Automated valuation models can price a home in seconds, but federal regulators now require mortgage originators and secondary-market issuers to run documented quality controls whenever they use an AVM for a covered credit or securitization decision. The rule took effect October 1, 2025, and it adds nondiscrimination to the older list of concerns around confidence, manipulation, conflicts and testing.
Mortgage lenders who ran automated valuation models without documented oversight lost the option on October 1, 2025. A federal rule now requires banks, credit unions and secondary-market issuers to build formal quality controls around every AVM used in a covered credit or securitization decision, including a new nondiscrimination standard. Nine months later, many institutions are still catching up.
An AVM Is an Estimate, Not an Appraisal
An automated valuation model is a computerized system which estimates a property’s value from data such as prior sales, property characteristics, comparable transactions, location and market trends. It is not a digital appraisal, and the distinction carries legal weight. Interagency appraisal guidance has stated for years an AVM result, by itself, does not meet minimum appraisal standards where an appraisal is required by law or policy.
Four categories serve different purposes, and conflating them creates risk for lenders: a consumer-facing estimate such as a Zestimate, a lender-grade AVM built for underwriting, an appraisal waiver or alternative valuation process, and a licensed appraisal completed by a certified professional. Zillow’s Zestimate belongs to the first category. Zillow itself describes the Zestimate as an automated estimate, not an appraisal, and says it cannot replace a professional appraisal. A national accuracy figure for a free consumer tool says little about how a lender-grade model performs on a rural refinance or an unusual property.
How the Models Produce a Value
AVMs pull from a wide range of inputs: public property records, deed and transaction history, tax assessments, listing data, square footage, lot size, age, bedroom and bathroom counts, and nearby comparable sales. Some systems fold in neighborhood characteristics, price indices and computer-vision analysis of listing photos.
Vendors build AVMs differently, and no single architecture dominates the industry. Common approaches include hedonic regression, repeat-sales models, comparable-sales algorithms, gradient-boosted trees and other machine-learning methods, and ensembles combining several estimates into one. A hedonic model paired with a repeat-transaction index is one common design, not an industry standard every vendor follows.
Most AVMs return more than a single number. Typical output includes a point estimate, a confidence score or forecast standard deviation, an estimated value range, and, in many products, a list of comparable properties or the factors driving the estimate.
Accuracy Changes With the Property and the Data
As of July 17, 2026, Zillow reported a nationwide median error rate of about 1.8% for on-market homes and about 7% for off-market homes. The gap illustrates how listing information and recent market signals can sharpen an estimate, but Zillow’s consumer-facing figures do not establish the performance of every lender-grade AVM.
A median error of 1.8% means half of eligible estimates land within 1.8 percentage points of the eventual sale price, and half land further off. The figure says nothing about how the model performs on any single property, including the one a loan officer is reviewing right now.
On-market models can pull in list price, listing descriptions and recent market activity, information which sharpens an estimate as a sale approaches. Off-market models work without list-price signals, so accuracy for a listed home is a poor proxy for a lender evaluating an unlisted refinance or a rural parcel.
Older figures circulating online, such as a claimed 2% to 4% on-market range and 5% to 10% off-market range for “top” AVMs, trace back to blended industry aggregates rather than a single verifiable study. Lenders evaluating a vendor should ask for the vendor’s back-tested numbers or an independent validation, not a figure pulled from a blog post.
A handful of terms show up in every serious accuracy conversation:
- Median absolute percentage error (MdAPE): the median percentage difference between a model’s estimate and the eventual sale price.
- Hit rate: the share of estimates landing within a defined band, such as 5%, 10% or 20% of sale price.
- Coverage: the share of properties for which the model can produce an estimate at all.
- Confidence score: a vendor-specific reliability signal; scales differ across providers and are not directly comparable.
- Forecast standard deviation: the estimated uncertainty around a point value.
The Five Controls in the Federal Rule
The OCC, the Federal Reserve Board, the FDIC, the NCUA, the CFPB and the FHFA jointly issued the rule. It applies specifically to mortgage originators and secondary-market issuers using an AVM to determine collateral value for a covered credit decision or securitization determination involving a consumer’s principal dwelling. Screening tools, internal monitoring dashboards and the estimate a real-estate agent shows a homeowner sit outside the rule’s direct reach.
