How Zillow Is Using AI To Improve Sale Home Value Estimates

There is no more precious right to Americans than phantasizing about the value of their home. In a country whose founding was a real estate deal - and the Louisiana Purchase, which more than doubled the country's size was yet another real estate transaction - every home owner believes the true value of their home is the highest amount anyone at a cocktail party ever said it was worth. 

But now Zillow is using AI - and even computer vision systems - in an attempt to regularize those value estimates. And good luck with that. JL 

Khari Johnson reports in Wired:

This year, inventory in the US housing market hit a record low while home prices hit a record high. Zillow previously determined the value of homes using 1,000 variations of algorithms derived for local markets. The company has also used computer vision systems that draw conclusions from photos of home listings. Now all prices nationwide will be decided by a single neural network. The new algorithm will reduce its price estimate errors by 11.5%. The new algorithm will more frequently update the value of 104 million properties in 23 US markets.

STORIES OF PEOPLE getting cash offers for their homes tens of thousands of dollars over asking price have become normal. This year, inventory in the US housing market hit a record low while home prices hit a record high. Redfin CEO Glenn Kelman recently highlighted the craziness with a tweet recounting the story of a home buyer who offered to name their first-born child after the seller—and was turned down.

As the US housing market began to overheat, in February Zillow began making initial cash offers to buy homes based on its price estimate. Now Zillow has updated its algorithm behind those estimates in a way the company says will make them more accurate—and allow Zillow to offer to buy more homes.

The owners of about 900,000 properties were initially eligible to receive automatic cash offers to purchase a home. Zillow chief analytics officer Stan Humphries said the change to its artificial intelligence will expand that pool by 30 percent. A company spokesperson said Zillow Offers can close sales in as little as one week.

Zillow previously determined the value of homes using nearly 1,000 variations of algorithms derived for local markets. Now all prices nationwide will be decided by a single neural network. Zillow says the new algorithm will reduce its price estimate errors by 11.5 percent for off-market homes in nearly 30 regions across the US. Compared with the previous version of the algorithm, errors shrank the most in Phoenix, followed by San Antonio, Tampa, and Houston.

Using the new algorithm, Zillow will more frequently update its estimates of the value of 104 million properties in 23 US markets. Around the time the company was founded in 2005, valuations were updated monthly. More recently, they’ve been updated several times a week; now, some estimates may be updated daily.

Zillow’s home estimates are popular topics of local conversation, particularly in hot housing markets. A recent Saturday Night Live skit compared surfing Zillow to phone sex and joked that, for people in their late thirties, “the pleasure you once got from sex now comes from looking at other people's houses.”

Zillow has been developing its home valuation algorithm for 15 years. A move toward a single neural network began in 2019 following a public competition to improve its estimates that drew more than 3,000 teams vying for a $1 million prize. Two out of three finalists in that contest harnessed approaches based on neural networks, a form of deep learning that Zillow says is more capable of recognizing the relationship between data used to estimate a home valuation.

For example, Humphries said the new approach is better able to grasp the value of waterfront property or how valuations are influenced by the size of nearby homes.

“The old approach would have difficulty in that county understanding the value of waterfront homes,” he said. “Neural networks that are trained on the entire country could take information from other parts of the country about the value of waterfront and apply it in a local geography, even if there weren't a lot of homes like that in that geography.”

Home value estimates are based on dozens of factors about a property, including square footage and location. Some listings also include data drawn from tax assessments and sales records. Since 2016, the company has also used computer vision systems that draw conclusions from photos of home listings.

By expanding its program to buy homes, the company hopes to bring the one-click sensibility of ecommerce into the real estate market, CEO Rich Barton said. “Eventually we want to have most houses in the country inside of that live-offer buy box,” he said.

The company also hopes to take advantage of what it calls “the great reshuffling,” as more people look to work from home more often, and millennials look to buy homes.

Billions of people look at Zillow listings every year, many, Barton said, to dream or engage in a bit of voyeurism. For them, Zillow tries to “engage or entertain,” but for people buying and selling homes, Zillow wants to sell more services. Beyond buying homes, Zillow is also increasing services like home loans, home insurance, and help going through the escrow process.