From last week's Seventh Circuit decision in Patel v. Zillow—an interesting case that I was fortunate enough to consult on (on Zillow's side)—the facts:
A Zestimate is an estimated value for real estate, available on the Zillow web site for about 100 million parcels. Zillow generates Zestimates by applying a proprietary algorithm to public data, such as a building's location, tax assessment, number of rooms, and the recent selling prices for nearby parcels. But because Zillow does not inspect the building, it cannot adjust for the fact that any given parcel may be more attractive and better maintained, raising its likely selling price, or the reverse. Zillow states that its median error (comparing a Zestimate with a later transaction price) is less than 6%, though the Zestimate is off by more than 20% in about 15% of all sales. Zillow informs users that none of the parcels has been inspected and that Zestimates may be inaccurate, though Zillow touts them as useful starting points.
Plaintiffs filed this suit after learning that the Zestimates for their parcels were below the amounts they hoped to realize. For example, Vipul Patel listed his home with an asking price of $1.495 million and contends that the Zestimate of $1,333,350 scared away potential buyers. Plaintiffs asked Zillow either to increase the Zestimates for their parcels or remove them from the database. When it declined to take either step, they filed this suit, under the diversity jurisdiction, invoking [among other things] the Illinois Uniform Deceptive Trade Practices Act, which forbids unfair or misleading trade practices....
The court rejected the deceptive trade practices claim, in an analysis related to the Illinois statute, but one that is parallel to the approach used in libel cases under the common law and under the First Amendment:
[T]he statute deals with statements of fact, while Zestimates are opinions, which canonically are not actionable. Plaintiffs want us to brush this rule aside because, they say, Zillow refuses to alter or remove Zestimates on request. This does not make a Zestimate less an opinion, however....
That Zillow sells ads to real estate brokers [also] does not affect the statutory analysis. Having labeled Zestimates as estimates (something built into the word "Zestimate"), Zillow is outside the scope of the trade practices act. Almost all web sites, like almost all newspapers and magazines, try to finance their operations by selling ads. That they do so without telling customers exactly what pitches are being made to potential advertisers does not convert a declared estimate into an inaccurate statement of fact....
And the court also noted (perhaps because the author of the opinion, Judge Frank Easterbrook, is a noted law and economics scholar):
[P]laintiffs are mistaken to think that the accuracy of an algorithmic appraisal system can be improved by changing or removing particular estimates.
Suppose plaintiffs are right to think that the Zestimates for their properties are too low. Removing them from the database would skew the distribution, because all mistakes that favored property owners would remain, not offset by errors in the other direction. Potential buyers would be made worse off.
Suppose instead that plaintiffs are wrong—that they have overestimated the value of their properties, while the Zestimates are closer to the truth. Then removing them from the database would not just skew the distribution but also increase the average error of estimates. Potential buyers of plaintiffs' properties would be deprived of valuable knowledge.
Finally, suppose that plaintiffs are behaving strategically—that they know the Zestimates to be accurate (or at least closer to the likely sales price than are plaintiffs' asking prices). Then removing their parcels from the database, or "correcting" the Zestimates to match plaintiffs' asking prices, would degrade the accuracy of the database as a whole without any offsetting benefits to the real-estate market. In general, the accuracy of algorithmic estimates cannot be improved by plucking some numbers out of the distribution or "improving" others in ways that depart from the algorithm's output. The process is more accurate, overall, when errors are not biased to favor sellers or buyers.