Even with an exceptionally good model, is there a minimum level of prediction error that we simply cannot eliminate?
An important distinction in residential property valuation is the difference between a property's underlying market value and the price at which it actually transacts. Two otherwise identical properties could sell for different prices because of negotiations, individual buyer preferences, seller urgency or other circumstances specific to the transaction.
An AVM aims to estimate market value, but it is typically trained and evaluated using observed transaction prices. Those prices contain an element of noise that may be impossible to predict from property characteristics alone.
This raises an interesting question: how much of an AVM's prediction error might be irreducible?
Measuring the Limits of AVM Accuracy
Measuring this is difficult. No two properties are exactly alike, and they rarely transact at precisely the same time. However, new-build developments offer an interesting opportunity to investigate. New-build properties within the same postcode, with the same property characteristics, are often relatively homogeneous. They are also less likely to differ because of extensions, renovations or variations in property condition. At MIAC, we explored this using new-build transactions across London. We identified groups of comparable properties that transacted within six-month periods and indexed their transaction prices to a common date using local property price indices. We then measured the dispersion of transaction prices within these groups.
The results were interesting:
- Across London, the median absolute percentage deviation from comparable-group median prices was approximately 6%.
- At borough level, this measure ranged from approximately 2% in Bexley to 17% in Westminster. The Boroughs with the highest prices generally saw the largest variation.
- New-build flats exhibited substantially greater price dispersion than detached, semi-detached or terraced houses.
These findings suggest that even among properties with very similar observable characteristics, transaction prices can differ materially. Of course, not all of this variation is necessarily irreducible. Some may arise from indexation error, while differences in characteristics we do not have enough data on such as views may also explain part of the dispersion, particularly for flats. Also this analysis is biased towards new builds. Nevertheless, the analysis provides an interesting starting point for understanding the limits of AVM accuracy.
How can we judge the accuracy of an AVM without taking into account this irreducible error floor? And perhaps more importantly, how close are today's AVMs to that theoretical limit?
The objective of an AVM is to maximise the proportion of properties valued within an acceptable error tolerance, for example ±5%. However, this analysis suggests that the irreducible valuation error in London may be approximately 6%. This raises an important question about the appropriateness of our accuracy criteria: if the underlying uncertainty in property values exceeds our target tolerance, we may be approaching the practical limits of AVM accuracy. Consequently, further improvements in model sophistication may deliver diminishing returns, and greater emphasis should be placed on understanding irreducible uncertainty, setting realistic performance expectations and maximising the proportion of properties that can be valued reliably within those constraints.
At MIAC, we are continually enhancing our AVMs, improving the data, modelling techniques and property characteristics we use, with the aim of getting as close as possible to the limits of achievable valuation accuracy. If you'd like to learn more about MIAC's AVMs and our approach to improving valuation accuracy, please get in touch.