Fair Lending Analysis

The process of identifying non-protected-class variables that function as close substitutes for protected characteristics like race, gender, or age. Common proxies in lending include ZIP code (correlated with race), credit product type (correlated with age), or language preference. Analysts use marginal effects analysis to isolate the statistical contribution of a suspect variable while holding all other factors constant. Removing a proxy should not degrade model performance if it lacks legitimate business justification.

Two distinct analytical lenses for fair lending compliance:

• Outcome-Based Testing: Analyzes final credit decisions (approve/deny, pricing) to detect statistically significant disparities. This is the classic disparate impact analysis.
• Process-Based Testing: Examines the operational steps leading to a decision, such as underwriting turnaround times, documentation request rates, or call center hold times, to detect subtle forms of disparate treatment during the customer journey.

When a model exhibits a statistically significant disparate impact, the legal framework requires a search for Less Discriminatory Alternatives (LDAs). This involves an iterative model rebuild process where the challenged variable is either removed, transformed, or replaced. The goal is to find a model specification that achieves the same business necessity (e.g., default prediction accuracy) while reducing the adverse impact ratio. Model documentation must record this search process for audit.

Fairness is not a one-time check but a continuous monitoring requirement. Production models must be tracked for fairness drift, where the adverse impact ratio degrades over time due to shifting population demographics or economic conditions. Automated dashboards track segmented model performance metrics (approval rates, false positive rates, APR margins) across protected groups, triggering alerts if disparities exceed predefined thresholds, aligning with SR 11-7 ongoing monitoring mandates.

A quantitative methodology that identifies facially neutral model features or decision rules that disproportionately and adversely affect a protected group. The adverse impact ratio—typically the 80% rule or a statistical significance threshold—is calculated by comparing the approval rate of a protected class against a control group. This testing is a cornerstone of Regulatory Technology (RegTech) automation for compliance with the Equal Credit Opportunity Act (ECOA) and Fair Housing Act.

A game-theoretic attribution method that decomposes a model's individual prediction into the additive contribution of each input feature. In fair lending, SHAP values are used to detect if protected class proxies—such as zip code or educational attainment—are exerting undue influence on credit decisions. This provides a consistent, mathematically grounded measure of local explainability for auditors and regulators reviewing adverse action reasons.

A human-interpretable statement identifying the minimal set of feature changes required to flip a model's adverse decision to a favorable one. For a denied mortgage applicant, a counterfactual might state: 'If income had been $5,000 higher, the loan would have been approved.' This technique is critical for generating compliant adverse action notices and answering the regulatory question of 'what would need to be different?' for protected class members.

A structured, transparent short document accompanying a deployed model that discloses its intended use, evaluation metrics, and performance benchmarks across different demographic cohorts. For fair lending, a model card explicitly reports false positive and false negative rates segmented by race, gender, and age, along with known ethical limitations. This artifact operationalizes Responsible AI (RAI) principles and supports compliance with emerging regulations like the EU AI Act.

The systematic tracking and analysis of instances where a human operator reverses or modifies a model's automated recommendation. In lending, override monitoring detects patterns where loan officers consistently override denials for one demographic group while upholding them for another—a potential indicator of disparate treatment. This Human-in-the-Loop (HITL) governance control ensures that manual intervention does not introduce unmonitored bias into the decision workflow.

A symmetric metric that quantifies the shift in a variable's distribution between a development sample and a production sample. In fair lending, PSI is monitored specifically for protected class distributions to detect if the applicant pool is shifting demographically in ways that could cause a previously fair model to develop disparate impact. A PSI above 0.25 signals a significant shift requiring immediate model re-validation.