Credit Risk and Banking Analytics

Introduction and Context

Credit risk is the possibility that a borrower or counterparty will fail to meet contractual obligations, causing financial loss to a bank. It is the largest traditional risk for many banking institutions because lending produces income but also exposes capital to default, recovery uncertainty, fraud, sectoral shocks and macroeconomic deterioration.

Banking analytics converts borrower, account, market and economic data into decisions about approval, pricing, monitoring, provisioning, capital and portfolio strategy. Effective credit-risk management therefore combines financial analysis, statistical modelling, expert judgement, governance and continuous monitoring.

This chapter develops an integrated framework based on four central quantities: probability of default (PD), loss given default (LGD), exposure at default (EAD) and expected loss (EL). It then applies these measures to credit scoring, loan pricing, provisioning, concentration analysis, stress testing and early-warning systems.

Learning Objectives

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

1. Explain the nature and sources of credit risk.
2. Distinguish credit scoring, rating and portfolio analytics.
3. Calculate and interpret PD, LGD, EAD and expected loss.
4. Evaluate borrower quality and loan-pricing decisions.
5. Analyse provisioning, concentration and stress-testing results.
6. Design useful early-warning indicators.
7. Assess model, ethical, regulatory and governance risks.

1. The Credit-Risk Decision Framework

A lending decision is not simply a prediction of whether a borrower will default. It is a sequence of decisions:

• Origination: Should the bank lend?
• Structuring: What amount, maturity, collateral and covenants are appropriate?
• Pricing: What interest rate compensates for expected loss, funding cost, capital and operating cost?
• Monitoring: Is the borrower’s risk changing?
• Remediation: Should the bank restructure, collect, sell or write off the exposure?
• Portfolio management: Is risk concentrated by sector, geography, product, group or borrower type?

Credit risk has several sources. Default risk concerns failure to pay. Migration risk concerns deterioration in credit quality before default. Recovery risk concerns uncertainty about collateral value, legal enforcement and timing. Counterparty risk is especially important in derivatives, securities financing and interbank transactions. Settlement risk arises when one party performs but the other does not.

A sound process separates the first-line business function, which originates loans, from independent risk oversight and internal audit. This separation reduces conflicts of interest and supports challenge of optimistic assumptions.

2. Core Credit-Risk Measures

Probability of Default

PD is the probability that an obligor will default over a specified horizon, commonly one year. Default must be defined consistently, for example by a material payment delinquency, insolvency, bankruptcy or distressed restructuring.

A one-year PD may be estimated as:

This historical estimate assumes that the observation population and economic conditions are sufficiently comparable. For a borrower-level model, PD is usually a forecast conditional on borrower characteristics and macroeconomic variables.

PD is not the same as delinquency rate. Delinquency may be temporary, whereas default is a formally defined credit event. PD is also distinct from lifetime default risk, which considers several future periods.

Loss Given Default

LGD is the proportion of exposure lost if default occurs, after considering recoveries and recovery costs. A basic form is:

where EAD is exposure at default and PV(Recoveries) is the present value of cash recoveries and collateral proceeds, net of collection, legal and liquidation costs. The discount rate should reflect the time value of money and the risk characteristics of recoveries, according to the applicable accounting or regulatory framework.

If recoveries equal or exceed exposure, economic loss may be zero or negative before other adjustments; however, operational definitions generally constrain LGD to an economically meaningful range. LGD depends on collateral quality, seniority, guarantees, jurisdiction, loan type, default timing and economic conditions.

Exposure at Default

EAD is the amount exposed to loss at the time of default. For a fully drawn term loan, EAD may be close to principal outstanding plus relevant accrued amounts. For a revolving credit facility, future drawings must be estimated.

CCF is the credit conversion factor, representing the expected proportion of an undrawn commitment drawn before default. This formula assumes that the undrawn portion is measured consistently and that a single CCF is suitable for the facility segment.

Expected Loss

For a single exposure, the commonly used expected-loss relationship is:

PD is dimensionless, LGD is a proportion, and EAD is measured in currency; therefore EL is measured in currency. The formula assumes that PD, LGD and EAD are appropriately aligned to the same horizon and exposure and that their conditional dependence is either modelled or acceptably approximated.

