Zillow's algorithm bet on home prices — and lost $420M in a quarter

Zillow, best known for its Zestimate home valuation tool, launched Zillow Offers in April 2018 as an iBuying business. The model was simple in concept: Zillow would use its algorithmic Zestimate to make instant offers to home sellers, buy the homes, make minor repairs, and resell them at a profit. The margin was supposed to come from the efficiency of the algorithm and the scale of the operation.

The problem emerged in 2021 when the US housing market became unusually volatile. Zillow's algorithm, which had been trained on historical data, could not keep up with rapid price changes. The company continued buying homes at prices the algorithm deemed fair, but the market was shifting faster than the model could adjust. By the third quarter of 2021, Zillow owned approximately 7,000 homes and was losing money on nearly every one. The Zillow Offers segment lost $420 million in Q3 2021 alone.

In November 2021, CEO Rich Barton announced that Zillow would shut down the iBuying business entirely. The company took a $304 million write-down on the unsold inventory and laid off 25% of its workforce, roughly 2,000 employees. Zillow's stock fell sharply in the following weeks. Barton later acknowledged that the company could not accurately forecast home prices. The failure was a stark demonstration that an algorithm trained on stable markets is not reliable in volatile conditions.

Shut down iBuying after the model failed in volatility, taking a $304M writedown and cutting 2,000 jobs to stop the $420M quarterly bleed.

Zillow's algorithm was trained on historical housing data and could not adapt to the rapid price swings of 2021, causing it to buy homes at prices that were no longer profitable.

The iBuying model required Zillow to hold inventory, which meant every pricing error was magnified by the cost of carrying thousands of unsold homes.

Zillow scaled the business aggressively before proving the algorithm worked in volatile conditions, accumulating 7,000 homes by the time the flaw was undeniable.

An algorithm is only as good as the market it was trained on. When the market regime changes, the model that worked yesterday becomes a liability today. Scale before proof magnifies the damage.