The Global Risks Alliance (GRA)
Credit scores, probability of default models and insurance pricing models decide who gets credit or cover and at what price. Machine learning can improve prediction and also introduce opacity, instability and proxy discrimination. These models also feed provisioning and capital. This page sets out how credit and pricing models fail, how their risks are measured, the data they depend on and who decides how they are used.
GRA · In development · service level R0
Credit scoring and pricing model risk is the risk that models deciding who gets credit or insurance, and at what price, are inaccurate, unstable or unfair. It arises from historical data, changing conditions and proxies, and feeds provisions and capital. Credit committees, validators, conduct supervisors and audit committees hold the decisions.
Where a portfolio's assets sit, and how much of its value is most exposed. Hypothetical assets, not an assessment.
Finance-specific automated decisions. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Models learn from historical lending and claims that reflect past decisions, economic conditions and exclusions. Applicants unlike past borrowers are scored with less reliable evidence.
Models fitted in stable periods can misjudge default or claims when rates, inflation or unemployment shift, which is when accurate risk estimates matter most.
Variables correlated with protected characteristics can produce disparate outcomes without using those characteristics directly. Law and supervisors constrain which factors may be used.
Probability of default and loss given default estimates feed expected credit loss provisions and capital. Systematic optimism understates both until losses arrive.
Each term as this page uses it. Where a standard-setting body defines the term, the wording follows that body.
| Term | Definition |
|---|---|
| Probability of default | The estimated likelihood that a borrower defaults within a set horizon, usually one year. |
| Calibration | The agreement between predicted probabilities and the rates actually observed. |
| Proxy discrimination | Unfair outcomes caused by variables that correlate with protected characteristics. |
Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Discriminatory power of scorecards | Gini coefficient or area under the curve | quarterly | Model monitoring | ESTIMATED |
| Calibration of probability of default | observed default rate against predicted, ratio | annual | Back-testing | ESTIMATED |
| Population stability index | dimensionless | monthly | Model monitoring | ESTIMATED |
| Approval rate differences across groups | ratio of approval rates | periodic | Fair lending analysis | ESTIMATED |
| Overrides of model decisions | share of decisions | monthly | Decision records | OBSERVED |
| Expected credit loss model adjustments | currency of post-model adjustments | quarterly | Financial statements | OFFICIAL |
The topic as a computational graph: named sources feed the methods, an evidence review gates the outputs, and the decisions stay with the authorities that hold them.
Fig. 02Nodes carry their type, two metric rows and a state; connectors show what feeds what; dashed frames name the owners; the timeline is the cadence. Design of the pipeline, not a record of a run.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Source | What it gives | Licence class |
|---|---|---|
| EBA guidelines on loan origination and monitoring | Expectations for credit decision models including automated models | Free, EU reuse terms |
| CFPB circulars on adverse action and complex algorithms | United States expectations for explaining credit denials | Public domain |
| HMDA data | United States mortgage application and outcome data | Public domain |
| Bank Pillar 3 credit risk disclosures | Probability of default back-testing and model outcomes | Public records, terms per bank |
| Actuarial and insurance supervisory publications on pricing models | Practice and expectations for pricing models | Free, terms per publisher |
AI and Financial Model Risk prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.
| Who decides | The question | Evidence this page serves |
|---|---|---|
| Credit risk and pricing committees | Which models to use and with what overlays | Performance, calibration and fairness evidence |
| Model validators | Whether models are fit for purpose | Back-testing and benchmark comparison |
| Conduct and fair lending supervisors | Whether model outcomes are lawful and fair | Outcome data and adverse action practice |
| Audit committees | Whether provisioning models and overlays are reasonable | Model outputs and adjustment rationale |
Credit and pricing questions combine accuracy, fairness and legality.
Credit models estimate the likelihood that a borrower will repay; pricing models estimate expected losses and costs. Both learn from historical data, which reflects past lending practices, economic conditions and social inequalities. A model can be accurate on average while performing poorly for groups underrepresented in the data, or can reproduce historical discrimination through correlated variables.
Credit models are especially exposed to economic regime change: relationships that held in an expansion may break in a downturn. Some jurisdictions classify AI used to assess the creditworthiness of individuals as high-risk, with specific obligations for providers and users.
| Method or data | What it shows | Limits |
|---|---|---|
| Discrimination and calibration testing | Accuracy and calibration across groups | Protected attributes may not be recorded |
| Fairness metrics | Disparities in outcomes or errors | Different metrics can conflict |
| Explanation methods | Main drivers of individual decisions | Approximate explanations of complex models |
| Through-the-cycle backtesting | Performance across conditions | Few downturns in the data |
The current service level is R0: information and enquiries. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.