For a portfolio:

This is an average-loss measure. Actual losses can be much higher because defaults are correlated, especially during recessions. Unexpected loss, rather than expected loss, is associated more closely with economic capital and capital adequacy.

Worked Example 1: Expected Loss

A bank has a business loan with an outstanding balance of $800,000 and an undrawn commitment of $200,000. The estimated one-year PD is 3%, LGD is 45%, and CCF is 50%.

First calculate EAD:

Then calculate expected loss:

The result is an estimated average one-year credit loss of $12,150, not a prediction that the bank will lose exactly that amount. If the loan is priced or provisioned using this estimate, the bank must also consider operating costs, funding, capital, tax, model uncertainty and profit requirements.

3. Credit Scoring and Rating Models

Credit Scoring

Credit scoring assigns a numerical score associated with the likelihood of default. Retail models often use application data, repayment history, income, employment, indebtedness, account behaviour and bureau information. A scorecard commonly groups variables into bands and assigns points to each band.

A logistic model can estimate default probability:

where p is estimated default probability, e is the base of natural logarithms, and z = β₀ + β₁x₁ + β₂x₂ + ⋯ + β_kx_k. β₀ is the intercept, β_j are estimated coefficients and x_j are explanatory variables.

The logistic function ensures that predicted probabilities lie between zero and one. Coefficients describe the effect of variables on log-odds, not directly on probability. A positive coefficient increases default odds when other variables are held constant, but the probability effect depends on the starting risk level.

Rating Models

Corporate and institutional borrowers are often assessed through an obligor rating, such as an internal grade or external rating category. Rating systems may combine quantitative ratios with qualitative factors, including management quality, industry position, governance, business model and access to funding.

Scoring is usually highly automated and suited to high-volume homogeneous portfolios. Rating is often more judgement-intensive and suited to larger, complex or less frequent borrowers. Both require clear definitions, sufficient data, validation and periodic recalibration.

Model Development and Validation

A responsible model-development process includes:

1. Define default and the observation horizon.
2. Assemble representative, accurate and legally permissible data.
3. Select variables using economic reasoning and statistical evidence.
4. Separate development, validation and out-of-time samples.
5. Test discrimination: whether higher-risk borrowers receive higher predicted risk.
6. Test calibration: whether predicted probabilities correspond to observed default frequencies.
7. Assess stability across time, products and demographic groups.
8. Document limitations and establish monitoring thresholds.

Common discrimination measures include the area under the receiver operating characteristic curve (AUC) and the Gini coefficient:

The relationship applies when AUC is expressed between 0 and 1. A high AUC does not prove that probabilities are well calibrated, nor does it establish causal relationships.

4. Pricing Credit Risk

A loan interest rate should compensate the bank for more than expected default loss. A simplified annual pricing identity is:

This is a managerial approximation, not a universal accounting identity. Rates also depend on maturity, liquidity, competition, optionality, collateral, taxes and interest-rate risk.

If expected loss is expressed as a percentage of EAD, then:

For the preceding example, the expected-loss rate is 0.03 × 0.45 = 1.35% of EAD. A bank charging less than this amount may still lend profitably if other benefits exist, but it should understand that the stated rate does not fully compensate for expected credit loss.

Risk-based pricing must not become disguised discrimination. Variables should be relevant to repayment risk, legally permitted and subject to fairness testing. Customers should receive explanations appropriate to applicable consumer-protection rules.

5. Provisioning and Credit-Loss Measurement

Provisioning recognises expected credit deterioration in financial statements. Many accounting frameworks use an expected-credit-loss approach. A simplified lifetime model is:

where t identifies a future period, PDₜ is the probability of default in that period under the chosen convention, LGDₜ is loss severity, EADₜ is exposure, and DFₜ is the discount factor. If the annual effective interest rate is r, a simple discount factor is DFₜ = 1 ÷ (1 + r)^t.

Care is required: PD terms must be defined as marginal period default probabilities or as conditional probabilities converted consistently. Simply adding cumulative PDs can double count default. Projections should incorporate reasonable and supportable forward-looking information, often through probability-weighted economic scenarios.

