Credit Risk Measurement and Rating Models

Introduction and Context

Credit risk is the possibility that a borrower or counterparty will fail to meet agreed contractual obligations, causing a financial loss to the lender. It is the central risk of commercial banking because loans and other credit exposures usually constitute the largest portion of a bank’s assets. Effective credit risk measurement supports four linked decisions: whether to lend, how much to lend, at what price, and how much capital and loss-absorbing capacity to maintain.

Modern credit management combines professional judgement with quantitative models. A model does not replace the credit officer. Rather, it provides a consistent method for analysing information, ranking risk, estimating losses, monitoring deterioration and allocating capital. Models must therefore be interpreted within the borrower’s industry, economic environment, facility structure, collateral position and relationship history.

Learning Objectives

By the end of this chapter, a student should be able to:

1. Explain expected loss and distinguish probability of default, loss given default and exposure at default.
2. Calculate expected loss using borrower and facility information.
3. Describe application scoring, behavioural scoring and internal rating systems.
4. Explain the logic and limitations of Altman-type distress models.
5. Interpret credit-rating transition matrices.
6. Distinguish point-in-time and through-the-cycle rating approaches.
7. Explain model validation, governance, ethics and responsible use of automated credit decisions.

1. The Architecture of Credit Risk Measurement

Credit risk measurement has three levels. First, the bank assesses the borrower’s intrinsic repayment capacity, including cash flow, leverage, liquidity, management quality and competitive position. Second, it assesses the transaction: maturity, seniority, guarantees, collateral, covenants, repayment schedule and product type. Third, it aggregates individual exposures into portfolio risk, considering concentration, correlation, sector conditions and macroeconomic scenarios.

A useful distinction is between default risk and loss risk. Default is an event, such as failure to pay a material obligation within a specified period. Loss is the economic consequence after recoveries, collateral realization, workout expenses and timing effects. A borrower can have a relatively high probability of default but produce a modest loss if the exposure is strongly secured; another can have a lower probability of default but generate a large loss if it is unsecured and highly drawn.

Credit risk systems commonly estimate risk over a chosen horizon, often one year for rating and capital purposes. Retail lending may also use monthly or lifetime horizons. The horizon must be stated because a five-year default probability is not interchangeable with a one-year probability.

2. Expected Loss and Its Components

For a single exposure, a simplified expected-loss relationship is:

Here, EL is expected loss in currency units; PD is probability of default over the specified horizon; LGD is loss given default, expressed as a proportion of exposure; and EAD is exposure at default in currency units. The formula assumes that PD, LGD and EAD are appropriately defined for the same exposure, horizon and default event. Since PD and LGD are proportions, the result has the same currency unit as EAD.

Expected loss is an average forecast, not a prediction that the exact amount will occur. Actual losses may be zero or very large. A bank therefore needs both expected-loss provisions and capital or other protection against unexpected loss.

A portfolio version is:

This additive form is suitable for expected loss when each exposure’s expected loss is estimated consistently. It does not, by itself, measure portfolio volatility or unexpected loss because correlations between defaults are not represented.

Worked Example 1: Basic Expected Loss

A bank has a business loan with an EAD of $2,000,000. The one-year PD is 3%, and estimated LGD is 45%.

The estimated one-year expected loss is $27,000. This does not mean the bank will necessarily lose $27,000. It means that, over many exposures with comparable characteristics, the average loss implied by the assumptions is approximately that amount.

Probability of Default

Probability of default is the likelihood that a borrower will enter a defined default state during a stated period. The definition may be regulatory, contractual or internal. Common indicators include a material payment delinquency, insolvency, bankruptcy, distressed restructuring or a judgement that the borrower is unlikely to pay in full.

PD may be estimated from historical default rates, statistical models, market information, rating agency data or a combination of methods. A one-year observed default rate can be calculated as:

This simple rate requires careful treatment of new accounts, prepayments, censoring, mergers and accounts that leave the portfolio before the observation date. For longer horizons, survival analysis may estimate the probability of default while an account remains active.

PD estimates can be:

• Point-in-time (PIT): responsive to current economic and borrower conditions.
• Through-the-cycle (TTC): designed to represent average risk across economic cycles.
• Marginal: probability of default during a particular period, conditional on survival to that period.
• Cumulative: probability of default by a specified future date.

For example, a cumulative five-year PD is not obtained by simply multiplying a one-year PD by five unless highly restrictive assumptions hold. Default probabilities generally change over time and surviving borrowers become a selected population.

