Towards Data Science

Credit Risk: What Are We Actually Measuring?

8.5内容质量

TL;DR · AI 摘要

信用风险建模中的PD、LGD、EAD参数是银行评估贷款损失和资本需求的核心工具,但其应用需结合数据质量与监管框架。

核心要点

  • PD(违约概率)是评估借款人未来违约可能性的关键指标
  • 巴塞尔协议要求银行使用EAD(风险敞口)计算资本充足率
  • LGD(违约损失率)受抵押品价值和回收率直接影响

结构提纲

按章节快速跳转。

  1. 通过银行贷款场景引出信用风险建模的核心问题

  2. 解释PD/LGD/EAD的统计学定义与业务含义

  3. 说明三大参数如何影响银行资本计算

  4. 讨论实际建模中数据质量与假设验证的难点

  5. 通过住宅贷款场景演示参数计算过程

思维导图

用一张图看清主题之间的关系。

查看大纲文本(无障碍 / 无 JS 友好)
  • 信用风险建模
    • 核心参数
      • PD(违约概率)
      • LGD(违约损失率)
      • EAD(风险敞口)
    • 监管框架
      • 巴塞尔协议
    • 应用挑战
      • 数据质量
      • 模型假设验证

金句 / Highlights

值得收藏与分享的关键句。

#信用风险#巴塞尔框架#金融建模#数据科学
打开原文

Credit Risk: What Are We Actually Measuring? | Towards Data Science

Data Science

Credit Risk: What Are We Actually Measuring?

Let’s start with a loan and work out what PD, LGD, and EAD tell us about losses and bank capital.

JUNIOR JUMBONG

November 2, 2026

15 min read

Prefer us on Google

Share

Photo by Markus Winkler on Unsplash .

Suppose you wanted to become a credit risk modeller, working on the parameters used in the Basel framework. That may not be your plan today. But we hope that understanding the job a little better will make you want to find out more.

You would soon discover how much it draws on: statistics, banking regulation, data, and the way credit works. Looking at everything you would need to learn, you might feel that even two years would not be enough to master all these concepts. Where do you begin?

We could explain every concept in detail. But you might end up collecting pieces of knowledge without understanding how they fit together. Knowing how to calculate a probability of default is only a start. You also need to understand what it measures, why a bank needs it, and which decisions it can inform.

So we will begin with the big picture. We will follow the reasoning from a banking question to the data, then to a model and the use of its results. The technical details will find their place as we go.

This first explanation will involve some simplifications. We will not cover every method, every requirement, or every possible situation. But simplifying an explanation should preserve how things work. We will keep the essential mechanisms and point out the limits of what we explain.

The aim is to help you understand how the different parts of the job fit together, so that you know what you are learning and why you are learning it. Becoming a specialist will take more than a few pages.

To understand the job, let us start with one of a bank’s most familiar activities: lending money.

A bank lends money. But how does it know it will get it back?

Imagine a simple situation. You want to buy an apartment, but you do not have enough money to pay for it. So you go to your bank. To assess your finances, the bank asks you a few questions. What is your income? What are your monthly expenses? Do you rent or own your home? Do you have a permanent employment contract?

After reviewing your application, it agrees to lend you €250,000, repayable over 25 years. You sign a contract committing you to repay the loan, together with interest, according to an agreed schedule. So far, nothing particularly complicated. But how can the bank know, today, that you will still be able to repay it in five, ten, or twenty years? It cannot. You could lose your job. Your income could fall. It could also rise.

The bank must therefore make a decision today without knowing the events that will affect your ability to repay. It faces uncertainty. That uncertainty also creates a risk: if you do not repay as agreed, the bank may lose money. This is credit risk.

The same applies to businesses. A company that borrows to build a factory or buy machinery may also run into repayment difficulties. When a borrower becomes unlikely to repay, or has certain material payments overdue, we may be dealing with a default. That is the general idea. Regulation sets out the criteria to apply: a late payment does not automatically amount to a default.

Now imagine thousands of loans. Some borrowers will repay everything. Others will go through a difficult period. Others will default. The bank expects to incur losses without knowing exactly who will default, when, or how much it will lose.

