Quick Reference: Fair Lending Requirements for AI Models

Quick Reference: Fair Lending Requirements for AI Models

This quick reference provides essential information on fair lending compliance for AI and machine learning models used in credit decisions. Keep this guide accessible for rapid reference during model development, validation, and ongoing monitoring.

---

Key Regulatory Framework

Primary Laws and Regulations

Law/Regulation	Scope	Key Requirement
Equal Credit Opportunity Act (ECOA)	All credit decisions	Prohibits discrimination based on protected characteristics
Fair Housing Act (FHA)	Mortgage and housing-related credit	Prohibits housing discrimination
Regulation B	Implements ECOA	Requires adverse action notices with specific reasons
CFPB Guidance	All creditors using AI/ML	"No fancy technology exemption" - all laws apply

Protected Characteristics

Prohibited Basis for Credit Decisions:

• Race
• Color
• Religion
• National Origin
• Sex (including sexual orientation and gender identity)
• Marital Status
• Age (if applicant has capacity to contract)
• Receipt of Public Assistance
• Exercise of Consumer Credit Protection Act rights

Critical Rule: Cannot use protected characteristics directly OR indirectly through proxies.

---

Testing Requirements

1. Disparate Treatment Testing

Definition: Intentional discrimination - treating applicants differently based on protected status.

Testing Approach: Review model features for prohibited characteristics Verify protected characteristics excluded from training data Test for "masking" or encoding of protected attributes Assess business rules for differential treatment

Red Flags:

• Different credit standards for different groups
• Protected characteristics in model inputs
• Segmented models by demographic groups

---

2. Disparate Impact Testing

Definition: Facially neutral practice with disproportionate adverse effect on protected class.

Three-Part Test:

Part 1: Prima Facie Case

• Does the model have statistically significant adverse impact on protected class?
• Use 80% rule as screening threshold
• Calculate approval/denial rates by protected class

Part 2: Business Justification

• Is there legitimate business necessity?
• Is the practice related to creditworthiness?
• Are there documented business reasons?

Part 3: Less Discriminatory Alternative

• Does a less discriminatory alternative exist?
• Would alternative achieve business objectives?
• Is alternative equally effective?

Key Metrics:

Metric	Definition	Threshold
Disparate Impact Ratio (DIR)	(Minority approval rate) / (Control group approval rate)	< 0.80 indicates potential issue
Standardized Mean Difference (SMD)	Difference in mean scores between groups / pooled std dev	> 0.25 warrants investigation
Adverse Action Rate Difference	Difference in denial rates between protected and control groups	> 10 percentage points is concerning

Testing Frequency:

• Pre-Deployment: Comprehensive testing before production use
• Quarterly: Ongoing monitoring of key fairness metrics
• Annual: Full bias audit with LDA analysis
• Post-Change: Retest after material model changes

---

Explainability Requirements

ECOA Adverse Action Notices

Requirements (Regulation B):

Specific Reasons Required

• Provide specific reasons for adverse action
• Reasons must be meaningful and accurate
• Generic statements insufficient ("low credit score")
• Must list principal reasons in order of importance

Prohibited Practices

• Cannot use "model score" as sole reason
• Cannot cite protected characteristics as negative factors
• Cannot provide misleading or inaccurate reasons

AI Model Challenge: Black-box models make generating specific reasons difficult. Solutions:

Approach	Description	Pros	Cons
SHAP Values	Explain individual predictions using Shapley values	Model-agnostic, local explanations	Computationally intensive
LIME	Local surrogate models for interpretability	Simple, intuitive	May not reflect true model behavior
Partial Dependence	Show feature effect on predictions	Easy to understand	Assumes feature independence
Feature Importance	Global importance of each feature	Quick assessment	Doesn't explain individual decisions
Surrogate Models	Interpretable model approximating AI model	Regulatory acceptable	Less accurate than original

Best Practice: Implement multiple explainability methods and document approach in model governance documentation.

---

Protected Class Identification

Methodologies for Non-Mortgage Credit

Challenge: Protected class data often unavailable for non-mortgage credit (unlike HMDA data for mortgages).

CFPB-Endorsed Approaches:

1. Bayesian Improved Surname Geocoding (BISG)

• Combines surname analysis with geographic information
• Estimates probability of race/ethnicity
• CFPB uses in fair lending examinations
• Industry-standard methodology

2. Self-Identification (Preferred)

• Collect demographic data voluntarily
• Only for monitoring purposes (cannot use in decisions)
• Ensures accurate protected class assignment

3. Geocoding

• Uses Census data by geography
• Less precise than BISG
• Acceptable when other methods unavailable

Important: Never use estimated protected class data in credit decisioning - monitoring purposes only.

