Automated Property Valuation Model with Market Feeds

The Challenge

A mortgage technology company needed an automated valuation model (AVM) to provide instant property value estimates for loan origination decisions. Their existing process relied on manual appraisals that cost $400–600 per property and took 7–14 days to complete. For refinancing and home equity line of credit (HELOC) products, this delay was causing 23% of applicants to abandon the process.

Our Approach

We built an ensemble AVM combining gradient-boosted trees (XGBoost), a geospatial feature engineering layer, and live market data feeds.

Feature Engineering: Each property is represented by 140+ features including physical characteristics (square footage, lot size, year built, room count), location features (school district ratings, walk score, crime index, distance to transit), market features (days on market for comparable sales, price-per-square-foot trends), and macroeconomic features (local employment rate, mortgage rate movements).

Comparable Sales: The model identifies comparable sales using a learned similarity metric rather than traditional appraiser rules. A separate embedding model maps each property into a 64-dimensional space where similar properties cluster together. The 20 nearest neighbours in this space, filtered by recency (< 6 months) and proximity (< 2 miles), serve as the comparable sales input to the valuation model.

Nightly Retrain: The model retrains every night on fresh MLS transaction data via an Apache Airflow pipeline. Training takes approximately 45 minutes on a single g4dn.xlarge instance. Model performance is evaluated against a holdout set of the previous day's closed transactions, and the model is only promoted to production if the median absolute percentage error (MAPE) stays below 4%.

API: Valuations are served through a FastAPI endpoint with sub-200ms response time. Each response includes the point estimate, a confidence interval, the top 5 comparable sales with photos and sale prices, and an explanation of the key value drivers.

Results