You may notice that these borrowers are very different: individuals, small businesses, large companies, and governments. It would be difficult to study all of them in the same way. The bank will therefore ask us, as its analysts, to work with more comparable groups. We might study a portfolio of residential mortgages or a portfolio of loans to large companies.

You will often encounter the terms retail for retail exposures, and non-retail for large corporate or SMEs exposures. Individuals generally fall within retail, but some small businesses can also qualify, subject to conditions. The distinction is therefore more involved than simply putting individuals on one side and businesses on the other.[1]

A portfolio remains a collection of loans whose risks can vary. We will see later how to distinguish levels of risk within that collection. For each portfolio, we need to distinguish two things. Expected losses are the losses we anticipate on average. Suppose they amount to €10 million over one year. That does not mean the bank will lose exactly €10 million every year.

Unexpected losses are losses above that average level. A crisis can make them larger: unemployment rises, businesses close, and several borrowers run into difficulties at the same time.

Expected losses enter the economics of lending and the provisioning process: the bank recognises estimated credit losses in its accounts under the applicable accounting rules. To absorb losses beyond that level, it needs a cushion of capital. This includes, among other things, money contributed by shareholders and profits the bank retains: resources that can absorb losses.[2]

But before working out how large that cushion should be, we need to measure the risk.

What do we need to estimate for each portfolio?

Let us return to our mortgage. To understand how much the bank could lose, we need to ask three questions. How likely is the borrower to default? This is PD, the Probability of Default . In the framework we will follow, it refers to a one-year horizon.

A PD of 2% does not mean that the borrower will repay 98% of the debt. It means that the estimated probability of entering default during the year is 2%. Across a large population of comparable borrowers, and provided the estimate is appropriate for the conditions being studied, this corresponds to an expected average default frequency of 2%. We do not expect exactly two defaults in every group of one hundred borrowers.

If the borrower defaults, what proportion of the exposure will the bank lose? This is LGD, or Loss Given Default , expressed as a proportion of the exposure. A default does not mean the bank loses its entire claim. It may recover some of the money, for example through collateral or a guarantee. Two loans can have the same PD and different LGDs. In practice, calculating LGD also involves the costs of recovery and how long recovery takes.

How much will be exposed when the default occurs? This is EAD, or Exposure at Default . It can differ from the original loan amount because of repayments already made or, for a credit line, further drawdowns.[3]

We now have three quantities: a probability, a loss proportion, and an amount of money. For an exposure that is not already in default, they combine in the regulatory calculation of expected loss in euros:[4]

Expected loss = P D × L G D × E A D . \text{Expected loss} = \mathrm{PD} \times \mathrm{LGD} \times \mathrm{EAD}. Expected loss = PD × LGD × EAD .

Take an example purely to illustrate the calculation: a PD of 2%, an LGD of 40%, and an assumed EAD of €250,000.

Expected loss = 0.02 × 0.40 × 250 000 = 2 000 euros . \text{Expected loss} = 0.02 \times 0.40 \times 250\,000 = 2\,000\ \text{euros}. Expected loss = 0.02 × 0.40 × 250 000 = 2 000 euros .

If default occurs, the estimated loss is €100,000, or 40% of the exposure. But default is uncertain, so the expected loss is €2,000. These are two different quantities. For a portfolio, we add up the expected losses of the individual exposures. Multiplying three averages calculated separately across the whole portfolio will not necessarily give the same answer.

You may be wondering where these estimates come from. There is no column in our database that tells us who will default in the future. We can, however, observe the past: a borrower’s circumstances at a particular date, followed by what happened over the next year. Those observations provide the starting point for our models.

Why are these estimates regulated?

A bank, particularly a large bank, plays several roles in the economy. It takes deposits, provides financing, and keeps payments moving. When a large bank fails, its difficulties can spread to others. Banks lend to one another and may be exposed to the same economic shocks. This is systemic risk.

