Executive Summary

Automated Valuation Models (AVMs) – algorithmic systems that estimate real estate values without (or with minimal) human judgment – have become central to contemporary collateral valuation and risk management in finance. In the past decade, their use has expanded significantly, driven by technological advances (big data, machine learning, cloud computing) and by industry demand for faster, cheaper, and more consistent valuations. This evolution has, however, raised important questions about AVM quality and model risk. Regulators and industry participants are acutely aware that an AVM’s predictive accuracy, transparency, and governance critically affect credit risk, lending decisions, and fair lending compliance. In response, U.S. federal agencies have finalized new rules (effective 2025) mandating quality controls on AVMs in residential mortgage valuation (Source: www.occ.treas.gov).

This report provides a detailed state-of-play as of 2026 regarding AVM quality and model risk. It synthesizes historical context, current practices, research findings, case studies, and future trends. Key findings include:

• Regulatory Push: The 2024 Interagency Final Rule (OCC, Fed, FDIC, CFPB, NCUA, FHFA) enshrines risk management standards for AVMs (FIRREA §1473(q), requiring lenders and securitizers to establish controls ensuring “a high level of confidence” in AVM estimates, protect data integrity, avoid conflicts of interest, perform random back-testing, and comply with fair lending laws (Source: www.occ.treas.gov) (Source: kpmg.com). The regulators emphasize that institutions “own the risk” of any AVM they use (Source: kpmg.com). This formalizes the importance of rigorous model validation and governance.

• AVM Accuracy and Metrics: Modern AVMs, especially those using machine-learning (e.g. random forests, gradient boosting, neural nets), can achieve substantial accuracy. Empirical studies report median absolute errors typically in the single-digit percentage range. For example, Zillow’s Zestimate historically had a 7.7% median absolute error on initial list price, improving to 3.6% when measured just prior to sale (Source: www.zillow.com). Independent studies (using large transaction samples) find AVM-based estimates often exceed traditional appraisals in unbiasedness, especially in rural areas where appraisals are upward-biased (Source: link.springer.com) (Source: link.springer.com). CoreLogic notes that data and model enhancements can improve the AVM’s “within-10% of sale price” accuracy by ~8–9% (Source: www.corelogic.com.au). However, AVM performance can vary: rural and heterogeneous markets (or highly volatile conditions) present more challenges, as AVMs require quality data and homogeneity (Source: www.rics.org).

• Model Risk and Uncertainty: Use of AVMs introduces model risk – the possibility of errors due to flawed model design, data issues, or misuse. Recognizing this, banks treat AVMs as any quantitative model, subject to standards (e.g. OCC’s Model Risk Management Guidance) requiring robust validation, backtesting, and documentation (Source: www.occ.treas.gov) (Source: link.springer.com). The recent literature emphasizes the need to quantify an AVM’s predictive uncertainty. A 2024 study introduced a framework to generate an “AVM Uncertainty Estimate (AVMU)” that correlates with actual error, allowing users to gauge confidence in each valuation (Source: link.springer.com) (Source: link.springer.com). Continuous monitoring and recalibration are now best practices, often using random sample testing of AVM estimates against subsequent actual sale prices (back-testing) as mandated by the rule (Source: www.occ.treas.gov) (Source: kpmg.com).

• Data Quality and Fairness: AVMs are only as good as their data. High-quality, up-to-date inputs (sales records, tax data, listings, geographic attributes, etc.) are essential (Source: www.rics.org). Data gaps (e.g. missing condition or retrofit information) can degrade accuracy and introduce biases. Researchers and regulators increasingly scrutinize AVMs for fair lending risks. Some studies (Urban Institute, 2024) have found that typical AVMs still yield larger percentage errors in majority-minority census tracts, suggesting residual effects of historical bias (Source: www.urban.org). Conversely, industry analyses (e.g. Veros 2022) claim no systematic racial bias in a professional AVM’s residuals (Source: www.veros.com). The interagency rule explicitly requires institutions to ensure AVM policies prohibit discriminatory outcomes (Source: www.occ.treas.gov) (Source: www.alstonconsumerfinance.com), aligning with the Biden Administration’s PAVE initiative on appraisal equity. As a result, model risk management programs must now explicitly address data and algorithmic fairness (e.g. excluding protected characteristics, auditing for disparate impact).

• Use Cases and Trends: AVMs today underpin many mainstream valuation processes. For routine retail mortgages, FannieMae and FreddieMac have expanded valuation waivers (automatic appraisals waivers) contingent on AVM results (Fannie Mae’s “Collateral Underwriter” flag, Freddie’s ACE), significantly reducing appraisal rates. By early 2025, appraisal waivers accounted for ~13–15% of GSE loan volume (Source: mtgefi.com), largely facilitated by AVMs meeting agency confidence criteria. Lenders (banks, fintechs, credit unions) continue to integrate AVMs into credit underwriting and portfolio monitoring. Commercial real estate and other asset classes, though more heterogeneous, are seeing growing AVM pilot projects where data allow. (Source: www.rics.org)

• Future Outlook: Looking ahead to 2026 and beyond, AVM quality is expected to improve via new data and AI. Advances in computer vision, natural language (mining listing descriptions), and geo-spatial analytics are augmenting AVMs with richer signals (Source: journals.plos.org) (Source: www.corelogic.com.au). Regulatory attention will remain high: updates to fair lending rules and accounting standards may tighten how AVMs are governed. Moreover, climate risk and alternative data (e.g. utility, transaction frequency) represent emerging inputs that could be integrated.Overall, the “state of play” is one of rapid adoption constrained by rigorous risk management – institutions must balance efficiency gains with careful controls and transparency.

This report elaborates these points in detail, drawing on academic research, industry reports, regulatory releases, and real-world examples. It provides an evidence-based analysis of AVM methodologies, quality metrics, model risk frameworks, use cases, and future developments, with extensive citations to the latest credible sources.

Introduction

Background on Automated Valuation Models

An Automated Valuation Model (AVM) is broadly defined as “any computerized model used by mortgage originators and secondary market issuers to determine the value of a consumer’s principal dwelling collateralizing a mortgage.” (Source: kpmg.com). Conceptually, AVMs replace or supplement traditional appraiser-driven valuations with algorithmic price estimates. They typically combine a hedonic pricing approach (using property attributes and comparable sales) with modern machine-learning techniques, and can incorporate diverse data (public records, multiple listing service (MLS) data, geographic information systems (GIS), and more).

The idea of algorithmic home valuation dates back decades, but it has gained traction with the proliferation of digital data and computing power. Notably, online real-estate platforms (e.g. Zillow, Redfin) popularized AVMs in the mid-2000s, offering Zestimate and Redfin Estimate to consumers. In parallel, banks and mortgage investors developed internal AVMs for collateral risk assessment (e.g. early risk models in the 1990s/2000s (Source: link.springer.com). The banking regulators anticipated the rise of AVMs: Section 1473(q) of the Dodd–Frank Act (2010) amended FIRREA to mandate that mortgage lenders adhere to quality standards for AVMs used in valuing residential collateral (Source: www.occ.treas.gov). That statute foreshadowed regulatory action and reflects the systemic importance of real-estate valuation in the financial system.

