Overview
Model name
HousingVest Valuator v0.7
Model type
Ensemble (gradient-boosted regression + transformer-based tabular model)
Purpose
Estimate property value (point + confidence interval) for residential real estate in target U.S. census tracts, with integrated bias-audit output
Intended users
Real-estate investment funds, family offices, CDFIs, Opportunity Zone funds, LIHTC syndicators, sophisticated retail investors, municipal housing-finance agencies
Not intended for
Formal lending appraisals (use a licensed appraiser); transactional valuations (use a licensed AVM compliant with FIRREA)
Training Data
Sources
County recorder transaction records (multi-jurisdiction, scalable to any U.S. county with public recorder data)
· MLS-fed sold transactions (where access available)
· Tax-assessor records
· FHFA House Price Index (county-level)
· U.S. Census Bureau ACS 5-Year Estimates (demographic overlay)
· FEMA National Flood Hazard Layer (risk overlay)
Time period
2018 – 2025
Geographic coverage (current)
South Florida pilot regions; architecture designed for nationwide county-level expansion
Performance
Mean Absolute Error (MAE) in target census tracts
6.7%
Industry benchmark MAE for general-purpose AVMs in comparable tracts
approximately 15%
Reduction in valuation error
approximately 56%
Coverage
Residential single-family, 2-4 unit, multifamily up to 12 units, and manufactured housing
Bias Audit
Framework
HUD Disparate-Impact Rule (24 C.F.R. § 100.500)
Compliance
Interagency AVM Quality Control Rule (June 24, 2024)
Methodology
Disparate Impact Ratio (DI Ratio) computed on each valuation against tract-level demographic composition (ACS 5-Year)
Threshold
DI Ratio < 0.80 triggers manual review
Output
Every valuation produces a downloadable Bias Audit Report PDF
Limitations
Geographic coverage
Currently limited to listed Florida counties and one South Carolina pilot
Accuracy degradation
Model accuracy degrades for properties with no comparable transactions within 1.0 mile within 12 months
Manufactured housing
Valuations have wider confidence bands due to lower transaction volume
New construction
Model is not appropriate for new-construction pre-completion valuations
Not a substitute
Model output is informational and decision-support — not a substitute for licensed appraisal
Responsible AI Principles
NIST AI RMF
Compliance with NIST AI Risk Management Framework
EO 14179
Alignment with Executive Order 14179 (Removing Barriers to American Leadership in AI, Jan. 23, 2025)
White House AI Action Plan
Alignment with the White House AI Action Plan (July 2025)
Privacy by design
No individual demographic data stored at user-record level
Audit logging
Every inference is logged (date, inputs, output, DI Ratio)
Version History
Current version
v0.7 — May 2026
Retraining cadence
Quarterly
Next retraining target
July 2026
This Model Card is published in accordance with the Responsible AI principles articulated in NIST AI RMF and EO 14179.