What Are Automated Valuation Models?
A practitioner’s guide to how AVMs work, when they are appropriate, and where they fit within the professional valuation framework.
The short version
An automated valuation model — universally abbreviated to AVM — is a statistical system that estimates the market value of a property without a physical inspection. It does this by analysing recorded transaction data, property characteristics, and local market conditions, then applying mathematical models to produce a value estimate with an associated confidence measure.
If you have ever checked your home’s estimated value on Zoopla or seen a lender approve a remortgage without sending a surveyor, you have encountered AVM output. The technology underpins a significant and growing portion of residential property valuation activity across the UK and internationally.
But AVMs are not magic boxes, and they are not replacements for professional judgement. Understanding what they can and cannot do is essential for anyone who uses property valuations in their work — whether you are a lender, valuer, broker, investor, or risk manager.
How an AVM differs from a physical valuation
A traditional physical valuation involves a RICS Registered Valuer visiting the property, inspecting its condition and characteristics, assessing the local market, and forming a professional opinion of value. The result is a written valuation report, typically compliant with the RICS Red Book (Global Standards), that carries the valuer’s professional liability.
An AVM takes a fundamentally different approach. It works from recorded data rather than direct observation. This gives it two major advantages and one significant limitation:
An AVM produces a result in seconds. A physical valuation typically takes one to three weeks from instruction to delivery, including scheduling, the site visit, and report writing.
An AVM can value an entire mortgage portfolio overnight. Physical valuations are inherently limited by the number of available valuers and the time required per inspection.
An AVM cannot see what is not in the data. Internal condition, recent renovations, structural defects, and micro-location factors (views, noise, aspect) are invisible to a model that has never walked through the front door.
This trade-off — extraordinary speed and scale against an inability to observe the unrecorded — is the central fact of AVM technology. Every decision about when to use an AVM and when to require a physical inspection comes back to it.
The three main AVM methodologies
Not all AVMs work the same way. The industry recognises three broad families of approach, each with different strengths. Most modern production AVMs use a hybrid approach that combines elements of all three.
Hedonic regression
The oldest and most theoretically grounded approach. A hedonic model decomposes a property’s price into contributions from individual characteristics: location, size, property type, number of bedrooms, age, energy efficiency, proximity to schools, and so on. Each characteristic is assigned a coefficient — effectively a price weight — estimated from historical transaction data. The property’s value is the sum of these weighted characteristics.
Strengths: transparent, explainable, grounded in economic theory. Weaknesses: assumes linear relationships, struggles with unusual properties, requires careful feature engineering.
Comparable sales adjustment
This approach mirrors what a surveyor does mentally: find recently sold properties similar to the subject, then adjust for differences. An AVM does this algorithmically, typically using distance metrics to identify the most comparable transactions, then applying systematic adjustments for differences in size, type, condition indicators, and time since sale. The final value is a weighted average of the adjusted comparables.
Strengths: intuitive, directly grounded in market evidence, handles local variation well. Weaknesses: requires dense comparable data, performance degrades in thin markets or for unusual property types.
Hybrid and machine learning models
Modern AVMs increasingly use machine learning techniques — gradient-boosted decision trees, neural networks, or ensemble methods — that can capture complex, non-linear relationships between property characteristics and value. These models learn patterns from millions of transactions that would be impossible to specify manually. They often combine hedonic-style feature analysis with comparable-sales-style spatial reasoning in a single unified framework.
Strengths: captures non-linear relationships, handles feature interactions automatically, often highest accuracy. Weaknesses: less inherently transparent (though modern explainability tools like SHAP analysis address this), requires large training datasets.
In practice, the distinctions are blurring. Most commercial AVMs now use ensemble or hybrid approaches that combine the interpretability of hedonic methods with the accuracy of machine learning. The Gadsden Valuations model, for example, is a gradient-boosted decision tree ensemble trained on 8.4 million transactions with 47 features across six categories. Read the technical summary for methodology detail.
