Here come the hackers: Airbnb’s pricing algorithm
Nov 9th, 2015 | By FCT
Ladies and gentlemen, they’re in the financial services room, hackers are—and the real estate marketplace is about to undergo a sea-change.
Here’s but one example: the online short-term rental giant Airbnb had a big, big customer service problem: its clients were all too often stumped when assigning a price for their accommodation rental listing.
When Airbnb’s hacker/engineers began to design a solution, the Price Tips feature, to their clients’ price-matching issues, they realized they’d tackled something far more complex than they thought. Clients either “guesstimated” a fee, ignoring real market values—or just surrendered and left the site. (eBay had similar problems like this early on, too.)
What the engineers were designing was an algorithm, a complex set of rules governing calculations in computer code, to automate what human beings can’t do: model an entire pricing system, in real time.
The process when the engineers began to design in dynamic pricing—pricing models sensitive to daily market conditions—they realized that something very interesting applied: the more reviews in a given listing, the more valuable the property was on that day, the opposite of Amazon or eBay review behaviours.
Other variables incorporated include seasonal demand (FIFA World Cup, say) or the one-time demand spike (a Lady Gaga concert), but, like mortgage properties, the Airbnb model accounts for similarity (properties roughly equivalent to yours), recency (short-term market changes) and, of course, location, location, location. The complex model had to reflect that if a client was the first to list their property in Nairobi, that listing isn’t going to be vetted by a potential renter the same way a similar property in, say, Prague.
But the pricing considerations were even more subject to distortion than the engineers realized at first. Prices in London south of the River Thames are significantly more than north of the river, for a given summer weekend. (Why? Far more greenspace and smaller, cleaner townscapes than in inner London.)
Call in the cartographers: the cartographers modelled highly accurate, hyperlocal (street-by-street) maps of Airbnb market cities, down to spatial, geographic and structural features like bus lines, canals and rivers, and surface roads, because access influences price. Everything in the immediate environment that affected value was mapped, weighed and then factored in or not.
And all these variables had to be open to calculations based on seasonal demand, not static, the same way online airline and hotel prices have been set by near-real-time data for decades. Last—and perhaps most importantly of all—the whole beast had to appeal to human beings as a user experience.
Once the pricing algorithms (layers of them, actually) underwent testing so that the system could begin to loop back on itself and “learn” from the eventual fate of a listing and then adjust the calculations appropriately, using a Big Data Platform. The entire system “learns” from all past behaviours it’s computed.
End result? Airbnb says that when a host selects a price that’s within 5% of the company’s suggestion, they’re nearly four times more likely to successfully book their property. If that’s not a clear model powerful enough to transform the residential real estate market, Airbnb has released the machine-learning platform on which it’s based, as an open-source tool. The open-source tool will give people in industries that have yet to embrace machine—like real estate—an easy entry point.
For the tech minded, the IEEE (that’s the pointy-headed engineers organization) has a great piece on Airbnb’s pricing algorithm development process.