Problem
Zillow estimates the market value (the Zestimate) of every U.S. home that is not listed for sale, more than 100 million. The old system chained four tasks: regional price movements, adjustment of historical sale prices, an ensemble of models on the adjusted prices, and a combination step. Each task used separate models for small fixed regions such as counties. The system was hard to maintain and could not be optimized jointly: the price-adjustment model was trained independently of the price-prediction models.
Example from the article: in the Zillow Prize (Kaggle, 2017–2019), a single neural network on raw data packed these separate processes into one model and was a key component of the winning solution.
Solution in brief
One neural network takes many years of home details, geography, tax assessments and sales and learns the whole valuation jointly. Location enters as embeddings of multi-resolution geographic tiles, and time as continuous trend and cyclical seasonal features. Quantile regression gives a price range, and the estimate is suppressed when the range is too wide.
Questions this case answers
- How to replace a pipeline of many regional models with one national model that still captures local effects and time dynamics
- How to encode high-cardinality location and continuous time as model inputs so that nearby places and neighbouring dates stay close
- How to decide when a point prediction is too uncertain to show to users
- How to estimate the market value of every home in a country with one model instead of thousands of regional ones?
Tools and methods
- Neural network with categorical embeddings — one national valuation model instead of regional pipelines.
- Multi-resolution geographic tiling — grid cells at several scales as categorical features (Google S2, Uber H3).
- Trend and seasonality decomposition of time — continuous time features without step changes.
- Quantile regression — price interval and a rule for suppressing uncertain estimates.
Google or a one-time email code. No password.