Uses and Motivations

Speed, cost, and scalability are the principal drivers for AVM adoption (Source: www.rics.org) (Source: www.rics.org). A fully automated valuation can be generated in seconds for any property with available data, dramatically reducing appraisal turnaround time. In mortgage origination, forbearance servicing checks, and portfolio monitoring, AVMs allow lenders to screen collateral values en masse. They also improve consistency across markets (fewer “human errors”) (Source: www.rics.org). The COVID-19 pandemic accelerated this trend: travel restrictions and social distancing constraints forced the adoption of desktop or hybrid appraisals, and AVM waivers, proving that valuations could be produced remotely without sacrificing prudence (Source: sf.freddiemac.com) (Source: www.corelogic.com.au). Major mortgage agencies (Fannie Mae, Freddie Mac) responded by rolling out “Collaboration Underwriting Evaluations” (CUE) tools and standardized desk/hybrid apartments, eventually enabling widespread appraisal waivers tied to AVM scores (Source: sf.freddiemac.com) (Source: mtgefi.com).

At the same time, AVMs offer risk mitigation benefits. Traditional appraisals have known limitations – studies document consistent upward bias (overestimation) especially in rural or volatile markets (Source: link.springer.com) (Source: link.springer.com) – likely due to appraiser incentives to confirm contract price. AVMs, being objective algorithms, are inherently free of appraisal commission bias. This has led to the view that AVMs could flag potential human bias or even serve as a corrective “second opinion” (for example, Freddie Mac suggests that an appraisal far below an AVM’s value be subject to review) (Source: www.veros.com). As one analyst notes, “an objective, cost-effective solution would be to use [a professional AVM] to check if appraised values are in agreement and, if so, deem them low risk for bias (Source: www.veros.com).” Models are also updated daily from market data, affording them immediate responsiveness to price movements that periodic appraisals cannot match.

Challenges and Concerns

Despite their promise, AVMs come with caveats that inform the quality and risk discussion. AVM estimates can be poor or volatile when data inputs are sparse or nonrepresentative. According to RICS (Royal Institution of Chartered Surveyors), “AVMs work best with widely traded, homogeneous assets; their performance degrades with increasingly heterogeneous assets or assets that are thinly traded” (Source: www.rics.org). In practice, that means standard suburban homes with ample comps are ideal; unique properties, condos with complex fee structures, or very-high-end homes may challenge AVM assumptions.

Data integrity is critical: AVMs rely on massive datasets, and “garbage in, garbage out” applies fully. High-quality inputs must be current, accurate, and comprehensive (Source: www.rics.org). This involves not just numeric variables (square footage, number of beds) but also elusive ones like property condition. Missing or outdated data can systematically degrade AVM reliability (e.g. if sold home attributes are not reported quickly). Therefore, transparency about data sources and rigorous data governance are core to AVM quality (Source: www.rics.org).

Another major concern is fair lending and bias. As discussed further below, AVMs must not perpetuate or exacerbate historical inequities. Since housing data often reflects segregation and lending disparities, algorithms can inadvertently encode bias. Indeed, recent research has found that AVM errors tend to be larger (as a percentage of value) in majority-Black neighborhoods, suggesting systematic underprediction of home values in those communities (Source: www.urban.org). This is a key part of model risk, since regulatory fair lending laws apply even to automated models. U.S. regulators now explicitly require AVM quality controls include measures to prevent discriminatory outcomes (Source: www.occ.treas.gov) (Source: www.alstonconsumerfinance.com).

Finally, operationally, AVMs are regularly updated (e.g. weekly or daily) which itself introduces risk (concept drift). Models must be continually monitored and revalidated, especially after rapid market changes. The integration within lending workflows also means end-users (loan officers, automated underwriting systems) may trust an AVM output without understanding its uncertainty. Managing this “model risk” (the risk of incorrect decisions due to model limitations) is a paramount theme for lenders.

In sum, AVMs are a powerful tool with proven benefits, but they demand rigorous attention to data, validation, governance, and fairness to ensure they truly enhance rather than impair credit risk management. The state of play in 2026 reflects a maturing field: sophisticated models in production, extensive oversight frameworks, and ongoing research into uncertainty and bias mitigation. The following sections unpack these dimensions in depth, with data and expert perspectives.

Historical Evolution and Current Adoption

Early Development of AVMs

The roots of automated valuation lie in statistical hedonic pricing models developed in academic economics. As housing transaction data became more available in the late 20th century, analysts began using regression-based models (e.g. Mundlak, RIP – Repeat Income, etc.) to estimate property values from characteristics. However, these early models were simplistic by today’s standards and mostly used by governments for tax assessment or research (Source: link.springer.com).

In the 2000s, two parallel developments spurred AVM adoption:

1. Consumer Internet Platforms: Companies like Zillow (founded 2006) and Redfin launched homeowner-facing AVMs (Zestimate, Redfin Estimate). Zillow’s public tech blog later reported that in 2016 their Zestimate had a median absolute percent error of 7.7% on initial listing prices (versus 4.2% for the listing agent’s price) (Source: www.zillow.com). This public data building broadened social awareness of “algorithmic home values,” but these systems were largely for consumer reference and not regulated.

2. Institutional Models: Mortgage lenders and servicers started in-house AVMs to support appraisal reviews and portfolio oversight. For example, Veros (now CoreLogic) developed AVM products in the 2000s for lenders, and model guidelines (e.g. OCC 2000-16 Model Validation) existed to ensure prudent use. However, AVMs were initially a complement – appraisals remained the legal valuation standard.

GSEs and Early Regulatory Interest

The Federal Housing Finance Agency (FHFA), overseeing Fannie Mae and Freddie Mac, along with banking regulators, watched AVMs cautiously. GSE selling guides gradually allowed appraisal waivers (no appraisal required) if an automated value is acceptable. These have evolved into the current CUE (Collateral Underwriter Evaluation) scores and ACE (Automated Collateral Evaluation) systems, which rank an AVM’s deviation from an “expected” value. Early in the 2010s, usage was limited, but over time these programs expanded. (Detailed stats on waiver volumes are provided later.)

Regulators also signaled oversight: in 2011 the OCC issued its “Sound Practices for Model Risk Management” (Source: www.occ.treas.gov), reinforcing that any quantitative model (including AVMs) requires robust lifecycle governance. Although not AVM-specific, it framed model validation, governance, and use standards. Over the decade, several interagency appraisal and evaluation guidelines (e.g. Interagency Appraisal & Evaluation Guidelines) mentioned AVMs as a tool for “evaluations” (less formal assessments) but required careful controls.

Pandemic Era and Surge in Use

The COVID-19 pandemic in 2020–2021 dramatically accelerated AVM adoption in mortgage collateral valuation. With in-person appraisals subject to restrictions, lenders and GSEs moved aggressively to digital alternatives:

• Desktop and Hybrid Appraisals: Lenders used exterior-only or hybrid appraisals, often supported by AVM pre-scoring. Fannie Mae and Freddie Mac co-developed standardized forms for these (Residential Appraisal Report (expansion of commonly known as 2095 form) and multi-appraisal).
• Appraisal Waivers via AVM: Automated underwriting systems (AUS) began offering more appraisal waivers. Borrowers with simple, low-risk profiles could get loans with just an AVM-derived value (and skip an appraiser entirely). Notably, Freddie Mac’s Automated Collateral Evaluation (ACE) allowed appraisal waivers on Purchase & Refi transactions meeting criteria (Source: sf.freddiemac.com).

