involve adjusting PD estimates
to reflect insights into the borrower’s financial condition, industry trends, or other
qualitative factors. Similarly, human judgment may be used to determine appropriate
credit limits, pricing, or loan covenants based on the overall risk profile of the borrower.

It is extremely important to ensure that no biases are incorporated into the modelling
process and that these expert opinions are well-defined, explained and monitored.

Business
Managing credit risk in banking from a business perspective involves balancing the
bank’s overall business strategy and objectives with the risks of default and loss. The
bank needs to ensure that it lends to creditworthy borrowers, prices loans appropriately,
and monitors borrowers for signs of financial distress.

The management of credit risk also involves making decisions about the types of loans
and borrowers the bank is willing to finance, such as small businesses, real estate
developers, or consumers. Each type of borrower presents unique risks that need to be
considered in the bank’s credit risk management practices.

Effective credit risk management is crucial for the success of a bank’s business
operations, as it enables the bank to generate profits while minimizing the risk of loss
due to defaults.

Business needs are very important in the world of credit risk. The business owners of
models in banks are not always technical. It can become a challenging aspect of a data
scientist or credit quant to explain technical terms or details of models to the business
owners (think of trying to explain to someone how a neural network or boosting model
works or how to calibrate correctly — to someone who never heard of these terms
before). It is therefore crucial to not only develop hard skills when you are a beginner but
to also work on soft skills daily as these become more important with how you grow in
the organization. Soft skills, such as communication, teamwork, and adaptability, are
essential for working effectively with others, managing conflicts, and navigating complex
situations in an organization.

Data scientists and quants are not always expected to solely focus on mathematical
computations and coding, but also to communicate their results appropriately with
higher-up business owners. In the end, a bank is a business that needs to sell a product.
Therefore, data scientists and quants must understand the business side of things and
not solely focus on technical aspects.

Statistical Concepts for Modelling & Validation

Regulation states that estimates must be based on the material drivers of the risk
parameters. This means that the explanatory variables in a credit model such as PD is
usually called a risk driver. In the process of selecting risk drivers for models, there’s a
potential danger of mistakenly attributing risk drivers that only apply to defaulted
obligors as relevant risk drivers for the entire portfolio. To prevent this, institutions
should implement measures to counteract overfitting and model misspecification. This
is especially crucial when there’s limited default data available for developing the model.

Typical risk drivers are: Demographic (income, employment status, etc), Credit Bureau
variables (credit scores, default or delinquency history, etc), Financial ratios and
variables (company structure, size, etc), Behavioral information (payment behavior,
utilization of credit, etc), Transactional variables (products, collateral, etc).

Model Lifecycle:
Data Preparation (missing value treatment, outlier analysis, etc).

Variable Selection (transforming of variables and choosing final important variables

for modelling).

Model Development (for PD/Scorecard build a binary modelling framework and for
LGD/EAD these are usually continuous).

Model Validation (data quality tests, discriminatory power analysis, stability

analysis, calibration analysis, qualitative analysis).

Independent Validation (audit).

Supervisory Approval.

Model Implementation.

Periodic Monitoring (use mostly same tests as in validation phase).

Post Implementation Validation: Backtesting and Benchmarking.

Model Redevelopment/Refinement (if any issue).

Data Preparation
Data Collection: This involves finding the appropriate data for the specific model that
needs to be built. This can be internal or external data.

Data Cleaning: This involves identifying and correcting errors, outliers or missing data
in the dataset. It’s important to ensure that the data is complete and accurate before
building a model.

Variable Selection
Risk Driver Creation: This involves finding or creating the variables that explains our
purpose, i.e. finding explanatory variables that explain the PD, for example financial
information.

Data Transformation: This technique involves converting raw data into a format that is
suitable for modelling. For example, converting categorical variables into numerical
values or normalizing data to make it more consistent. Feature scaling is important here.

Feature Engineering: This involves creating new features or variables from the existing
ones to improve the performance of the model. For example, creating a debt-to-income
ratio variable or a variable that indicates how long a borrower has held their current job.

Sampling Techniques: These techniques are used to balance the dataset by oversampling
the minority class (defaulters) or undersampling the majority class (non-defaulters).
This is done to avoid model bias towards the majority class. This is rarely done, but in
smaller portfolios this might be a great option to utilize ML sampling techniques.

Risk Driver Selection: This involves using statistical variable selection techniques such as
forward or backward selection to obtain the most important variables that explain our
purpose. Clustering techniques can also be used or principal component analysis.
Regularization can also play an important role here. We will look deeper into these
techniques later.

Model Development
This involves building statistical models for our purpose (predicting PD or LGD/EAD).
Different models exist for different purposes:

PD: Binary Classification — Logistic Regression (GLMs/Logits/Probits), Markov-

Chains, Decision Trees, Support Vector Machines, Random Forests, Boosting or
Survival models (Neural Networks are not extremely famous in credit risk so don’t
expect to see it in any bank as they have extreme interpretability issues and barely out-
performs traditional and ML methods. The same holds for using voting classifiers, which
is basically stacking models together and using them to define a majority vote scenario).
These models can also be used in Early Warning Systems and Acceptance modelling.

