The Future of Real Estate Machine Learning: Risks and Opportunities
Risk professionals who analyze real estate portfolios must overcome complex modeling and liquidity risk challenges. How are disruptive technologies changing the real estate modeling game?
Wednesday, November 24, 2021
By Tony Hughes
Financial risk managers must assess a broad spectrum of threats, including the liquidity risks connected to substantial real estate investments. For many reasons, it’s hard to build valuation models that are precise enough to surmount such extreme liquidity constraints, but companies are increasingly using machine learning (ML) to try to tilt the playing field.
The results, to date, have been mixed. Just a few days ago, for example, Zillow announced that it was shuttering Zillow Offers - its entrant in iBuying, a nascent real estate sector in which buyers (“iBuying companies”) use automated, ML-driven valuation models to provide instant all cash offers to property sellers.
The company, like its competitors, was trying to address a critical problem - a house can be bought by a consumer within hours or days, but can sometimes take months or years to sell. If data and analytics can be used to solve the associated illiquidity, iBuying companies will be providing an enormously valuable service to the market.
As Zillow has discovered, however, the math behind these calculations is devilishly complex. The company taking on a seller’s property must be able to cope with the illiquidity while taking on significant capital costs. They must also be able to mitigate potentially devastating risks associated with flaky local market conditions that may leave them “carrying the can” indefinitely.
The value of the iBuyer service will be highest for homeowners whose properties are the most difficult to sell through traditional channels, and I can personally attest to financial risks associated with houses that are difficult to market.
Back in 2018, my family and I waited nine months to sell a property listed at 20% below the Zillow "Zestimate." After our initial optimism waned, we would have jumped at a lowball offer from Zillow Offers or one of its competitors. Alas, such an offer never materialized and we were forced to wait for a real family to imagine themselves resident in our former family home.
iBuyers vs. Insiders
A further complication for iBuyers stems from the fact that the highest bidders are subject to the winner’s curse. If there are 10 bidders for a given property, the one with the highest estimation of value will invariably win the auction. If a company ends up buying 1,000 properties in a competitive marketplace, the curse suggests that they will have paid too much for every single one of them.
To overcome all these hurdles, the iBuyer needs to have something in their arsenal that no-one else does. It may be the best data, the best models, the best intuition, or the best luck. If they plan to fix and flip, they need to be able to accurately determine the most cost-effective interventions and be able to get them done with greater alacrity, and at lower cost, than anyone else.
I’ve known a few individuals who have done very well flipping houses. They normally live in cities with dynamic markets, have an eye for a bargain, and are handy enough to be able to do their own renovations.
For bigger projects, they have a network of local contractors and can often get work done at “mate’s rates.” Networks with local realtors give them early access to the most committed sellers, so they can often win properties without engaging in a competitive bidding process. They are thus able to exorcize the dreaded winner’s curse from the equation.
ML Modeling in the Time of COVID-19
For a corporation to disrupt this enterprise, the models it uses to predict prices have to be extremely precise. The coronavirus, however, has turned real estate markets inside out, so models trained on pre-pandemic data now struggle to provide the precision needed for success.
In previous public 2021 disclosures, Zillow’s management viewed a 200 basis-point forecast error as the breakeven rate, but the company’s projections missed on the high-side by 576 points in Q2 and on the low-side by between 500 and 700 points in Q3. As a result, Zillow Offers was left holding thousands of valuable properties that it saw no prospect of selling for a gain.
All this happened, moreover, in the context of strong aggregate price growth, and one suspects the end for Zillow Offers would have come much sooner had overall prices instead been falling.
The aggregate COVID-19 real estate statistics mask some major ructions in underlying relative house prices. Urban locations that have historically been the most dynamic are now performing relatively meekly. Indeed, of late, the biggest gains have occurred outside the big cities, where newly-freed remote workers can secure a more favorable lifestyle at a lower price per square foot.
The spoils from housing investment have thus been spread more broadly, potentially making it harder for iBuyers to position themselves optimally in the marketplace.
Many of us are interested in whether ML methods are more or less robust to change than humble humans. There is no particular reason why the relative performance of robots will be enhanced by stability. ML tools have difficulty coping with structural change, but so do most people.
Interestingly, within days of Zillow’s announcement that it was shuttering Zillow Offers, OpenDoor - a competitor following a very similar game plan - reported stellar Q3 financials. In this time of ML-driven real estate modeling, it is natural to wonder why Zillow Offers failed amid the chaos of COVID-19 while OpenDoor has, to date, succeeded.
OpenDoor‘s algorithm may simply have been flexible enough to cope with the changes wrought by the horrendous virus. Alternatively, its continued success could be the result of manager-initiated override protocols.
Only an insider would be able to answer this critical question. But what we can say with some confidence is that OpenDoor’s success suggests that ML-driven, analytics-based real estate investment can successfully continue in the context of the post-pandemic reality.
Tony Hughes is an expert risk modeler for Grant Thornton in London, UK. His team specializes in model risk management, model build/validation and quantitative climate risk solutions. He has extensive experience as a senior risk professional in North America, Europe and Australia.