Dazzled by Wizardry, Federal Mortgage Regulators Ignore Zillow Debacle

Secretary of Defense Robert McNamara was a committed technocrat. Under his direction, a team of policy advisors descended into the Pentagon’s cavernous basement in 1967. They fed punch cards into the basement’s IBM mainframe computers with everything that could be quantified about the Vietnam War. Numbers of ships, tanks, transport helicopters, gunships, fixed-wing aircraft, artillery, troop strength, machine guns, ammo. They queried the computers, “What year will we win in Vietnam?” They went away on a Friday and let the computers work on the problem all weekend. When they returned the following Monday, there was one card in the output tray. It said, “You won in 1965.”

While computing power might have “solved” the Vietnam War on paper, the results clearly didn’t square with reality.

Fast-forward to 2023. Six federal agencies are seeking public comment on a newly-proposed rule designed to “ensure” the credibility and integrity of computer models used in automated real estate valuations. This rulemaking is one more sign that federal bureaucrats are all in on a whacky plan to use technical wizardry to tease out the value of individual properties across the country. The wager? The nation’s $12 trillion residential mortgage market, much of which is now backstopped by the full faith and credit of the U.S. Treasury.

These algorithmic valuations replace trained human appraisers with knowledge of local real estate markets. The so-called “black box appraisals” rely on mathematical formulas and Big Data to produce an opinion of value. The problem? The models aren’t able to think like buyers and sellers. They don’t pick up design flaws, emotion-based style preferences, ambient noise levels, smells, street parking issues, market tastes, perceived quality of school districts and responsiveness of local government. It’s too much to ask of a machine.

Rohit Chopra, director of the Consumer Financial Protection Bureau, rightly sounded the alarm in a recent blog post – the models have the potential to do great harm.

The evidence is there for anyone wishing to look. In a disastrous bet made by Zillow – one in which the company staked its future on investing in residential real estate based on its own algorithms – the company lost $32 billion in market capitalization from February to November 2021. What did Zillow learn about its “Zestimates” when its own money was at stake?

Zillow, a company that gained popularity through its online listings, bet the farm on Zillow Offers, a home-flipping venture that relied on algorithms to identify homes the company would purchase. The plan was to make minor renovations and sell them off quickly for a profit. In a colossal about-face, after a write-down of half a billion dollars, the company announced the closure of the subsidiary, despite it being the primary source of revenue but not profitability. Zillow also reduced its workforce by approximately 2,000 employees, constituting a quarter of its staff.

The company disregarded internal concerns that its algorithms were causing it to overpay for homes, former and current employees told the Wall Street Journal in November 2021.

More than a decade earlier, in one of the first bombshells of the financial crisis, Standard and Poor’s Rating Services abruptly announced it was putting 612 mortgage-backed securities on “CreditWatch negative” and that it expected most to soon be downgraded because of high delinquency and foreclosure rates. Its algorithmic valuation model for the securities was known to spit out investment-grade ratings on junk-quality mortgage-backed securities, after which they could be purchased by pension funds.

Moody’s Investors Service, another purveyor of algorithmic valuations, quickly dropped a similar bombshell. It announced it would be downgrading 399 securities and placing an additional 32 securities on review for possible downgrade. Two days later, Fitch Ratings followed suit, signaling big problems with its own automated valuation models. Many of the once “investment grade” loans were in the portfolios of Wall Street players like Bear Stearns, Citigroup, JPMorgan, Merrill Lynch and Morgan Stanley. Others were held by public employee pension funds as far away as Norway.

In 2009, Calpers, the largest public pension fund in the United States, sued Moody’s, Standard & Poor’s and Fitch for assigning their highest ratings to securities that later suffered enormous losses. The suit would not be settled until 2015. This part of the crisis traces directly to the flawed automated valuations of homes whose mortgages were packaged up in the securities.

The rule now being proposed by six vast federal bureaucracies blithely assumes the holy grail is already in hand – the ability to quality-control imperfect algorithms that can never simulate the workings of a complex market involving millions of buyers and sellers. Until this holy grail is discovered, appraiser and podcaster Phil Crawford believes the automated models will introduce something he calls “data cancer” into the increasingly nationalized U.S. mortgage market. He defines “data cancer” as an echo-chamber effect arising from the proliferation, over time, of distorted values as they work their way into sales data, thereby buttressing ever-more distorted values.

“To paraphrase Hemingway’s quote on bankruptcy,” said Crawford, “the resulting crash will happen gradually, then suddenly.”

The agencies involved in the federal rulemaking include the Federal Housing Finance Agency; the Consumer Financial Protection Bureau; the National Credit Union Administration; the Federal Deposit Insurance Corporation; the U.S. Department of the Treasury; and the Federal Reserve System.

Swedish economist Assar Lindbeck famously called rent control “the most efficient technique presently known to destroy a city—except for bombing,” Were professor Lindbeck still alive, he might now have included the automated valuation of a city’s real estate.