Credit Risk Analytics

Introduction and Context

Credit risk is the possibility that a borrower or counterparty will fail to meet contractual obligations, causing a financial loss to the lender. It is the dominant risk for many banks, consumer-finance firms, insurers, fintech lenders, leasing companies, and corporate treasury functions. Credit risk analytics converts borrower, facility, collateral, behavioral, and macroeconomic information into estimates that support underwriting, pricing, provisioning, capital allocation, collections, portfolio management, and stress testing.

A modern credit-risk framework must answer four related questions:

1. Will the obligor default? This is measured primarily by probability of default (PD).
2. How much will be lost if default occurs? This is measured by loss given default (LGD).
3. How much will be outstanding at default? This is measured by exposure at default (EAD).
4. What loss should be expected over a specified horizon? This is summarized by expected loss (EL).

The central identity is:

Here, PD is a probability between 0 and 1, LGD is the proportion of exposure lost after recoveries, and EAD is a monetary amount. The product therefore has monetary units. The identity is exact when the three quantities are defined consistently for the same obligor, facility, horizon, and default definition. In practice, portfolio expected loss is usually obtained by summing or averaging segment-level losses, while unexpected loss requires a model of dependence among borrowers.

Learning Objectives

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

• distinguish PD, LGD, EAD, expected loss, and unexpected loss;
• construct and evaluate credit scores and rating systems;
• model default using classification and survival-analysis techniques;
• estimate recoveries and drawdowns for retail and corporate facilities;
• use migration matrices and transition-based forecasts;
• explain credit portfolio models and systematic risk;
• address class imbalance, reject inference, data drift, and economic cycles;
• evaluate governance, fairness, validation, and regulatory implications.

1. The Architecture of Credit Risk Measurement

1.1 Default and the observation horizon

A default definition must be operational before a model can be developed. Common indicators include a material payment delinquency, bankruptcy, restructuring motivated by financial distress, or an assessment that the borrower is unlikely to pay in full. Regulatory and accounting definitions may differ from internal management definitions, so the definition should be documented explicitly.

A one-year PD is the probability of default during the next twelve months conditional on information available at the observation date. A lifetime PD is the probability of default over the remaining contractual or expected life. These are not interchangeable. A borrower can have a low one-year PD but a materially higher lifetime risk.

For an obligor i, define the default indicator:

The conditional PD is:

where X_i represents borrower, facility, and environmental information. The estimate applies to the population and time period represented by the development sample and must be recalibrated when those conditions change.

1.2 PD, LGD, EAD, and expected loss

PD measures likelihood. It may be point-in-time (PIT), responding strongly to current economic conditions, or through-the-cycle (TTC), designed to be more stable across the business cycle.

LGD measures the fraction of exposure not recovered. If EAD is 100, recoveries are 60, and direct and indirect collection costs are 5, then the economic loss is 45 and LGD is 45%.

Recoveries and costs should be discounted to the default date using a specified rate. If discounting is ignored, the calculation is a simplified accounting approximation rather than an economic LGD.

EAD is the amount exposed when default occurs. For a fully drawn term loan, EAD may be close to the outstanding balance. For a revolving facility, future drawings before default must be estimated.

For facility i:

where CCF is the credit conversion factor, expressed as a fraction. This formula is a practical approximation; it assumes the undrawn amount is measured consistently and that one CCF adequately summarizes future utilization.

Expected loss is an average loss over a defined horizon, not a worst-case loss. For a portfolio of n exposures:

This assumes the component estimates are appropriately aligned. It does not imply that actual loss will equal EL in any particular period.

1.3 Unexpected loss and concentration

Unexpected loss reflects dispersion around expected loss. It depends on default volatility, LGD volatility, exposure uncertainty, and correlations among obligors. Two portfolios with identical EL can have very different risk if one is diversified and the other is concentrated in a single industry, geography, or borrower group. Portfolio models therefore require dependence assumptions, not merely individual PDs.

2. Credit Data and Risk Segmentation

Credit analytics begins with a data architecture rather than an algorithm. Relevant data may include application variables, bureau histories, transaction behavior, repayment patterns, collateral values, financial statements, market prices, covenant breaches, relationship information, and macroeconomic variables.

A robust dataset must preserve the information available at the decision date. Using later information creates look-ahead bias. Analysts should also define observation windows, performance windows, censoring rules, treatment of restructurings, and the unit of analysis. A model developed at account level may not be valid at obligor level when one customer holds multiple facilities.

