A property technology company needed an automated valuation model (AVM) that could estimate residential property values for 50 million properties across 200 metropolitan areas. The AVM needed to produce valuations within 5% of actual sale price for 70% of properties to meet lender requirements for use in mortgage origination.

The Architecture

The system combines structured property data, comparable sales analysis, and market trend modeling.

Property feature store - A central feature store maintains attributes for each property: physical characteristics (square footage, bedrooms, bathrooms, lot size, year built, garage, pool), location features (school district ratings, crime rates, walkability scores, proximity to transit and amenities), tax assessment data, and previous sale history. Data is sourced from county assessors, MLS feeds, and third-party data providers. Updates arrive daily.

Comparable sales engine - For each subject property, the system identifies the most relevant comparable sales (comps) using a similarity model that weights geographic proximity, physical similarity, sale recency, and market segment. The model learns optimal comp selection weights from historical valuation accuracy, producing better comps than distance-only selection.

Valuation model - A gradient boosted model produces the property value estimate using three feature groups: property physical features, comparable sale adjusted values, and market trend features (price index for the property’s submarket, days-on-market trends, inventory levels). The model outputs a point estimate and a confidence interval.

Confidence scoring - Each valuation receives a confidence grade (high, medium, low) based on data completeness, comp availability, and model uncertainty. Properties with recent renovations, unique characteristics, or sparse comparable sales receive lower confidence grades. Lenders use the confidence grade to determine whether additional appraisal is needed.

Key Lessons

Data quality was the dominant accuracy driver. The single biggest accuracy improvement came not from model architecture changes but from improving the underlying property data. Square footage discrepancies between tax assessor records and MLS listings were common. Building a data reconciliation pipeline that resolved these discrepancies improved median accuracy by 3 percentage points.

Hyperlocal markets required hyperlocal models. A single national model performed poorly because real estate markets are intensely local. A 1,500-square-foot home might be worth $200,000 in one city and $800,000 in another, but beyond that, neighborhood-level factors (specific school attendance zones, flood zones, view corridors) created variation that metro-level features could not capture. The team built metro-specific model variants that shared a common architecture but were trained on local data.

Recent renovations were the biggest error source. The model had no way to know that a property had undergone a $100,000 kitchen renovation since its last sale. Properties with recent renovations were consistently undervalued. Adding permit data (building permits for renovation work) as a feature improved accuracy for renovated properties by 8 percentage points.

Market transition periods degraded accuracy. During rapid market shifts (interest rate changes, pandemic-era volatility), the model’s training data lagged reality. Increasing the weight of recent comps and adding market momentum features (3-month price change rate) improved responsiveness to market transitions.

Results

The AVM achieved within-5% accuracy for 72% of properties, meeting lender requirements. Median absolute percentage error was 3.8% across all properties. Valuation processing time was under 2 seconds per property, enabling real-time use in mortgage origination workflows. The system reduced appraisal costs by 40% for lender clients by identifying properties where a full appraisal could be waived based on AVM confidence.