Freddie Mac public reports in early 2023 emphasize this shift: they describe appraisal waivers and inspection-based waivers (ACE+PDR) as established tools for reducing cost and delay while focusing appraiser effort on complex cases (Source: sf.freddiemac.com) (Source: sf.freddiemac.com). In interviews, Freddie executives noted that these innovations “addressed painfully long wait times for traditional appraisals” during the volume surge (Source: sf.freddiemac.com).

Meanwhile, technological improvements continued apace. Providers like CoreLogic consistently retrain AVMs with new data (including non-traditional inputs like auction listings or text analysis) to adapt to sharp market changes (Source: www.corelogic.com.au) (Source: www.corelogic.com.au). For instance, CoreLogic reported adding over 50,000 new data points in 2023, improving its model’s performance by up to 9% (Source: www.corelogic.com.au). Such iterative model refinement has become standard practice among leading AVM vendors, ensuring AVMs remain “in tune with volatile markets” (Source: www.corelogic.com.au).

Market Penetration and Usage Statistics

By the mid-2020s, AVMs are widely embedded in mortgage lending and other real estate sectors. While precise market share data is proprietary, several indicators illustrate penetration:

• Lending Industry Surveys: In bond prospectuses and 10-K filings, large banks often report back-testing statistics for AVMs. For example, portfolio filings note that loans underwritten with AVMs undergo random audits of values, comparing AVM estimates to subsequent appraisals or sales (Source: growthmarketreports.com). (SEC filings from institutions like Ally Financial mention these processes, showing institutional reliance on AVMs.)
• GSE Loan Delivery Data: Public data on Fannie Mae and Freddie Mac deliveries shows that a growing fraction of loans utilize AVM waivers. Mortgagetech news (MtgeFi) reported that in March 2025, roughly 13.5% of GSE loan balance was delivered using appraisal waivers (Source: mtgefi.com), and an additional 2.05% using inspection-based waivers (leveraging an AVM + site inspection) (Source: mtgefi.com). The combined usage of waivers (AVM-based or inspection AVM) was about 15.5% of loans by balance (Source: mtgefi.com). This marks a clear increase from early-2020 levels when waivers were rare vs. nearly half of nonbank lenders now routinely use them (Source: mtgefi.com).
• Global Market Growth: Market research firms project robust growth. For instance, a 2024 industry report estimated the global AVM market at over $4 billion in 2023 and forecasted double-digit annual growth (reaching perhaps $10+ billion by 2030). While these numbers vary by source, the consensus is that AVMs are one of the fastest-growing segments of proptech (real estate technology).

In housing markets outside mortgages, AVMs appear in insurance underwriting (flood/hurricane claims estimation), commercial RE portfolio management, and as the basis for indices (e.g. Zillow U.S. Home Value Index). Retail consumers (homebuyers/sellers) increasingly expect an AVM number as part of any online home search. Meanwhile, appraisal firms are integrating AVMs into desktop-review workflows.

In summary, by 2026 AVMs have moved from experimental to mainstream. Major lenders incorporate AVMs into origination decisions, and regulators have responded accordingly. The next sections delve into the technical and risk management aspects of these models and their governance.

Technical Foundations of AVMs

Methodologies and Algorithms

At their core, AVMs implement statistical or machine-learning models that predict a property’s market value given observable inputs. Broad methodological categories include:

• Hedonic Regression: A traditional approach where price is modeled as a linear (or log-linear) function of features (sqft, beds, age, location fixed effects, etc.). Referred to as hedonic pricing models (HPM), these are transparent and explainable but can struggle with nonlinearities and interactions. They are still used as benchmarks and in some AVMs, especially where interpretability is valued (Source: journals.plos.org) (Source: link.springer.com).

• Ensemble ML Models: Most modern AVMs use ensemble decision-tree methods like Random Forests, Gradient Boosting Machines (XGBoost, LightGBM, etc.), or ensemble neural networks. These capture complex nonlinearities and interactions without requiring a priori feature engineering. For example, Tapia et al. (2025) employed LightGBM and convolutional neural net feature extraction for images, achieving superior precision over hedonic models (Source: journals.plos.org). Random forests are especially popular: one study reported an out-of-sample R² of 0.722 and RMSE of 0.297 for a random forest AVM, significantly improving on a basic hedonic regression (Source: link.springer.com). However, these gains often come at the cost of computational complexity and opacity (interpreting millions of trees is hard).

• Neural Networks & Deep Learning: Researchers have explored deep neural nets (fully connected or convolutional) for AVMs, especially to process images or large feature sets. Such models can ingest novel inputs (satellite imagery, street-view images, text from descriptions) that conventional models cannot easily use. For example, Tapia et al. (2025) found that adding CNN-based image features to an AVM substantially reduced error margins (Source: journals.plos.org). Another study used CNNs tied with hedonic inputs to account for geographic variation, showing promise in capturing subtle spatial patterns (Source: link.springer.com). These approaches are cutting-edge and mostly in research or pilot stages, as they require extensive data and careful training.

• Hybrid Models: Some systems combine techniques. For instance, a layered approach might use hedonic regression for base-level value, then adjust via a machine-learning model for residual patterns, or use ensembles of different model types. Meta-modeling (stacked generalization) has been shown to improve uncertainty estimation: Schulz et al. (2024) combined multiple uncertainty estimators via voting regressor and stacking, finding this yielded the best alignment of predicted vs actual uncertainty (Source: link.springer.com).

The choice of technique involves trade-offs: more complex ML models tend to yield higher raw accuracy (R² and lower RMSE) but are harder to explain. Transparency is a non-trivial requirement: some stakeholders worry about “black-box” outputs. Explainability tools (SHAP values, partial dependence) are often used to attribute importance to features and ensure the model reasoning is sensible (Source: journals.plos.org). Nevertheless, regulatory guidance generally accepts black-box models if validation is rigorous, but demands mechanisms for understanding risk and bias from them.

Data Inputs and Features

AVMs exploit a rich variety of data:

• Property Attributes: Physical characteristics (square footage, bedrooms, bathrooms, lot size, year built, construction type, etc.), which usually come from assessor records or MLS. Some advanced AVMs also use textual descriptions or categorical info (e.g. architectural style, presence of pool, basement finish level).

• Comparables (“Comps”): Nearby recent sale prices of similar properties form the traditional heart of valuation. AVMs quantify this via features like distance-weighted sale prices, or by including dummy/location effects.

• Geospatial Data: Location coordinates allow calculation of distance to amenities (schools, parks), neighborhood demographics, walkability indices, flood zones, and clustering. Satellite imagery or environmental data can add value (e.g. soil stability, tree cover, proximity to water). These “geospatial attributes” can improve accuracy versus a pure address code (Source: www.rics.org).

• Temporal Data: Market-level trends by zip code or city, or macroeconomic variables (interest rates, unemployment) may be included implicitly (via time dummies) or explicitly as predictors.

• Market & Administrative Data: Market indices (Case-Shiller type), tax assessments, price histories, permits, owner occupancy rates, etc. For portfolios, lenders also track price changes in loan pools over time.

Quality and recency of data matter immensely. RICS cautions that AVM performance depends on data recency, provenance, and consistency (Source: www.rics.org). Having the latest listings and sales is crucial: CoreLogic’s enhancements (e.g. ingesting undisclosed sales and auction data (Source: www.corelogic.com.au) exemplify how new data can materially improve model fit.

