“This Is Code Red”: Here’s How Zillow’s Home-Flipping AI Triggered Judgement Day For Real Estate Firm
Nearly three weeks ago, real estate firm Zillow Group Inc. announced during earnings that it was halting it’s AI-powered house-flipping program, firing 25% of its staff, and would be writing down losses of more than a half-billion dollars on the value of its remaining housing portfolio.
The resulting carnage brought the company’s market cap from a peak of $48.35 billion in February to around $16 billion.
The program, Zillow Offers, was originally supposed to make money primarily from transaction fees and services such as title insurance – not from flipping homes for huge profits. And according to a new analysis from The Wall Street Journal, the company’s algorithm just “didn’t seem to understand the market.”
The first quarter delivered home-sale profits that were more than twice as high as anticipated…
By the summer, it had the opposite problem, the company later acknowledged. It was paying too much money for homes, and buying too many of them, just when price increases were starting to slow. -WSJ
Zillow’s algorithm was designed to predict what a home would be worth in a few months, and offer a cash sum to the seller that would allow the company to make minor upgrades and repairs before flipping it for a tidy sum. Home sellers benefited from an instant cash offer which cut out real-estate agents, along with listing and showings during the Covid-19 pandemic. Zillow’s original plan was to make no more than 2% profit so homeowners wouldn’t feel cheated.
The entire model relied on the algo’s ability to predict future home prices based on its already-successful ‘Zestimate’ – which the company said in recent years was accurate within a median error rate of 2 percentage points.
So what went wrong?
According to WSJ, Zillow went beyond the limits of their technology in an industry where buyers are ultimately following personal tastes, emotional attachments and other intangible factors.
Compounding their missteps, Zillow – while understaffed – scrambled to play catch-up with competitors and disregarded internal concerns that the company was overpaying for homes according to current and former employees. The company also suffered from supply-chain issues and labor shortages that hindered its ability to quickly renovate and flip homes – in what is described as the “breaking point for a business that executives once predicted would generate $20 billion in annual revenue.”
In order to try and account for aesthetics, Zillow incorporated photographic analysis into its home pricing metrics in 2019, which included factors such as natural light, the quality of interior finishes, and ‘curb appeal.’ It also hired over 100 pricing analysts to double-check the algo’s figures by looking at recent comps in the area, which was aimed at reducing the risk of overpaying – but made it harder to quickly flip a growing portfolio of homes in major metropolitan areas.
Then, an unexpected surge in home prices and real estate activity during the pandemic caused their algo to go haywire, as buyer preferences began to change in unusual ways.
“That shift in buyer preferences is extremely hard for a machine-learning model to incorporate,” said BiggerPockets VP of data and analytics, Dave Meyer.
In the spring, Zillow execs and managers huddled for a tense meeting, according to a person who attended. While the company was making more money than anticipated on flips, they were on track to miss their annual target for the number of homes purchased – falling behind top competitor, Opendoor.
“This is code red,” said Joshua Swift, senior vice president of Zillow Offers, during the virtual meeting.
Following the meeting, Zillow hatched “Project Ketchup,” designed to speed up the volume of home purchases by purchasing more homes on an accelerated basis – and making offers which were well above what the company’s own algorithm predicted as a fair market value, according to people familiar with whjat happened.
“Our observed error rate has been far more volatile than we ever expected possible,” said CEO Rich Barton. “And makes us look far more like a leveraged housing trader than the market maker we set out to be.”
Sam Chandan, dean of New York University’s Schack Institute of Real Estate, said the complexity of the housing market makes it difficult to predict home prices months in advance. Some factors that have an impact on a home’s value are hard to capture with algorithms.
“The system may capture that there are three bedrooms, but does it capture that they are laid out in a way that makes sense?” he said. -WSJ
In Phoenix, home to one of the company’s largest portfolios, the median price Zillow paid for homes rose from $351,000 in May to $475,000 in September – at a time when competitors were easing off the throttle on their own purchases, according to Mike DelPrete, a real-estate tech strategist at the University of Colorado Boulder.
Fast forward two months later, and Barton now says the company expects to sell its homes at a 5-7% loss, down from an average profit of nearly 6% in the second quarter. The company halted its future home purchases for the remainder of the year in mid-October, and notably shut down Zillow offers for good earlier this month.
“We determined that further scaling up Zillow Offers is too risky, too volatile to our earnings and operations, too low of a return on equity opportunity and too narrow in its ability to serve our customers,” he said.
Did Zillow Offers’ AI become sentient and decide to nuke its creator? One can never know these days.
Thu, 11/18/2021 – 17:20