Provisioning and regulatory capital are related but not identical. Provisions address expected accounting losses; capital frameworks also address unexpected losses, solvency and risk-absorption capacity. Management should reconcile, rather than confuse, these purposes.

6. Portfolio Analytics

Concentration Risk

Diversification reduces risk only when exposures are not strongly correlated. Concentration can arise from:

• a single borrower or connected group;
• industry or commodity;
• geography or sovereign;
• collateral type;
• product, maturity or funding source;
• common economic drivers.

A simple exposure share is:

This measure describes size but not correlation or loss severity. A small segment with very high correlation may be more dangerous than a larger diversified segment. Banks therefore supplement exposure shares with sector stress tests, portfolio models, large-exposure limits and correlated-obligor analysis.

Portfolio Migration

A transition matrix reports the probability that an exposure moves from one rating grade to another over a period. For each starting grade, transition probabilities should sum to 1, allowing for all destination grades, including default. Migration analysis identifies deterioration before default and supports capital, pricing and watch-list decisions.

Risk-Adjusted Performance

A basic risk-adjusted return measure compares income after expected loss with capital employed:

Definitions must be consistent across business units. RAROC is useful for comparing transactions, but it depends on model assumptions and should not replace strategic, conduct or liquidity analysis.

7. Stress Testing

Stress testing asks how credit losses and capital would change under adverse but plausible conditions. A scenario may include recession, unemployment, interest-rate increases, currency depreciation, property-price falls or commodity-price shocks.

A practical stress-testing sequence is:

1. Define a coherent macroeconomic scenario.
2. Translate macroeconomic variables into PD, LGD and EAD changes.
3. Recalculate defaults, recoveries, provisions, income and capital.
4. Identify vulnerable sectors, products and borrowers.
5. Consider management actions and their timing.
6. Compare results with risk appetite, limits and regulatory requirements.

A severe scenario is not necessarily a forecast. It is a decision tool. Stress tests should include sensitivity analysis, because a single model may understate uncertainty. For example, falling property prices may simultaneously increase PD, reduce collateral recoveries and increase utilisation of credit lines.

8. Early-Warning Indicators

Early-warning indicators (EWIs) identify deterioration before formal default. They may be quantitative or qualitative.

Indicators should be timely, reliable and linked to action. A warning without an assigned owner, investigation deadline and escalation rule has little control value. Thresholds should be calibrated to avoid both excessive false alarms and delayed intervention.

9. Illustrative Case: Meridian Bank’s Commercial Property Portfolio

Background

Meridian Bank has a $500 million commercial-property portfolio concentrated in office buildings. The portfolio was originated when vacancy rates were low and property values were rising. Interest rates then increased, refinancing became more expensive and office demand weakened.

Problem

The average reported PD remains modest because many borrowers are current on payments. However, covenant breaches, falling valuations and maturity walls indicate that risk may be understated. Senior management must decide whether to increase provisions, restrict new lending and engage borrowers early.

Analysis

The risk team segments loans by location, building quality, loan-to-value ratio (LTV), maturity and tenant concentration. It observes that a loan with an outstanding balance of $10 million has an estimated stressed EAD of $10.5 million, stressed PD of 8% and stressed LGD of 50%.

The team also identifies that several borrowers mature within twelve months. A refinancing shock could raise PD beyond the current model estimate. Falling collateral values increase LGD, while unused credit facilities may be drawn, increasing EAD.

Approach and Solution

Meridian introduces a watch list based on vacancy, debt-service coverage, covenant breaches, valuation decline and refinancing difficulty. It commissions updated valuations, reviews borrower cash-flow forecasts and contacts high-risk borrowers before maturity. New exposure limits are imposed on highly concentrated sub-sectors. Provision estimates are updated under baseline, adverse and severe scenarios.

Outcome

The bank does not assume that current payment performance proves low risk. Instead, it uses forward-looking information and separates performing status from credit quality. Some facilities are restructured with stronger covenants, while others receive higher provisions and tighter collateral requirements. The portfolio is gradually diversified, and management receives monthly EWI reporting.

