Comparative Fair Housing Analysis (CFHA™)
Fair Housing Determines Eligibility. Performance Determines Selection.
The Interagency AVM Quality Control Standards require institutions to ensure their AVM programs comply with applicable nondiscrimination laws. AVMetrics delivers the industry’s first independent, comparative Fair Housing evaluation of AVMs — so you can demonstrate, with evidence, that every model you deploy was evaluated for Fair Housing eligibility before it was selected.
Stage 1
Stage 2
The Fifth Factor
A New Kind of Requirement
The Interagency AVM Quality Control Standards require institutions using AVMs to adopt quality control processes designed to:
The first four requirements concern governance, controls, and testing — disciplines AVMetrics has provided independently since 2005. The fifth is different. It is not about how a model is built or controlled. It is about outcomes: whether an AVM produces materially different results across protected and non-protected populations.
A model can be highly accurate overall and still produce different outcomes across demographic groups. It can look acceptable across a county while behaving differently within specific neighborhoods. That is why the fifth factor cannot be satisfied by performance testing alone.
Interagency AVM Quality Control Standards
- Ensure a high level of confidence in AVM results
- Protect against the manipulation of data
- Seek to avoid conflicts of interest
- Require random sample testing and reviews
- Ensure compliance with applicable nondiscrimination lawsThe fifth factor
Why Attestations Aren’t Enough
Vendor Governance Answers Half the Question
AVM providers invest heavily in responsible model development — input restrictions, protected-variable exclusions, internal validation, and fair lending attestations. That work matters. But vendor testing and attestations do not replace an institution’s responsibility for understanding and governing the AVMs it uses.
Independent evaluation has long played a role in AVM governance. Institutions need to understand how models perform against appropriate benchmarks, where limitations exist, and whether results remain reliable across different markets, property types, and uses.
Fair Housing adds another dimension to that responsibility: how does the model behave across different populations, and are less-disparate alternatives available?
A model can measure itself; it cannot tell you whether a less-disparate alternative exists.
No single AVM provider can fully answer that, because no provider can independently compare its outcomes using a common population, benchmark, and methodology.
Regulatory expectations increasingly require exactly that comparison — evidence that your institution evaluated alternatives and can document why the models you use were chosen.
Disparate Impact Takes Many Forms
One Model Can Disadvantage a Population in Different Ways
Disparate impact is not a single phenomenon, and no single statistic provides a complete view of potential disparity. A model can produce different outcomes across populations in several distinct ways:
- Less accurate valuations
- Valuation errors may simply be larger, on average, for one population than another.
- Lower hit rates
- One population may receive acceptable valuation outcomes less often — meaning fewer properties are valued within a defined performance tolerance.
- Higher variance
- Even with similar averages, valuations for one population may be more dispersed and less predictable, making individual results less reliable.
- Systematic over- or undervaluation
- A model may consistently value one population’s homes high and another’s low — a directional bias with direct consequences for borrowers and lenders alike.
A model can look equitable on one of these dimensions while exhibiting meaningful disparity on another. Two models with nearly identical overall accuracy can produce materially different Fair Housing results. That is why CFHA™ tests for disparate impact in multiple ways, rather than relying on any single measure.
Importantly, the goal is not to flag every difference — or even every statistically significant one, since at the sample sizes involved in AVM testing almost any difference registers as “significant” — but to identify differences that are meaningful: large enough to matter for borrowers, and for the governance decisions institutions must defend.
The AVMetrics Approach
Eligibility First. Performance Second.
AVMetrics’ Comparative Fair Housing Analysis (CFHA™) is a two-stage governance framework that deliberately separates Fair Housing assessment from performance optimization.
Stage 1
Fair Housing Eligibility Assessment
Each AVM is independently evaluated for Fair Housing outcomes within a defined geography. The question at this stage is not which model performs best — it is whether each model’s outcomes support its use, and under what conditions.
