Fair housing and anti-discrimination checks in scoring models

In the EU, the General Data Protection Regulation (GDPR) restricts processing of "special category" data (race, ethnicity, religion, health) and grants a right to explanation for automated decisions under Article 22. The EU's AI Act (phasing in through 2026) classifies credit and tenancy scoring systems as "high-risk," triggering mandatory bias testing and documentation obligations.

HUD issued specific guidance in 2024 clarifying that tenant screening algorithms using criminal or eviction records can trigger disparate impact liability if not narrowly tailored. That guidance is the direct ancestor of the checks below.

Where bias hides: proxy variables

A proxy is a variable that correlates tightly with a protected class even though it looks neutral.

Common real estate proxies:

| Variable | Proxies for |

|---|---|

| Zip code | Race, national origin |

| First name | Race, national origin, sex |

| Eviction record | Race (eviction rates are not uniform across groups, per Eviction Lab, Princeton) |

| Criminal record | Race (well documented in US arrest data disparities) |

| Source of income (housing vouchers) | Race, familial status, disability |

| Employment gaps | Sex (maternity), disability |

| Social media / rental history "soft" data | National origin, religion |

None of these need to be intentionally included to cause harm. A rent-pricing model trained on "market comparables" can absorb historical redlining patterns baked into decades-old price data. Redlining was the 1930s practice of denying mortgages in maps of minority neighborhoods, and its price effects persist in comps datasets today.

The core audit checklist

Regulators and fair housing auditors typically run four checks. Treat this as your working checklist.

1. Correlation scan

Test every input feature against protected-class membership (where legally obtainable, often via Bayesian Improved Surname Geocoding, BISG, a method the CFPB itself uses to estimate race/ethnicity when it isn't directly recorded). Flag any feature correlated above a chosen threshold (commonly 0.3 to 0.4 in practice, treat as illustrative, not a legal bright line).

2. Disparate impact ratio (the "four-fifths rule")

Borrowed from EEOC employment guidance but widely applied in housing: compare approval rates across groups.

Worked example:

• White applicants approved: 720 out of 900 → 80%
• Black applicants approved: 320 out of 500 → 64%
• Ratio: 64 / 80 = 0.80

A ratio below 0.80 (80%) is a common trigger for further scrutiny, not automatic illegality, but it shifts the burden to justify the practice.

3. Feature importance audit (post-hoc explainability)

Even without a banned variable, check which features drive decisions using tools like SHAP (SHapley Additive exPlanations). If "zip code" or "distance to a specific neighborhood" ranks high in a rent-pricing model, that's a red flag worth documenting and possibly removing or reweighting.

This surfaces exactly what regulators ask for in discovery: "show us what the model actually weighted."

4. Counterfactual / matched-pair testing

Take two synthetic applicants, identical except for a name signaling different ethnicities (a classic paired-testing method HUD itself uses in fair housing audits, historically with human testers, now increasingly simulated). If the score changes meaningfully, the model is proxying on the name or on features correlated with it.

Governance: who signs off

A defensible model needs a paper trail, not just a clean output.

• Model documentation: training data sources, feature list, exclusion of protected variables, version history.
• Pre-deployment fairness testing: run the four checks above before launch, not after a complaint.
• Human review threshold: automatic denials should route borderline cases to a human, a practice increasingly expected under both ECOA adverse action requirements and the EU AI Act's human oversight mandate.
• Adverse action notices: under ECOA/Regulation B, rejected applicants have a right to know the principal reasons. A black-box model that can't produce a reason is a compliance liability, not just an ethics one.

Property technology vendors like RealPage and CoreLogic have faced litigation and regulatory scrutiny (including a 2024 DOJ antitrust action against RealPage, distinct from but adjacent to fair housing claims) over algorithmic tenant screening and pricing tools, which is why documented audit trails are now a procurement requirement for many large landlords, not just a best practice.

Running the check yourself: a mini workflow

1. Pull the model's feature list. Cross-reference against a proxy table like the one above.

2. Compute approval or pricing outcomes by protected-class proxy group (using BISG or self-reported data where available).

3. Apply the four-fifths rule. Flag ratios below 0.80.

4. Run SHAP or an equivalent explainability tool on the top decision driver.

5. Document everything. If HUD or the CFPB asks, the absence of a test is itself treated as a governance failure.

For a deeper technical reference on fairness metrics applicable across sectors, the NIST AI Risk Management Framework is a solid, free, non-sector-specific starting point that housing auditors increasingly cite.

🎬 [VIDEO: "How Algorithms Can Discriminate (and What to Do About It)" - youtube.com/results?search_query=algorithmic+bias+housing+fair+lending - search for recent explainer content from CFPB or academic channels covering proxy discrimination in lending and housing models]

Key Takeaways

• Disparate impact law (FHA, ECOA, *Inclusive Communities*) means intent doesn't matter: an outcome that falls unevenly on a protected class can be illegal even if no protected variable was ever used.
• Zip code, name, eviction history, and criminal record are the classic proxy variables in real estate scoring models; each one needs a documented correlation check before deployment.
• The four-fifths rule (comparing approval rates across groups) is the fastest first-pass test; anything below an 80% ratio needs justification and documentation.
• Explainability tools like SHAP and matched-pair counterfactual testing are what regulators and auditors actually request in practice, not abstract fairness talk.
• Governance failure is itself a liability: undocumented models, missing adverse action reasoning, or absent pre-deployment testing are treated as red flags independent of the outcome.