Learning Points

The case illustrates that PD, LGD and EAD interact dynamically. Concentration, maturity and collateral correlations can create losses even when arrears are initially low. Early engagement is usually more effective than waiting for default.

10. Governance, Ethics and Technology

Credit analytics requires data governance: clear ownership, lineage, access controls, quality checks, retention policies and documented transformations. Personal data must be collected and used lawfully. Sensitive attributes and proxy variables require particular care.

Artificial intelligence and machine learning can identify nonlinear patterns in transaction, cash-flow and alternative data. However, complex models may be difficult to explain, unstable under economic change or biased by historical lending decisions. Banks should compare predictive gains with explainability, fairness, cybersecurity and monitoring costs. Human oversight remains essential for exceptions, hardship cases and unusual borrowers.

Model risk arises from incorrect specifications, poor data, implementation errors, changing relationships and misuse outside the development population. Controls include independent validation, challenger models, back-testing, override governance, stress testing, performance monitoring and a formal model inventory.

Good practice includes:

• defining default consistently across systems;
• aligning PD, LGD, EAD and horizon assumptions;
• distinguishing expected loss from unexpected loss;
• using conservative and evidence-based collateral values;
• monitoring drift and population changes;
• documenting overrides and adverse decisions;
• testing fairness and explainability;
• integrating analytics with accountable human decisions.

Challenges and Limitations

Credit models face several limitations. Defaults are relatively rare, producing class imbalance and estimation uncertainty. Historical data may not represent future recessions. Macroeconomic relationships can change after policy, technology or consumer-behaviour shifts. Recoveries are uncertain and correlated with economic conditions. Data may be incomplete, delayed or biased toward existing customers.

A model can rank borrowers accurately while producing poorly calibrated probabilities. Conversely, a calibrated model may not discriminate well between individual borrowers. Therefore validation must examine discrimination, calibration, stability, sensitivity, operational implementation and business usefulness.

Analytics should support—not replace—sound credit judgement. A bank that follows a score mechanically may miss fraud, structural change or a borrower’s credible recovery plan. A bank that overrides models excessively may destroy their consistency and predictive value.

Chapter Summary

Credit risk is the possibility of loss caused by failure or deterioration of a borrower or counterparty. Banking analytics supports origination, pricing, monitoring, provisioning, capital and portfolio decisions. The core measures are PD, LGD and EAD, whose product estimates expected loss: EL = PD × LGD × EAD.

Credit scoring commonly uses borrower and behavioural data, while rating systems combine quantitative and qualitative assessment. Logistic models, discrimination testing and calibration analysis are useful, but models require representative data, validation and governance. Pricing should account for funding, expected loss, operating costs, capital and profit. Provisioning uses forward-looking expected-credit-loss estimates, often across multiple periods and scenarios.

Portfolio analysis must address concentration, migration and correlation rather than relying only on average risk. Stress testing explores adverse but plausible conditions, while early-warning indicators support intervention before default. Ethical use requires lawful data practices, fairness, explainability and accountable human oversight. The most effective credit-risk function integrates quantitative models with expert judgement, strong controls and continuous monitoring.

Key Terms

PD: Probability that an obligor defaults over a specified horizon.

LGD: Proportion of exposure lost after recoveries and recovery costs following default.

EAD: Exposure outstanding when default occurs.

Expected loss: Probability-weighted average credit loss.

Unexpected loss: Loss above expected loss arising from variability and adverse outcomes.

Credit scoring: Statistical or rule-based assignment of risk scores to borrowers.

Rating: Credit-quality classification assigned to an obligor or facility.

Default: A formally defined credit event indicating failure or likely failure to pay.

CCF: Credit conversion factor applied to an undrawn commitment.

Collateral: Asset or security supporting repayment and potential recovery.

Calibration: Agreement between predicted risk probabilities and observed outcomes.

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

AUC: Area under the receiver operating characteristic curve.

Provision: Accounting recognition of anticipated credit loss.

Concentration risk: Loss risk arising from large or correlated exposures.

Stress testing: Analysis of portfolio performance under adverse scenarios.

