Independent testing at national scale

Every optimization decision starts with knowing how each AVM actually performs where you lend. We test continuously throughout the year. Each quarter, we collect 500,000 to 700,000 recent residential sales nationwide, scrub and deduplicate them, and measure every participating model against them.

Most AVMs now draw on listing data, so a model can look accurate simply because it has already seen the list price. Our Predictive Testing Methodology (PTM®) evaluates each model’s estimate from before the property was listed, so it measures what the model knows on its own. That is the accuracy that matters when you value a property that isn’t on the market: a refinance, a home equity line, or a loan already in your portfolio.

Benchmark heat map of U.S. counties shaded by the number of properties tested, from 1–10 (dark green) to 10,000+ (dark orange); coverage is densest in California, Arizona, Florida and the Northeast.
Benchmark heat map: properties tested per U.S. county.

What we measure

Testing built around your portfolio

We can scope testing to the counties you lend in and the AVMs you actually use. We can also test internal or proprietary models against the same benchmarks we use for commercial AVMs.

For AVM developers

Most commercial AVM providers participate in our testing and receive independent results on how their models perform.

Fair Housing eligibility comes first

Traditional AVM governance often begins with performance. Modern AVM governance begins with Fair Housing eligibility. Only after Fair Housing considerations have been evaluated should performance optimization occur.

The interagency AVM quality control standards require institutions to address compliance with nondiscrimination laws alongside accuracy. Our Fifth Factor testing evaluates AVM outputs for disparate impact, so you can confirm that each model is eligible for use before you rank it for performance.

Eligibility precedes excellence.

Learn about Fifth Factor testing →

Rankings: the Model Preference Table®

A Model Preference Table® (MPT®) ranks AVMs by county and, within each county, by property type and price tier. Our MPTs are built from current testing results using a proprietary scoring algorithm, refreshed every quarter, and can be scaled to your footprint or to the AVMs you license.

An MPT also documents your model selection in advance. Interagency guidance has long cautioned against “value shopping”: trying several AVMs and choosing the highest value. A pre-set, test-based ranking shows examiners that model selection is driven by reliability, not by the result.

County1st choice2nd choice3rd choice
Maricopa, AZModel 34Model 13Model 10
Pima, AZModel 2Model 13Model 10
Yavapai, AZModel 2Model 34Model 10
Los Angeles, CAModel 15Model 2Model 13
Orange, CAModel 34Model 15Model 10
Sacramento, CAModel 2Model 10Model 34
San Diego, CAModel 15Model 2Model 10
Anonymized MPT excerpt.

Cascades: from rankings to decisions

An MPT tells you which AVM to try first. A cascade decides whether to accept the answer. Each estimate is checked against the model’s own confidence measures, and if it falls short, the cascade moves to the next model in the ranking. A custom cascade goes further, building your institution’s risk tolerances into that logic.

AVMModel Preference TableCascadeCustom cascade
Value estimateIncludedIncludedIncludedIncluded
Ranking of AVMsNot includedIncludedIncludedIncluded
Decision logicNot includedNot includedIncludedIncluded
Your institution’s risk tolerancesNot includedNot includedNot includedIncluded

Calibrated to how each model actually performs

Confidence scores and FSDs are only useful if they predict accuracy. We use correlation analysis to measure how closely each model’s self-reported confidence tracks its actual error, and we set acceptance thresholds accordingly. We also set price floors and ceilings at the price levels where a model’s accuracy decays, so that it isn’t used where it performs poorly.

Built around your risk policy

A custom cascade can apply tighter limits in some regions or above certain property values, and different standards for different uses, such as a home equity decision versus a portfolio review. Confidence thresholds are also how an institution puts into practice the first quality control factor in the interagency AVM standards: a high level of confidence in the estimates its AVMs produce.

Independent by design

Our rankings and cascades are built from cross-model testing and are not tied to any AVM provider’s financial interests.

Want a cascade like this without building and maintaining it yourself?

ComplianceTRACK®: our testing, ready to use

ComplianceTRACK puts AVMetrics’ independent testing to work inside your valuation process, without the cost of building and staffing a testing program of your own. It is available directly from AVMetrics or through select platform partners, and it is delivered by API.

Model selection

A cascade built on our Model Preference Tables, with acceptance thresholds calibrated to each model’s confidence scores and FSDs, and price floors and ceilings where accuracy decays.

County-level accuracy

The distribution of PPE10 across the models we test, for every county we can measure, so you can see where your primary AVM falls below the field.

Refreshed every quarter

Testing runs year-round. Rankings, cascade logic, and county data are updated quarterly.

Documented

Reporting that records how models were selected and how they performed, ready for your model risk and exam files.

See it for your county.

Our interactive demo shows how the AVMs we test performed in a county you choose: the spread of PPE10 across the anonymized models we test there, from the weakest to the best.

Other ways to work with us

Custom MPT and cascade design

Run your own AVM contracts and decision engine? We’ll design Model Preference Tables and cascade logic around your footprint, AVM licenses, and risk policy, and deliver them to run in your own systems, with documentation for model risk management and examiners.

Start a conversation →

AVM benchmarking data

For institutions with their own analytics teams: a standardized, quality-controlled dataset of recent sales and the values participating AVMs returned for them, so you can run your own analyses and build your own model preference frameworks.

Start a conversation →

Tell us how you use AVMs today.

Which AVMs do you use, in which markets, and for which decisions? Tell us, and we’ll show you how those models compare where you lend and what an optimized approach could look like.