When AVMs are appropriate
The question is never “is an AVM good enough?” in the abstract. It is always “is an AVM appropriate for this specific use case, this specific property, at this level of confidence?”
Where AVMs work well
- • Portfolio revaluation — lenders holding thousands or millions of mortgages need regular revaluation for capital adequacy and risk management. Physical inspection of every property is neither practical nor proportionate. Basel 3.1, expected to take effect in January 2027, will require portfolio revaluation at scale, making AVM capability essential for regulated firms.
- • Remortgage and product transfer — where the borrower is already on the lender’s book, the property was recently inspected at origination, and the loan-to-value ratio is conservative, an AVM provides a proportionate and cost-effective valuation.
- • Desktop valuation support — a Registered Valuer conducting a desktop review can use AVM output as one input among several, applying professional judgement to assess whether the model estimate is reasonable for the specific property.
- • Cross-checking and quality assurance — AVMs can flag physical valuations that appear anomalous relative to market evidence, providing an independent check on the valuation panel.
- • Market monitoring — tracking value movements across a portfolio to identify properties approaching the 10% revaluation threshold or entering negative equity territory.
Where physical inspection is required
- • Purchase lending at high LTV — where the borrower has limited equity, the lender needs confidence in both value and condition that only physical inspection can provide.
- • Non-standard construction — properties built with unusual materials, unconventional layouts, or significant modifications that are not captured in standard data sources.
- • Low-confidence AVM output — when the model itself signals that it lacks sufficient comparable evidence or data completeness to produce a reliable estimate, a physical valuation is the appropriate fallback.
- • Regulatory or contractual requirement — some lending contexts, court proceedings, and insurance claims require a formal RICS Red Book valuation by regulation or contract.
The best practice is not to choose between AVM and physical inspection but to use the right tool for the right situation. A well-designed AVM includes a confidence measure that signals when it is and is not appropriate to rely on the model output. For more on how RICS and UK regulation govern this decision, see our dedicated knowledge article.
AVMs as tools, not replacements
A common misconception is that AVMs replace surveyors. The regulatory framework does not support this interpretation, and neither does good practice.
The International Valuation Standards (IVS 105, 2025 edition) explicitly permit automated valuation models as tools within a professional valuation framework. Section 105.20 addresses the use of models and technology-assisted approaches, positioning them as inputs to professional judgement rather than substitutes for it. The valuer remains responsible for the valuation opinion; the AVM provides evidence and analysis that the valuer can accept, adjust, or override.
Under the RICS Red Book, PS 1.6 specifically addresses automated valuation models and allows a “specialist or service organisation” to provide the AVM. The Registered Valuer who uses that AVM output within a written valuation retains professional responsibility. This is no different in principle from a valuer using comparable evidence from a data provider — the tool provides data, the professional provides judgement.
The practical effect: an AVM handles the data-intensive work — processing millions of transactions, adjusting for time, identifying comparables, quantifying uncertainty — freeing the valuer to focus on what they do best: applying local knowledge, assessing condition, and exercising professional judgement on factors the data cannot capture.
Confidence and accuracy
AVM accuracy is not a single number. The industry uses several complementary metrics to describe how well a model performs. Two are particularly important and appear in most AVM quality assessments.
PE10 — Percentage within ±10%
PE10 measures the proportion of valuations where the model’s estimate fell within 10% of the actual transaction price. It answers a straightforward question: if you use this AVM, how often will it be roughly right?
Academic research (Kirchmeyer, Matysiak, Rossini) has established several PE10 thresholds. A PE10 above 50% is considered the minimum for mortgage lending use. Above 70% is considered acceptable for routine purposes. Above 80% is considered strong by international standards. These thresholds were originally defined against surveyor valuations, which are themselves estimates. AVMs benchmarked against actual sale prices — a harder test — will show lower PE10 figures for equivalent real-world accuracy.
MdAPE — Median Absolute Percentage Error
MdAPE is the median percentage by which the model’s estimate differs from the actual price. Using the median rather than the mean prevents a handful of extreme outliers from distorting the overall picture. Lower is better.
