Zillow writes down the houses its model told it to buy

The bet inside every offer

Zillow Offers was a trader in houses. It ran an iBuying business, meaning it bought homes directly from sellers at a price the company itself set, made light repairs, and resold them.¹ It set that price with an automated valuation model, software that estimates what a house is worth from its features and recent nearby sales. The model did not only estimate today's value. It forecast what the house would sell for months later, after Zillow had bought and fixed it, and Zillow offered a number below that forecast so the resale would clear a margin.

Every offer was therefore a bet. Zillow quoted a price, bought the asset, and tried to sell it higher. A seller reads that offer as a price. It was a position: Zillow now owned a house it believed was worth more than it had paid, and the belief would be settled only when a buyer turned up. On November 2, 2021, Zillow said the bet had gone wrong and it was closing the business, cutting its workforce by about 25%.¹ The chief executive, Rich Barton, gave a one-sentence reason: "the unpredictability in forecasting home prices far exceeds what we anticipated."¹

The homes were worth less than Zillow paid

The model priced homes above what they would later sell for. Zillow said so directly: it had been "purchasing homes at higher prices than our current estimates of future selling prices."² In the third quarter of 2021 it wrote down the value of the homes it still owned by $304 million.¹ A write-down is an admission on the books that an asset is worth less than its recorded cost. Zillow ended that quarter holding 9,790 houses in inventory and another 8,172 under contract to buy, roughly eighteen thousand homes it was on the hook for.²

This was not a rounding error against the quarter. The Homes segment lost $421.6 million before taxes in the three months,¹ and the whole company, profitable in its other segments, still posted a $339 million net loss.³ The clearest picture of the model going wrong is in the per-home economics: the profit Zillow expected on a typical home swung about 1,200 basis points, twelve percentage points, from a gain in the second quarter to an expected loss of 500 to 700 basis points by the fourth.²

The losses kept coming as the book unwound. Zillow guided to a further $240 million to $265 million of losses in the fourth quarter, on homes it had already agreed to buy.¹ By the time the program was fully wound down, the Homes segment had lost $881 million before taxes for 2021, and the company overall lost about $528 million.⁴ Those are three different numbers, and coverage often blurs them. The $304 million was the third-quarter write-down. The $240-to-265 million was the next quarter's expected loss. The $881 million is the full-year segment loss, a broader figure that includes operating costs, not only the markdown.

The market repriced Zillow the same day. Its shares fell about 11.5% on November 2 and another 6.4% after hours, by press accounts.⁵ Zillow itself disclosed only the "approximately 25%" workforce cut; outside estimates of the headcount ranged from about 1,600 to 2,000 people, depending on which base the outlet used.⁵

The loss stayed hidden until the marking

A house on Zillow's books had two values. One was what Zillow paid for it, the number recorded at purchase. The other was what a buyer would actually pay, which nobody knew until the house sold. While a home sat in inventory it was carried at cost, so a house Zillow had overpaid for looked, on the balance sheet, exactly like a good buy. The loss existed the moment the model overpaid. It simply was not visible yet.

It became visible when Zillow marked the homes to what they would fetch, writing inventory down to the "lower of cost and net realizable value," the accounting rule that forces an asset held for sale down to its expected sale price.³ That is why $304 million landed at once rather than trickling in over the sales. And the position was large and slow to exit. Zillow's home inventory on the balance sheet had grown to $3.76 billion by the end of September 2021, up from $491 million a year earlier.³ A trader holding a mispriced stock can sell it in seconds. Zillow held billions of dollars of mispriced houses, each of which had to be listed, shown, and closed one at a time, over months. The size of the book set the size of the loss, and the illiquidity of houses set how long it would take to climb out.

The forecast could not keep pace

The model was fit to one stretch of the housing market and run into a different one. A forecast learned from how prices had moved through 2020 and early 2021 was used to price homes as the market's pace changed underneath it. That gap between the world a model learned and the world it runs in is distribution shift, the same failure a tumor detector shows when it is moved to a hospital it never trained on. Barton described it without the term: the unpredictability of prices "far exceeds what we anticipated."¹ The model's error was not a bug. It computed exactly what it was built to compute; the prices it had learned from stopped describing the prices it now faced.

Two things made the gap worse. The first is that the model was imperfect even in calm conditions. Zillow's own Zestimate, the public valuation behind its offers, carried a median error of 7.26% on off-market homes, the kind Zillow was buying, against 2.65% for homes already listed.⁶ On a typical American home, 7.26% is tens of thousands of dollars in either direction, before the market moves at all. Zillow's own annual filing called Zillow Offers "an unproven business model that relies on our ability to accurately value homes and manage inventory."⁶

The second is the choice to win volume. Zillow did not only misread the market; it raised its offers to buy more houses. The independent real-estate analyst Mike DelPrete found that Zillow's buy-to-list premium went negative, meaning it was paying more to acquire homes than it later listed them for, while its competitor Opendoor, buying through the same cooling market, kept that premium positive.⁷ Opendoor priced through the same forces without the same losses, which shows the failure was Zillow's pricing, not iBuying as a category. As DelPrete put it, Zillow "didn't see those brake lights" as the market slowed.⁷

There is a deeper reason a valuation model overpays as it scales. Economists writing up work at Stanford point to adverse selection. When Zillow bid at or above its model's estimate for ordinary homes, the sellers most willing to accept were the ones whose houses were worth less than the model thought, the lemons the model could not see. "There are a lot of features and attributes of a home that the model doesn't capture," the Stanford account notes, and Zillow "started bidding more for cookie-cutter homes than what their algorithmic model predicted," then moved into harder-to-value homes "hoping their algorithmic valuation model was accurate."⁸ The write-downs they tally come to about $569 million, consistent with Zillow's own $304 million plus the guided fourth quarter.⁸

The trade did not end with houses

The same trade runs anywhere a model sets a price and the world it is pricing keeps moving. Automated valuation and dynamic-pricing systems now set apartment rents, insurance premiums, advertising bids, and credit limits. Each is a forecast fit to the recent past and deployed into a present that can shift. When it shifts faster than the model expects, the prices go wrong the way Zillow's did.

One further risk is worth naming precisely, because the record around Zillow does not support the loose version of it. When a system is big enough to set prices, its own prices can begin to move the market it is trying to read. That is a feedback loop, and automated pricing at scale carries the risk in general. It was not the documented cause of Zillow's losses. The one academic study of a Zillow feedback loop examined the consumer-facing Zestimate rather than the iBuying business, found the effect short-lived and self-correcting, and drew no connection to Zillow Offers' losses.⁹ So the feedback loop belongs on the watch list for any price-setting model, kept separate from what actually sank Zillow: a forecast that could not keep pace, and a book too large and too slow to unwind once it was wrong.