Automated Property Valuation Model Specification Template

An AVM is an automated estimate based on data models. An appraisal is a licensed professional's opinion of value based on physical inspection and market analysis. AVMs are faster and cheaper ($0-5 per estimate vs. $300-600 for an appraisal) but less accurate for unique properties. In the US, federally regulated mortgage lenders are required to use licensed appraisals for most transactions over $400,000, though AVMs can be used for lower-value transactions and portfolio monitoring under recent regulatory changes.

For a statistical AVM to produce reliable estimates, you need at minimum 3 years of transaction history with at least 50-100 comparable sales within a reasonable radius of the subject property. In dense urban markets, this is easy. In rural areas with few transactions, your model will have low confidence. Start with markets where you have the most data density, then expand. Use the [product metrics framework](/glossary/aarrr-pirate-metrics) to define minimum data coverage thresholds before entering a new market.

If property valuation is a core differentiator for your product, build it. If it is a supporting feature (e.g., an estimated value badge on listings), license from a provider like HouseCanary, CoreLogic, or ATTOM. Building from scratch requires a data engineering team, a data science team, and 6-12 months of iteration. Licensing costs $0.10-2.00 per valuation depending on volume and provider. Most startups license first and build later once they have enough proprietary data to justify the investment.

Start by excluding protected characteristics (race, ethnicity, religion, national origin) from model inputs. Then test for proxy bias: zip code, school district, and neighborhood features can correlate with race and produce discriminatory outcomes. Run disparate impact analysis comparing median estimate accuracy across demographic groups. Document your bias testing methodology. Engage a third-party fair lending auditor if your AVM is used in lending or insurance decisions.

In active markets, retrain weekly to capture price trends. In slower markets, monthly is sufficient. Additionally, trigger retraining when model accuracy metrics degrade past your defined threshold (e.g., MdAPE increases by more than 1 percentage point). Always maintain the ability to roll back to the previous model version if a new training run produces worse results. ---