They didn't eliminate their fees (fee is the wrong term to use). Their model was built, maybe this changed, on being within 200bps of breakeven. Obviously, they only bought when the model would say: this will make money. Or are you saying they looked at the model, the model says you will lose money, and they decided to do it...that makes no sense, even for SV.

Flip this around, are you saying that if the model was correct they wouldn't have made money? The problem was the model saying something was a good buy when it wasn't. The model was bad. Sellers do have good information, at least better than Zillow.

Generally, this is a misconception about how things like quant investing actually work (this was an attempt to apply quant investing to housing). Some people, usually people without actual market knowledge, view quant systems as providing greater information. In reality, most quant systems are just responding to changes in liquidity. The amount of actual fundamental information these systems provide is very minimal, and will always be beaten by a knowledgeable human. The reason why is simple: there is a huge amount of private, non-quantifiable information with these domains (and this is true in investing and property, doing this in resi housing is nonsensical).

I have seen fundamental quant investing work but only when you combine quantitative work with a knowledgeable human. I have seen the same thing in sports betting syndicates too (it does vary though, in some games quantitative data does capture more of the relevant information and machines can beat humans in those instances...but if there is substantial private, non-quantifiable information then it stops working).

This is hard for people to accept because lots of people spend lots of time and effort at university being taught that ML is effective. But ML is only as good as the information you put in. The demise of value factor investing is a perfect example: collect a ton of PHd quants and finance professors, they start doing fundamental investing but without doing any research themselves, and it has done nothing but haemorrhage cash. It takes an extraordinary amount of education to supress common sense here.

You have to understand the domain. You have to understand the information you are putting in. Zillow did neither, they thought ML would save them.

Look up their “project ketchup”. Their managers overrode the models and cut both fees and reno cost to win more deals. The WSJ and Business Insider wrote about this. I was at Zillow for many years and the insiders I know tell me the articles are correct but just lacking some nuance.

Many people leap to their own reasons why Zillow offers failed but the most proximate cause really does seem to be management and operational failure.

Saying that management bought at prices higher than model is not the same thing as saying they bought houses expecting to lose money. All that was said was that management increased the prices they would pay and changed the model so they could pay more. Nothing validates the model (again, this is a common-sense conclusion given the informational disparity that Zillow was at).

Right, they didn’t expect to lose money. They saw they were only closing 10% of deals and wanted to take a higher share from opendoor. They probably thought the market was going up fast and their models were too slow.