An MdAPE in single digits is considered strong performance for a national AVM. For context, asking prices on property portals typically differ from eventual sale prices by 5–15%, and RICS surveyor valuations are typically within 5–10% of the price eventually paid. An AVM with an MdAPE of 7–8% is operating within the same margin as human professionals, but at a fraction of the cost and time.
Beyond aggregate accuracy, a well-designed AVM provides a per-property confidence score. This is critical because aggregate metrics mask significant variation. A model might achieve 60% PE10 nationally, but 70% in dense urban markets and 40% in rural areas with thin transaction volumes. The confidence score tells the user how much weight to place on the specific estimate they are looking at.
The European AVM Alliance defines a common 0–7 confidence scale (Forecast Standard Deviation) that allows comparison across AVM providers. Our accuracy page publishes live metrics including PE10, MdAPE, and segmented breakdowns by property type, price band, and region.
A brief history of AVMs
Automated valuation is not new technology. Its roots go back over half a century.
The first AVMs appeared in the United States in the 1960s and 1970s, developed by local government agencies for property tax assessment. Counties needed to value every residential property in their jurisdiction for tax purposes — a task that demanded scale long before anyone used the word “fintech.” These early models were simple hedonic regressions, but they established the principle that statistical methods could produce useful property valuations.
US mortgage lending adopted AVMs in the 1980s and 1990s, initially for quality control (checking physical appraisals against statistical estimates) and gradually for direct use in lower-risk lending decisions. The government-sponsored enterprises Fannie Mae and Freddie Mac played a significant role in establishing AVM acceptance criteria and quality standards.
The UK came later. While some lenders experimented with AVMs in the early 2000s, adoption accelerated significantly after the 2008 financial crisis. The crisis exposed weaknesses in physical valuation processes — particularly panel management conflicts, valuer shopping, and the difficulty of maintaining quality at scale during lending booms. AVMs offered consistency, auditability, and independence from the conflicts that had plagued panel valuation.
Today, AVMs are used routinely by UK mortgage lenders for remortgage, product transfer, and portfolio monitoring. The regulatory framework has matured substantially: RICS PS 1.6 addresses AVMs directly, IVS 105 provides international guidance on valuation models, the PRA’s SS1/23 sets expectations for model risk management, and the EAA’s ESSVM provides quality standards specific to statistical valuation methods.
The next major regulatory milestone is Basel 3.1, expected to take effect in January 2027, which will require banks to revalue their property collateral portfolios at scale for capital adequacy purposes. This is expected to further accelerate demand for reliable, transparent AVM services.
How Gadsden Valuations fits this picture
We built Gadsden Valuations as a modern AVM that addresses the specific requirements of the current UK regulatory environment. Here is how our approach maps to the principles described above.
Methodology
A gradient-boosted decision tree ensemble (the hybrid/ML approach), trained on 8.4 million Land Registry transactions and enriched with EPC, census, school, flood risk, and house price index data. Forty-seven features across six categories.
Validation
Walk-forward backtesting against actual Land Registry sale prices — not surveyor valuations. The current test set is 295,026 properties. All metrics are published on our accuracy page.
Confidence
Every valuation receives a three-tier confidence classification (High, Medium, Low) mapped to the EAA’s Forecast Standard Deviation scale. Properties with insufficient data are declined rather than valued with false confidence.
Regulatory alignment
Designed with reference to IVS 105, RICS Red Book PS 1.3 and PS 1.6, PRA SS1/23, and EAA ESSVM. Output is intended as input to written valuations supervised by a Registered Valuer, not as a standalone replacement for professional valuation.
Data provenance
All training data sourced from government open data under Open Government Licence v3. No scraped portal data, no asking prices, no self-reported property details.
See how our AVM performs
We publish live accuracy metrics from walk-forward backtesting against real Land Registry sale prices. No cherry-picking. No simulated data. Real predictions against real outcomes.