Uncertainty and Confidence Measures

Most AVMs output a single point estimate. However, understanding and communicating prediction uncertainty has become a focus of research. Recent studies have proposed methods to quantify the error bounds of AVM predictions:

• Direct Uncertainty Estimation: Methods like quantile regression forest or conformal prediction can provide prediction intervals. Schulz et al. (2024) evaluated six uncertainty estimators (including direct loss modeling, bootstrap, quantile regression) and then calibrated them so that their predicted confidence matched observed errors (Source: link.springer.com). They showed that ensemble “voting” and “stacked” meta-estimators gave the most reliable uncertainty estimates (Source: link.springer.com).

• AVM Uncertainty Score (AVMU): The 2024 study introduced an “AVM Uncertainty” estimate, which they calibrated so that higher AVMU implied higher actual error (Source: link.springer.com). Practically, a lender could use AVMU to flag valuations that are particularly uncertain before the actual sale.

• Confidence Score or Coverage: Industry AVM providers often supply a confidence score or probability that the true value lies within a certain range. The OCC’s final rule aligns with this: institutions must ensure “a high level of confidence in the estimates produced by AVMs” (Source: www.occ.treas.gov). In practice, back-testing an AVM (comparing predicted values vs actual sales ex-post) is used to compute metrics like hit-rates (e.g. “what fraction of AVM values fall within ±10% of sale prices”). CoreLogic’s report cites improvements in “within-10% accuracy” as a performance metric (Source: www.corelogic.com.au).

In summary, modern AVMs strive not just to predict a central estimate, but to quantify the precision of that estimate. This makes the model’s output more actionable in risk frameworks: for example, loans where the AVM’s confidence is low might require an enhanced appraisal.

Assessing AVM Quality and Performance

Measuring an AVM’s quality involves multiple dimensions: accuracy, bias, coverage, and reliability over time. In practice, the most common metrics used by both academia and industry include:

• Median Absolute Percentage Error (MdAPE) or Mean Absolute Error: How far, on average, the predicted value is from the actual sale price. Zillow reports MdAPE (which they call “median absolute percent error”) of 7–8% (Source: www.zillow.com) for their estimates at listing.

• R-squared and RMSE: Statistical fit measures; e.g. a Random Forest AVM achieved out-of-sample R² ≈0.7224 (with RMSE=0.2971) in one U.S. study (Source: link.springer.com), outperforming a baseline hedonic.

• Coverage Metrics: Fraction of valuations that lie within given tolerance (e.g. ±5%, ±10% of actual price). CoreLogic regards “within-10% of sale price” rates (later described as 8–9% improvement) (Source: www.corelogic.com.au).

• Bias/Calibration: The systematic tendency to under- or over-predict. Some models gamble high (overestimate market values); others might be conservative. The Zillow analysis noted that bureaucratic list prices tended to skew high, whereas Zestimates were symmetrically distributed (Source: www.zillow.com). Appraisals historically skew upward (90% confirmed or exceeded sale price) (Source: link.springer.com) (Source: link.springer.com), whereas well-calibrated AVMs strive for neutrality.

• Coverage/Availability: The percentage of properties the AVM can actually value (i.e., has enough data). A national AVM may not cover all rural or new-build segments equally, leaving "no value" in data-poor areas. Regulators emphasize “coverage” as a quality control (i.e., the model must be monitored for gaps).

• Churn and Stability: How volatile the AVM is to small data changes. A highly erratic model indicates overfitting.

Empirical Findings on Accuracy

Numerous studies have analyzed AVM performance versus appraisals:

• General Accuracy: Research largely finds that well-tuned AVMs can match or surpass traditional appraiser accuracy, particularly in aggregated statistics. For example, the Freddie-vs-Appraiser study (Urban Institute) revealed appraisals often come in ~5% above contract price, whereas an optimized AVM tends to have lower mean error (implying AVMs might show less bias) (Source: link.springer.com) (Source: link.springer.com). In rural areas especially, AVMs outperformed appraisals in unbiasedness (Source: link.springer.com) (Source: link.springer.com).

• Influence of Model Complexity: Linear models (hedonic, lasso, ridge, elastic net) often plateau in performance. Schulz et al. found that elastic net (a linear ensemble method) performed between lasso and ridge, but still did not greatly outperform the simple hedonic baseline (Source: link.springer.com). Tree-based ML models (random forests, boosting) added significant value: in their data, random forest increased R² by ~6.1 percentage points over baseline and reduced RMSE (Source: link.springer.com). This is consistent with broad consensus that non-linear methods capture complex market relationships that linear models miss, yielding material accuracy gains. IFR, in the PLOS Chile case, LightGBM + CNN outperformed spatial hedonic pricing, especially when visual/text features were added (Source: journals.plos.org) (Source: journals.plos.org).

• Geographic and Market Effects: AVM accuracy is not uniform across all markets. The RICS analysis notes that as assets become more heterogeneous or thinly traded (e.g., rural farmland, unique luxury homes, or rapidly changing urban markets), AVM performance declines (Source: www.rics.org) (Source: www.rics.org). Indeed, rural areas (fewer comps) historically saw appraisals go high, but even AVMs can struggle if there are too few sales. A multi-year analysis (Fannie/Freddie 2012–2016) found ~25% of rural appraisals were >5% above sale price, versus 12.7% in urban; an AVM applied to those loans reduced the bias substantially (Source: link.springer.com). In practice, lenders often restrict AVM-based waivers to properties in well-covered areas for this reason.

• Time Stability: Lenders perform periodic re-testing of their AVMs to ensure they hold up over time. For example, CoreLogic’s iterative back-testing in the press release (Source: www.corelogic.com.au) demonstrates industry practice of segregating out-of-sample data. Models tuned to Q4 2021 data were retested on 12-month holdouts, leading to empirically measurable improvements before deployment (Source: www.corelogic.com.au).

• Consumer AVMs: Independent testers have examined consumer AVMs. A 2019 analysis noted Zillow’s typical error (median) was about $18,000 nationally (Source: realestatedecoded.com). While raw dollar error depends on region, relative metrics like percentage error provide context. The 2017 Zillow research note collected ~2 million sales from 2016 and found the distribution of Zestimate vs sale price was symmetric, unlike list prices which skewed high (Source: www.zillow.com) (Source: www.zillow.com). Overall, they reported ~84% of Zestimates were within ±20% of final sale price (Source: www.zillow.com), compared to 92% for list prices.

Bias and Fairness in Valuations

AVM model quality must be evaluated not just on overall accuracy but on justice and equity. Appraisals have long exhibited racial disparities, with minority homeowners often receiving lower appraised values. AVMs were proposed by some as a way to reduce human bias, since they do not “see” race (though they do see correlated proxies). Key research points:

• Urban Institute (2023): This study found that even after improving data, the percentage error of AVM valuations remained higher in majority-Black neighborhoods (Source: www.urban.org). That suggests AVMs, given the same inputs, still underestimate or deviate more on such properties. Possible causes include omitted variables (e.g. condition not captured) and historical market segregation patterns. The authors caution that AVMs do not automatically eliminate systemic fire-hosing of minority areas (Source: www.urban.org). This confirms that AVM model risk includes fairness risk.

