I recall in one of the previous crashes it turned out a firm had a model that couldn't handle homes going down in value. A linear model is superior to that.

This is basically how human-lead appraisals work in most markets.

There was a Kaggle Competition in case you missed it. At least here you can see what kind of features they are using and what models people employed. See https://www.kaggle.com/c/zillow-prize-1

Zillow Offers team was isolated from the Zestimate team. They didn't share algorithms.

AFAIK this is false; Offers began transacting at the Zestimate price in several markets earlier this year

Could be stale information. I talked to them a while back. Still, transacting at the Zestimate price doesn't mean they're privy to the Zestimate algorithms.

What could possibly go wrong with that approach? :-)

In fairness, I would think humans include nearby home prices as one of their parameters as well.

Human/realtor appraisals are often very poor. They will take 3-4 nearby & similar homes and basically add/subtract the differences from the home they are comparing. Then they basically just average the adjusted sale prices. They might add a bit onto that price since prices go up over time.

So this has some obvious issues: * Areas without a ton of very similar houses that have also sold recently will basically have no 'comps' to use. * It's really easy to keep identifying differences (pros/cons) until the adjusted prices equal each other but it's hard to know if all of the meaningful differences have actually been identified. * It gives average homes and average buyers a huge advantage. This model assumes that the housing market is hot but not hot enough that someone will pay 10-15% more for some specific feature. Anything unique to a home/property is only going to be worth a fraction of the time & money it would cost to add. Anything super common (kitchen remodel, finished basement, etc.) can actually add 100% or more of it's cost to the houses value because the buyer is paying with 5x or more leverage so paying a bit extra to have it included. This is also why it's often better to fix a few things as the seller than give a discount on the sale price - the buyer will often pay a premium to get things move-in ready since it doesn't impact their monthly costs significantly.

A Broker Price Opinion (BPO) is typically set by appraisal, comparable sales and comparable listings. Depending on how cookie cutter the neighborhood is, there are many different levels of "appraisal" from drive by photos to a deep inspection.

lol ... this infact is the sample problem which Andrew Ng explains in his famous Machine Learning course on Coursera!

Wait, can you elaborate? I’m sincerely interested in going over that material.

should move to xgboost.

Maybe some reinforcement learning algorithm running as well, where the maximum reward is winning the bid. Hence the algorithm just goes and lays waste on other bidders. Gotta get that infinite reward!

I doubt RL is involved. That would be soooo risky.

Bidding algorithms can model the price of winning the auction, you can maximize profit without fancy RL