Segmentation is valuable when relationships differ materially across groups. Typical segments include:

Segmentation should be justified statistically and economically. Excessive segmentation creates small samples, unstable estimates, and inconsistent governance.

3. Scoring and Rating Systems

3.1 Logistic credit scoring

For a binary default outcome, logistic regression models the conditional probability:

with linear predictor:

The coefficients β describe changes in log-odds, holding other variables constant:

The model assumes a correctly specified relationship between predictors and log-odds, conditional independence is not required among predictors but multicollinearity can make coefficients unstable, and observations used for estimation should represent the target population. Categorical variables require an appropriate reference category; missing values must be handled without using future information.

A scorecard often transforms log-odds into a more interpretable scale. If a score increases when risk decreases, a generic form is:

A common design fixes the score increase associated with doubling the odds at a specified number of points. The exact constants A and B are business conventions, not universal statistical quantities.

3.2 Evaluation

Discrimination asks whether risky borrowers are ranked above safer borrowers. The area under the receiver operating characteristic curve (AUC) is the probability that a randomly selected defaulter receives a worse risk ranking than a randomly selected non-defaulter, subject to tie conventions. The Gini coefficient is often defined as:

Calibration asks whether predicted probabilities match observed default frequencies. A portfolio with predicted PD of 4% should experience approximately 4% default under comparable conditions, allowing for sampling variation. Calibration can be assessed by bins, reliability plots, calibration intercept and slope, and statistical tests, although no single test is decisive in large samples.

Accuracy alone is inappropriate for rare default events. A model predicting “non-default” for every account may have high accuracy while being useless. Analysts should examine precision, recall, specificity, sensitivity, AUC, precision-recall curves, expected profit or loss, and stability across segments.

3.3 Corporate ratings and expert judgment

Corporate ratings combine quantitative indicators such as leverage and interest coverage with qualitative information such as governance, competitive position, refinancing risk, and management quality. Expert overlays can be useful when data are sparse or structural breaks occur, but overrides should be documented, monitored, and back-tested. A rating is a risk classification, not a guarantee of repayment.

4. Survival Analysis and Time-to-Default

Binary models discard timing information. Survival analysis estimates the distribution of time until default and handles right-censoring, where an account has not defaulted by the end of observation.

Let T be time to default. The survival function is:

The cumulative default probability by time t is:

The hazard rate, when defined as a continuous-time quantity, is the instantaneous default intensity conditional on survival to t:

The relationship between hazard and survival is:

For a proportional-hazards model:

where h₀(t) is the baseline hazard and βᵀX_i is the covariate effect. The proportional-hazards assumption means that hazard ratios between two borrowers remain constant over time, conditional on the model. If this is implausible, time-varying effects, discrete-time hazard models, accelerated-failure-time models, or flexible machine-learning survival models may be preferable.

Survival models are especially useful for lifetime expected loss, refinance risk, prepayment, cure behavior, and competing events. A borrower may prepay, default, or remain active; treating all exits as default can bias estimates.

5. LGD and EAD Modeling

5.1 LGD

LGD depends on collateral, seniority, guarantees, jurisdiction, workout strategy, default type, and economic conditions. Recovery data are often highly skewed, with many near-zero or near-one outcomes. Models may use fractional regression, beta-type approaches for interior values, two-part models for mass points, or decision trees and gradient boosting with suitable controls.

A workout-based LGD should include realized recoveries, time to recovery, legal and collection costs, and discounting. A simple example illustrates the calculation. Suppose EAD is 1,000,000 monetary units, undiscounted recoveries are 650,000, and collection costs are 50,000. If the present value of recoveries is 620,000 and the present value of costs is 45,000:

Thus LGD is 42.5%, and the loss amount is 425,000. The result assumes recoveries and costs are measured on the same valuation date and that EAD is correctly defined.

5.2 EAD and revolving credit

EAD for revolving products requires a utilization forecast. If current drawn balance is 40,000, undrawn commitment is 60,000, and estimated CCF is 35%:

CCF may depend on utilization, account age, delinquency, limit management, payment behavior, and macroeconomic stress. A single average CCF can conceal important nonlinear behavior, especially when borrowers draw heavily immediately before default.

6. Migration Matrices and Rating Dynamics

A migration matrix records probabilities of moving from one rating state to another over a specified period. States may include AAA, AA, A, BBB, default, and exit. For a one-year transition matrix P, element p_{ij} is the probability of moving from state i to state j during one year. Each row must sum to one when all possible destinations are included.

