Automated valuation models (AVMs) are computer generated algorithms that analyze various data points to produce an estimate on the current value of a home or property. Typically, online visitors type in a property address and the AVM uses regression models to form an estimate of that property’s market value. Data analyzed can include the age of a home, market values, trends, historical data, property features and more. An AVM is designed to provide instant valuation for a property using various data sources and allows I-buyers such as Opendoor, OfferPad, and Zillow Offers to make instant offers to purchase homes.
The most popular and widely known AVM is Zillow’s signature valuation tool, which they call the Zestimate. People love Zestimates! Last month, Zillow announced the Zestimate was being upgraded to increase the model’s accuracy.
We want to know if AVMs really work.
Well, Zillow, the leader in the space since they first launched the Zestimate in 2006 employs over 100 data scientists, so you would kind of hope it does. However, the numbers tell a different story. According to data published by Zillow itself, before the recent algorithm upgrade, only 60.5% of Zestimates pegged property sales prices to within a 5% margin of error, 76.3% within a 10% margin of error, and 86.6% within a 20% error. That level of error makes the Zestimate pretty useless to actual buyers and sellers, and probably the reason for the upgrade.
So does it work now?
Umm. Kind of. The accuracy we mentioned above has improved significantly, but only for properties that are being marketed. Why? Because the major change that Zillow made to the algorithm was including the seller’s listing price as a variable in their calculation.
They needed 100 data scientists to uncover that there’s a correlation between list price and sales price?
That pretty much defeats the purpose of an AVM, relying on list price as an indicator of value.
We love the idea of automated valuation. At Compound, we are working on our own models to build an accurate AVM for urban markets with vertical housing stock (think high-rise apartment buildings.) Notwithstanding the hype surrounding terms like AI, machine learning, and neural networks, let’s be brutally honest with what actually exists: the automated valuation model is a work in progress that may take some time to come to fruition.