LGD/EAD: Formulas or fixed values (provided by regulation), Regressions (Linear,

Trees, Tobit, Beta, etc), Survival models, ML (Support Vector Machines, Random
Forests or Boosting).

Model Validation
Typically, models in the development stage are also validated through the use of unseen
data. Usually, the data is split into a training and a test set. The model is trained on the
training set and validated on the test set. Various cross-validation or bootstrapping
techniques can be used. External model validation happens in a different model lifecycle
than the train-test split method.
There are usually a few stages in validating the above models:

Data Quality: Here we validate if the data used in the model development cycle was
accurate, complete, and reliable. The question we ask is, was our sample data used to
train the model representative of the current use of the model. Examples of tests to use:
Tukey’s Fence, box plots, scatter plots, etc.

Discriminatory Power: Here we check how well our model distinguish/discriminate

between events (defaults/cures) and non-events (non-defaults/non-cures). Otherwise
stated as, the model’s ability to differentiate between good (non-defaulters) and bad
(defaulters) clients. For now, I will provide a list of important metrics we can look at. In
the future, I will go deeper into these techniques.

For PD/Binary Classification models:

Confusion Matrix, Sensitivity & Specificity analysis.

Receiver operating characteristic (ROC) curve and the area underneath (area under
ROC, or AUROC, or AUC) or Cumulative Accuracy Profile (CAP).

Accuracy Ratio / PowerStat.

Lorenz Curve.

Gini Coefficient (for PD/Cure models this is used industry wide).

Somer’s D (concordance vs discordance).

Pietra Index.

Kolmogorov-Smirnoff Statistic (KS).

Mann-Whitney Test Statistic.

For LGD/EAD/Continous Regression models:

Variation Measures (Mean Absolute Deviation, R-Squared, etc).

Rank Correlations (Spearman, Pearson’s or Matthew’s).

Loss Capture Ratio.

When assessing discriminatory power, it’s important to be aware of a few limitations.
Firstly, since defaults are considered random events, the results of discriminatory power
measures can also be considered random variables. Therefore, any comparisons made
should be subject to statistical tests and confidence intervals.

There is a great list in Scikit-Learn that is available freely to use when doing these
validations:
https://scikit-learn.org/stable/modules/model_evaluation.html

Calibration Accuracy: Here we check how close the predicted values are to the actual rate
of events. We want our model to calibrate to as close to reality as possible.

For PD/Binary Classification models:

Binomial Test.

Vasicek Regulatory Test.

Brier Score (for PD/Cure models this is used industry wide).

Hosmer and Lemeshow Test (HL).

Jeffrey’s Prior Test.

Learning Curves.

Calibration Curves.

For LGD/EAD/Continous Regression models:

T-test.

Loss Shortfall.

Deviance and Residual Tests.

Stability: Here we check whether internal or external environmental changes will impact
the model. This test how stable the parameters are over time. The typical test here is,
H0: The distribution of the model is stable vs H1: It is not stable and changes too rapidly
over time.

For PD/Binary Classification models:

 Pearson’s Chi-Squared Test.

Population Stability Index or System Stability Index (PPI or SSI).

Mann-Kendall Test.

For LGD/EAD/Continous Regression models:

Kolmogorov-Smirnoff Statistic (KS).

Anderson-Darling Test.

Mann-Kendall Test.

I will zoom into the above methods with examples & code in the future. The biggest
challenge in analysing these techniques is to find a dataset.

Conclusion
In the above I have tried to give an overview of IFRS9 and the complete model lifecycle
with an introduction to the important statistical concepts needed for model
development and validation. In a future post, I will provide a full example of converting
TTC to PIT PDs for IFRS9 calculations. I will first dive deeper into model building and
validation techniques with specific datasets.

Thank you for reading! I hope you found this article interesting and helpful (please feel
free to share with me). I am always trying to improve my writing and knowledge sharing.
If you know of any tools or tips please let me know in the comments.

You can find the references I used throughout the post attached as links to the specific
topics.

For now, I have no newsletter or active mail list so best is to add me on LinkedIn. Here I’ll
let you know if I have any new posts. Otherwise, best is to give me a follow on Medium
directly.

If you enjoy reading articles like this and want unlimited access to my articles and all
those supplied by Medium, consider signing up using my referral link below.
https://medium.com/@willempretorius/subscribe

You can also buy me a coffee if you’d like :).

https://www.buymeacoffee.com/wlpretorius

Disclaimer: The views expressed in this article are my own and do not represent a strict outlook or the
view of any corporation.

Credit Risk Management Data Science Machine Learning

About Help Terms Privacy

Get the Medium app