If the initial state distribution is a row vector v₀, the distribution after k periods under time-homogeneous Markov assumptions is:

The Markov assumption means the next state depends on the current state rather than the complete history. This is convenient but often unrealistic: ratings may exhibit duration effects, issuer-specific persistence, and cycle dependence. Empirical matrices should distinguish downgraded, upgraded, defaulted, and withdrawn exposures carefully. Matrix multiplication can produce lifetime default probabilities, but it does not by itself estimate LGD or EAD.

7. Credit Portfolio Models

7.1 Dependency and systematic factors

Individual PD estimates do not determine portfolio loss distribution. Defaults may rise together because firms share macroeconomic, industry, geographic, or supply-chain exposures. A one-factor latent-variable model represents asset quality as:

where Z is a common systematic factor, ε_i is an idiosyncratic factor, both are commonly standardized normal variables in the Gaussian version, and ρ_i is the asset correlation parameter. Default occurs when Y_i falls below a threshold chosen to match PD_i. This framework is analytically convenient but depends on distributional and correlation assumptions.

Portfolio loss is:

Simulation of Z and ε_i produces a loss distribution. Quantiles of that distribution support economic capital and stress analysis. Tail measures are sensitive to correlation, concentration, parameter uncertainty, and model specification.

7.2 Economic capital and stress testing

Expected loss is generally priced or provisioned, whereas economic capital is intended to absorb unexpected loss at a selected confidence level. A conceptual unexpected-loss measure is:

where VaR_α(L) is the α-quantile of portfolio loss. VaR is not universally preferred; expected shortfall, the average loss beyond a high quantile, may better reflect tail severity. Neither metric is meaningful without a clearly defined horizon, confidence level, loss definition, and dependence model.

Stress testing imposes adverse but plausible paths for unemployment, interest rates, GDP, inflation, property prices, exchange rates, or industry variables. A stressed PD can be estimated through a macroeconomic satellite model, scenario-conditioned transition matrix, or expert-supported adjustment. Stress tests should distinguish baseline forecasts from severe scenarios and should report uncertainty rather than false precision.

8. Class Imbalance, Bias, and Model Risk

Defaults are often rare, creating class imbalance. Appropriate responses include stratified sampling for development, class-weighted loss functions, careful threshold selection, and precision-recall analysis. Oversampling the minority class can assist estimation, but predicted probabilities must be recalibrated to the true population prevalence. Synthetic observations should not be allowed to leak across training and validation periods.

Temporal validation is usually more informative than random splitting because credit portfolios evolve. A typical design trains on earlier periods, validates on a subsequent period, and tests on a later out-of-time sample. Population stability, characteristic stability, calibration drift, and performance by segment should be monitored.

Reject inference arises because outcomes are observed mainly for approved applicants. The approved sample is not random. Methods such as augmentation, parceling, or explicit experimentation can be considered, but each relies on assumptions that should be documented. Treating rejected applicants as non-defaulters is generally unjustified.

Model risk also arises from missing data, unstable definitions, data leakage, target leakage through collections actions, excessive feature engineering, and unchallenged vendor scores. Explainability is important for adverse-action notices, internal challenge, customer communication, and bias assessment. A more complex model is not automatically a better model.

9. Economic Cycles and Point-in-Time Design

Default behavior is cyclical. In expansionary periods, delinquency may fall and collateral values may rise; in recessions, unemployment, refinancing stress, and business failures may increase. A PIT model incorporates current and forecast conditions and is responsive for underwriting and near-term provisioning. A TTC model smooths cyclical variation and may be useful for strategic capital planning. Neither is intrinsically superior; the choice must match the decision use.

A simple macroeconomic specification is:

where logit(PD) means ln[PD ÷ (1 − PD)]. The signs and magnitudes must be estimated rather than assumed, and relationships may be nonlinear or lagged. Scenario forecasts should preserve economically plausible variable paths and avoid extrapolating beyond the historical range without expert justification.

10. Illustrative Case: Meridian Bank’s Small-Business Portfolio

Background. Meridian Bank has 20,000 small-business revolving facilities. Management observes rising utilization and wants to improve one-year expected-loss estimates and early-warning processes.

Problem. The existing model uses a static application score, applies a portfolio-average LGD, and assumes EAD equals current drawn balance. It therefore understates risk for borrowers that draw down before default and fails to distinguish secured from unsecured facilities.

Analysis. The analytics team defines default as 90 days past due, bankruptcy, or a distressed restructuring within twelve months. It creates an observation date, a twelve-month performance window, and an out-of-time test period. It develops:

• a logistic PD model using application and monthly behavioral variables;
• a CCF model segmented by current utilization and delinquency status;
• an LGD model incorporating collateral, seniority, recovery timing, and workout cost;
• a macroeconomic overlay using unemployment and interest-rate scenarios.

