Estimating a One-Year Probability of Default

A probability of default (PD)The probability that a defined borrower or exposure defaults within a specified horizon under stated assumptions.Open full glossary entry ↗ is the estimated chance that an obligorA borrower or counterparty whose credit quality or default risk is assessed.Open full glossary entry ↗ will meet a defined default event over a stated horizon, conditional on information available at the measurement date. A one-year PDA PD measure over a one-year horizon, with the default definition, population, and conditioning stated.Open full glossary entry ↗ refers to default during the next 12 months. It is meaningful only when the population, unit of analysis, default definition, observation date, and outcome horizon are specified. Basel’s IRB framework uses a one-year PD for corporate, sovereign, and bank exposures; a model score alone does not establish that a PD meets regulatory requirements.

Define default before building labels. The definition may include both days-past-due criteria and unlikeliness-to-pay events, subject to applicable policy and jurisdiction. Apply it consistently to development, validation, and monitoring data. Link related accounts to obligors where the target is obligor default, set the observation date before the outcome window, and exclude features that were not available on that date. These choices determine what the model learns.

Create one observation per eligible obligor at a defined date, with predictor values reconstructed as they were known then. Mark the target as one when the defined default occurs within the following 12 months, and zero only when the full horizon is observed without default. Recent observations whose outcome windows have not matured are censored; treating them as non-defaults biases measured default rates downward. Document exclusions, restructures, cures, and repeat observations.

Split samples by time where possible: develop on earlier periods, tune on a later period, and reserve a fully out-of-time period for final validation. Keep all observations from the same obligor together when repeated records could otherwise leak across splits. Review missingness, outliers, category changes, and population coverage before fitting. A statistically strong model can still be unreliable if the historical data does not represent the intended use population.

A common scorecard approach estimates a linear predictor, or logit, from approved features: z = β₀ + β₁x₁ + … + βₖxₖ. Convert it to a probability with the logistic function: PD = 1 / (1 + exp(−z)). Coefficients describe conditional associations in the fitted sample; they are not causal effects. Feature transformations, binningGrouping continuous values or categories into intervals for analysis or model use.Open full glossary entry ↗, regularization, and variable selection should be justified, stable, and checked for leakage and unreasonable relationships.

For illustration only, suppose an intercept of −3.2 and two feature contributions of +0.35 and +1.10 produce a logit of −1.75. The logistic transform gives a raw PD of about 14.8%. Those values are synthetic and say nothing about a real borrower or portfolioA defined group of credit exposures analyzed together under stated inclusion rules.Open full glossary entry ↗. The guide page includes a matching browser-based Python cell that you can edit and run locally using synthetic values.

DiscriminationA model's ability to rank cases with different outcome risk, distinct from probability calibration.Open full glossary entry ↗ asks whether higher-risk obligors tend to rank above lower-risk obligors. CalibrationAgreement between predicted probabilities and observed event frequencies for a matched population and horizon.Open full glossary entry ↗ asks whether predicted probabilities agree with realized frequencies over the stated horizon. Assess calibration on matured, representative observations using multiple views: observed versus predicted default rates by risk band, observed-to-expected totals, calibration plots, and suitable probability scores. Include confidence intervals or sample counts so unstable estimates in small segments are visible.

If a model ranks well but systematically overpredicts or underpredicts, investigate changes in the population, default definition, sample weighting, economic conditions, and development-to-application differences. Any recalibration should preserve the intended target and be documented, independently reviewed, and monitored. A single portfolio-level adjustment can conceal offsetting errors across segments, so inspect relevant product, vintage, geography, and risk bands where data supports them.

Use complementary measures rather than relying on one headline statistic. ROC-AUCArea under a receiver operating characteristic curve, summarizing ranking behavior across binary-classification thresholds.Open full glossary entry ↗ or Gini can summarize ranking; Brier scoreThe mean squared difference between binary outcomes and predicted probabilities; it reflects multiple aspects of probabilistic prediction quality.Open full glossary entry ↗ and log lossA probabilistic scoring rule that penalizes the negative log probability assigned to the observed class.Open full glossary entry ↗ assess probability accuracy; observed-to-expected and calibration plots reveal systematic bias. Track data qualityThe degree to which data is fit for a defined use, assessed through properties such as completeness, validity, consistency, uniqueness, accuracy, and timeliness.Open full glossary entry ↗, missingness, feature and score distributions, approval or selection effects, and the maturity of outcome windows. Define escalation limits using portfolio size, uncertainty, risk appetiteThe amount and types of risk an organization is prepared to accept while pursuing its objectives.Open full glossary entry ↗, and intended use; there is no universal threshold that is appropriate for every portfolio.

Compare current performance with the development and validation baselines, investigate material shifts, and record remediation decisions. Monitor at a frequency suited to the product’s risk and data availability. A one-year outcome requires time to mature, so use interim indicators carefully and do not present them as completed one-year validation. Maintain versioned data definitions, code, coefficients, calibration steps, approvals, and limitations so results can be reproduced.

A Basel IRB one-year PD and an IFRS 9The International Financial Reporting Standard for financial instruments, including classification, measurement, hedge accounting, and impairment requirements.Open full glossary entry ↗ expected-credit-loss estimate answer different questions. IFRS 9’s 12-month ECLThe portion of lifetime ECL associated with default events possible within the 12 months after the reporting date; it is not simply cash shortfalls occurring only in that year.Open full glossary entry ↗ represents the lifetime cash shortfalls associated with default events that are possible in the 12 months after the reporting date, weighted by the probability of those defaults; it is not simply the cash shortfall expected to be paid during the next 12 months. IFRS 9 ECL measurement also requires probability-weighted, reasonable and supportable information, including relevant forward-looking conditions.

Do not treat a regulatory PD as an IFRS 9 provision input without assessing the target definition, measurement date, horizon, scenario conditioning, and the other ECL components. PD, loss given default, exposure at default, discounting, staging, and scenario weightingThe probability weights assigned to alternative economic scenarios in a probability-weighted estimate; weights should be justified and sum to one.Open full glossary entry ↗ must be aligned to the applicable accounting policy. This guide is educational and does not replace institutional model governance, supervisory requirements, or accounting advice.

IFRS 9 Financial Instruments, paragraphs B5.5.17 and B5.5.43: https://www.ifrs.org/content/dam/ifrs/publications/pdf-standards/english/2021/issued/part-a/ifrs-9-financial-instruments.pdf