Valuation Office Agency: Automated Valuation Model
The automated model is being used to support the valuation of 1.5m domestic properties for a Council Tax revaluation in Wales in 2028.
Tier 1 Information
1 - Name
Automated Valuation Model
2 - Description
The Automated Valuation Model is series of models to:
- predict domestic property prices in Wales at a set point in time in line with Council Tax valuation assumptions
- identify the closest comparable properties that have sold
- adjust the estimates to account for the residuals on the sale comparable, and place into Council Tax Bands
- score the reliability of these estimates (again using the residuals on comparable sales)
- prioritise batches of properties for manual review
3 - Website URL
https://www.gov.uk/guidance/how-model-assisted-valuation-delivers-large-scale-property-valuations
4 - Contact email
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
Valuation Office Agency
1.2 - Team
Valuation Analytics
1.3 - Senior responsible owner
Joint SROs:
- Head of Valuation Analytics
- Head of Valuation for Council Tax Reform and Mass Appraisal.
1.4 - External supplier involvement
No
Tier 2 - Description and Rationale
2.1 - Detailed description
The Assisted Valuations Model (AVM) consists of a series of models to:
- predict domestic property prices in Wales at the valuation date in line with Council Tax valuation assumptions; this is a regression model utilising a Gaussian-Markov random field approach to estimate spatial impacts.
- identify the closest comparable properties that have sold; from a comparable model that captures valuers’ judgement on what makes properties similar, weighting different components.
- adjust the estimates to account for the residuals on the sale comparable, and place into Council Tax Bands; optimising the relevant weights between the regression model and comparable residuals.
- score the reliability of these estimates (again using the residuals on comparable sales); identifying and weighing the key factors that predict reliability.
- prioritise batches of properties for manual review; utilising the reliability and proximity to band margins.
2.2 - Scope
The model is being used to support the valuation of 1.5m domestic properties for a Council Tax revaluation in Wales in 2028.
This tool was designed and implemented to provide first pass valuations to aid council tax banding decisions and to enable valuer resource to target and undertake the most difficult or complex decisions that most likely need the intervention of experienced valuers.
The model estimates can also be used to aid in the design of policy.
2.3 - Benefit
Our early assessments of the benefits of using mass appraisal include significant reductions in cost and timescales, as well as benefits to customers through greater consistency and quality of valuations. Initial estimates place the reduction in cost of a revaluation by one-third, compared to carrying out a manual valuation. Moreover, by providing an initial valuation the VOA can prioritise cases for manual valuation and can better target our professional valuer resource to where they will have the greatest impact, providing more rewarding roles for our staff.
2.4 - Previous process
Prior to this tool there was no other tools or technology used. Valuers were deployed to make manual banding assessments.
2.5 - Alternatives considered
Deployment of valuers to undertake manual banding assessments. This is significantly more inefficient and costly.
During model development a series of approaches were tested. A base mixed model was initially developed and then alternative methods considered and tested against this. The Gaussian-Markov Random Field performed best in testing and was computationally reasonable to implement.
Tier 2 - Decision making Process
3.1 - Process integration
The VOA has adopted a model assisted valuation approach, to help target its valuer resource when valuing properties at scale.
VOA valuers underpin each stage of the model assisted approach.
First, they ensure the property data and sales used are accurate. They then collaborate with analysts to build and refine the AVM.
Valuers sample model valuations to ensure the model is accurate and to understand where valuer reviews should be targeted.
Groups of properties, with similar characteristics in the same location, are then reviewed. The focus of valuers is on reaching the right Council Tax band for each property.
There is a minimum and maximum property value for each band and valuers will also focus on checking those properties at the band margins to ensure a property is placed in the correct Council Tax band.
3.2 - Provided information
Valuers use comparable evidence to make their valuation decisions. The estimated values are provided to operational staff, along with the property attributes that led to this valuation and the sales evidence in a bespoke Valuation Tool. Using this information estimates can be verified and adjusted as required. A more thorough explanation of how the model arrived at its value is available to a central team of valuers and analysts, who sense check the estimates.
3.3 - Frequency and scale of usage
There are 1.5m domestic properties in Wales, with the model providing a first pass valuation for the vast majority.
The next revaluation will be in 2028, with the valuation phase starting from September 2026, with a five year revaluation cycle thereafter.
3.4 - Human decisions and review
The model is assured by a central team of valuers and analysts, who determine the model structure and variables that should be included. Operations then conduct targeted assurance of batches of properties, adjusting the council tax bands as required. Operations further undertake conduct consistency checks between batches and on batches not selected for targeted reviews.
3.5 - Required training
Operational staff are given a high level overview of the modelling as they utilise the output. However training focusses on how to use the Valuation Tool to utilise sales evidence to verify and adjust as necessary the estimate.
The core modelling team of analysts and valuers have had specific training on the modelling techniques and how the models work, including with external spatial modelling experts.
3.6 - Appeals and review
Taxpayers have a right to challenge their Council Tax Bands allocation when the list is first published and further informal routes are available after this. The Revaluation in Wales is planned for April 2028.
Tier 2 - Tool Specification
4.1.1 - System architecture
The model and data pipelines are built in R on databricks in the VOAs Data and Analytics Platform.
4.1.2 - Phase
Production
4.1.3 - Maintenance
Cyclical review in line with Revaluation cycles. Following work to prepare for a potential 2025 revaluation we are learning the lessons and implementing further developments for the 2028 Revaluation
4.1.4 - Models
The AVM consists of a series of models to:
- predict domestic property prices in Wales at the valuation date in line with Council Tax valuation assumptions; this is a regression model utilising a Gaussian-Markov random field approach to estimating spatial impacts.
