As the project lead for the latest point release of the Zestimate algorithm, I wanted to take a second to write up a bit about what we here at Zillow Data Science are excited about in the latest update to our favorite estimate. Along with this release, we have also recalculated the historical Zestimates for all homes going back to January 2008 with the new method to ensure consistency with the latest numbers.
Making the best estimate of home value even better
To start off, compared with the previous version of the Zestimate our national accuracy for Q3 2016 improves from a median absolute error (MAPE) of 5.4% to 4.5%. In other words, now over half of all homes sold in the country sell for a price less than five percent from our estimate of the home’s value! We also decreased the number of “big misses” the Zestimate makes by decreasing the percentage of homes selling for more than 20% above or below their Zestimate from 13.6% to 10.3%. While there are many tweaks, bug-fixes, and improvements in this latest update to the Zestimate, there are two particular areas that lead to the significant improvements in accuracy.
More raw data means better estimates
First, we added a new public records data provider to the mix of data we ingest. Adding this new provider helped in two ways: it increased our overall number of homes for which we have facts like the number of bedrooms and finished living area. Additionally, for areas where we already had records coverage, it means that we are able to get updates to these important facts faster. This work impacted accuracy everywhere but by far the most noticeable impact will be for folks in West Virginia. Statewide, MAPE dropped from 8.9% over the last three months with the old algorithm to 4.8% under the new one thanks to major increases in public records coverage statewide.
New Construction gets more attention
We started 2016 with our accuracy on new construction listings as one of the most glaring weaknesses for the Zestimate. After much work in April, we shipped a specific new construction model in Z5.3 and reduced our MAPE from 18% to 12% nationally for those listings. Nice but still not awesome. With today’s release though, we’re projecting our MAPE on new construction listings to drop to about 8% nationally. To put these accuracy numbers in context, this is about as accurate as human appraisers are for newly built homes.