> is that they failed to put in adequate margin for adverse selection. Adverse selection turns prediction variance, which you might naively expect to "average o

> is that they failed to put in adequate margin for adverse selection. Adverse selection turns prediction variance, which you might naively expect to "average out" when using a well-calibrated model at scale, into a bias towards underperformance. Presumably, they trusted that their model does a more accurate assessment of price and future potential than a person themselves. If you had an excellent model, why _not_ would it balance out? My theory is that most people would consult various sources and professionals-who-assess-homes (whatever they are called) and this the likelihood of Zillow having a better assessment than the average person is just low.