Loss Given Default

Loss given default is the proportion of EAD not recovered after default. A simplified economic-loss formula is:

Recoveries can include cash collections, collateral proceeds, guarantees and sale proceeds. PV means present value, which discounts future amounts to the default date. Workout costs include legal, servicing, valuation and enforcement expenses. The formula assumes that recovery cash flows and costs are measured consistently and discounted using an appropriate rate.

If recoveries are $700,000 on an EAD of $1,000,000 and workout costs are $50,000, all measured at the default date:

LGD is affected by collateral type and value, loan seniority, jurisdiction, insolvency law, guarantees, time to recovery and economic conditions. In a recession, collateral prices may fall and recoveries may take longer. This is known as downturn LGD or recovery cyclicality.

Exposure at Default

Exposure at default is the amount owed or expected to be owed when default occurs. For a fully drawn term loan, EAD may be close to the outstanding principal plus eligible accrued amounts. For revolving facilities, EAD must consider future drawings before default.

A simplified revolving-credit estimate is:

CCF is the credit conversion factor, representing the expected proportion of the undrawn commitment that will be drawn before default. If a borrower has drawn $400,000, an undrawn commitment of $300,000 and a CCF of 60%:

Actual EAD models may depend on utilization, limits, borrower behaviour, product type, cancellation rights and the time between observation and default.

3. Credit Scoring Models

Credit scoring assigns a numerical score to an applicant or account using measurable characteristics. The score is then mapped to an estimated PD or decision category. Scoring supports consistency, speed and evidence-based decision-making, especially in high-volume retail and small-business portfolios.

Application Scoring

Application scoring is performed when credit is requested. Variables may include income, debt-service burden, employment stability, repayment history, age of the account relationship, loan-to-value ratio and requested amount. The model should use information available at or before the decision point; using later information creates data leakage and overstated performance.

A common statistical approach is logistic regression. It estimates the probability of an event, such as default, using predictor variables:

Here, p is estimated PD; β₀ is the intercept; β_j are coefficients; x_j are borrower or facility variables; and exp is the exponential function. The output is between 0 and 1. Logistic regression assumes that the log-odds of default are linearly related to the predictors unless transformations or interactions are included.

The log-odds relationship is:

A positive coefficient increases default odds as the variable increases, holding other variables constant. Interpretation requires attention to variable coding. For example, a positive coefficient for debt-to-income ratio would generally be economically plausible, while a coefficient’s sign alone does not establish causation.

Other approaches include decision trees, random forests, gradient boosting and neural networks. These may improve predictive power but can be less transparent, more sensitive to data drift and harder to explain to customers, auditors and supervisors. Model selection should therefore consider discrimination, calibration, stability, explainability and operational constraints—not accuracy alone.

Behavioural Scoring

Behavioural scoring evaluates existing customers using observed account behaviour. Variables may include payment punctuality, utilization, overdraft frequency, deposit patterns, transaction volatility, cash-flow coverage and recent delinquency. It supports credit-limit management, early-warning systems, collections prioritization and retention decisions.

Behavioural scores must avoid unfairly treating temporary financial stress as permanent inability to pay. They should also distinguish a fall in deposits caused by seasonal business activity from a structural deterioration in repayment capacity.

4. Internal Rating Systems

An internal rating system assigns borrowers to ordered grades, such as Grade 1 through Grade 10, or labels such as AAA through CCC. Each grade should correspond to a defined risk meaning and, where possible, a calibrated PD range.

A robust rating process normally separates:

1. Obligor rating: the borrower’s general creditworthiness.
2. Facility rating: the risk of a particular loan after considering seniority, collateral and guarantees.
3. Rating assignment: quantitative analysis combined with qualitative judgement.
4. Rating approval: independent review and appropriate authority limits.
5. Monitoring: review triggered by time, risk events or material changes.

For corporate borrowers, analysis may include audited financial statements, leverage, profitability, liquidity, cash-flow adequacy, industry structure, management, ownership, governance and stress performance. A rating override may be permitted where documented evidence is not captured by the model, but overrides should be tracked and tested. Frequent overrides in one direction may indicate model weakness or poor underwriting discipline.

A rating scale should be monotonic: worse grades should not have lower observed default rates than better grades over a sufficiently large and relevant sample. It should also be stable enough to support pricing and limit decisions, while responsive enough to identify deterioration.

