What if an AI model could estimate property prices more accurately — while also explaining why it reached each valuation? That is the central question behind this research on automated valuation in England and Wales.
Using property transaction data from 2011 to 2019, the authors develop a high-fidelity Automated Valuation Model (AVM) using AutoGluon, an ensemble machine-learning framework. The study is not only about prediction accuracy. It also asks whether an AI valuation can be transparent and useful for real decisions such as mortgage lending, investment analysis, and housing policy.
Why accuracy is the foundation
A central argument of the paper is that predictive accuracy is not just a desirable feature. It is the foundation for everything that comes after it.
If a model's price estimates are unreliable, then its uncertainty intervals will also be difficult to trust. Its explanations may be technically interesting, but they will not provide much practical value. The research therefore treats accuracy as a prerequisite for two important functions: measuring how uncertain a valuation is, and explaining which factors influenced the prediction.
The model achieves over 96% accuracy — a significant improvement over the 70–85% typical of traditional valuation approaches. This is not just a statistical win. In property markets, where a valuation can affect borrowing, investment, taxation, and public policy, that gap matters enormously.
A basic AVM gives a predicted value. But a single number can create a false sense of precision. This model also produces prediction intervals — showing a range in which the actual transaction price may fall. A valuation of £400,000 is very different if the likely range is narrow versus if the model considers values between £320,000 and £480,000 plausible. For mortgage lenders and investors, uncertainty is not a weakness to hide — it is part of the information needed to assess risk.
Making the model explainable
Machine-learning models are often criticised as black boxes. They may produce accurate predictions, but users can struggle to understand the reasoning behind them. To address this, the researchers use SHAP values (SHapley Additive exPlanations) to identify how different property characteristics contribute to each individual valuation.
The analysis highlights factors such as property size, location, and energy efficiency. This allows users to ask not only "What is this property worth?" but also "Why does the model think it is worth that amount?"
This transparency is what separates a useful AI tool from a technically impressive but practically opaque one.
Location still matters — but differently
One interesting aspect of the research is that the AI model does not replace the importance of location. Instead, it learns to represent location in a more detailed and context-sensitive way.
Property values are shaped by local markets, and the effect of a property characteristic can vary between regions. A feature that adds value in one area may have a different effect somewhere else. This is one reason machine learning can be useful in real estate — it can identify non-linear relationships and interactions that are difficult to capture with simpler models.
Who should pay attention
| Audience | What this means for you |
|---|---|
| Mortgage lenders | Prediction intervals give you risk-adjusted valuations rather than point estimates — better for underwriting decisions |
| Property investors | SHAP analysis reveals which features are driving value in specific markets — useful for portfolio analysis and acquisition screening |
| Surveyors & valuers | AI augments professional judgement rather than replacing it — the model shows its reasoning, giving you better evidence to work with |
| Policymakers | The UK's Valuation Office Agency is already developing AVMs for large-scale property valuation — this research shows what good practice looks like |
The bigger picture
AI may change how property is valued, but the most useful systems will not simply produce faster estimates. They will show their uncertainty, explain their reasoning, and help professionals make better decisions.
That is the real promise of explainable automated valuation: not removing human judgement, but giving it better evidence to work with. An AVM should not be treated as a complete replacement for professional expertise. Its usefulness depends on data quality, model monitoring, appropriate uncertainty estimates, and human oversight.
What makes this research important is that it moves the discussion beyond the usual question of whether AI is more accurate. Accuracy matters, but it is only the beginning. For an AI valuation model to be trusted, users also need to know how confident the model is, which factors influenced the result, how performance varies across markets, and when the model may be outside its reliable range.
That is why the combination of AutoGluon, prediction intervals, and SHAP analysis matters. It creates a more complete valuation tool — one that aims to be accurate, informative, and accountable.
Explore more on AI and real estate
Sherry Xu's research and books cover AI, PropTech, and investment strategy in the built environment.