Aside from the long list of unobservable attributes your ML algorithm can't learn, the biggest problem would be the small size of the dataset available, especially in the hottest markets (few sellers).
Many neighborhoods (not cities, or zip codes - too much variance in god that large) have on the order of single digits a dozen sales per year, meaning that in some cases the most recent 'comps' can be a month or two old. In a fast rising or dropping market, there's nowhere near enough data to train an algorithm before the prices have changed again.
Many neighborhoods (not cities, or zip codes - too much variance in god that large) have on the order of single digits a dozen sales per year, meaning that in some cases the most recent 'comps' can be a month or two old. In a fast rising or dropping market, there's nowhere near enough data to train an algorithm before the prices have changed again.