External Control Layer

A home-buying algorithm buys more homes than the business can carry

Zillow’s home-buying business used algorithmic pricing to make offers at scale, bought more homes than it planned at prices it could not recover, and shut down.

When
Third quarter of 2021; wind-down announced on 2 November 2021
Who
Zillow Group (Zillow Offers)
System
Algorithmic home valuation and offer pricing
Action
Zillow Offers made thousands of purchase offers, each within the pricing model’s logic, that together built an exposure the business could not carry.
Outcome
Harm. About $304 million written down in the third quarter, $240 million to $265 million more in expected losses, about a quarter of the workforce cut, and the business closed
In the record
Entry 9, marked G5 Blind to the total

What happened

Zillow Offers made cash offers on homes using algorithmic pricing and resold the homes it bought. In the third quarter of 2021 it bought 9,680 homes, more than planned, at prices above what it later expected to recover. It attributed the extra purchases to higher-than-anticipated conversion rates. Zillow wrote down about $304 million of inventory and expected further losses of $240 million to $265 million. It announced that it would wind the business down and cut about 25% of its workforce. Its chief executive, Rich Barton, said, “We’ve determined the unpredictability in forecasting home prices far exceeds what we anticipated.”

Zillow’s own account attributes the result to the difficulty of forecasting prices and to buying more homes than planned. How fast to buy was a business decision as well as a model output. The AI Incident Database lists the episode as Incident 149.

What allowed the action to take effect

Whatever limits existed did not slow the buying before losses reached this scale. Each offer could look reasonable on its own while the portfolio drifted, and a check that looks at one action at a time does not detect that kind of failure.

An illustrative control

Assumptions. Offers go out only through the layer. Realized resale prices and the purchase plan are available to it as evidence.

Rule
rule:       exposure-and-drift
applies to: offer.submit
hold when:  purchases this week exceed [110%] of plan in the market
            inventory exceeds the market's exposure cap
            resale prices over [30] days run more than [5%] below model estimates
authority:  investment committee
action
offer.submit, single-family home, $412,000
policy
acquisitions v9, rule exposure-and-drift
evidence
30-day resale prices 6.8% below model; weekly volume 118% of plan
decision
Hold
authority
investment committee
A decision record the rule above would produce. It is not from the incident.

Without predicting the market, the layer detects that outcomes have stopped matching the model and that volume has passed the plan, and it holds the next offer for people who can decide whether to continue.

Limits of this control

Forecasting error is a business risk that no control layer removes. The layer enforces the limits that leadership sets but does not choose them. If a company decides to buy more aggressively, the layer records the new limits and enforces them.

Controls that apply

Sources

Facts come from the sources listed. To correct a fact, see how to send a correction.