It is not often that you hear of AI algorithms failing in the market, but that is essentially what happened in the case of real estate company Zillow and its house-buying estimator algorithm, Zestimate. The headline writers and tweeters were quick to feed off the wounded animal, asking whether this was the beginning of the end for AI (typical tweet: “A significant strategic blunder! Everything in our world cannot be managed with algorithms parsing big data! They failed to think through the implications of buying all those homes. A perfect HBS case study on bonehead! @zillow heads should roll!”).
But, like the informed readers we are, we need to look behind the headlines to see exactly what happened with Zillow. The algorithm was part of a business that bought and sold (flipped) houses in the US using price estimates provided by the AI model. It looked for properties that were under-valued, bought them, did them up, and sold them off at higher price. At least that’s what was meant to happen. What did it for the algorithm was the massive uncertainty that entered the market with the pandemic.
Back in March 2020, I wrote about how the huge skewing of trend data that the pandemic brought will play havoc with predictive algorithms, and this is a prime example of that. The predictions simply become less accurate. But there is also another line to this story, because blaming the algorithm (which, remember, is just some clever maths) is the easy thing to do. What is not often considered is the role the humans played in all of this (which is, of course, everything). Another tweet that asks the right sort of questions was from a property entrepreneur: “It would be fascinating to find out where in the stack Zillow’s failure lives. Was it incorrect use of ML? Too much trust in ML? Aggressive management that wouldn’t take ‘we aren’t ready’ for an answer? Wrong KPIs?”. Notice how none of these questions are blaming the algorithm?
A piece from Wired manages to take a deeper dive and concludes that it was probably parts of all of these things. For example, the heart of the buying activity was in Phoenix, where there are plenty of ‘cookie cutter’ homes that are easier to price. Once you try and take that model to New York then many of the assumptions and training data are no longer relevant – the variability in the types of homes makes predicting much, much harder, before you then introduce pandemic impacts. It also seems that Zillow expected the algorithm to do everything, including estimating the amount of capital required for the business, when all it did was estimate prices of individual houses.
So there should probably be a Harvard Business School case study made out of this, but the lessons are clearly about how to use the algorithm in the right way – understand what it can and cannot do, model extreme events to know its limitations, don’t try to extend it to areas where it will not perform, and, most importantly, never trust it too much.