- identify the closest comparable properties that have sold; from a comparable model that captures valuers’ judgement on what makes properties similar, weighting different components.
- adjust the estimates to account for the residuals on the sale comparables, and place into Council Tax Bands; optimising the relevant weights between the regression model and comparable residuals.
- score the reliability of these estimates (again using the residuals on comparable sales); identifying and weighing the key factors that predict reliability.
- prioritise batches of properties for manual review; utilising the reliability and proximity to band margins.
Tier 2 - Model Specification
4.2.1 - Model name
Automated Valuation Model
4.2.2 - Model version
Oct-23
4.2.3 - Model task
To predict property values based on sales data from the property attributes and location
4.2.4 - Model input
Property attributes and location
4.2.5 - Model output
A value estimate for each domestic property in Wales with a set of comparable sales. An understanding of the reliability of these estimates and, accounting for this and proximity to band margins, prioritisation of batches of properties for valuer review
4.2.6 - Model architecture
The AVM consists of a series of models to:
- predict domestic property prices in Wales at the valuation date in line with Council Tax valuation assumptions; this is a regression model utilising a Gaussian-Markov random field approach to estimating spatial impacts.
- identify the closest comparable properties that have sold; from a comparables model that captures valuers’ judgement on what makes properties similar, weighting different components.
- adjust the estimates to account for the residuals on the sale comparables, and place into Council Tax Bands; optimising the relevant weights between the regression model and comparable residuals.
- score the reliability of these estimates (again using the residuals on comparable sales); identifying and weighing the key factors that predict reliability.
- prioritise batches of properties for manual review; utilising the reliability and proximity to band margins.
4.2.7 - Model performance
The model has undergone statistical testing throughout its development, aligned to the International Association of Assessing Officers (IAAO) AVM standards, including on ratio studies, which test for vertical (across the value range) and horizontal (across different property types) equity.
The AVM has been assured externally:
The Centre for Appraisal Research and Technology (CART), experts in mass appraisal, provided advice during the model development.
The VOA commissioned the International Association of Assessing Officers (IAAO), as a recognised world authority on AVMs, to carry out a review of the model development process.
The IAAO’s report on the model concluded: “These findings are more than satisfactory and should lead the VOA to have confidence in the quality of the new valuation project conducted in Wales.”
4.2.8 - Datasets
Sales Pipeline & Live properties
4.2.9 - Dataset purposes
Sales Pipeline to train the model, live properties estimate values for all properties
Tier 2 - Data Specification
4.3.1 - Source data name
Sales Pipeline - to train and test the model Live list pipeline to run the model on all properties
4.3.2 - Data modality
Tabular
4.3.3 - Data description
Sales Pipeline - provides cleaned sales data, property attributes (at the time of sale) and location in order to train and test models Live List - provides property attributes and location of all properties to apply models to
4.3.4 - Data quantities
Sales Pipeline c.200 columns and c.200k rows Live List c110 variables and c.1.5m rows
Sales Data was partitioned into 4 samples: Train 1, Train 2 and Test - 30% each and 10% holdout. Primarily train1 and 2 were combined into a train set and then test used to assure, but some cross validation was conducted (e.g. train1 and test to train, testing on train 2)
4.3.5 - Sensitive attributes
Sales Pipeline includes names of buyer/seller and solicitor. This is to help identifying sales that are not of open market value.
4.3.6 - Data completeness and representativeness
Sales data is overall representative of the live list, though contains proportionately more new build properties
4.3.7 - Source data URL
N/A
4.3.8 - Data collection
Sales pipeline combines Welsh Land Transaction Tax (WLTT) records from HMRC with Land Registry Price Paid Data. Property attributes are collected and maintained by VOA staff. Plot size was included using Ordnance Survey National Geographic Database and Land Registry Title Polygons. Locations used VOA data, Ordnance Survey Address Base Premium and the Office for National Statistics National Postcode Lookup
4.3.9 - Data cleaning
Sales data is cleaned to remove non-open market transactions identifiable from the data available. Further sales verification is conducted by VOA staff where deemed necessary, including on model outliers.
4.3.10 - Data sharing agreements
Model outputs, as well as final bandings, are shared with the Welsh Government.
4.3.11 - Data access and storage
Data Analytics Platform access is limited to those analysts requiring it. Data is retained in line with VOA policy.
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessment
DPIA Completed.
The Welsh Government has conducted impact assessments of the Revaluation as a whole.
The VOA has adopted a model assisted valuation approach, to help target its valuer resource when valuing properties at scale. VOA valuers underpin each stage of the model assisted approach.
5.2 - Risks and mitigations
Risk the model did not meet acceptable standards - in house testing and independent IAAO ratios study mitigated this. This included an assessment of both horizontal equity (e.g. across different types of properties) and vertical equity (across the range of values)
Risk of insufficient expertise to produce and maintain an AVM - we have built up a successful partnership between analytical and valuation professionals in establishing an in-house valuation modelling capability. The speed in which this capability has been established, and our innovative approaches has been recognised by the Institute for Revenue, Rating and Valuation (IRRV), winning their Excellence in Innovation Award. With the capability established, the VOA will continue to explore the potential of model assisted valuations in other areas.
Risk of insufficient technology to produce and utilise the AVM - AVM developed on our existing DAP (Data & Analytics platform) a, demonstrating the capability. Work underway to build the AVM into our new strategic IT system (VOS - Valuation Operating System).