The algorithm that powers Zillow’s signature home valuation tool — the Zestimate — is getting an upgrade. Now more accurate than ever, the new Zestimate uses computer vision to analyze photos of a home to understand not just its facts and figures, but its quality and curb appeal. The Zestimate also now incorporates real-time data for homes listed for sale, including its listing price and how many days it’s been on the market.
For context, when Zillow first launched the Zestimate in 2006, it marked the first time that homeowners could find an instant, free estimate of their home’s value. It empowered people with information and fundamentally changed the dynamics in the real estate industry. We’ve worked over the last 13 years to improve it, from adding new features and more data sources to enabling homeowners to edit their home facts. Thanks to these improvements over time, I’m proud to say that the Zestimate now has a median error rate of less than 2 percent for homes listed for sale, meaning half of all Zestimates fall within 2 percent of the home’s eventual sales price.
From a technical standpoint, incorporating computer vision and advanced machine learning models into the Zestimate algorithm enables us, for first time, to give consumers a more quantitative accounting of the qualitative aspects of their home. We’ve always known that photos provide consumers a rich source of information about a home’s quality. Yet before recent advances in machine learning, there was no way for computers to look at photos of a home and get the same information that humans do. Now, we’ve taught the Zestimate to discern quality by training convolutional neural networks with millions of photos of homes on Zillow, and asking them to learn the visual cues that signal a home feature’s quality. For instance, if a kitchen has granite countertops, the Zestimate now knows — based on the granite countertop’s pixels in the home photo — that the home is likely going to sell for a little more.
This release also marks the first time we are directly evaluating a home’s list price, listing description, and how many days it’s been on the market, as part of the Zestimate’s calculations. As part of our data science team’s work to uncover new ideas for improving the Zestimate, we explored different ways for incorporating list price into the Zestimate. Overall, we saw how the list price and other real-time listing data helped our algorithm pick up on signals about a homeowner and agent’s listing strategy, and what the homeowner believes their home is worth. Evaluating these signals enabled us to better estimate what a home would sell for.
Most importantly, we now incorporate these signals into the Zestimates for homes on the market in real time — making it the most updated, accurate estimate of an on-market home’s value we’ve ever provided. Think about what that means for consumers to get an up-to-the-second snapshot of what’s going on in the local market and how that impacts their home. For comparison, when we first launched the Zestimate, listings were updated monthly as new information became available. That’s a big improvement!
Incorporating these new technologies and data sources not only adds another dimension of sophistication to the Zestimate’s infrastructure, it helps provide a more accurate home value estimate for millions of homes across the country, particularly those currently for sale.
As I reflect on the journey that culminated in today’s release, I’d be remiss not to mention that the new Zestimate is the product of a push that began in 2015 when we began leveraging cloud-based services for data storage and computation. We adopted a new data language and infrastructure and, in 2017, threw open the doors to new ideas through the Zillow Prize, a two-year data science competition that drew more than 3,800 entrants from 91 different countries. Innovative ideas from the winning Zillow Prize team are already being incorporated into the new Zestimate, including sophisticated machine learning techniques to identify and account for special home features, like the number of bathroom fixtures; location details like commute times; and a home’s proximity to a park or freeway.
Today’s release represents another milestone along that journey: an algorithm that leverages machine learning across our data-rich platform to provide consumers with the most accurate information about a home’s estimated value. It also embodies the incredible brain power, hard work, and collaboration among ourdata engineering, data science, AI and machine learning teams, who are constantly finding new ways to innovate.
While we’re excited about this release, it’s by no means the last stop. Our multiple data sources, accumulated over the past 13 years, and the data underlying the Zestimate provide Zillow with both a comprehensive representation of the U.S. housing stock and an incredible data advantage. As we continue to refine the Zestimate algorithm, we’ll work toward even more hyper-local and hyper-specific models capable of discerning how different housing markets value certain home features.
Our commitment to accuracy, combined with 13 years of experience valuing millions of homes across the country, is the bedrock of our business. It’s also a key factor as we ramp up our business of changing the way people buy and sell homes throughZillow Offers, our home buying service that allows sellers in eligible markets to skip showings, repairs and open houses by selling directly to Zillow.
I’m energized by our work to both evolve the Zestimate and fundamentally change the way people buy and sell homes. We’ll continue to generate, gather, and integrate innovative technology and ideas so that consumers have the confidence that our home valuations are the best place to start their home journey.
Originally published on the Zillow Group Tech Blog (medium.com/zg-tech-blog/) on June 27, 2019.