PIT and TTC Ratings

PIT ratings react strongly to current conditions and may migrate frequently. They are useful for provisioning, early warning and current credit decisions. TTC ratings seek to reduce cyclical migration and may be useful for longer-term portfolio comparison and capital planning. Neither approach is universally superior. The appropriate approach depends on purpose, horizon, regulation and the institution’s risk appetite.

5. Altman-Type Distress Models

Altman-type models use financial ratios to distinguish financially distressed firms from healthier firms. The classic Z-score model was developed for a particular historical sample of manufacturing companies and should not be treated as a universal probability-of-default model.

A commonly cited original form is:

where:

The ratios are dimensionless because numerator and denominator are measured in compatible currency units. Higher values generally indicate stronger liquidity, accumulated profitability, operating performance, solvency and asset utilization. Historical interpretation often associates a high score with a lower distress risk and a low score with a higher distress risk, but cut-offs depend on the model version, industry and sample.

Variants have been developed for private firms and non-manufacturing firms. Analysts must use the appropriate version and not substitute book equity for market equity without recognizing that the interpretation changes. The model is best used as a screening or supplementary tool alongside cash-flow analysis, current information and qualitative assessment.

Limitations include historical data, accounting-policy differences, survivorship and sample-selection bias, sector mismatch, inflation, changing business models and the fact that distress classification is not identical to legal default. A high score cannot guarantee repayment.

6. Transition Matrices

A credit-rating transition matrix shows the proportion of obligors that move from one rating grade to another over a specified period. Rows represent the initial grade and columns represent the grade at the end of the period. Each row should sum to 100%, allowing for rounding, and may include a default or withdrawal state.

For an A-rated portfolio of 10,000 borrowers, the matrix implies approximately 100 defaults over the period under the stated conditions. It does not predict exactly which borrowers will default.

Transition matrices support rating migration analysis, stress testing, pricing, provisioning and portfolio modelling. A one-year matrix can be multiplied by itself to produce a multi-period matrix only under assumptions such as a consistent transition process and compatible treatment of default and withdrawals. In practice, transition probabilities may depend on the economic cycle, sector, geography, borrower type and rating methodology.

Matrices may be cohort-based, following accounts present at the beginning of a period, or duration-based, estimating transitions while accounting for time spent in each state. Analysts must define treatment of withdrawn, merged, prepaid and defaulted accounts. Otherwise, results may be biased.

7. Model Validation and Performance Assessment

Model validation is an independent process that determines whether a model is conceptually sound, performs adequately and remains appropriate for its use. It is not limited to checking whether a model produces a score.

Discrimination

Discrimination is the ability to rank higher-risk accounts above lower-risk accounts. Common measures include the area under the receiver operating characteristic curve (AUC) and the Gini coefficient:

The Gini coefficient is dimensionless. A value of zero indicates no ranking power, while higher values indicate stronger discrimination. AUC and Gini measure ranking, not whether predicted probabilities are numerically accurate.

Calibration

Calibration compares predicted default probabilities with observed default frequencies. If a group of accounts is assigned a PD of 4%, approximately 4% should default over the relevant horizon, subject to sampling variation and stable definitions. Calibration should be tested overall and by grade, segment, product, geography and time period.

A model may discriminate well but be poorly calibrated. For example, it may correctly rank borrowers but assign PDs that are systematically too low. Calibration can be affected by economic changes, portfolio growth, underwriting changes, default-definition changes and data quality.

Stability and Back-Testing

Validation should include out-of-time and, where possible, out-of-sample testing. Back-testing compares forecasts with subsequent outcomes. Population stability measures whether the distribution of model inputs or scores has changed. Material drift may require recalibration, redevelopment or a temporary overlay.

Validation should also examine missing data, coding errors, rejected applications, class imbalance, data leakage, extreme values, override rates and performance across protected or vulnerable groups. Stress testing asks how PD, LGD and EAD might change under adverse unemployment, interest-rate, exchange-rate, commodity-price or property-price scenarios.

Governance

A sound governance framework includes documented purpose and scope, data lineage, independent validation, approval authority, version control, change management, monitoring thresholds, issue escalation, inventory of models and periodic review. Model risk is the risk of making poor decisions because a model is incorrectly designed, implemented, interpreted or used outside its purpose.

8. Applied Case: Northstar Equipment Finance

The following is a fictional illustrative case.

Background

Northstar Bank is considering a $5 million revolving facility for a medium-sized equipment distributor. The current drawn amount is $3.2 million, and the undrawn commitment is $1.8 million. The borrower’s internal rating has moved from Grade B to Grade C after declining margins and slower receivables collection.

