Mind Matters Natural and Artificial Intelligence News and Analysis
devastated kitchen
Devastated kitchen in a demolition house
Photo licensed via Adobe Stock

Zillow’s House-Flipping Misadventure

People know more about their homes and neighborhoods than algorithms could possibly know

Zillow has become the largest real estate website in the United States, with detailed data on millions of homes including square footage, number of bedrooms, number of bathrooms, year built, and price and tax history. If the home is or has been offered for sale or rent, there are also interior and exterior photographs and more.

Zillow uses its massive database to provide proprietary computer-generated estimates of market values and rental values. When Zillow went public in 2011, its IPO filing with the SEC boasted that,

[O]ur algorithms will automatically generate a new set of valuation rules based on the constantly changing universe of data included in our database. This allows us to provide timely home value information on a massive scale. Three times a week, we create more than 500,000 unique valuation models, built atop 3.2 terabytes of data, to generate current Zestimates on more than 70 million U.S. homes.

Wow! Ten years later, there are even more data and more homes in its database.

Part of Zillow’s appeal is voyeuristic. We can see how much our relatives, friends, and neighbors paid for their homes. We can track what is happening to the value of our homes. Around 200 million people go Zillow’s website each month and most are digital looky-loos.

Zillow’s original business model relied on advertising revenue from realtors, builders, and lenders but it has attempted to capitalize on its data and brand name by expanding into a variety of real-estate-related ventures, including a house-flipping business called Zillow Offers that was launched in 2018. The idea was that Zillow could use its market-value estimates to make a cash offer to homeowners who want a quick-and-easy sale, net of a service fee that averaged around 7.5 percent (compared to a typical realtor’s commission of 5 to 6 percent). After a check to see if any major repairs were needed, the offer would be finalized. If accepted, Zillow would use local contractors to pretty the house up (minor repairs and maybe a new carpet and a fresh coat of paint) and then sell the house for a fast profit. It was industrial-scale house flipping — and it seemed like a good plan.

Some appreciate the convenience and trust Zillow’s reputation. It can be a real pain to keep a home tidy enough for showings and to have to leave your house every time a realtor wants to come by with a potential buyer. Not to mention dealing with possible thefts and snoopy burglars. Zillow bought thousands of houses before belatedly recognizing the inherent problems.

When Zillow Offers was launched, a spokesperson told stockholders that, 

We’ve taken a lot of prudent measures to mitigate and minimize risk here. The most obvious one is that we will see issues coming because of consumer demand trends and data that we have on the housing market. And we can adjust our purchasing and we can adjust our selling commensurate with market conditions.

Basically, we have the most data, so we have the best market-value estimates.

However, it is often better to have good data than more data. A timeless aphorism is that the three most important things in real estate are locationlocationlocation. A second, related aphorism is that all real estate is local. Data on homes thousands, hundreds, or even a few miles away are not necessarily relevant and possibly misleading. Even homes a few hundred feet apart can sell for quite different prices because of proximity to various amenities and disamenities, such as schools, parks, metro stations, noise pollution, and overhead power lines.

Even seemingly identical adjacent homes can sell for different prices because buyers and realtors, but not algorithms, might know, for example, that one house had been renovated recently or that someone had been murdered in one house.

I’ve seen homes sell for 50 to 100 percent over Zillow’s estimated market value because they were designed by a famous architect or owned by a celebrity. After the sale, Zillow increased the estimated market values of nearby homes because its algorithms did not know that the nearby homes had not been designed by famous architects or owned by celebrities.

Algorithms also have trouble valuing homes with quirky layouts — indeed, recognizing that a layout is quirky — or taking into account that the house next door has beer bottles, pit bulls, and cars on blocks in the front yard. I’ve seen homes in my neighborhood sell for a premium without going on the market because the buyers were eager to live next-door to their children and grandchildren or wanted to live on the street where they grew up. Algorithms would conclude that the value of my home went up too. I know it did not.

The fundamental problem for computer algorithms is asymmetric information. People know more about their homes and neighborhoods than algorithms could possibly know. They can consequently take advantage of the algorithms’ ignorance. Suppose that a really good algorithm gives estimates that are unbiased and generally within 5 percent of the actual market value. It might seem that the overestimates and underestimates will balance out, with the algorithm paying fair prices, on average.

However, informed sellers (who are also likely to have talked to local realtors) need not accept algorithmic generated offers. If the algorithmic offer is too high, the seller may well accept. If the algorithmic offer is too low, the seller is more likely to use a local realtor. On average, the algorithm will pay too much for the homes they buy.

After a sale, the algorithmic buyer may discover how hard it is to find local contractors who are honest, competent and inexpensive and, if the home is not sold quickly, how expensive it can be to pay for insurance, home-owner-association fees, and interest on the loan used to finance the purchase.

Despite the high service fees and the booming real estate market, Zillow Offers managed to lose money — lots of money — because its algorithms couldn’t take into account everything that homeowners and realtors know. Zillow paid too much for too many houses.

The CEO admitted that the company was “unintentionally purchasing homes at higher prices.” He also said that, “We’ve determined the unpredictability in forecasting home prices far exceeds what we anticipated” and “Fundamentally, we have been unable to predict future pricing of homes to a level of accuracy that makes this a safe business to be in.”

On November 2, Zillow announced that it was closing down its home-flipping misadventure, accepting a half-billion dollars in losses, and eliminating 25 percent of its workforce.

No surprise, Zillow’s stock price has plummeted and the lawsuits have started….


Gary N. Smith

Senior Fellow, Walter Bradley Center for Natural and Artificial Intelligence
Gary N. Smith is the Fletcher Jones Professor of Economics at Pomona College. His research on financial markets statistical reasoning, and artificial intelligence, often involves stock market anomalies, statistical fallacies, and the misuse of data have been widely cited. He is the author of The AI Delusion (Oxford, 2018) and co-author (with Jay Cordes) of The Phantom Pattern (Oxford, 2020) and The 9 Pitfalls of Data Science (Oxford 2019). Pitfalls won the Association of American Publishers 2020 Prose Award for “Popular Science & Popular Mathematics”.

Zillow’s House-Flipping Misadventure