Zillow Makes It Less Important to See The House You’ve Just Bought

Credit: Pexels.com

Credit: Pexels.com

The Early Days of The Zestimate

“There were orange extension cords everywhere in the office”

Now valuing over 100 millions homes every day, when Zillow first began developing it’s Zestimate 12 years ago, it started somewhat smaller: just 43 million homes. Coupling this with the desire to create up to a 15 year history of a home’s value, this called for analysis of 7.7 billion data points.

Recently quoted in Popular ScienceZillow’s Chief Analytics Office, Stan Humphries, recalls the early days of generating that much computing power, claiming, “There were orange extension cords everywhere in the office.” After more than a decade of work, Zillow now runs millions of machine learning models daily that produces a Zestimate with what they claim is an average error rate of 4.3% (when comparing to the sold price).

What’s always been an issue with the Zestimate is the fact it uses “structured data” which simply highlight the common elements of homes: square feet, number of beds and baths, and location data. Now, Zillow is ramping up to leverage image recognition models that can look deeper into the interior of a home, something the company hopes will better control for what home buyers truly look for in the next home and better estimate home valuations .

Supercharged Zestimates

Simply put, Zillow’s new models look at images of the home to determine additional features not always captured by Zillow that likely include white kitchens, amount of natural light, and closet spaciousness, among many others to quantify the ‘quality’ of the house.

The Zestimate has always been at risk of being stale. Two similar homes, one gutted, and one refinished, would not be reflected in the Zestimate. For this reason, it’s important agents understand the Zestimate in order to educate clients. Timing will still be an issue with the Zestimate, but newly listed homes could feature more accurate Zestimates as could homeowners who actively update the pictures of their home. Zillow has not mentioned this, but I’m beginning to wonder if folks will treat their home’s “page” on Zillow like a Facebook profile, updating pictures as they remodel?

Zillow claims the additional parameters passed through the models increases the Zestimate by 15%. I’ll admit this is a pretty good start, despite the fact they’re likely using past sold data to confirm, and could potentially be fitting the model (for geeks like me, we know the true test comes in out-of-sample testing). However, I’ll get really excited when Zillow’s machine learning algorithms can identify MDF from real wood trim, age of appliances, or other more challenging aspects for image recognition. It would be nice for agents to simply be able to snap photos of a listing and classify everything required to pipe into the MLS listing.

What Does this Mean for Real Estate Agents?

Undoubtedly, artificial intelligence applications in the real estate industry are still in the early innings of development and use, but agents will surely benefit from more demand as they’re better able to cater to their client’s needs, and more efficiently.

Thus far Zillow has only introduced these models it’s King County backyard, but expects to broaden it’s application over time. Whether it’s Zillow or someone else, machine learning will continue to automate away the mundane tasks of agents, and eliminate those who get by simply doing the minimum. Accepting and adopting these technologies will help agents get back to doing more for their clients. As Humphries told Popular Science, “Teaching a computer to appreciate curb appeal is truly artificial intelligence.” Guiding your clients to make the best decision possible and wade through the computer-enabled data is not. That’s real.