• Industry Study (Veros 2022): In contrast, Veros (an AVM vendor) analyzed Chicago ZIP codes and reported no evidence of racial bias in their AVM’s undervaluation rates (Source: www.veros.com). They found no correlation between undervaluation and neighborhood race/ethnicity proportions. They argue that professional AVMs, carefully designed, can serve as impartial comparators to flag anomalies. Veros suggests using AVMs to sanity-check appraisals, and escalate if large discrepancies appear (Source: www.veros.com).

• Regulatory Focus: Responding to concerns, regulators require AVM programs to explicitly guard against bias. The final rule adds as a quality control factor that AVMs comply with all applicable unfair-discrimination/nondiscrimination laws (Source: www.occ.treas.gov) (Source: www.alstonconsumerfinance.com). Regulatory commentary links this to the White House PAVE Task Force on appraisal equity (Source: www.alstonconsumerfinance.com). The interagency guidance clarifies institutions must establish procedures to ensure AVMs do not produce values reflecting discriminatory bias (Source: www.alstonconsumerfinance.com). In other words, the model risk management framework must integrate fairness testing (e.g. analyzing errors by cohort) and remove any proxies for protected status.

• Ongoing Research: As of 2026, the debate is unresolved. Several research initiatives (FHFA, CFPB, academia) continue to investigate algorithmic bias in valuation. Proposed approaches include adjusting AVM training data (debiasing), or explicitly correcting AVM outputs, though these are complex given housing market context. Most experts agree that human oversight remains needed: flagged valuations (AVM vs appraiser big gaps) should trigger detailed review, combining algorithmic and expert judgment.

In summary, unbiasedness is a core desired attribute of AVM quality, yet ensuring it is non-trivial. Institutions must monitor model residuals and auditing performance across subgroups, and regulators will require documentation of these processes going forward.

Model Risk Management and Quality Controls

Given the predictive nature of AVMs, they are subject to model risk – the risk of adverse outcomes from model error. Banking regulators treat any significant quantitative model under their purview. The key elements of an AVM model risk framework include:

Governance and Policy

Financial institutions must incorporate AVMs into their overall model risk governance frameworks. According to industry guidance [10†L27-L36], a sound model risk program covers the full lifecycle: model development, implementation, use, validation, and monitoring. For AVMs specifically, this means:

• Establish policies and procedures that define acceptable use cases (e.g. which loans use AVM valuations, how results are reviewed).
• Define roles and responsibilities: e.g. model owners (who updates and recalibrates the AVM), validators (independent reviewers who test it), and business users (loan officers or credit officers using AVM outputs).
• Maintain documentation (model description, methodologies, assumptions) and change logs for the AVM (especially important for ML models where retraining is frequent).

Regulators demand that AVM control systems be scaled to the size and complexity of the institution (Source: www.occ.treas.gov) (Source: kpmg.com). Small banks might have simpler programs, whereas large banks (or those securitizing loans) need more formal validation teams.

Validation and Testing

The core of model risk control is validation: independent testing of the AVM before and during its use. Key validation activities include:

• Back-testing: Randomly sample loans where an AVM estimate was used and later replaced by an appraisal or had a sale, comparing the two values to assess error. This quantifies performance metrics (error distribution, bias, coverage). The OCC rule explicitly requires “random sample testing and reviews” of AVM estimates (Source: www.occ.treas.gov).
• Benchmarking: Compare the AVM against alternative models or external benchmarks periodically. For instance, banks might rerun a simpler hedonic model monthly to see if the machine-learned AVM still provides material added accuracy (Source: link.springer.com).
• Stress Testing: Evaluate AVM performance under hypothetical market scenarios (e.g. sudden price drops, interest rate shocks). This is analogous to stress testing credit models – one might feed outdated market data to simulate obsolescence and measure degradation.
• Independent Reviews: For third-party AVMs (vendor-supplied), institutions still must validate them themselves. Vendors often provide proprietary AVMs, but banks are required to audit model selection, data handling, and outputs. As KPMG notes, regulators will expect stringent third-party oversight as part of AVM MRM (Source: kpmg.com).

Throughout these tests, special attention is paid to the five quality control factors mandated by law:

• Estimate Confidence: ensure statistically that the AVM's expected accuracy meets a high standard (e.g. median error thresholds).
• Data Integrity: verify that inputs are reliable and checks exist to prevent data manipulation (which could artificially inflate or deflate valuations) (Source: www.occ.treas.gov).
• Conflict Avoidance: guard against scenarios where AVM design might benefit a party (e.g. if a lender owns the AVM vendor, RSI).
• Independence: maintain separation between AVM developers and approvals, so models are not rubber-stamped.
• Fair lending compliance: include bias audits and procedural compliance with nondiscrimination laws (Source: www.alstonconsumerfinance.com).

Adjustments and Controls

A unique aspect of AVM model risk is that institutions use or adjust AVM outputs rather than always taking them at face value. For example, Freddie Mac’s Selling Guide may require the underwriter to consider AVM output alongside other factors; if the AVM suggests a value significantly lower than an appraisal, the lender must either reconcile or require additional appraisal work. The final rule clarifies that AVMs used in cooperation with third parties (e.g. lenders following GSE appraisal waiver conditions) still fall under quality controls (Source: kpmg.com).

Importantly, the rule excludes AVMs only when used by licensed appraisers to develop a formal appraisal (Source: www.alstonconsumerfinance.com). In other words, if an appraiser is preparing or reviewing an appraisal report, the AVM they use is not subject to the same quality controls (since that process is governed by USPAP standards, not this regulation (Source: www.alstonconsumerfinance.com). However, any lender or issuer using AVM values for loan approvals, securitization, or waivers must comply (Source: www.alstonconsumerfinance.com) (Source: www.alstonconsumerfinance.com).

Regulatory Landscape

United States: As discussed, the key recent policy is the AVM Quality Control Final Rule (Bulletin 2024-17 by OCC, jointly with other agencies) (Source: www.occ.treas.gov). It becomes effective in 2025 and instructs institutions to create formal AVM policies as part of their existing governance frameworks. Notably, it does not “approve” specific AVM methodologies, but enforces a process framework. The regulators have stated they will still examine institutions’ use of AVMs under existing guidance (e.g. Interagency Appraisal Guidelines, Appendix B) and now under the new rule framework (Source: www.alstonconsumerfinance.com).

Parallel to this, Federal Housing Administration (FHA) and VA have their own valuation rules: FHA historically prohibited appraisal waivers, but in 2023 proposed limited AVM-based waivers for streamline refinances. Other U.S. regulators (e.g. California Department of Real Estate) have also begun to consider rules on algorithmic home valuations to protect consumers, although concrete regulations are still emerging.

Europe: The European Union does not yet have AVM-specific regulations, but valuation standards implicitly cover them. The RICS guidance in the UK (April 2022 Insight Report) is influential: it emphasizes that AVMs range from pure automation to hybrid models, and calls for transparency in methodologies (Source: www.rics.org). In continental Europe, Basel regulations require prudent collateral evaluation under CRR/CRD IV, which likely covers AVMs under “internal models” for credit risk. The European Banking Authority (EBA) has not issued AVM rules per se, but banks must report property valuations under IFRS 13 (fair value) or local accounting standards, so AVMs used internally must still meet fair-value measurement quality.