Problem

The credit committee must estimate one-year expected loss and decide whether to renew the facility. A behavioural model indicates increased utilization and two recent late payments. The borrower provides inventory as collateral, but market values have weakened.

Analysis

The bank estimates a one-year PD of 8%, a CCF of 50%, and LGD of 55% after considering collateral haircuts and workout costs.

The rating migration matrix also indicates that Grade C borrowers have a material probability of moving to default or a lower grade. However, the committee does not rely on the matrix alone. It reviews aged receivables, inventory turnover, supplier terms, cash-flow forecasts, management plans and covenant headroom.

Approach and Solution

The bank renews only $4 million, introduces a borrowing-base limit linked to eligible inventory and receivables, requires monthly reporting, and sets an early-warning trigger for further utilization or overdue receivables. It also reassesses LGD because collateral values may fall during an industry downturn. The decision is documented, independently reviewed and scheduled for an earlier rating review.

Outcome and Learning Points

The case demonstrates that PD, LGD and EAD are connected but distinct. Reducing the commitment may reduce EAD; stronger collateral controls may reduce LGD; tighter monitoring may identify deterioration earlier and improve the decision process. The expected loss is an input to pricing and provisioning, not the sole determinant of approval. A responsible decision combines model evidence, current information, collateral analysis and professional judgement.

9. Ethical, Regulatory and Technological Considerations

Credit models influence access to housing, education, employment-related finance and business capital. Ethical concerns arise when models use discriminatory proxies, poor-quality data or opaque automated rules. Variables such as location, occupation, device characteristics or transaction patterns may unintentionally correlate with protected characteristics. Institutions should assess disparate outcomes, provide legally appropriate explanations and maintain human review for contested or exceptional decisions.

Regulatory expectations commonly emphasize sound definitions of default, conservative estimation, data quality, independent validation, stress testing, documentation and sufficient capital or provisions. Exact requirements vary by jurisdiction, so institutions must follow applicable banking, consumer-protection, privacy and anti-discrimination law.

Artificial intelligence can improve pattern recognition, cash-flow analysis and early-warning systems, but it introduces risks involving explainability, data drift, privacy, cyberattack, vendor dependence and feedback loops. A model trained on historical approvals may reproduce past exclusion. Good practice includes purpose limitation, representative data, explainability appropriate to the decision, human oversight, monitoring, challenger models and clear accountability.

10. Best Practices for Credit Risk Measurement

Effective practice includes:

• Define default, horizon, exposure and recovery consistently.
• Match the model to the decision and borrower segment.
• Separate development, approval, use and validation responsibilities.
• Combine quantitative scores with documented qualitative analysis.
• Monitor discrimination, calibration, stability and fairness.
• Re-estimate assumptions when portfolio or economic conditions change.
• Treat collateral conservatively and include realization costs and time.
• Track overrides and investigate systematic patterns.
• Use stress testing rather than relying only on average conditions.
• Preserve data lineage, audit trails and version-controlled documentation.
• Communicate model limitations to credit officers and decision-makers.

Chapter Summary

Credit risk measurement converts borrower and facility information into structured estimates that support lending decisions, pricing, monitoring and capital planning. Expected loss is commonly expressed as EL = PD × LGD × EAD, where PD measures the likelihood of default, LGD measures the proportion lost after default, and EAD measures the amount exposed when default occurs. The components must share a consistent horizon and default definition.

Application scoring evaluates new borrowers, while behavioural scoring evaluates existing-account conduct. Internal rating systems classify obligors and facilities into ordered risk grades and may use point-in-time or through-the-cycle perspectives. Altman-type models provide useful financial-distress screening but have important historical and sector limitations. Transition matrices describe migration between rating states and support portfolio analysis.

A credible model must be validated for discrimination, calibration, stability, data quality and appropriate use. Governance, ethical controls, regulatory compliance and human judgement are essential, especially when artificial intelligence is used. Models improve consistency, but they remain estimates subject to uncertainty, changing conditions and misuse.

Key Terms

Credit risk: The possibility of loss because a counterparty fails to meet obligations.

Default: A defined failure or expected failure to meet contractual repayment obligations.

Expected loss: Average loss estimated for a specified exposure and horizon.

Probability of default (PD): Likelihood of default during a specified period.

Loss given default (LGD): Proportion of exposure lost after recoveries and costs.

Exposure at default (EAD): Amount expected to be outstanding when default occurs.

Credit conversion factor (CCF): Expected conversion of an undrawn commitment into exposure.

