Every property has a price. But unless you've just sold it, that price is a guess. Estate agents estimate. Banks commission surveyors. Online tools spit out a number with little explanation of how they got there.
We wanted to do better. So we built a machine learning valuation model trained on nearly 8 million real UK property transactions and over 2 billion individual datapoints. It now produces instant valuations for 19 million UK properties, and on average, it's accurate to within 8% of the actual sale price.
Here's how we did it, and more importantly, how we know it works.
What goes into a valuation?
A property's value isn't determined by a single factor. It's shaped by hundreds of overlapping signals: the size of the building, the street it sits on, what similar homes sold for last month, whether there's a train station nearby, even the energy rating on the boiler.
Our model considers 254 features for every single property. These are drawn from six major data sources:
- Land Registry: 12 years of actual transaction prices across England and Wales
- Energy Performance Certificates (EPC): floor area, construction age, property type, heating systems, energy efficiency ratings, and room counts
- UK House Price Index: regional and local market trends, including 1-year and 5-year price momentum
- Transport data: distance to the nearest train station and rail link
- School data: proximity to primary and secondary schools
- Property characteristics: 46 individual features including bedrooms, bathrooms, garden type, parking, period features, and more
But the real magic happens with the comparable analysis.
168 features from comparable sales alone
The single biggest predictor of what a property is worth is what similar properties nearby have recently sold for. This sounds simple, but doing it well across 19 million properties is an enormous challenge.
For every property, we find comparable sales at two levels: properties on the same street and properties in the wider postcode sector. We then rank those comparables three different ways:
- Size view: prioritises properties with the most similar floor area
- Balanced view: weights size and recency equally
- Recency view: prioritises the most recent sales
Each view captures a different angle on value. A house that's the same size matters. A house that sold last month matters. Our model gets all three perspectives for both street and sector comparables, and it learns which to trust most depending on the situation.
This comparable engine alone generates 168 features per property: individual comparable prices, price-per-square-metre, market-adjusted values, recency, size differences, and aggregate statistics. It's the backbone of every valuation we produce.
The machine learning engine
At its core, our model is a machine learning system that has studied nearly 8 million real property sales. It builds thousands of layers of decision logic, each one refining the predictions of the last, until it arrives at the most accurate estimate possible for each property.
We made three important design choices:
Trained to minimise real-world error. The model is optimised to get as close as possible to the actual sale price in pounds, not just on average, but for every individual property. This makes it resilient to unusual sales like probate, auction, and repossession transactions that could otherwise throw off the numbers.
Tested against the future, not the past. We don't shuffle the data randomly. We train on older sales and test exclusively on the most recent 5 months of transactions. The model has to predict prices it has genuinely never seen before. This is the most honest test of real-world accuracy.
Calibrated against real market conditions. After the initial prediction, we apply three layers of calibration: adjusting for recent house price movements in the local area, correcting for known tendencies at different price levels, and fine-tuning based on how much data is available for each property.
The results
We tested the model on 68,800 actual property sales from the last 5 months. These are real transactions the model never saw during training. No cherry-picking, no synthetic data, no tricks.
Overall accuracy
| Metric | Value |
|---|---|
| Average error | 7.93% |
| Median error | 4.74% |
| Average error in pounds | £25,799 |
| Median error in pounds | £13,558 |
| Variance explained (R²) | 93.6% |
Half of all valuations land within £13,558 of the actual sale price. The average error is about 8%, but the median is under 5%, meaning most predictions are significantly better than the headline number.
How often are we close?
| Threshold | % of valuations |
|---|---|
| Within 1% | 19.2% |
| Within 2% | 29.3% |
| Within 5% | 51.6% |
| Within 10% | 75.4% |
| Within 15% | 87.1% |
| Within 20% | 93.0% |
Three out of four valuations land within 10% of the actual sale price. 93% are within 20%. And nearly one in five nail it to within 1%.