---

Less Discriminatory Alternative (LDA) Analysis

LDA Requirements

CFPB Expectation: Actively search for and implement less discriminatory alternatives.

Analysis Steps:

Step 1: Identify Alternatives

• Different model architectures
• Alternative feature sets
• Modified decision thresholds
• Bias mitigation techniques

Step 2: Test Effectiveness

• Compare disparate impact of alternatives
• Assess business objective achievement
• Evaluate credit risk accuracy
• Test operational feasibility

Step 3: Document Decision

• Document all alternatives considered
• Explain why each was/wasn't selected
• Justify final model choice
• Retain analysis for examination

Debiasing Techniques to Consider:

Technique	Description	When to Use
Reweighting	Adjust training data weights by protected class	Class imbalance in training data
Threshold Optimization	Set different thresholds by group to equalize outcomes	Disparate impact at decision boundary
Fairness Constraints	Add fairness metrics to model optimization	During model training
Post-Processing	Adjust scores after model prediction	Quick fix for existing model
Feature Removal	Eliminate features with proxy discrimination	High correlation with protected class

---

Documentation Requirements

Essential Documentation

Model Development Phase: Data sources and preparation procedures Feature selection rationale and testing Training process and hyperparameter tuning Pre-deployment bias testing results Explainability method selection and testing LDA analysis and decisions Management and board approval

Validation Phase: Independent validation report Fair lending testing results Sensitivity analysis Limitation documentation Recommendations and management responses

Ongoing Monitoring: Quarterly fairness metric tracking Model performance monitoring Drift detection results Incident reports and resolutions Annual comprehensive bias audit

Examination Preparation: Model inventory with fair lending risk assessment Governance policies and procedures Validation and monitoring documentation Testing methodologies and results Issue tracking and remediation evidence

---

Common Violations and Penalties

Recent Enforcement Actions

Enforcement Trends:

• State AGs increasingly active (e.g., Massachusetts $2.5M settlement)
• CFPB scrutinizing AI models in examinations
• Focus on inadequate bias testing and LDA analysis
• Penalties for explainability failures

Violation Categories:

Violation Type	Example	Potential Penalty
Disparate Impact	Model disproportionately denies protected class	$100K - $5M+
Inadequate Testing	No bias testing or monitoring	Enforcement action, penalties
Explainability Failure	Generic adverse action reasons	Per-violation penalties
Proxy Discrimination	Features correlate with race/ethnicity	Significant fines, restitution

---

Pre-Deployment Checklist

Before deploying AI credit model:

• Protected characteristics excluded from model inputs
• Proxy discrimination testing completed
• Disparate impact analysis shows compliance (DIR > 0.80)
• LDA analysis documented and alternatives tested
• Explainability method validated and tested
• Adverse action reason generation verified
• Ongoing monitoring plan established
• Governance committee approval obtained
• Board notification completed
• Documentation package finalized

---

Emergency Response: Bias Detected

If bias identified post-deployment:

Step 1: Immediate Actions

• Suspend model use (if material bias)
• Notify senior management and governance committee
• Assess scope of impacted consumers
• Document issue comprehensively

Step 2: Investigation

• Conduct root cause analysis
• Determine when bias emerged
• Identify affected decisions
• Assess restitution requirements

Step 3: Remediation

• Implement bias mitigation techniques
• Revalidate model post-remediation
• Retest for disparate impact
• Update governance controls

Step 4: Restitution (if required)

• Identify harmed consumers
• Calculate damages
• Provide remediation per regulatory guidance
• Document process for examination

---

Key Takeaways

No Technology Exemption: All fair lending laws apply to AI models regardless of complexity

Proactive Testing Required: Quarterly monitoring and annual comprehensive bias audits mandatory

LDA Analysis Critical: Must actively seek less discriminatory alternatives

Explainability Non-Negotiable: ECOA requires specific, accurate adverse action reasons

Documentation Essential: Comprehensive documentation critical for examination defense

Continuous Vigilance: Fair lending compliance is ongoing, not one-time

---

Need fair lending testing or AI bias audit? RegVizion provides comprehensive fair lending compliance services, including disparate impact testing, LDA analysis, and explainability assessment. Contact us for expert support.

Download Resources:

• Fair Lending Testing Checklist PDF
• LDA Analysis Template
• Protected Class Estimation Guide