Il prossimoProssima guida
RANSAC per l'adattamento robusto di modelli
Tecnico
GUIDA TECNICA
A probability-of-default (PD) model estimates the chance that a borrower meets a defined default event over a stated horizon.
Ascolta questa guidaNarrato dall'IA · ~3 min
Its meaning depends on the default definition, population, data period and intended use; a score is an estimate, not a certainty about an individual.
Probability of default is one component of credit risk. In the Basel internal-ratings framework, PD is associated with a borrower grade and a defined default event; for corporate, sovereign and bank exposures, the reference horizon is one year. Other uses can define a different population or horizon, so a PD number is meaningful only when those choices are stated. It estimates the chance of default within the specified setup; it does not determine that a particular borrower will default. A model may rank borrowers by risk or produce estimates intended to align with observed default frequencies. Common approaches include logistic regression, survival methods and machine-learning models, but the algorithm alone does not define the target or validate the result. Data should match the use population as closely as possible, and features must be available at the point when the score is used. Defaults are relatively infrequent in many portfolios, so random splits can conceal time drift or leak information. Use time-aware validation where appropriate, compare to a simple benchmark, and check calibration as well as ranking performance. Banks using Basel IRB approaches face specific supervisory requirements for default definitions, data, estimation and validation. These are not universal requirements for every company that builds a credit score. Model risk guidance also emphasizes validation and limitations. State whether the output supports research, portfolio monitoring or lending decisions, and apply relevant consumer-protection, privacy and fair-lending rules. Do not describe PD as a guarantee or use a validation metric as a complete decision policy.
A bank estimates one-year default risk for a portfolio under its approved internal-rating framework.
A modeler checks that features were available before the prediction date to prevent future information leaking into training.
A credit-risk analyst compares predicted and observed default rates across score bands and time periods.
A reviewer studies how economic downturns and policy changes affect model performance.
Default models will continue to incorporate new data sources and techniques, while economic cycles and lending products change the meaning of past outcomes. Revalidate when the portfolio, default definition, policy or data pipeline changes. Model explanations and fairness analysis may receive more attention, but no one metric resolves all lending decisions. Keep the score tied to a clear target and a governed use. Economic conditions, underwriting policy and borrower mix may change faster than historical training data. Re-estimate only under controlled governance, comparing new and old versions on appropriate holdouts. Preserve enough documentation for independent validation and supervisory review where applicable.
Newsletter gratuita
Tre storie verificate sull'IA ogni mattina nei giorni feriali, scritte in un inglese semplice. Gratuito per sempre, senza pubblicità.
Un'email ogni giorno feriale. Disiscriviti con un clic. Non vendiamo né condividiamo mai il tuo indirizzo.
Mettiti alla prova
Feedback immediato su ogni risposta e un certificato condivisibile con un ID verificabile una volta superato un corso.
Sostieni l'istruzione gratuita sull'IA. AI Understanding è un'organizzazione non profit 501(c)(3) — nessun annuncio, nessun paywall, mai. Fai una donazione
A probability-of-default (PD) model estimates the chance that a borrower meets a defined default event over a stated horizon. Its meaning depends on the default definition, population, data period and intended use; a score is an estimate, not a certainty about an individual.
The guide defines PD as a probability tied to an event and horizon.
The guide says PD is meaningful only with a target event and horizon.
Basel CRE32 describes one-year PD for corporate, sovereign and bank exposures.
The guide says features must exist before the prediction timestamp.
The guide warns random splits can mask temporal changes and leakage.
Continua a imparare
Altre guide selezionate per questo argomento
Il prossimoProssima guida
RANSAC per l'adattamento robusto di modelli
Tecnico