Zillow’s Zestimate improves by 20% in Phoenix
Zillow today launches significant upgrades to its Zestimate home valuation model. The changes allow the algorithm to react more quickly to current market trends and improve the national median error rate to 6.9% — an improvement of nearly a full percentage point for more than 104 million off-market homes. These upgrades have improved the accuracy of the nearly 20,000 Zestimates available for homes in Phoenix by 20%.
The new Zestimate algorithm leverages neural networks, the latest machine learning approach, and incorporates deeper history of property data such as sales transactions, tax assessments and public records, in addition to home details such as square footage and location.
Neural networks are artificial intelligence systems that imitate how the human brain works. They are able to map hundreds of millions of data points efficiently, drawing connections among inputs and using the relationships formed to produce or predict an output. In the case of the Zestimate algorithm, the neural network model correlates home facts, location, housing market trends and home values.
As a result of this update, the Zestimate can now react more quickly to dynamic market conditions, providing homeowners with a more accurate estimate [prediction] of a home’s current value. In addition, transition to a neural network-based model will reduce Zestimate processing time.
“Since we introduced the Zestimate in 2006, we have never stopped innovating in order to provide consumers with the most accurate home valuations,” said Dr. Stan Humphries, Zillow chief analytics officer and creator of the Zestimate. “The new architecture we’re debuting today represents another significant step forward in our efforts to harness big data to create more certainty for consumers, which leads to better decisions.”
Fifteen years ago, the Zestimate gave people instant access for the first time to an estimated value for millions of homes across America for free. Over the past decade and a half, Zillow has released multiple major Zestimate algorithm updates as well as incremental improvements between major upgrades, and now calculates valuations for more than 104 million homes across the country.
As a result of the company’s increasing confidence in Zestimate accuracy, in February Zillow began using the Zestimate as a live, initial cash offer through its home buying program, Zillow Offers. The Zestimate is an initial cash offer on about 900,000 eligible homes across 23 markets. With this latest update and increased Zestimate accuracy, the number of homes eligible for a cash offer will likely increase by 30%.
Applying a neural network model to a national real estate dataset was an innovation used by the winning team of Zillow Prize, the two-year, $1 million data science competition that included more than 3,800 teams from 91 countries working to improve the Zestimate. One member of the team, Jordan Meyer, is now a senior applied scientist at Zillow and works on home valuations for Zillow Offers.