How do I ensure my borrower risk evaluation with a credit decisioning engine is compliant with fair lending rules?

To make a borrower risk evaluation compliant with fair lending rules, design the credit decisioning engine so it is consistent, explainable, monitored for disparate impact, and not using prohibited characteristics directly or indirectly. In practice:

1) Avoid prohibited bases and obvious proxies

Do not use protected characteristics directly, such as:

• Race, color, national origin
• Sex, religion
• Marital status
• Age (except where permitted and handled correctly)
• Disability
• Receipt of public assistance
• Exercise of consumer rights

Also be careful with proxy variables that can closely track protected status, such as:

• ZIP code or geolocation
• School attended
• Language preference
• Employment history patterns
• Device data or behavioral signals that may correlate strongly with protected classes

2) Use only legitimate, business-justified variables

Your model should rely on factors that are:

• Predictive of credit risk
• Permissible under applicable law
• Documented as relevant to the lending decision

Common examples include:

• Income and income stability
• Debt obligations
• Credit history
• Payment performance
• Loan-to-value ratio
• Cash flow, where allowed
• Verified employment, where relevant

3) Build explainability into the decisioning engine

You should be able to explain:

• What factors influenced the outcome
• Why a borrower was approved, declined, or routed for manual review
• The main adverse action reasons, if applicable

Keep:

• Model documentation
• Feature lists and exclusions
• Scorecard logic or model interpretation artifacts
• Version history of rules and thresholds

4) Test for disparate impact and fairness

Regularly evaluate outcomes across protected classes, where legally and operationally appropriate, using approved methodologies and privacy safeguards.

Common checks:

• Approval rates by group
• Pricing/terms by group
• False positive/false negative rates
• Cutoff sensitivity
• Adverse action reason distribution
• Manual override patterns

If you see disparities, investigate whether they are caused by:

• A legitimate credit-risk factor
• Data quality issues
• Proxy effects
• Threshold design
• Human override bias

5) Validate and govern the model

Put the engine under formal model risk management:

• Pre-deployment validation
• Independent review
• Ongoing monitoring
• Periodic revalidation
• Change management for model updates
• Audit logs of decisions and overrides

If you use machine learning, ensure:

• Training data is representative
• Target leakage is prevented
• Drift is monitored
• Features are stable and permissible
• Humans can review edge cases

6) Ensure adverse action compliance

If a borrower is denied or receives less favorable terms, provide:

• Specific, accurate adverse action reasons
• Reasons tied to actual decision factors
• Notice within required timeframes
• A process to correct errors and reconsider where needed

7) Watch for human-in-the-loop bias

If loan officers or underwriters can override the engine:

• Standardize override criteria
• Log all overrides
• Review for patterns by protected class
• Train staff on fair lending obligations

8) Check vendor and data-source risk

If you use third-party data or an external decision engine:

• Conduct due diligence on data provenance and bias
• Contractually require compliance and audit rights
• Test outputs independently
• Don’t assume “vendor-approved” means fair-lending compliant

9) Align legal, compliance, and data science teams

Fair lending compliance is not just a technical problem. Involve:

• Compliance/legal counsel
• Credit policy
• Data science/ML engineering
• Fair lending specialists
• Internal audit

10) Document everything

Maintain a clear record of:

• Business purpose
• Feature selection rationale
• Fairness testing results
• Validation findings
• Remediation actions
• Monitoring reports

Practical checklist

Before deployment, confirm:

• No prohibited variables used
• No obvious proxies without justification
• Model is predictive and explainable
• Fairness testing shows no unexplained disparities
• Adverse action reasons are correct
• Overrides are controlled and monitored
• Governance and audit trail are in place

If you want, I can help you with:

1. a fair lending compliance checklist for a credit decisioning engine,
2. a model governance policy template, or
3. a list of fairness tests and metrics to run on borrower risk models.