Other Jurisdictions: In countries like Canada and Australia, regulators have emphasized model risk in banking and insurance but not singled out AVMs yet. Industry associations (e.g. Canadian Appraiser Guilds, Australasian property boards) are debating how to integrate AVMs with professional appraisals. Globally, the overarching trend is acknowledgement that AVMs are part of valuation practice and require oversight, mirroring the U.S. stance.

In the banking crisis context, accurate collateral values are critical for lenders’ risk-weighted asset calculations (Basel II/III). It is conceivable that future regulatory capital guidelines will explicitly consider AVM use; already, Basel III implicitly assumes reliable collateral values for secured lending. This will add to the impetus for high-quality AVM models.

AVMs in the Appraisal and Lending Process

Complementing Appraisals

Although AVMs have grown, they have not replaced appraisals entirely. In most mortgages, a licensed appraiser still performs a full inspection and report. Instead, AVMs often complement appraisals in several ways:

• Triangulation / Cross-check: A lender might obtain both an appraisal and AVM, comparing the two. If they diverge by more than a set threshold (e.g. 5-10% of value), a quality control review is triggered. This flags potential errors from whichever source.

• Gatekeeping for Waivers: As noted, AVMs underlie appraisal waivers. If the AVM (or a combined AVM and quick exterior inspection) indicates property is “safe”, the lender may skip a formal appraisal. This streamlines low-risk loans (e.g., simple fixed-rate refinances) while reserving appraisers for complex deals (Source: sf.freddiemac.com) (Source: mtgefi.com).

• Post-Loan Monitoring: Servicers use AVMs to monitor collateral values on existing portfolio loans (for delinquency triage or loss mitigation). Changes in AVM values can prompt workouts or mark-to-market.

• Pricing and Indexing: Companies and investors use AVMs to generate home price indices (Zillow HVI, CoreLogic indexes) or to price MBS securities. In such cases, monthly or quarterly AVM aggregates substitute for expensive sales-based measures.

Appraisers themselves are integrating AVMs. Some appraisal firms have developed proprietary AVM tools to speed desk reviews, or use roadside photos (via vendors) to inform assessments. RICS highlights a trend toward AVM hybrid workflows: valuations with partial automation plus professional judgment (Source: www.rics.org). One article describes a UK valuer firm building its own AVM “where will this be most useful and how might valuers work with AI?” (Source: ww3.rics.org), reflecting how professionals are adapting, not ceding their role.

Nevertheless, pure AVM usage (with no human review) remains cautious except for targeted waivers. Consumers in many jurisdictions have legal rights to an appraisal. The regulatory separation (excluded rule scope) shows that appraisals by credentialed professionals are still seen as the “gold standard,” and AVMs as tools for non-appraisal valuations/evaluations (Source: www.alstonconsumerfinance.com).

Appraisal Waivers and Automated Collateral Evaluation (ACE)

A major development is the Appraisal Waiver Program of the GSEs. Under certain automated underwriting findings, FHA-approved lenders can obtain VA-familiar, conditional appraiser waivers. As the data from MtgeFi 2025 shows, this has become significant:

• In March 2025, $6.13B of loans (13.5% of GSE-received volume) used appraisal waivers (Source: mtgefi.com).
• An additional $931M (2.05%) used inspection-based waivers (AVM plus a brief exterior inspection by a notary-type person) (Source: mtgefi.com).
• Combined, about 15.5% of loans employed a waiver solution (Source: mtgefi.com).

These are striking numbers: once marginal, AVM-backed waivers now comprise a non-trivial segment of originations. Notably, nonbank mortgage companies are leading users: nearly 80% of nonbanks have used some AVM waiver (Source: mtgefi.com) (vs ~70% of banks). Within channels, the correspondent and retail channels show high adoption. Major players like Rocket Mortgage and United Wholesale (UWM) drive much of this growth, as they can integrate automated processes end-to-end (Source: mtgefi.com).

It is expected that these waiver proportions will continue rising, subject to market conditions. Even banks like Wells Fargo and Chase have expanded their use of ACE/VA waivers advertised on their loan officer portals. Given that average appraisal costs per loan are $500–600, each successful waiver saves substantial expense and time. The FHA announced in late 2024 that it too will launch an AVM waiver option (on a pilot basis), acknowledging the trend.

For each waived loan, the originator must still ensure collateral adequacy. This often means higher documentation standards: e.g., requiring an AVM score above a threshold (representing high confidence), or obtaining property data reports. Freddie Mac’s new ACE+PDR program couples an AVM with a Property Data Report (an exterior inspection by a licensed professional), bridging between full appraisals and pure AVMs (Source: sf.freddiemac.com) (Source: mtgefi.com).

In short, AVMs have transformed secondary-market valuation policy. They create an “on-ramp” for automation in underwriting, but with layered controls (random re-checks, additional data checks) to mitigate risk. The industry view is that this technology, carefully managed, can better allocate appraisal resources (human expertise to complex or borderline cases, automation on routine ones) (Source: sf.freddiemac.com).

Comparison: AVMs vs Human Appraisers

It is instructive to compare qualitative differences:

• Speed: AVM (seconds) vs Appraisal (~1-2 weeks). Advantage: AVM.
• Cost: AVM (nearly free incremental) vs Appraisal ($300–700).
• Coverage: AVM (any property with data) vs Appraisal (any property on request). Ambiguous: AVMs may fail without data, appraisals cost/time.
• Accuracy/Bias: As discussed, AVMs can systematically differ from appraisals; emotionless vs judgement-based. Appraisals may capture unique features (interior condition, upgrades) that AVMs miss.
• Transparency: Appraiser provides reasoning and photos; AVM is algorithmic (though output may include comparable listings). Regulators often prefer appraisals for traceability.
• Consistency: AVMs apply same logic globally; appraisers vary. CoreLogic touts “removing human errors” as a benefit (Source: www.rics.org).
• Auditability: Both are audit-able, but in different ways—appraisals follow USPAP guidelines; AVMs follow model validation protocols.

Professionals emphasize that SVPs (subjective valuer perceptions) still matter: e.g. imminent listings could be overpriced but market-susceptible factors (noise, traffic) seen by an appraiser on-site. AVMs have no field view (though some use street images or drone data, this is nascent). As RICS notes, hybrid approaches will likely prevail (Source: www.rics.org), with valuers using AVMs as one tool.

Case Studies and Industry Examples

Freddie Mac and GSE Initiatives

Freddie Mac’s published insights provide a case of institutional AVM strategy. In their January 2023 memo, Freddie emphasizes “modernization” of valuation to balance risk with efficiency (Source: sf.freddiemac.com) (Source: sf.freddiemac.com). They note that prior to COVID, only two options existed (full appraisal or appraisal waiver). Afterwards, they developed desktop and hybrid appraisals and the ACE waiver (Source: sf.freddiemac.com).

Notably, Freddie states: “ACE appraisal waivers use models and algorithms to forgo an appraisal altogether” (Source: sf.freddiemac.com). They also announced new “inspection-based waivers” (ACE+PDR) in summer 2023 (Source: sf.freddiemac.com), approving walk-only exterior inspections to combine with AVM scores. This move was explicitly aimed at freeing appraiser capacity and focusing effort on complicated assignments (Source: sf.freddiemac.com).