It becomes easier to see why risk measurement is regulated. If a bank underestimates its credit risk, it may also underestimate the capital it needs. The Basel framework sets international prudential standards, which are incorporated into the rules applicable to banks, including in Europe. Regulation and supervision govern definitions, permitted methods, data quality, and the use of models.[1]

Three parts of the job therefore come together. Regulation specifies what must be measured and under which conditions. Statistics allows us to build models from data and estimate risk parameters. The bank’s needs give those estimates a purpose : calculating capital requirements, measuring risk, managing portfolios, and, depending on the model, making lending decisions.

These parts constantly interact. A statistical method can be useful without meeting a regulatory requirement. A compliant model can be difficult for its users to understand. Our job as modellers is to make these requirements work together. That is what makes the role broader than simply building a model. We still need to understand which quantities the bank can estimate for itself.

How do we estimate a PD?

For the regulatory calculation of credit risk, we will focus on two broad approaches: the standardised approach and the Internal Ratings-Based approach, or IRB.[1] Under the standardised approach, prudential rules prescribe the risk weights. These coefficients give exposures more or less weight according to their risk. The bank must classify its exposures and apply the corresponding rules. It does not use its own PD, LGD, and EAD estimates in the way the IRB approach does. That still leaves plenty of work for the analyst.

Under the IRB approach, the bank uses internal estimates of certain parameters, subject to meeting the applicable requirements and obtaining the necessary supervisory permissions. For certain non-retail asset classes, there is a foundation approach, F-IRB, and an advanced approach, A-IRB. Under the foundation approach, the bank estimates PD and uses supervisory values or methods for the other components. Under the advanced approach, where permitted, it also estimates LGD and EAD. There is no foundation variant for retail: the IRB approach relies on internal estimates of PD, LGD, and EAD.[1]

We will now use PD to follow the reasoning. How do we get from data to this probability? Four broad steps will help us understand the process. Define default. Before counting defaults, we need to say what counts as one. The regulatory definition shapes the dataset and the estimation that follows.

Build the dataset. We choose a coherent population, such as mortgage borrowers. For each observation, we keep the information available at the date being studied: income, indebtedness, payment history, and loan characteristics. We then check whether the borrower, who was not in default at the start, defaults during the following twelve months. The information used to assess the borrower must be available at the starting date. We cannot use something we will only learn later.

We need several years of observations. For PD estimation, the Basel framework includes requirements for historical data covering at least five years, with data representative of the relevant economic conditions. Five favourable years will not necessarily tell us the whole story about risk.[5]

Data is rarely perfect. Suppose that after dealing with missing information, only 70% of the observations are usable. Having fewer observations is only part of the problem. If the excluded borrowers are also the most vulnerable, our sample may give an overly favourable picture of the portfolio. We therefore need to investigate what is missing and how that affects the estimate.

Differentiate risk. We build a score that ranks borrowers according to their risk. Logistic regression is a common tool for studying a binary outcome, such as default or no default. Other methods are available.

We can then group borrowers into risk classes. You will encounter the term grade for non-retail ratings and pool for groups of sufficiently similar retail exposures. Borrowers within a group should have sufficiently comparable risks, while the groups should allow us to distinguish different levels of risk.[5]

You can now see how this differs from defining the portfolio: first we chose the loans to study; then we distinguished the levels of risk within them. Calibrate the probabilities. Ranking borrowers well is not enough. Knowing that one group is riskier than another does not tell us whether its PD is 2%, 5%, or 10%.

To understand the starting point, imagine a group of 1,000 borrowers who are not in default, all followed for one year. If 20 default during that year, the observed default rate is 2%. But that result describes that group during that year. Estimating a regulatory PD also requires us to use the historical data, seek an estimate consistent with the long-run average, and address shortcomings in the data or method. Adjustments and margins of conservatism are part of that work. A rate observed in a single year does not automatically become the PD we should use.[5]

I have presented these steps for building a score on a retail portfolio in a series of articles:

  • How to Define the Modeling Scope of an Internal Credit Risk Model
  • Exploratory Data Analysis for Credit Scoring with Python
  • Handling outliers and missing values in borrower data
  • A Practical Guide to Measuring Relationships between Variables for Feature Selection in a Credit Scoring.
  • How to select variables robustly in a scoring model
  • A practical guide to categorization in credit scoring
  • How to Validate That Your Variables Tell a Consistent Risk?
  • How to Train a Scoring Model in the Age of Artificial Intelligence
  • Turning model coefficients into a 0–1000 score, with risk classes and stability checks

The last article presents a division into six risk classes or six grades. The default rates observed within those classes help illustrate the move from a score to a probability. They provide a teaching example; regulatory calibration requires the additional work we have just discussed.

