Explainable machine learning valuation in thin housing markets: evidence from Newcastle upon Tyne and Sunderland
Dr Victor Joes Obasse
Lakeshore Management College
Plain English summary
This study tests how accurately open British transaction data can value homes in two thin, lower priced housing markets, and whether machine learning improves on standard methods without bias against cheaper homes. It is the first full empirical test of the method behind ValuInsight AI.
The data were HM Land Registry Price Paid Data for Newcastle upon Tyne and Sunderland between 2018 and 2024, covering 51,899 standard residential sales. Sales from 2018 served only as history, so the modelling sample ran from 2019 to 2024 with 44,841 sales. Every feature was built only from sales that happened before the one being valued.
The models were tested on all 6,857 sales of 2024, refitted each quarter so that no model ever saw the sale it was valuing. The ensemble of gradient boosting and random forests reached a median absolute percentage error of 12.6 per cent, against 20.6 per cent for a hedonic regression and 19.8 per cent for a comparable sales baseline, a cut in typical error of about a third.
That is useful but modest. Fewer than half of the ensemble estimates, 41.4 per cent, fell within 10 per cent of the price, and its coefficient of dispersion of 18.0 exceeds the IAAO guideline of 15. The method suits screening and triage, not lending decisions on its own.
Accuracy rested on fine grained local price history: the postcode unit, the street, and the dwelling’s own previous sale. Terraced houses and flats were the hardest to value, new builds the easiest. Where a previous sale was on record the median error fell from 13.6 to 9.2 per cent, and accuracy held up in the thinnest markets.
Fairness depends on how the question is asked. Grouped by sale price the model looks strongly regressive, overvaluing the cheapest fifth of homes by about 12.8 per cent. Grouped instead by predicted value the ensemble’s ratios stay between 0.98 and 1.01, and it was the least regressive of the five methods on either measure.
The paper is careful about what it does not show. It does not reach the accuracy a lender needs without an inspection, does not show that interest rates improve valuation over a quarter, and does not show that the method generalises beyond Tyne and Wear.
Key figures from the test
- • 51,899 Land Registry standard sales studied, 44,841 of them modelled, 2018 to 2024
- • 6,857 sales of 2024 valued strictly out of time, quarter by quarter
- • 12.6% median error and £34,006 median absolute error
- • 41.4% of estimates within 10% of the sale price
- • About a third lower error than a hedonic regression or a comparable sales baseline
- • 9.2% median error where the property had sold before in the data
- • Coefficient of dispersion 18.0, above the IAAO guideline of 15
Out of time accuracy, all 2024 sales (n = 6,857)
| Method | Mean abs. error | Median error |
|---|---|---|
Ensemble (ValuInsight AI method) 0.35 random forest, 0.65 gradient boosting | £34,006 | 12.6% |
Comparable sales baseline Recent local sales of the same type | £52,707 | 19.8% |
Hedonic regression Type, tenure, new build, sector and month effects | £54,645 | 20.6% |
Rolling origin by quarter, each model refitted on all sales from January 2019 to the day before the quarter. Gradient boosting alone came within £138 of the ensemble, so the ensemble adds stability rather than accuracy.
What the evidence supports
- •Tree ensembles built on open transaction data cut typical valuation error by about a third against a hedonic regression or a simple comparable approach on the same inputs.
- •The gain comes mainly from fine grained price history: the postcode unit, the street and the dwelling’s own previous sale.
- •Accuracy holds in thin postcode markets, where fewer than 10 comparable sales exist (median error 12.8 per cent, on a small sample of 165 sales).
- •The ensemble is the least regressive of the methods tested, and much of its apparent regressiveness is an artefact of grouping by sale price.
What the paper does not support
- •It does not match the accuracy a lender needs for a decision without inspection. A coefficient of dispersion of 18.0 and 41.4 per cent within 10 per cent fall well short.
- •It does not show that macroeconomic data, including Bank Rate, improves valuation over a horizon of weeks.
- •It does not show that the method generalises beyond Newcastle upon Tyne and Sunderland.
Where errors concentrate
- •Terraced houses and flats are the hardest to value, the most varied stock where size and condition differ most within a street.
- •New builds are the easiest, because developers sell similar units at similar prices.
- •A previous sale on record lowers the median error from 13.6 to 9.2 per cent.
- •Sunderland is valued slightly less accurately than Newcastle, because an error of the same size in pounds is a larger share of a cheaper price.
What drives the valuations
The models rely most on the price history of the same postcode unit, then property type, the dwelling’s own previous sale, and street and sector premia. That is what a valuer would expect. Calendar time and district carry no importance once the six month market anchor is in place, and adding Bank Rate changed the mean error by £40, with a confidence interval spanning zero.
Stated limitations
- 1.No floor area, room count, condition, plot size or energy rating in the price data. Much of the remaining error is size and condition within a street.
- 2.Repeat sales were identified only within 2018 to 2024, so 77.8 per cent of 2024 sales had no earlier price on record.
- 3.Two districts in one region. Markets with different stock, such as London flats or rural detached houses, may behave differently.
- 4.Category B transactions were excluded, including repossessions and many buy to let purchases, so the models are untested on them.
- 5.The thinnest market group held only 165 test sales, so that finding is provisional.
- 6.Permutation importance gives a global view only. Individual valuation explanations through SHAP values were not computed.
- 7.The target is the recorded price, which can depart from market value through incentives or atypical motivation.
The model has no floor area or condition data, was tested in two districts only, and is suitable for screening rather than lending decisions. Indicative estimates only.
Obasse, V.J. Explainable Machine Learning Valuation in Thin Housing Markets: Evidence from Newcastle upon Tyne and Sunderland, England. Analysis code and cleaned data available from the author on request.
Contains HM Land Registry data © Crown copyright and database right 2025. This data is licensed under the Open Government Licence v3.0.