Early-warning indicator: Signal of possible credit deterioration before default.

RAROC: Risk-adjusted return relative to economic capital.

Model risk: Risk of adverse outcomes from incorrect or inappropriate model use.

Review Questions

5 Short Answer Questions

1. What is the difference between PD and LGD?
2. Why can EAD exceed the current drawn balance?
3. State the expected-loss formula and define its components.
4. What is the purpose of model calibration?
5. Give four examples of early-warning indicators.

5 Long Answer Questions

1. Explain how credit scoring and corporate rating systems differ in data, method and application.
2. Discuss the role of PD, LGD and EAD in loan pricing and provisioning.
3. Explain why concentration and correlation matter in portfolio credit risk.
4. Describe a complete credit-model development and validation process.
5. Evaluate the benefits and limitations of artificial intelligence in credit-risk management.

3 Analytical / Application Questions

1. A facility has a $4 million drawn balance, a $1 million undrawn commitment, a 40% CCF, PD of 5% and LGD of 35%. Calculate EAD and one-year expected loss.
2. A bank’s office-property portfolio has low arrears but falling valuations and many loans maturing within one year. Design an early-warning and stress-testing programme.
3. A model has a high AUC but systematically underestimates defaults during recessions. Analyse the likely problem and recommend governance actions.

Multiple Choice Questions with Answers

Question 1

Which formula represents expected loss for a single credit exposure?

A. PD + LGD + EAD

B. PD × LGD × EAD

C. PD ÷ LGD × EAD

D. EAD − PD − LGD

Correct Answer: B. PD × LGD × EAD

Question 2

A loan has a drawn balance of $800,000, an undrawn commitment of $200,000 and a CCF of 50%. What is EAD?

A. $800,000

B. $850,000

C. $900,000

D. $1,000,000

Correct Answer: C. $900,000

Question 3

If PD is 4% and LGD is 40%, what is the expected-loss rate as a percentage of EAD?

A. 0.16%

B. 1.6%

C. 4.4%

D. 10%

Correct Answer: B. 1.6%

Question 4

What does calibration primarily assess?

A. Whether predicted probabilities correspond to observed default frequencies

B. Whether a bank has sufficient liquidity

C. Whether collateral is legally enforceable

D. Whether interest rates exceed inflation

Correct Answer: A. Whether predicted probabilities correspond to observed default frequencies

Question 5

Which measure is most directly associated with a model’s ability to rank risky borrowers above safer borrowers?

A. AUC

B. CCF

C. LTV

D. RAROC

Correct Answer: A. AUC

Question 6

Why may a revolving facility have an EAD greater than its current drawn balance?

A. The borrower may draw part of the undrawn commitment before default

B. LGD is always added to the balance

C. PD automatically increases the principal

D. Collateral is recorded as a liability

Correct Answer: A. The borrower may draw part of the undrawn commitment before default

Question 7

Which situation is an example of concentration risk?

A. A portfolio contains many small, unrelated exposures

B. A large share of lending is to one highly correlated industry

C. A borrower pays before the due date

D. A model includes an intercept

Correct Answer: B. A large share of lending is to one highly correlated industry

Question 8

What is the primary purpose of stress testing?

A. To guarantee the exact future loss

B. To examine portfolio performance under adverse but plausible conditions

C. To eliminate all model assumptions

D. To replace financial statements

Correct Answer: B. To examine portfolio performance under adverse but plausible conditions

Question 9

Which is an early-warning indicator of possible borrower deterioration?

A. Falling deposits and repeated returned payments

B. A stable repayment history with improving cash flow

C. A lower data-storage cost

D. A completed internal audit with no findings

Correct Answer: A. Falling deposits and repeated returned payments

Question 10

Why should expected loss and unexpected loss not be treated as identical?

A. Expected loss is an average estimate, whereas unexpected loss reflects adverse variability above that estimate

B. Expected loss is always zero, whereas unexpected loss is always known

C. Expected loss applies only to collateral, whereas unexpected loss applies only to deposits

D. They are identical under every credit framework

Correct Answer: A. Expected loss is an average estimate, whereas unexpected loss reflects adverse variability above that estimate