Covered institutions must design controls aimed at five outcomes:
- A high level of confidence in the estimates produced.
- Protection against manipulation of data.
- Avoidance of conflicts of interest.
- Random sample testing and independent review.
- Compliance with applicable nondiscrimination laws.
The fifth factor is the newest. Dodd-Frank named the first four when Congress passed the act in 2010. The agencies added nondiscrimination during rulemaking. The addition folded a fair-lending requirement into what had been a data-quality framework.
The agencies chose a principles-based structure over a fixed formula. No rule text sets a minimum accuracy percentage, a mandated sample size or a specific fairness metric. The flexibility lets a five-branch credit union scale its controls differently than a national bank securitizing billions in mortgages every quarter, but it also means a vendor’s marketing scorecard alone is not evidence of compliance.
Where Bias Can Enter the System
The nondiscrimination factor exists because a model trained on skewed inputs can reproduce, and sometimes amplify, the patterns baked into the data. Federal regulators have not concluded every AVM discriminates. The concern is narrower and more mechanical: training labels shaped by historically unequal markets, data quality gaps between neighborhoods, proxy variables correlated with protected characteristics, uneven model coverage, and feedback loops connecting valuation, lending and future transaction data.
Compliance, legal and model-risk teams generally combine several checks rather than relying on one metric: comparing AVM estimates against the arm’s-length sale prices which follow, breaking out performance by geography and property type, examining full error distributions instead of a single average, tracking coverage and no-hit rates, testing outcomes across lawful demographic and proxy groupings, reviewing data lineage, documenting overrides, and monitoring drift over time.
The rule text does not name census-tract analysis as a required test. Census-tract review can be a useful method inside a broader program, but the regulation stops short of prescribing it, or any single geographic unit, as mandatory.
How to Evaluate an AVM Provider
Push your vendor past the sales deck. Ask for:
- Scope: the property types and transaction types the model is built and validated for.
- Validation history: the length of the training and validation periods, and whether an independent party reviewed the results.
- Segmented accuracy: MdAPE, hit rates and coverage broken out by geography and property class, not a single national average.
- Confidence calibration: how the confidence score maps to observed outcomes over time.
- Fair-lending testing: what disparity testing runs, at what cadence, and what happens when a gap turns up.
- Data lineage: where inputs come from, how frequently they refresh, and how fast a correction propagates.
- Change management: what triggers revalidation after a model or data update.
- Manipulation and conflict controls: who can influence an output, and how the vendor prevents it.
- Contract terms: audit rights and incident-notification requirements.
- Fallback logic: what happens automatically when a property returns a low-confidence score or no estimate at all.
How to Use AVMs Safely
You reduce exposure by building guardrails around the model, not just around vendor selection. Set a confidence threshold below which a human reviewer takes over. Escalate unusual properties, rural parcels and recent renovations to manual review by default. Validate vendor output against your historical data before trusting it at scale. Run more than one model in parallel for higher-risk transactions. Monitor for drift as local markets shift. Log every override with a reason attached. Match the valuation method to the transaction instead of defaulting to whichever tool returns an answer fastest.
Compliance Checklist
| Control area | Question to ask |
| Intended use | Is the AVM approved for this transaction and property type? |
| Accuracy | How does it perform by region, property class and market condition? |
| Confidence | What does the score mean, and is it calibrated against observed outcomes? |
| Data integrity | How does the provider prevent, detect and correct manipulated or erroneous data? |
| Conflicts | Does any party benefit from a higher or lower value? |
| Testing | What random sample tests and independent reviews occur? |
| Nondiscrimination | How are disparities detected, investigated and remediated? |
| Change management | What triggers revalidation after a model or data change? |
| Vendor oversight | What audit rights and incident notices does the contract provide? |
| Human escalation | When does the workflow require a different valuation method? |
Scale Requires Governance
AVMs will keep growing more central to mortgage lending because they deliver something appraisals cannot: speed, consistent coverage and a lower cost per file. The 2025 rule does not challenge the value AVMs deliver. It challenges the assumption a fast estimate is automatically a sound one. Lenders who treat AVM output as a governed input, checked against data quality, fairness and human judgment, will meet the standard. Lenders who treat a score as a self-validating answer will find out the hard way: speed and reliability were never the same thing.
