Valuation Group
How the group cut their AVM override rate from 31% to 12%
A quarterly-updated comparable dataset was causing outlier valuations at scale. Live data rebuilt the model's credibility.
The challenge
Their automated valuation model had taken 18 months to build. Their team of RICS valuers had designed it carefully, calibrated it against hundreds of manual valuations, and built a confidence-scoring system that flagged when the model needed human review. In theory, it was exactly what desktop valuation should look like.
The problem was the data underneath it. The comparable sales dataset that powered the model was updated quarterly — curated manually, validated by the team, and fed into the model as a bulk update. In a stable market, the 90-day lag was manageable. But in a market that moved — a rate change, a planning announcement, a localised supply shock — the model was working from stale evidence before the ink was dry on the update.
The symptom was a 31% manual override rate: nearly one in three valuations required a human to step in and correct the model's output. That number made scale impossible. It meant the AVM wasn't automating workload — it was creating audit work. Clients noticed the inconsistency. The team's confidence in their own product was eroding.
Sarah Mitchell, Head of Analytics, had been tracking the override rate for two quarters. The correlation was clear: overrides spiked when comp data was more than 6 weeks old. The solution was obvious — live comparable data. The question was finding a source that could handle 3,000+ valuations a year at a cost that made the AVM economics work.
"Our AVM was only as good as its last data update. Now we're pulling live comps for every single valuation. The quality improvement was immediately visible to our clients."
The solution
the group rebuilt their AVM data pipeline to pull live Homedata comparables for every single valuation — replacing the quarterly bulk update with real-time data at point of need.
The architectural change was significant but straightforward. Instead of a model that ingested a periodic bulk dataset, the group rebuilt the comparable evidence layer as a live API call. Every time the AVM generates a valuation, it calls the comparables endpoint for that UPRN — pulling up to 200 bedroom-matched sales within half a mile, ranked by proximity, from the complete Land Registry dataset.
The comparables endpoint (/api/comparables/{'{uprn}'}/) is supplemented by outcode-level price trends to provide the directional market context the model needs to weight recent transactions appropriately. The property record endpoint confirms the floor area — a critical calibration input that had previously required manual verification on a significant proportion of instructions.
Sold history gives the model 30 years of transaction data for each subject property — enabling it to factor in the individual property's price trajectory, not just the surrounding market. Together, these four data sources replaced everything the group had previously maintained as a curated internal dataset.
The model's confidence score improved immediately. The override rate fell from 31% to 12% within the first month. The group was able to expand into new postcodes where they'd previously avoided operating because their comp coverage was thin.
/api/comparables/{'{uprn}'}/
Up to 200 comps, 0.5mi PostGIS radius, bedroom-matched, distance-ranked. The AVM evidence core.
/api/v1/price_trends/?outcode=
Median prices, transaction volumes, YoY change — the directional market signal the model weights comps against.
/api/v1/properties/{'{uprn}'}/
Bedrooms, property type, floor area — critical calibration inputs for AVM model accuracy.
/api/v1/property_sales/?uprn=
30 years of Land Registry data for the subject property — individual price trajectory context.
The results
Integration
13 API calls per valuation: comparables (10) + price trends (1) + property record (1) + sold history (1). All four calls are made in parallel at valuation time. The API responses feed directly into Their Python-based AVM scoring model as structured inputs. At 3,000+ valuations a year, the group runs approximately 39,000 API calls per month — well within the Growth tier. The rebuild of the data pipeline layer took Their engineering team four weeks; the model retraining on live data took a further two weeks before it reached production confidence levels.
"Our AVM was only as good as its last data update. Now we're pulling live comps for every single valuation. The quality improvement was immediately visible to our clients."
Read more customer stories
See how estate agents, surveyors, CRM providers, and proptech companies use Homedata.
All customer stories →Get your free API key
100 calls/month free. No credit card. Access live comparables, price trends, and property records.
Start building freeFree tier · No credit card required