The Global Risks Alliance (GRA)
Customers, staff and supervisors need to understand why a model made a decision, and decisions must not discriminate unlawfully. Explainability methods give reasons for individual outputs and overall behaviour; fairness testing compares outcomes across groups. Both have limits, and law sets duties such as giving reasons for credit denial. This page sets out how explainability and fairness are assessed, the evidence used and who decides.
GRA · In development · service level R0
Explainability and fairness in financial AI concern whether firms can give reasons for model decisions and whether outcomes treat groups lawfully and fairly. Explanation methods are approximate and fairness criteria conflict, so choices must be justified. Boards, developers, regulators and compliance functions hold the decisions.
How satellite radar turns a flood into an exposure update. Synthetic scene; no real event or place.
The working screen behind this subject: its structure, terms and states. Names are placeholders and values illustrative.
Explainability and fairness · Model inventory · Committee 3
| MD-169 PD retail | Tier 1 | 19 months | 4 | Overdue |
| MD-140 LLM complaints triage | Tier 1 | never | 0 | Not validated |
| MD-400 IFRS 9 staging | Tier 1 | 9 months | 1 | Approved |
| MD-386 marketing propensity | Tier 3 | 17 months | 1 | Approved |
Interpretation, bias and accountability. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Methods such as feature attributions explain a complex model approximately. Different methods can give different reasons for the same decision, and none guarantees the true cause.
Equal approval rates, equal error rates and calibration across groups cannot all hold at once when base rates differ. Choosing a criterion is a policy and legal decision, not a technical default.
Credit laws require specific reasons for adverse decisions. Generic reasons or ones unrelated to the actual driver do not meet the duty, whatever the model's complexity.
Without explanation and outcome testing, unfair patterns surface only through complaints, litigation or enforcement, after many customers are affected.
Each term as this page uses it. Where a standard-setting body defines the term, the wording follows that body.
| Term | Definition |
|---|---|
| Feature attribution | A method that estimates how much each input contributed to a model's output. |
| Adverse impact ratio | The ratio of favourable outcome rates between a group and a reference group. |
| Adverse action notice | A notice telling an applicant the specific reasons credit was denied or terms worsened. |
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 |
|---|---|---|---|---|
| Adverse impact ratio | ratio of favourable outcome rates between groups | periodic | Fairness testing | ESTIMATED |
| Error rate differences across groups | difference in false positive or negative rates | periodic | Fairness testing | ESTIMATED |
| Explanation stability | agreement between explanation methods | per validation | Validation testing | ESTIMATED |
| Adverse action notices citing model reasons | share of declines with specific reasons | periodic | Compliance review | DECLARED |
| Complaints alleging unfair automated decisions | count | quarterly | Complaints data | REPORTED |
| Enforcement actions on algorithmic discrimination | count | annual | Regulator records | 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 |
|---|---|---|
| CFPB guidance on adverse action notices and AI | United States duty to give specific reasons | Public domain |
| MAS FEAT principles and Veritas methodology documents | Fairness assessment methods for financial services | Free, MAS terms |
| IAIS Application Paper on the supervision of AI | Fairness, transparency and redress expectations for insurers | Free, IAIS terms |
| EIOPA report on governance principles for ethical and trustworthy AI | Insurance sector AI fairness and governance | Free, EU reuse terms |
| Academic literature on fairness definitions and impossibility results | Formal analysis of fairness criteria | Free to read, 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 |
|---|---|---|
| Boards and conduct committees | Which fairness criteria apply to each use | Legal requirements and outcome testing |
| Model developers and validators | Which explanation methods are reliable for a model | Stability and fidelity testing |
| Conduct and equality regulators | Whether outcomes are lawful | Outcome data, complaints and testing evidence |
| Compliance functions | Whether customer explanations meet legal duties | Notice reviews and model reason codes |
The questions concern understanding, justice and responsibility.
Simple models such as scorecards are directly interpretable. Complex models, such as gradient-boosted trees and neural networks, require post-hoc explanation methods that approximate their behaviour; these approximations can be unstable or misleading. Explainability needs differ by audience: a validator needs technical depth, a customer needs a clear reason.
Bias arises from historical data, unrepresentative samples, proxy variables and feedback loops in which model decisions shape future data. Fairness has several mathematical definitions, such as equal outcomes, equal error rates or equal calibration, which generally cannot all be satisfied at once. Choosing among them is a value judgement that should be made explicitly and accountably.
| Method or data | What it shows | Limits |
|---|---|---|
| Interpretable model design | Inherent transparency | May sacrifice some predictive performance |
| Post-hoc explanation methods | Feature contributions to outputs | Approximate and sometimes unstable |
| Fairness testing across groups | Disparities in outcomes and errors | Requires group data that may be unavailable |
| Impact assessment | Consequences for affected people | Relies on foresight and engagement |
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.