Application scoring: Risk scoring performed when a credit application is made.

Behavioural scoring: Risk scoring based on observed account behaviour.

Internal rating: A bank-assigned risk grade for an obligor or facility.

Point-in-time rating: Rating responsive to current conditions.

Through-the-cycle rating: Rating designed to represent average risk across cycles.

Altman Z-score: A ratio-based financial-distress screening model.

Transition matrix: Table showing movements between rating states.

Discrimination: A model’s ability to rank higher-risk cases above lower-risk cases.

Calibration: Agreement between predicted probabilities and observed event rates.

Back-testing: Comparison of forecasts with later actual outcomes.

Model risk: Risk arising from incorrect model design, implementation or use.

Override: A documented departure from a model-generated rating or decision.

Stress testing: Analysis of performance under adverse hypothetical conditions.

Review Questions

Short Answer Questions

1. Define PD, LGD and EAD.
2. Why is expected loss not the same as an exact forecast of an individual loan’s loss?
3. Distinguish application scoring from behavioural scoring.
4. What information does a transition matrix provide?
5. Why must a credit model be calibrated as well as discriminatory?

Long Answer Questions

1. Explain the expected-loss framework and discuss the main factors affecting PD, LGD and EAD.
2. Compare point-in-time and through-the-cycle internal rating systems.
3. Explain the construction, interpretation and limitations of Altman-type distress models.
4. Discuss the role of model validation in credit risk governance.
5. Evaluate the benefits and risks of using artificial intelligence in credit scoring.

Analytical / Application Questions

1. A bank has a drawn balance of $1.5 million, an undrawn commitment of $500,000, a CCF of 40%, a PD of 5% and an LGD of 50%. Calculate EAD and expected loss.
2. A model has strong AUC but systematically predicts PDs below observed default rates. What model property is deficient, and what corrective actions should the bank consider?
3. A borrower’s rating deteriorates, utilization rises and collateral values fall. Analyse how these developments may affect PD, EAD and LGD, and recommend appropriate credit actions.

Multiple Choice Questions with Answers

Question 1

A loan has a PD of 4%, an LGD of 50% and an EAD of $1,000,000. What is its expected loss?

A. $2,000

B. $20,000

C. $200,000

D. $500,000

Correct Answer: B. $20,000

Question 2

Which component measures the amount expected to be outstanding when a borrower defaults?

A. PD

B. LGD

C. EAD

D. AUC

Correct Answer: C. EAD

Question 3

A facility has a drawn balance of $400,000, an undrawn commitment of $300,000 and a CCF of 60%. What is EAD?

A. $420,000

B. $580,000

C. $600,000

D. $700,000

Correct Answer: B. $580,000

Question 4

Which statement best describes behavioural scoring?

A. It assesses only the borrower’s original application form.

B. It uses observed account conduct to assess existing customers.

C. It measures collateral prices only.

D. It replaces all internal ratings.

Correct Answer: B. It uses observed account conduct to assess existing customers.

Question 5

What does a transition matrix primarily show?

A. Interest-rate movements on loans

B. Movements between credit-rating states over time

C. The legal ownership of collateral

D. The bank’s income statement

Correct Answer: B. Movements between credit-rating states over time

Question 6

The formula Gini = 2 × AUC − 1 is used to assess what property?

A. Model discrimination

B. Collateral recovery cost

C. Exposure maturity

D. Loan amortization

Correct Answer: A. Model discrimination

Question 7

A model assigns accurate rankings but predicted PDs are consistently lower than actual default rates. Which property is weakest?

A. Calibration

B. Collateralization

C. Seniority

D. Utilization

Correct Answer: A. Calibration

Question 8

Which is a major limitation of the original Altman Z-score model?

A. It cannot use financial ratios.

B. It was developed from a specific historical and sectoral sample.

C. It directly measures future interest rates.

D. It requires no accounting information.

Correct Answer: B. It was developed from a specific historical and sectoral sample.

Question 9

If recoveries are $700,000, workout costs are $50,000 and EAD is $1,000,000, what is LGD using the chapter formula?

A. 15%

B. 30%

C. 35%

D. 75%

Correct Answer: C. 35%

Question 10

Which practice is most appropriate for responsible credit-model governance?

A. Allow unlimited undocumented overrides.

B. Use model outputs without validation.

C. Maintain independent validation, monitoring and documented limitations.

D. Exclude all human review from automated decisions.

Correct Answer: C. Maintain independent validation, monitoring and documented limitations.