In this example, Grade 1 is the riskiest, with an observed default rate of about 58%. What does that number tell us? Roughly 58 out of every 100 borrowers in that grade defaulted during the period studied. Can we simply use 58% as the grade’s regulatory PD? Not automatically. We must check whether the observations cover the right horizon, represent long-run default experience, and provide enough reliable information. Expert judgement and knowledge of the lending business can support justified adjustments. We must also include appropriate margins of conservatism to account for estimation uncertainty.

Each borrower can then be assigned a risk class (or grade) and its associated PD. But producing an estimate does not, by itself, authorise its use. We still need to check the model, document its limitations, and arrange for ongoing monitoring.

How do we check and monitor the model?

A model is built from observations of the past, and the world changes. Feynman reminded us to be wary of our ability to fool ourselves. Modellers face that difficulty too.[6] Three activities help us deal with it.

Validation. We assess the model using, among other things, data that was not used to build it, preferably from a later period. We examine its ability to distinguish levels of risk, the reliability of its probability estimates, and the stability of its results over time and across populations.

A difference between estimated PDs and observed defaults does not automatically mean the model is wrong. Observations fluctuate. We need to consider the sample size, the number of defaults, the economic period, and what the calibration is intended to represent. A PD designed to reflect a long-run average is not expected to reproduce each year’s observed default rate exactly.

Independent validation examines the choices and results from a perspective separate from that of the development team. Internal audit also forms part of the control framework, with a different role.[5]

Documentation. A model that we cannot explain is difficult to check. We record the data used, the assumptions, the methodological choices, their limitations, and the reasons for making them. Someone else should be able to understand the work, reproduce it, and challenge it.

Monitoring. Once the model is in production, we monitor its performance, changes in the portfolio, and data quality. We regularly compare estimates with observations. Using defined criteria, we determine whether to correct data, revisit the calibration, or rebuild the model.[5]

This work continues throughout the model’s life. As you might expect, our job does not end when the algorithm works. We now have estimates and a way to check them. We still need to understand how they feed into the calculation of capital requirements.

How do the results feed into regulatory capital calculations?

Let us return to unexpected losses. The bank needs to absorb losses above their average level. We must therefore consider much less favourable situations. In the model underlying the IRB risk-weight functions, PD enters the calculation of a default rate conditional on a very unfavourable economic situation. We will call it WCDR here, for Worst Case Default Rate .

The confidence level is 99.9% over a one-year horizon. Under the model’s assumptions, this corresponds to a loss level that would be exceeded in only 0.1% of cases. It is neither an individual borrower’s probability of default nor a prediction of when the next crisis will occur.[2]

To understand the mechanism, we can write the following before the adjustments that apply to different exposure categories:

K ≈ ( W C D R − P D ) × L G D . K \approx (\mathrm{WCDR} - \mathrm{PD}) \times \mathrm{LGD}. K ≈ ( WCDR − PD ) × LGD .

Here, K K K is a proportion of the exposure. To obtain the corresponding capital amount, which we will call C C C , we multiply by EAD:

C = K × E A D . C = K \times \mathrm{EAD}. C = K × EAD .

Return to our example: a PD of 2%, an LGD of 40%, and an EAD of €250,000. Suppose, purely to illustrate the calculation, that WCDR is 10%. We have chosen that rate for the example. In practice, it follows from the regulatory function, including PD and the prescribed correlation. That correlation represents borrowers’ common sensitivity to economic conditions.

K ≈ ( 0.10 − 0.02 ) × 0.40 = 0.032. K \approx (0.10 - 0.02) \times 0.40 = 0.032. K ≈ ( 0.10 − 0.02 ) × 0.40 = 0.032.

