AVM research preview — Seattle

This is a v1 MVP hedonic regression. Its job is to prove we can stand up a from-scratch AVM pipeline using only free public data. Its job is not to predict house prices accurately. The label source is Zillow ZHVI at the ZIP level, which means every building in ZIP X gets the same target value. A model with a ZIP one-hot in the design matrix can therefore "memorize" the label trivially. The headline MAPE of <1% is a degenerate accuracy number, not a real one. It tells us the pipeline runs — nothing more.

What we actually learned

• dist_mi came out positive — farther from downtown = more valuable. This is a real artifact of the sample: downtown Seattle (98101) has the LOWEST ZHVI ($594k) because it's mostly condos, while the residential hills (98115 Wedgwood, 98105 U-District) are $1M+. The model is correctly learning the pattern but attributing it to distance instead of housing-type. Spatial-lag in v2 will fix this.
• age came out positive too — older buildings worth more, because Madison Park / Wedgwood pre-WWII SFR dominates the high-ZHVI ZIPs.
• age_missing = 1 adds +$38,612 — buildings without year tags correlate with higher-ZHVI areas (a story about OSM volunteer coverage, not real estate).
• The five highest-|t| features are all ZIP dummies. ZIP one-hots ate all the variance. This is the predicted failure mode.

Path to a real AVM (v2)

• Replace ZIP-level ZHVI label with per-property sale prices from King County Assessor (public record, free).
• Use OSM features to predict price deviation from ZIP ZHVI, not ZHVI itself. Forces the model to learn structure / location.
• Add spatial-lag (Moran's I, k-nearest neighbor weights) — "block-next-door" effect.
• Add Census ACS tract demographics, transit/amenity proximity.
• Walk-forward validation: hold out one ZIP, hold out one year.
• Fairness audit per CFPB AVM Quality Control rule (effective Oct 2025).