House price volatility is localized to the point that sample sizes are too small. Prices are generally neighborhood-based and the sales in a neighborhood are too infrequent. One quick divorce sale, or one over-eager buyer can skew the data enough to foil this approach.

As an anecdotal example, my development has 115 homes. In the last 12 months, 3 have changed hands. The first sold at 5% over the Zestimate, the second, 10% over, the third 15% over. How would the Zillow algorithm predict the price of the next sale?