AVMs for a Proportional Property Tax

An AVM is a statistical or algorithmic system that estimates property values using historical transactions, property attributes, and location data. In taxation, AVMs operate within mass appraisal systems (IAAO, 2017) that apply consistent methods across entire jurisdictions with ongoing statistical validation.

For a proportional property tax, the AVM must produce a point estimate of each property's capital value, not just assign it to a band. Two broad model classes are used:

What matters is the overall system, not any single model. A working framework ties together data collection, regular revaluations, quality checks, and a credible appeals process.

Regression vs machine learning

Regression models remain common in tax systems because their coefficients are transparent and legally defensible. ML models, especially gradient boosting, are increasingly used where data is rich.

The choice involves a genuine tension: ML models tend to produce lower valuation error, but regression models are far easier to explain and defend at appeal. In practice, most jurisdictions manage this by layering ML on top of a regression baseline, gaining accuracy while retaining an interpretable core.

Spatial modelling

Location is the primary driver of value. Systems incorporate it through market zones, coordinate-based features, proximity and accessibility metrics, and environmental variables such as flood risk and noise. More advanced systems directly model spatial relationships rather than relying on crude zoning.

All mature tax valuation systems are hybrid, combining model outputs with human review. No jurisdiction we are aware of relies on fully automated valuations for tax purposes.

Human intervention focuses on outliers and unique properties, sales validation, and legal compliance. The balance is itself a design choice: more automation delivers consistency and lower unit cost; more human oversight catches errors the model cannot see and strengthens public legitimacy.

Even highly automated systems, such as Cook County (Illinois), retain substantial manual validation. A proportional tax covering all residential properties in England would almost certainly follow this hybrid pattern.

How often valuations are updated involves a three-way tension between fairness, cost, and political acceptability. Outdated valuations create inequities between properties whose values have moved at different rates; but frequent updates mean more visible shifts in individual tax bills, generating political resistance and driving up appeal volumes.

England's Council Tax valuations date from 1991; a proportional replacement would need to establish a regular cycle from the outset.

The Netherlands offers perhaps the closest analogue to what a proportional property tax in England would require: annual capital-value assessments of the entire housing stock, centrally audited, with a mature appeals process.

Property valuation for tax is hard for reasons that no algorithm fully solves: properties are heterogeneous (every home is to some degree unique), prices vary sharply across space, attribute data is incomplete, and models risk systematic bias against particular areas or property types.

Under a proportional tax these problems matter directly, because valuation accuracy determines how fairly the burden is shared.

Tax AVMs are evaluated using ratio studies (IAAO, 2013), not standard ML metrics:

Key risks include geographic or socio-economic bias, property-type mispricing, model drift, and data errors; the last of these is often more consequential than model choice.

A proportional tax would generate significantly more appeals than the current band system, because individual valuations create more grounds for dispute.

Appeals serve a dual function: they correct data errors and model weaknesses, and they confer legitimacy on the system as a whole. But a generous appeals process also carries real administrative cost, and high volumes can delay assessments and create uncertainty for local authority budgets.

International experience shows that systems which separate factual corrections (wrong floor area, missing extension) from valuation disputes (disagreement over market value) operate more efficiently (Almy, 2014). Factual corrections can often be resolved quickly and fed back into the model; valuation disputes require formal adjudication.

Designing the appeals process is as important as designing the model.

Data quality is the binding constraint. Three categories matter most:

Model reliability falls in low-liquidity segments such as rural areas, high-value properties, and unusual types; here the need for human review is greatest.

Transaction volumes

England has good transaction price coverage via HM Land Registry. Volume, however, is uneven:

• Around 1.0-1.2 million annual residential transactions in normal conditions
• Sharp cyclical variation driven by tax and interest rate changes
• Thin turnover in rural areas and at the top of the market

Transaction volumes alone would support modelling for mainstream residential properties. But an AVM also needs property attributes; size, type, condition, and tenure; linked to each address, and this data is where England falls short. The feasibility question is less about sale prices and more about whether a reliable attribute dataset can be assembled.

Data availability

The Netherlands maintains a unified property register with standardised attributes linked to every address. No equivalent public dataset exists in England.

However, the Valuation Office Agency (VOA) maintains its own property records, including dwelling type, area, age, bedrooms, bathrooms, parking, and plot size. These records are not publicly available but do constitute a working attribute dataset; as the VOA's Welsh AVM demonstrates.

The VOA's Welsh AVM: a working precedent

The VOA has already built and deployed an AVM for the 2028 Council Tax revaluation in Wales, covering 1.5 million domestic properties. The model uses a Gaussian Markov Random Field to estimate continuous spatial variation in property values across Wales, rather than relying on fixed geographic zones.

It was externally assured by the International Association of Assessing Officers (IAAO), who concluded that the results were "more than satisfactory".

The approach is explicitly "model assisted": the AVM generates first-pass valuations, then professional valuers review batches of properties, focusing on those near band boundaries and those flagged as unreliable. The VOA estimates this reduces revaluation costs by roughly one-third compared with purely manual valuation.

This is significant for England because it demonstrates that the VOA already holds workable property attribute data and has built the institutional capability to run a mass appraisal system. The question is whether these can scale.

The Welsh AVM is the most relevant precedent for any future English system. It confirms that the VOA can build a credible AVM using existing data and in-house expertise.

But it does not settle the question of whether the same approach can deliver the accuracy needed for a proportional tax across a property stock 16 times larger and considerably more diverse.

A proportional property tax in England could take several forms, each with different AVM requirements: