patrick.net

Choosing homes there is the job of Amherst’s highly automated purchasing system. In its 19th-floor office on New York City’s Madison Avenue, a dozen buying specialists screen leads on their workstations, delivered by a proprietary program called Explorer, an offshoot of the software Dobson developed to price mortgages. Each morning, the team gets alerts on newly listed homes that meet its price range and geographic criteria—around 1,400 listings a day.

For each “first cut” listing, Explorer estimates the costs of renovation. This is machine learning at work: The estimate is based on Amherst’s experience with homes of similar age and size in the same or nearby neighborhoods. In an older home, this might include replacing the HVAC system; for one whose listing photos suggest wear and tear, it might include a new roof. (Team members help the software make that call.) Explorer has become so precise, Negri says, that the actual renovation costs average within 5% of the estimates.

Explorer also runs a separate calculation, finding three homes being rented within a two-mile radius that are close in age, size, and bed-and-bath specs to the newly listed home. Machine learning helps the software estimate what each house would rent for based on these “comps.” Explorer then churns out an estimated “rental yield”—the net rent after such expenses as taxes and maintenance, divided by all-in cost.

If that yield meets Amherst’s target (which Fortune estimates is between 5% and 6%), the team will make an offer. About 20% of each day’s listings qualify; Amherst bids on those candidates no more than 12 hours after they’re first listed, making all-cash offers. Around 10% of its offers—on roughly 30 homes a day—get accepted and go to contract.