- Forecasting
- Real Estate
- Zestimate
Zillow’s Forecasting Model: Zestimate
How COVID reshaped the King County housing market, and how Zillow’s Zestimate keeps pace.
The coronavirus pandemic raised significant questions about its potential impact on the housing market, particularly in King County, Washington, which includes the major cities of Seattle and Bellevue. Market indicators at the time showed reduced buyer interest amid stock market volatility, low inventory levels, and restrictions from stay-at-home orders. While some buying and selling activity continued, uncertainty remained high regarding the duration of lockdowns, the scale of the pandemic, and its broader effects on employment and consumer confidence.
Real estate plays a central role in the economy, influencing household net worth and mobility. Seattle, as an early housing market hotspot in the United States, provides valuable data for observing trends that may influence other regions.
In April 2020, Seattle experienced notable shifts: the average home sale price reached approximately $815,000, with a sharp decline in pending sales that did not close. However, committed buyers showed resilience, and a significant portion of sellers achieved prices above the asking amount. This dynamic suggested strong commitment from active participants despite broader market uncertainty, while limited inventory helped support price levels.
Zillow developed the Zestimate in 2006 as an algorithmic tool to estimate home values. The model initially covered around 43 million homes in its database. To maintain these valuations, it processed several terabytes of data monthly through approximately 34,000 statistical models. The median absolute percentage error at the time was about 14 percent.
Subsequent improvements significantly enhanced accuracy. The updated algorithm expanded coverage to 110 million homes, including smaller cities, and reduced the median absolute percentage error to roughly 5 percent. This progress resulted from access to vastly larger datasets and more sophisticated modeling techniques. Zillow increased model frequency from monthly runs of 34,000 models to approximately 11 million models processed nightly.
The Zestimate algorithm has undergone major updates in 2008, 2011, and 2018, with ongoing refinements. A key advancement involved greater geographic granularity. Rather than relying solely on county-level or state-level data, the model now prioritizes finer sub-county divisions. In larger counties with sufficient transaction volume, it identifies homogeneous groups of homes at the sub-county level to train more precise models.
For the remainder of 2020, Zillow and similar platforms projected a roughly 15 percent decline in home sales for the full year, with steeper drops expected in the second quarter due to pandemic-related disruptions. Home prices were forecast to show modest gains, supported by limited inventory, though this balance was expected to face pressure as economic effects lingered.
When developing or refining post-pandemic housing models, factors such as regional inventory levels, buyer commitment, employment impacts, and local COVID-19 case trends are important considerations.