Key takeaways
- Zillow wrote down about $304 million of homes in the third quarter of 2021 and announced the end of Zillow Offers on 2 November.
- Sellers accept a cash offer when it beats their own estimate, so an automated buyer mostly wins the houses it overpaid for.
- Bloomberg reported that staff overlays raised the algorithm’s offers by up to about 7% to hit volume goals, enlarging the losing slice.
- Opendoor sold 5,988 homes in the same quarter at a 10.3% adjusted gross margin, so the 2021 market alone cannot explain Zillow’s loss.
- For a home that is not listed, treat any Zestimate as roughly 7% either way, and be most wary when it looks generous.
Why did Zillow Offers fail only four months after Zillow announced that its home-pricing model had become more accurate? In June 2021 the company launched the Neural Zestimate, with a median error of 6.9% on homes not listed for sale. On November 2 it said it would close its home-buying business. It wrote down about $304 million of houses and cut about a quarter of its staff.
The answer matters if you have ever looked up your home’s Zestimate before setting an asking price or making an offer. By the end you will know how wide a range to put around it, and which direction of error should worry you more. An automated price is most dangerous in the cases where it looks most generous. Zillow learned that with its own money.
The usual retelling says the algorithm was bad at pricing houses. The filings point somewhere more useful. Zillow’s error on the average home was modest. The losses came from which homes it ended up owning, and from a decision inside the company to push its offers higher.
What Zillow Offers was built to do
Zillow Offers started in 2018. Zillow made a cash offer to a homeowner, with the price set by an algorithm built on the Zestimate. If the owner accepted, Zillow bought the house, renovated it and put it back on the market. Companies that work this way are called iBuyers, short for instant buyers.
The ambition was large. In 2019 Zillow told investors it aimed to buy about 5,000 homes a month within three to five years, as reported by Forbes. That is roughly 1% of US residential sales, and it would have meant around $20 billion a year in revenue for the segment.
The plan needed a model that could price a house well enough to buy it, pay for renovation and fees, and still make money. In June 2021 Zillow said the model had improved: its launch announcement gave a median error of 6.9% for off-market homes across more than 104 million properties. A median error is the middle value, so half of all homes are estimated closer than that and half are further off.
What happened in the second half of 2021
Through 2021 Zillow bought faster than it sold. The value of homes on its balance sheet rose from $491.3 million at the end of 2020 to $1,169.6 million by June 30, 2021. Three months later it stood at $3,758.2 million. Zillow’s third-quarter 10-Q filing gave the cause directly: home purchases were outpacing the sale of homes.
Source: Zillow Group 10-Q filings, Q2 and Q3 2021, SEC EDGAR
Sales rose from 1,965 homes in the first quarter of 2021 to 3,032 in the third, at average prices climbing from $356.7K to $386.8K. Purchases grew faster still, and the houses waited. Zillow’s third-quarter release put Homes revenue at $1.2 billion against an outlook midpoint of $1.45 billion, “due primarily to renovation and resale capacity constraints”.
As reported at the time, KeyBanc Capital Markets sampled 650 homes Zillow owned and found 66% listed below what Zillow had paid, by 4.5% on average. YipitData found that Zillow had cut prices on nearly half its September 2021 listings. In Phoenix, about 250 of its listings averaged 6% below the purchase price. Zillow reportedly tried to sell about 7,000 homes in bulk to institutional investors for about $2.8 billion.
The end came with the third-quarter results on November 2, 2021. Zillow wrote down its inventory because it had bought homes “at higher prices than the company’s current estimates of future selling prices”. It expected a further $240 million to $265 million of losses in the fourth quarter, mostly on homes it still planned to buy. The Homes segment lost $422 million before tax in the third quarter, and about 2,000 jobs went, roughly 25% of the workforce, as CNN reported.
Chief executive Rich Barton gave the official finding in the release: “the unpredictability in forecasting home prices far exceeds what we anticipated”. The statement leaves out who chose which homes to buy, and who chose the price.
Get the next one by email
Physics, engineering and the people behind them. No spam, unsubscribe any time.
