Home Price Estimators

Predicting a house price from its size and location is the textbook example of linear regression. Real-estate websites do it at national scale.

Zillow launched the Zestimate in 2006, giving homeowners a free estimate that used to require an agent or appraiser. Such automated valuation models use:

• Features of the home: floor area, bedrooms, bathrooms, age, lot size (feature engineering).
• Location, which often matters most.
• Recent sales of comparable nearby homes, tax assessments and listing information.

Early versions used regression and tree ensembles like gradient boosting. In 2021, Zillow switched to a neural network version and reported a median error of 6.9% for homes not on the market and 1.9% for listed homes, which have an asking price as a strong hint.

“Median error” hides the tails: half of all estimates are off by more than that, and unusual homes can be far off. Zillow learned this the hard way when, in 2021, it closed its business of buying homes based on its own price predictions after large losses in a fast-moving market. Good regression metrics and an honest view of uncertainty matter as much as the model.