Steve, based on your above comment I can't tell if you are in favor of ML or not. There is no doubt real estate pricing is quantifiable. Just comes down to having the right data and the right inputs.
There is little doubt in my mind a sophisticated model like a neural network with the right inputs would crush Zillow's price estimate.
The key is engineering the right features from the data so that you are capturing / quantifying meaningful signal. For example, if I were designing such a model and had all the data available to me, I would include distance to a highway as a feature, school quality, or even creative variables like the count of liquor stores in close proximity, income of zip code, average age of neighbors (to know whether they're adults and poor or maybe just students)... These are all important indicators of neighborhood character or quality that probability get factored into the final sale price (consciously or not).
The thing about ML is that it's just mathematical modeling that uses a fairly simple optimization technique. It's only as good as the inputs you provide it, so the trick to winning with ML is to think outside the box to produce quality data signals and to have lots of data records (the more the better).