Why Zillow Offers failed: intended, actual, mechanism
| Area | Intended | Actual | Mechanism | Source |
|---|---|---|---|---|
| Pricing accuracy | Median error 6.9% on off-market homes (June 2021) | 66% of 650 sampled Zillow listings priced below purchase price, 4.5% below on average | Sellers accept the offers that beat their own estimate, so accepted offers lean high | Zillow press release, 2021-06-15; KeyBanc, as reported Nov 2021 |
| Offer discipline | Algorithm sets the cash offer | Staff “overlays” raised offers, reported at up to about 7% | Higher offers win more sellers, and the extra sellers are the ones being overpaid | Bloomberg, 2021-11-08 |
| Volume | About 5,000 homes a month within 3 to 5 years | 3,032 homes sold in Q3 2021; inventory $3,758.2M on Sep 30, 2021 | Purchases outpaced sales, so more money waited on future prices | Forbes, 2019; Zillow Q3 2021 10-Q |
| Resale speed | Renovate and resell promptly | Q3 Homes revenue $1.2B vs $1.45B outlook midpoint | Capacity limits kept homes on the books for months while prices cooled | Zillow Q3 2021 release |
| Financial result | A profitable segment near $20B a year | $304M write-down; $422M Q3 segment pretax loss; $240M to $265M more expected in Q4 | Losses concentrated in homes bought at the highest prices | Zillow Q3 2021 release |
Why did Zillow Offers fail when Opendoor did not?
A seller takes a cash offer when it beats what they think the house is worth. They know things the model cannot see, like a cracked foundation or a noisy road. Suppose the model’s errors are balanced, too high on some houses and too low on others by similar amounts.
The owners of underpriced houses decline and sell elsewhere, so those errors never reach Zillow. The owners of overpriced houses say yes. A balanced error across the whole market becomes a one-sided error across the homes Zillow actually bought.
Economists call this adverse selection, which means the deals you get are chosen by the side that knows more. At auctions the same effect is called the winner’s curse: the bidder who wins is often the one who guessed highest. A buyer that wins most often when it overpays does not need a bad model to lose money. Any model with some error will do, as long as the other side decides which errors to accept.
Zillow then made the losing slice bigger. Bloomberg reported that an initiative called Project Ketchup added “overlays” to the algorithm’s offers. Staff could raise offers by up to about 7% to win more sellers and meet volume goals, and pricing experts were told to stop questioning the algorithm’s valuations. I think this was the heaviest single cause, because it pushed in the one direction the winner’s curse already favoured.
An iBuyer must predict the price three to six months ahead, long enough to buy, renovate, list and close. US price growth peaked around mid-2021 and then cooled, while capacity limits kept homes on Zillow’s books longer. By September 30 that bet was $3.76 billion.
Other iBuyers faced the same market in the same quarter. Opendoor sold 5,988 homes in the third quarter of 2021 on revenue of $2,266.4 million, with an adjusted gross margin of 10.3%, according to its quarterly supplement. Offerpad sold 1,673 homes on revenue of $540.3 million and reported gross profit of $53.1 million, about 9.8%, as Inman reported. Opendoor’s figure is adjusted, so compare the two loosely.
Source: Opendoor Q3 2021 8-K supplement (SEC EDGAR); Offerpad via Inman, 2021-11-10; Zillow Q3 2021 release (SEC EDGAR)
The filings do not say why Opendoor held up, so this is my reasoning. Its offers carried fees and a spread that left room for error, its homes turned over faster, and it pushed volume less hard. Zillow spent its margin to buy volume. Our Challenger O-ring temperature autopsy follows the same pattern, with a safety margin people trusted until it was used up.
What the winner’s curse means for your Zestimate
Zillow publishes the median error of the Zestimate. For homes on the market it is about 1.9%, and for homes not listed it is about 6.9% to 7%. The on-market figure is small partly because the model can see the list price. For a house you have not yet listed, the off-market figure is the one that applies.
On a $400,000 home, 7% is about $28,000 in either direction. That is the median case, so half of homes are further off than that.
Treat any AI valuation of an unlisted home as a range of roughly 7% either side. Before you anchor an asking price, get recent comparable sales or an appraisal, because a local agent or appraiser sees the house and the model does not. One published figure can hide a wide spread, as the two different numbers for AI energy per query showed.
Be most careful when the algorithm’s number is higher than you expected. If you are selling and a cash buyer’s offer comfortably beats the Zestimate, ask what that buyer knows about your market. If you are buying and plan to bid up to the Zestimate, remember that your bid wins most often where the estimate runs high. Your position is the same as Zillow’s was, only without a balance sheet to absorb the loss.
Zillow’s model was reasonably accurate on the average home. Zillow never bought the average home. It bought the homes whose owners said yes, and an owner says yes most readily when the offer is too high.
One story like this, most days
Written by a CERN physicist. No spam, unsubscribe any time.