Suppose one facility has current drawn exposure of 80,000, undrawn commitment of 120,000, PD of 4%, CCF of 50%, and LGD of 45%. Its EAD is:

Its expected loss is:

The previous method, using current drawn exposure and a 40% portfolio LGD, would estimate 0.04 × 0.40 × 80,000 = 1,280. The revised estimate is higher because it recognizes likely future drawings and the facility-specific recovery assumption. The difference is not evidence that the new model is correct by itself; it must be validated against outcomes and reviewed for data quality.

Approach and outcome. Back-testing shows improved ranking and better calibration in the out-of-time sample. Monitoring identifies that the model underpredicts defaults for firms in construction during a regional downturn. Management introduces an industry stress indicator, tighter review triggers, and a quarterly override committee. The bank uses the model for risk-based pricing and collections prioritization, while retaining human review for thin-file and strategically important businesses.

Learning points. The case demonstrates that PD, LGD, and EAD are complementary rather than substitutable; that behavioral information can materially improve retail and small-business models; and that governance, segmentation, and stress testing are as important as predictive accuracy.

11. Governance, Ethics, and Regulation

Credit models influence access to finance, pricing, limits, and collection intensity. Governance should include clear ownership, independent validation, version control, documented assumptions, approval thresholds, performance monitoring, and an effective challenge process. Material model changes require controlled implementation and retrospective evaluation.

Fairness requires attention to legally protected characteristics and proxy variables. Excluding a protected attribute does not guarantee fairness if geography, occupation, device information, or transaction behavior acts as a proxy. Analysts should test disparate impact where legally and operationally appropriate, investigate performance differences across groups, and balance predictive performance with explainability and lawful treatment.

Privacy principles include purpose limitation, data minimization, access controls, retention limits, and secure handling of bureau and transactional data. Automated decisions should provide appropriate explanations and avenues for review. Generative AI can assist documentation, coding, and analyst productivity, but it should not be trusted to invent data, silently change model logic, or make ungoverned credit decisions. Reproducibility, audit trails, and human accountability remain essential.

12. Best-Practice Implementation Sequence

A defensible implementation commonly follows this sequence:

1. define default, horizon, population, and decision use;
2. inventory and reconcile data sources;
3. establish an observation-performance framework without leakage;
4. segment only where economic and statistical evidence supports it;
5. develop benchmark models before complex alternatives;
6. evaluate discrimination, calibration, stability, fairness, and business value;
7. validate out of time and under adverse scenarios;
8. document assumptions, limitations, overrides, and implementation controls;
9. monitor drift, realized loss, overrides, and data quality;
10. recalibrate or redevelop when evidence shows material deterioration.

Chapter Summary

Credit risk analytics decomposes lending risk into PD, LGD, and EAD, whose aligned product provides expected loss. Logistic regression and scorecards support borrower ranking and probability estimation; survival analysis adds time-to-default and censoring; LGD models capture recoveries and costs; EAD models capture utilization and future drawings. Migration matrices describe rating movement, while portfolio models incorporate dependence and systematic risk to estimate loss distributions and unexpected loss.

Effective practice requires careful default definitions, representative data, temporal validation, calibration, treatment of class imbalance, and explicit handling of economic cycles. PIT and TTC approaches serve different purposes. Model governance, fairness, privacy, explainability, stress testing, and monitoring are not supplementary concerns: they determine whether analytics can be used responsibly in credit decisions.

Key Terms

Probability of default (PD): Probability that an obligor defaults within a specified horizon.

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

Exposure at default (EAD): Exposure outstanding when default occurs.

Expected loss (EL): Average predicted monetary credit loss.

Unexpected loss: Loss variation above expected loss, often associated with capital needs.

Credit scorecard: Statistical system that ranks applicants or accounts by credit risk.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Discrimination: Ability to rank higher-risk cases above lower-risk cases.

AUC: Area under the receiver operating characteristic curve.

Gini coefficient: AUC-based discrimination measure, commonly 2 × AUC − 1.

Hazard rate: Instantaneous default intensity conditional on survival.

Survival function: Probability that default has not occurred by a specified time.

Censoring: Incomplete observation of an event time, commonly because an account remains active.

Credit conversion factor (CCF): Estimated fraction of undrawn commitment drawn before default.

Migration matrix: Matrix of probabilities for transitions among rating states.