These developments reflect a strategic shift in valuation workflows; Freddie’s public posture underscores that technology-enabled valuation is not just a fad but core to their risk-management strategy. They are advocating industry change, warning that reverting to 2019 practices would leave the mortgage system unprepared for future disruptions (Source: sf.freddiemac.com).

Small Lenders and Automated Underwriting Systems

Smaller lenders and fintechs have often been earlier adopters of AVM-driven processes. Mortgagee and mortgage technology vendors have developed end-to-end loan origination platforms where an automated underwriting system (like Fannie’s Desktop Underwriter or Freddie’s Loan Product Advisor) embeds AVM triggers. For example, lenders using Fannie DU may get an “AVM Waiver Offer” if the DU finds a high-quality collateral candidate (DU Certification 10/10). Similarly, Freddie’s Loan Advisor can automatically approve a waiver under ACE.

Industry publications (such as Mortgagenewsdaily.com) reported exponential growth in waiver volume by lenders like Rocket Mortgage, PennyMac, and NewRez in 2024 (Source: mtgefi.com). These lenders leverage AVM waivers heavily—Rocket and UWM grew inspection-waiver usage 100% year-over-year (Source: mtgefi.com). This shows that firms with vertically integrated pipelines (loan application to delivery) capture more of the efficiency gain.

One example: a regional bank might commit to an AVM-backed process for standard 30-yr fixed refinances up to 80% LTV, using an AVM and property data report instead of a full appraisal. Such a program would involve underwriting guidelines, specified AVM tools (often vendor-sourced), training for underwriters on interpreting AVM confidence scores, and a quality assurance loop to spot-check outcomes.

International Perspective

Outside the U.S., AVM adoption varies. In Australia, CoreLogic (formerly RP Data) provides AVMs as a professional tool. Lenders often use a default AVM when an appraiser appointment fails, or for small top-up loans. CoreLogic’s 2023 note underscores their effort to improve AVM performance during the recent downturn (Source: www.corelogic.com.au) (Source: www.corelogic.com.au). They measure success by known metrics (within ±10% of sale), indicating a mature usage environment where such targets are meaningful.

In the United Kingdom, the Royal Institution of Chartered Surveyors (RICS) has been actively guiding members. Their 2021–22 “AVM Roadmap and Insight” reports [34][36] stress that AVMs are used in valuations of residential property portfolios (e.g. for fund valuations or bank collateral oversight). RICS emphasizes that even when using 100% automated workflows, professionals must ensure transparency and governance. An RICS-regulated firm called Yuvoh is noted as developing its own AVM software, reflecting the entrepreneurial interest in AVMs in the UK (Source: www.rics.org).

Moreover, some European banks have developed their own models for cross-border collateral verification (given mortgages often end up held by multinational investors). Countries with less robust sales data (e.g. parts of Continental Europe) have lagged, but even there the push for harmonized digital valuation is growing (for example, the EU is discussing digital registries of property transactions which would feed AVMs).

Zillow and Consumer Big Data

While not part of institutional lending, Zillow’s experience in AVMs offers instructive lessons. In 2019, Zillow reported that the typical national Zestimate error was $18,600 (it used a geographies approach) (Source: realestatedecoded.com). Their technology blog (2017) emphasized context: the Zestimate’s MdAPE was 7.7%, symmetric around zero error, and for 84% of homes the Zestimate was within ±20% of final sale price (Source: www.zillow.com). By benchmarking against list prices, Zillow demonstrated that an AVM’s average “accuracy” is not far off from what realtors achieve at listing time (Source: www.zillow.com).

Zillow’s research also highlights how AVMs can sometimes outperform even agents: in 35% of transactions the Zestimate was closer to final sale price than the initial list price (Source: www.zillow.com). This suggests AVMs can serve as a sanity check on pricing decisions. However, Zillow’s own retreat from iBuying (Zillow Offers shut down in 2021) is a cautionary tale: algorithmically buying houses relies on AVM accuracy in predicting future sale prices. Market volatility exposed gaps in their models, underscoring the risks in over-relying on AVMs without adequate risk buffers.

Data Analysis and Evidence

Throughout the sections above, we have referenced extensive data and analyses. For clarity, key quantitative findings are summarized here:

• AVM vs List Price (Zillow, 2016 data): Initial list price median error = 4.2%; Initial Zestimate median error = 7.7% (Source: www.zillow.com). At sale, errors shrink to 2.7% vs 3.6%. Thus, Zestimates had somewhat higher median error, but still within single digits. In 68% of homes, Zestimate was within 10% of final price (Source: www.zillow.com).

• Rural Appraisal Bias: Studies using millions of transactions show appraisals are biased high ~90% of time (Source: link.springer.com), with >25% of rural appraisals at +5% above sale (Source: link.springer.com) (vs ~13% in urban) (Source: link.springer.com). Automated estimates (ML models) reduced this bias significantly.

• Model Performance (Molinaro, JRE 2019): Random forest AVM R²=0.7224, RMSE=0.2971 (Source: link.springer.com), outperforming linear models. (Note: R² and RMSE vary by dataset scale; these were normalized errors of log-prices.)

• CoreLogic AVM Improvements: With new data, CoreLogic increased “within 10% of sale” accuracy by 8–9% (Source: www.corelogic.com.au) and reduced over-valuations by 2–3% (Source: www.corelogic.com.au) compared to prior model versions. (These gains were on a holdout test of one year of data.)

• GSE Waiver Volumes (Mar 2025): 84.5% of GSE loans used appraisals; 13.5% used AVM waivers; 2.05% used inspection waivers (Source: mtgefi.com). Inspection waivers accounted for 18.7% of all waivers by loan count (Source: mtgefi.com).

• Lender Adoption: Over the past 2 years, 79.5% of nonbank lenders delivered at least one loan via an appraisal waiver; 45.9% via inspection-based waiver (Source: mtgefi.com). For banks, 69.0% used AVM waivers at least once (Source: mtgefi.com).

• Mistakes and Oversight: FHFA and CFPB reports (2021) have documented hundreds of instances where appraisers used prohibited discriminatory language or criteria in reports. This prompted focus on valuation equity – under which AVMs are now considered. The new AVM rule’s nondiscrimination factor is explicitly aimed at countering “the potential for AVMs to produce property estimates that reflect discriminatory bias” (Source: www.alstonconsumerfinance.com).

These data highlight both the promise (quantitative precision and efficiency) and the pitfalls (biases, error rates) of AVMs. They form the evidence base for our analysis and recommendations.

Discussion: Implications and Future Directions

Implications for Lenders and Regulators

The rise of AVMs has significant operational and regulatory implications. Lenders see AVMs as a way to streamline origination and manage collateral more efficiently, hence they are incorporating them heavily. However, to maintain portfolio quality, lenders must also invest in robust model risk management: qualified validation teams, continuous monitoring, and compliance functions. The AVM guidance means boards and management must treat AVMs with the same seriousness as other critical models (e.g. credit models, liquidity models).

Regulators have likewise signaled that AVM reliability is now a supervisory focus. Examiners will probe whether banks follow the new quality control standards, like performing random back-tests and documenting anti-bias measures (Source: www.occ.treas.gov) (Source: www.alstonconsumerfinance.com). Nonbank lenders (mortgage companies) are also caught, because the rule applies to “originators” who engage in a credit decision or securitization. The net effect is to raise the baseline of due diligence on all automated valuations.