The capital proportion in this simplified example is therefore 3.2%, giving:

C ≈ 0.032 × 250 000 = 8 000 euros . C \approx 0.032 \times 250\,000 = 8\,000\ \text{euros}. C ≈ 0.032 × 250 000 = 8 000 euros .

We still need to connect this amount to RWA, or Risk-Weighted Assets . RWA express exposures as amounts weighted according to their risk. Two loans of the same amount can therefore produce different RWA. The minimum total capital requirement is 8% of RWA, before buffers and other additional requirements. The reciprocal of 8% is 12.5. That is where the coefficient in the following relationship comes from:[7]

R W A = 12.5 × K × E A D = 12.5 × C = 12.5 × [ ( W C D R − P D ) × L G D ] × E A D \mathrm{RWA} = 12.5 \times K \times \mathrm{EAD} = 12.5 \times C = 12.5 \times [(\mathrm{WCDR} - \mathrm{PD})\times \mathrm{LGD}] \times \mathrm{EAD} RWA = 12.5 × K × EAD = 12.5 × C = 12.5 × [( WCDR − PD ) × LGD ] × EAD

In our simplified example:

R W A ≈ 12.5 × 8 000 = 100 000 euros . \mathrm{RWA} \approx 12.5 \times 8\,000 = 100\,000\ \text{euros}. RWA ≈ 12.5 × 8 000 = 100 000 euros .

You can check it yourself: 8% of €100,000 is indeed €8,000. The complete regulatory formulas depend on the exposure category and may include a maturity adjustment, related to the length of the commitments. The LGD used must also meet the applicable prudential requirements, including those concerning downturn conditions. Our calculation explains the mechanism; applying the framework requires the full set of rules.[8]

What do we understand now, and what is left to learn?

A bank lends without knowing the future. To measure its credit risk, it seeks to estimate three quantities: the probability of default, the proportion lost if default occurs, and the amount exposed.

Those estimates draw on data, statistical methods, and prudential rules. We have followed the construction of a PD: define default, build the dataset, differentiate risk, and then calibrate the probabilities.

We have also seen why a model must be validated, documented, and monitored. Finally, we connected the risk parameters to RWA and capital requirements.

You can now see how the different parts of the job fit together. When you study a statistical method, a calibration requirement, or a data problem, you will have a clearer idea of the question that work is meant to answer.

If our colleagues read this article, they will notice the details we have left out. There are plenty: estimating LGD and EAD, portfolios with few defaults, margins of conservatism, model changes, and many other topics.

But we wanted to start simply, so that newcomers could understand the job and, we hope, feel encouraged to join us. There are many challenges to work on, particularly around data, artificial intelligence, and the way modelling practices evolve.

The first thing is to know which question we are trying to answer. The details can then find their place.

References

[1] Basel Committee on Banking Supervision, CRE30 — IRB approach: overview and asset class definitions , especially paragraphs 30.1, 30.19–30.22, and 30.32–30.42.

[2] Basel Committee on Banking Supervision, An Explanatory Note on the Basel II IRB Risk Weight Functions , 2005, sections 2, 4.2–4.4, and 5.1. This note is used to explain the model’s economic mechanism.

[3] Basel Committee on Banking Supervision, CRE32 — IRB approach: risk components .

[4] Basel Committee on Banking Supervision, CRE35 — IRB approach: treatment of expected losses and provisions , paragraphs 35.2–35.3.

[5] Basel Committee on Banking Supervision, CRE36 — IRB approach: minimum requirements to use IRB approach , especially sections 3, 5, 7, and 8.

[6] Richard Feynman, Cargo Cult Science , Caltech commencement address, 1974. The passage in section 6 is a paraphrase, rather than a direct quotation.

[7] Basel Committee on Banking Supervision, RBC20 — Calculation of minimum risk-based capital requirements , paragraph 20.1.

[8] Basel Committee on Banking Supervision, CRE31 — IRB approach: risk weight functions , and CRE36, paragraphs 36.83 onwards.