The house-pricing model was accurate on average. Average was not the problem.

Zillow Offers bought houses directly from owners, using the company's own valuation models to decide what to pay, then renovated and resold them. Volume scaled fast: Zillow sold 5,337 homes in 2020 and 15,436 in 2021 [2].

On 2 November 2021 Zillow's board decided to wind the business down. The company told the SEC the decision was made "in light of home pricing unpredictability, capacity constraints and other operational challenges", and that the wind-down would take several quarters and cut roughly 25% of Zillow Group's workforce [1].

The same filing disclosed a $304.4 million inventory write-down for the third quarter alone, the result of "purchasing homes during the third quarter at higher prices than Zillow Group's current estimates of future selling prices after selling costs", with a further $240–265 million of charges expected on homes it was still contractually obliged to buy [1].

The full-year figure was worse. In its FY2021 10-K, Zillow recorded write-downs "totaling $407.9 million" against inventory, describing the cause as "unintentionally purchasing homes at higher prices" than its own estimates of future selling prices [2].

The 10-K is unusually candid about the mechanism. Among the risks it lists: "Our pricing model may not account for submarket nuances — for example, the location of a home on a hill or in a building — which could have a significant impact on price" [2].

The valuation model behind this was not a failure as an estimate. As a free public number that helps millions of people orient themselves in a housing market, it was and remains genuinely useful — being roughly right about a lot of houses is exactly the job. What changed was not the model's accuracy but the consequence attached to it: the same estimate that is helpful as a suggestion becomes a purchase order when a company buys the house. Nothing about the number had to get worse for the outcome to [1][2].

An estimate that is unbiased on average can still be ruinous, because the errors are not symmetric in their consequences. When the model bid too low, Zillow simply did not buy the house and lost nothing. When it bid too high, Zillow bought it — every one of them. The business systematically acquired the houses its model was most wrong about, and the write-down is the sum of those errors. Nothing in the average accuracy of the model would have revealed this; only asking what happens on the wrong tail would have [1][2].