For borrowers and consumers, the proliferation of AVMs may mean faster closings and possibly lower costs (if lenders pass along savings). On the other hand, transparency and fairness concerns mean that regulators (via PAVE, CFPB, HUD) will be closely watching complaint rates and appraisal challenges. If a borrower suspects an AVM undervalued their home, they might request a traditional appraisal under CFPB’s Regulation Z “PAL” (property appraisal disclosure) rules. Market education is needed: many consumers still misunderstand Zestimates as guarantees, whereas banks use certified appraisers for formal lending decisions.

Technological Trends

Looking to the near future (2026 and beyond), we expect several trends:

• Data Enrichment: AVMs will ingest ever-richer data. Beyond traditional records, inputs like automated-condition signals (smart home data on maintenance), IoT data (energy usage patterns), and even social media trends (e.g. neighborhood sentiment indexing) could be explored. Some companies are already experimenting with computer vision: analyzing satellite or street-level imagery to detect renovations or curb appeal factors.

• Artificial Intelligence Advances: Generative AI techniques might be used to synthesize additional comparable sales data when actual comps are sparse, or to simulate market conditions. However, explainability remains a concern, so hybrid models (combining neural networks with interpretable layers) may be the norm. We might see AVMs that provide narrative explanations or confidence heatmaps of how each input affects the output.

• Cross-Asset AVMs: The technology will extend to other asset classes. Commercial real estate (CRE) AVMs are emerging, though they will require property-specific data (rental rates, expense reports). Land and agricultural AVMs (e.g. farmland valuations) are nascent fields. Regulators may eventually set guidelines for these as well once usage grows.

• Climate and Resilience Factors: A big emerging requirement is to incorporate climate risk (flood zones, wildfire risk) into property values. AVMs could either adjust for hazard (discounting at-risk properties) or produce alternative valuations under stress scenarios. Banks will want to know “climate-adjusted LTVs,” which suggests integrated models linking real estate AVMs with hazard models.

• Blockchain and Data Lines (speculative): Some envision a future where property transactions and characteristics are logged on immutable ledgers. If realized, AVMs would have immediate access to a continuous stream of verified sales and property events. In such a scenario, AVMs might update dramatically in real time, altering how credit decisions are judged.

Future of Model Risk Regulation

On the regulatory side, the recent U.S. rule is just the beginning. Over the next few years, we may see:

• International Harmonization: U.S. agencies may advocate for global standards on AI/AVMs through Basel Committee or Financial Stability Board. For instance, the FSB’s AI principles might be supplemented with AVM-specific checklist items for banks globally.

• Explicit Capital Requirements: If AVMs enable more automated underwriting, regulators might factor in model risk for capital calculations. For example, Basel 3.1 (scheduled 2025) revises real estate risk weights; if banks rely heavily on AVMs, regulators might scrutinize whether AVM-backed loans should attract higher risk weights unless validated by appraisals.

• Consumer Protection Laws: Legislatures might enact laws requiring AVM disclosure to consumers (much like the Apple Card fiasco led to disclosure rules). Already New Jersey has proposed legislation requiring any automated home valuation given to a consumer (like Zillow's Zestimate shown in listings) to come with an error range. Similar ideas are in discussion in several states (e.g. New York).

Case Study Revisited: Hypothetical Lender

Let’s consider Acme Bank, a mid-sized U.S. mortgage lender, and how AVM quality/risk applies:

• AVM Deployment: Acme has purchased a third-party AVM that claims 85% accuracy within ±10%. They integrate it with their AUS to auto-approve certain mid-risk applications (e.g. 80% LTV purchase loans on SFRs meeting criteria).

• Quality Controls: Acme establishes a policy that any loan closed based solely on an AVM value undergoes yearly spot-check: 10% of such loans get an appraisal post-closing, to compare the AVM vs actual sale price (which CFO expects to be within 5%). They tag any AVM that shows >10% discrepancy, feeding that data back to the vendor for model tuning.

• Bias Checks: Their compliance team runs quarterly analyses: grouping AVM errors by location race demographics and income. If systematic bias emerges (e.g. underestimation in certain zip codes), they would adjust training weights or exclude problematic variables. They also ensure their loan officers explain to borrowers that an AVM was used and note the contact for question (for lending transparency).

• With Regulatory Change: Upon the 2024 AVM rule effective date, Acme updates model risk policies to explicitly address the five factors. They document how they protect data (segregated databases, audit trails) and avoid conflicts (not taking AVM consulting from any party with a vested interest). They train staff on the requirement to “own the risk” (Source: kpmg.com) – meaning management acknowledges AVM is part of credit risk model, not a “black box exception.”

• Outcomes: Suppose in 2026 Acme’s random audits show 2% of AVM-only loans had losses where LTV was actually under-calculated due to AVM undervaluing property. Because of its controls, Acme identifies this flaw quickly and orders updated AVM parameters from the vendor. Without these controls, last-resort resolution might have been more costly (claim adjustments, reputational harm).

This scenario underscores that AVM value is unlocked only with diligent oversight. It also highlights how 2026 lenders must be savvy in linking quantitative model management to traditional credit risk processes.

Conclusion

The landscape of collateral valuation is undergoing a transformation. Automated Valuation Models (AVMs), once a novel adjunct, are now embedded in mainstream lending, appraisal review, and even regulatory frameworks. By 2026, AVMs are expected to handle an ever-larger share of day-to-day valuation tasks, providing speed and consistency. However, this expansion carries substantial model risk that banks must actively manage.

Key takeaways:

• Quality Matters: Metrics like error rates, confidence intervals, and bias measures must be rigorously tracked. A robust AVM program continuously updates models, backtests results, and integrates enhancements (as seen with industry efforts to leverage more data (Source: www.corelogic.com.au).

• Even Good Models Have Limits: AVMs excel in uniform markets but falter in heterogeneous or under-recorded areas. Every AVM has a known error distribution and outlying cases. Stakeholders must be aware of these limits (e.g. rural coverage gaps, unusual property types) (Source: www.rics.org).

• Governance is Critical: The new regulatory requirements enshrine a principle that lenders “own the risk” of AVMs (Source: kpmg.com). This means accountability flows to senior management, requiring tight controls on vendor models, validation teams, and exception processes.

• Ethics and Bias: AVMs must be part of the fair lending conversation. While they remove subjective prejudices inherent in human appraisals, they can inadvertently perpetuate systemic inequities. Vigilance and research into algorithmic fairness in real estate remain essential (Source: www.urban.org) (Source: www.veros.com).

• Future Forward: Emerging AI methods promise incremental quality gains (e.g. image-based features (Source: journals.plos.org). But as models get more complex, the demand for explainability and robust validation only grows. The industry will likely coalesce around a hybrid vision: leveraging powerful AVMs under the guardrail of human oversight and regulation.

This report has provided an in-depth analysis of AVM quality and model risk, citing current research and examples. As one senior risk officer put it, “The key is striking the balance between innovation and control” (Source: www.kroll.com). In 2026 and beyond, that balance will be tested by new market conditions and technologies. The institutions that succeed will be those that combine AVM-driven insights with sound model governance, ensuring that the efficient valuations of tomorrow come with the trust and reliability demanded by regulators and stakeholders.

All factual claims and data above are supported by sources, as cited in the text.