Point-in-time (PIT): Risk estimate responsive to current economic conditions.

Through-the-cycle (TTC): Risk estimate designed to be more stable across economic cycles.

Class imbalance: Unequal prevalence of default and non-default observations.

Reject inference: Estimation challenge caused by missing repayment outcomes for rejected applicants.

Model risk: Risk of adverse consequences from incorrect model design, data, implementation, or use.

Review Questions

5 Short Answer Questions

1. Define PD, LGD, and EAD and explain why their horizons must be aligned.
2. What is the difference between discrimination and calibration?
3. Why can current drawn balance understate EAD for a revolving facility?
4. What does a migration matrix describe?
5. Why is accuracy an inadequate primary metric for rare-default prediction?

5 Long Answer Questions

1. Explain the construction, assumptions, evaluation, and governance of a logistic credit-scorecard model.
2. Compare point-in-time and through-the-cycle PD models and identify appropriate uses for each.
3. Explain how survival analysis improves on a binary default model for lifetime credit-risk estimation.
4. Discuss the sources of dependence in a credit portfolio and the implications for unexpected loss.
5. Design a complete validation and monitoring framework for a small-business credit model.

3 Analytical / Application Questions

1. A facility has current drawn exposure of 75,000, undrawn commitment of 125,000, CCF of 40%, PD of 3%, and LGD of 50%. Calculate EAD and expected loss. State the assumptions behind your calculation.
2. A bank’s model has strong AUC but consistently predicts 2% PD for a segment that realizes 5% default. Diagnose the problem and recommend corrective actions.
3. A recession scenario raises unemployment, increases utilization, and reduces collateral values. Explain separately how the scenario should affect PD, EAD, and LGD, and identify governance controls needed before applying the changes.

Multiple Choice Questions with Answers

Question 1

Which expression represents expected loss for a single credit exposure?

A. PD + LGD + EAD

B. PD × LGD × EAD

C. PD ÷ LGD × EAD

D. LGD ÷ (PD × EAD)

Correct Answer: B. PD × LGD × EAD

Question 2

A facility has drawn exposure of 40,000, an undrawn commitment of 60,000, and a CCF of 35%. What is its estimated EAD?

A. 40,000

B. 51,000

C. 61,000

D. 100,000

Correct Answer: C. 61,000

Question 3

What does calibration primarily assess in a PD model?

A. Whether riskier borrowers receive lower scores

B. Whether predicted probabilities match observed default frequencies

C. Whether all predictors are normally distributed

D. Whether the portfolio has zero correlation

Correct Answer: B. Whether predicted probabilities match observed default frequencies

Question 4

In survival analysis, the survival function S(t) represents:

A. The probability of default exactly at time t

B. The probability of no default beyond time t

C. The loss recovered after default

D. The exposure outstanding at time t

Correct Answer: B. The probability of no default beyond time t

Question 5

Which issue is most directly associated with reject inference?

A. Future drawings on revolving accounts

B. Missing repayment outcomes for applicants who were not approved

C. Discounting recovery cash flows

D. Correlation among corporate obligors

Correct Answer: B. Missing repayment outcomes for applicants who were not approved

Question 6

If AUC is 0.75, what is the commonly defined Gini coefficient?

A. 0.25

B. 0.50

C. 0.75

D. 1.50

Correct Answer: B. 0.50

Question 7

What is the principal purpose of a credit migration matrix?

A. To estimate recovery costs directly

B. To show probabilities of movement among rating states

C. To calculate a borrower’s income

D. To eliminate economic-cycle effects

Correct Answer: B. To show probabilities of movement among rating states

Question 8

Why can a model with high accuracy still be ineffective in a portfolio with rare defaults?

A. Accuracy cannot be calculated for binary outcomes

B. Predicting non-default for nearly everyone can produce high accuracy but miss defaults

C. Accuracy is identical to LGD

D. Rare defaults always imply perfect calibration

Correct Answer: B. Predicting non-default for nearly everyone can produce high accuracy but miss defaults

Question 9

A loan has EAD of 1,000,000, present-value recoveries of 620,000, and present-value collection costs of 45,000. What is LGD?

A. 38.0%

B. 42.5%

C. 57.5%

D. 66.5%

Correct Answer: B. 42.5%

Question 10

Which statement best describes a point-in-time PD model?

A. It is designed to ignore current economic conditions

B. It responds to current and forecast economic conditions

C. It always produces the same PD across the cycle

D. It models only collateral recovery values

Correct Answer: B. It responds to current and forecast economic conditions