What happened
It was reported that in November 2021, Zillow announced it was shutting down Zillow Offers, its business of buying houses, renovating them and selling them for a profit.
The company had relied heavily on algorithms to estimate property values and guide purchasing decisions. As US house prices rose, Zillow bought homes aggressively, expecting to sell them later at higher prices.
But the market began to change. Price growth slowed, renovations took longer and properties remained unsold. Homes purchased at optimistic prices were no longer worth what Zillow had expected.
On 2 November 2021, Zillow announced an inventory write-down of about 304 million USD for the third quarter, with further losses expected. The total anticipated write-downs exceeded 500 million USD.
The company abandoned Zillow Offers and announced plans to cut approximately 2,000 jobs, around a quarter of its workforce.
Chief executive Rich Barton acknowledged that forecasting house prices had proved far more unpredictable than the company anticipated.
What an auditable version would have shown
The important question was not simply whether Zillow’s algorithm could estimate a property’s value. It was whether those estimates remained reliable as the housing market changed.
An auditable system would have recorded the expected resale price, renovation costs and time needed to sell each property when the purchase was approved.
Those predictions could then have been compared with actual sales and holding costs.
A MetricRecord could track whether the gap between forecasts and reality was growing. A ConductRecord could preserve the assumptions behind purchasing decisions and identify when those assumptions needed reviewing.
Zillow had extensive data and operational reporting. The public record does not establish that it lacked these particular checks, only that its forecasting and business assumptions ultimately proved inadequate.
Where the gap was
Zillow’s mistake was not simply trusting an algorithm. It was expanding a business that depended on predictions remaining accurate in a changing market.
Headlights identifies overconfidence in the decision to scale a model-dependent business while exposing it to increasingly uncertain market conditions.
There is a familiar human pattern here. When something works repeatedly, people begin to believe it will continue working. Success becomes evidence of reliability, even when the conditions responsible for that success are changing.
Zillow’s growing inventory made the consequences more severe. Each additional purchase increased its exposure if the forecasts proved wrong.
The algorithm was part of the problem, but so were the business decisions surrounding it. The company chose how aggressively to buy, how much risk to accept and when to stop.
This behavioural interpretation is Headlights’ analysis, not a formal finding about the motivations of Zillow’s executives.
What governance should have looked like
Zillow needed clear limits on how much risk the business could take when its forecasts became less reliable.
That meant regularly comparing predicted resale prices and holding periods with actual outcomes, across different cities and market conditions.
A MetricRecord could highlight growing forecasting errors. A ConductRecord could document decisions to increase purchasing volumes despite changing market conditions.
Pre-agreed thresholds could trigger slower purchasing, smaller offers or a review by senior management when forecasts consistently missed their targets.
The point is not that an algorithm should predict every market movement. No forecasting system can.
The point is that a business must recognise when its predictions are becoming unreliable and respond before the losses become too large.
Failure Pattern: Zillow expanded its algorithm-supported home-buying business during a rising property market. When conditions changed, its forecasts proved inadequate, contributing to major write-downs and the closure of Zillow Offers.
Governance Principle: organisations using predictive models for financial decisions must continually compare forecasts with actual outcomes and reduce exposure when performance deteriorates.
The reference implementation of MetricRecord and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Zillow’s Q3 2021 Shareholder Letter (Q3 earnings release, 2 November 2021)
- What Went Wrong with Zillow? A Real-Estate Algorithm Derailed Its Big Bet (Wall Street Journal)
- Zillow, facing big losses, quits home-flipping business (New York Times)
- The Algorithm That Could Have Saved Zillow (Bloomberg)
- Zillow Group Securities Litigation, Case No. 2:21-cv-01551 (W.D. Wash.)