No systematic bias
A model that's consistently too high or too low isn't useful to anyone. Ours shows almost no directional bias:
- Average signed error: just -£398 (essentially zero on a median house price of around £280,000)
- 50.1% valued above the sale price, 49.9% below: a near-perfect split
The model isn't optimistic or pessimistic. It's just as likely to estimate slightly high as slightly low.
Accuracy by price band
| Price band | Avg error | Within 10% |
|---|---|---|
| £100k to £150k | 14.2% | 64.7% |
| £150k to £200k | 9.0% | 71.2% |
| £200k to £250k | 7.8% | 76.1% |
| £250k to £300k | 6.9% | 78.6% |
| £300k to £400k | 6.5% | 79.3% |
| £400k to £500k | 6.5% | 78.5% |
| £500k to £750k | 6.8% | 76.4% |
| £750k to £1M | 7.0% | 75.1% |
| £1M to £2M | 10.8% | 56.8% |
| £2M+ | 11.5% | 58.1% |
The model is at its sharpest in the £250k to £750k range, the heart of the UK housing market, where accuracy sits between 6.5% and 6.9%. It performs well across the full spectrum, though naturally the extremes are harder: cheaper properties often involve unusual circumstances, and luxury homes tend to be one-of-a-kind.
Where the model really shines
When we have rich data for a property (full EPC, measured floor area, plenty of recent comparable sales nearby) accuracy gets even better:
| Segment | Avg error | Within 10% |
|---|---|---|
| All properties | 7.93% | 75.4% |
| Data-rich properties | 7.13% | 78.7% |
| Core market (£200k to £500k, data-rich) | 6.18% | 81.3% |
| 3+ recent sales on the same street | 6.1% | 82.8% |
| 2+ previous sales on record | 7.0% | 79.2% |
For the typical UK property (a three-bed semi in the £200k to £500k range with an EPC and recent nearby sales) the model achieves a 6.18% average error with over 81% of predictions within 10%.
Consistent across every property type
| Type | Avg error | Within 10% |
|---|---|---|
| House | 7.9% | 75.9% |
| Flat | 7.9% | 75.7% |
| Semi-Detached | 7.9% | 75.2% |
| End-Terrace | 7.7% | 75.9% |
| Mid-Terrace | 7.5% | 76.4% |
| Detached | 8.4% | 74.5% |
| Bungalow | 8.4% | 70.6% |
| Maisonette | 8.7% | 78.6% |
No property type is a blind spot. Whether you're looking at a terraced house in Manchester, a flat in London, or a detached home in the Cotswolds, the model delivers consistent results.
What matters most to the model
We can look inside the model to see which factors carry the most weight. The answer lines up with how property professionals think about value:
- What similar properties nearby have sold for: this dominates the model's thinking, exactly as you'd expect
- Specific recent comparable sales: individual transactions on the same street and in the same postcode sector, adjusted for market movement
- Council tax band: a strong signal, since local authorities have already assessed every property's relative value
- The property's own sale history: what it last sold for, projected forward using local market trends
- Property type and structure: detached vs semi vs terrace, number of bedrooms, floor area
- Local market conditions: house price trends, sales activity, and price momentum in the area
The top 10 factors account for over half of every valuation. But the remaining 244 features collectively contribute the other 45%. This is where the model picks up on subtler value signals like underfloor heating, period features, views, solar panels, and proximity to good schools.
What this means for you
This model powers every property page on our site. When you look up an address, the price you see isn't pulled from a simple average or a rough price-per-square-foot calculation. It's the output of a machine learning model trained on nearly 8 million actual sales, considering 254 dimensions of data, calibrated against the very latest market conditions.
It's not a formal valuation (no algorithm can replace a surveyor walking through the front door). But as an instant, data-driven estimate of what a property is likely worth today, we believe it's among the most accurate publicly available in the UK.
Every valuation on our site includes the data behind it: comparable sales, EPC data, transaction history, and local market trends. We show our working because we think you deserve to understand the number, not just see it.
Our valuations are updated regularly as new Land Registry data is published. All accuracy figures are based on forward-looking validation against the most recent 5 months of real transactions that the model never saw during training.