Stage 2
Operational Performance Optimization
Only after eligibility has been evaluated are models ranked on accuracy, precision, coverage, stability, and confidence-score behavior, then incorporated into geography-specific Model Preference Tables (MPT™), cascades, and routing logic.
The objective of Stage 1 is not to determine legal discrimination. It is to identify meaningful outcome differences and determine whether less-disparate operational alternatives are available — and to document that determination.
“Traditional AVM governance often begins with performance. Modern AVM governance begins with Fair Housing eligibility.”
How the Analysis Works
Independent. Comparative. Built for Governance.
The complete CFHA™ methodology is patent pending. At a high level, the analysis is:
- Independent
- Testing is performed by AVMetrics using independently controlled benchmark data. No AVM provider controls the data, the test design, or the evaluation criteria.
- Comparative
- Competing AVMs are evaluated side by side on a common methodology — providing a consistent basis for identifying whether a less-disparate alternative exists.
- Multi-metric
- Because disparate impact takes different forms, CFHA™ evaluates multiple complementary dimensions of model behavior rather than relying on any one measure.
- Magnitude-based
- At AVM sample sizes, trivial differences can be “statistically significant.” CFHA™ measures whether differences are large enough to matter for governance decisions.
- Geographically sensitive
- Analysis is performed at a granular geographic level, with data-sufficiency safeguards, so that localized disparities are not masked by broad averages — while governance decisions remain practical to deploy.
Detailed methodology documentation is provided as part of a CFHA engagement and is subject to appropriate confidentiality protections.
What You Receive
Independent, Documented Evidence for AVM Governance
CFHA™ engagements produce documentation designed to support institutional governance, model risk management, audit, and regulatory reviews:
- Fair Housing eligibility determinations for each AVM, by geography
- Comparative AVM analysis on a common, independent methodology
- Geography-specific Model Preference Tables (MPT™) of eligible models
- Deployment, routing, and cascade recommendations
- Plain-language governance narratives explaining each determination and its rationale
- Ongoing monitoring reports covering performance and Fair Housing outcomes
- Independent challenge documentation supporting model-selection decisions
The result is not simply more testing. It is the ability to document and explain why each model was selected, what alternatives were evaluated, and how Fair Housing considerations were incorporated into your institution’s AVM governance processes.
Why AVMetrics
The Independent Standard Since 2005
AVMetrics has spent more than 20 years building the industry’s independent AVM testing infrastructure: the Model Repository Database (MRD™), Predictive Testing Methodology (PTM®), and Model Preference Tables (MPT™). We test the vast majority of major commercial AVMs — over 20 models — on data no single vendor controls.
CFHA™ extends that same independent-testing philosophy to the fifth AVM Quality Control Standard — compliance with applicable nondiscrimination laws — providing institutions with independent, comparative evidence to support their Fair Housing and AVM governance processes.
FAQ
Common Questions
Does CFHA™ determine whether an AVM is discriminatory?
No. CFHA™ is not a legal determination. It identifies meaningful outcome differences across populations and whether less-disparate operational alternatives are available — the evidence institutions need to support their own compliance and governance decisions.
Our AVM vendors already provide fair lending attestations. Isn’t that sufficient?
Attestations describe how a model was designed. They cannot describe how it behaves in production relative to competing models. The two are complementary — and supervisory expectations increasingly encompass both.
Can a model be appropriate in one market and not another?
Yes. Eligibility is evaluated by geography, because AVM behavior varies across markets. A model whose outcomes support its use in one county may warrant restrictions, additional review, or exclusion in another.
Is this a one-time assessment?
No. Markets, models, and data change, and Fair Housing outcomes can change with them. CFHA™ is designed as ongoing governance, with monitoring and periodic reevaluation.
Eligibility precedes excellence.
Before you optimize AVM performance, make sure the models you deploy belong in the running. Talk to AVMetrics about independent Fair Housing evaluation for your AVM program.