Algorithmic Macroprudential Risk, and a legal case.
According to The New York Times (23 August), The Justice Department “filed an antitrust lawsuit on Friday against the real estate software company RealPage, alleging its software enabled landlords to collude to raise rents across the United States.” I am not an expert in law or anti-trust, but there is another (systemic risk management) angle to this that NYT missed, and worth spelling out.
First, here’s how the Times summarized the case:
RealPage’s software, YieldStar, gathers confidential real estate information and is at the heart of the government’s concerns. Landlords, who pay to use the software, share information about rents and occupancy rates that is otherwise confidential. Based on that data, an algorithm generates suggestions for what landlords should charge renters, and those figures are often higher than they would be in a competitive market, according to allegations in the legal complaint. By Danielle Kaye, Lauren Hirsch and David McCabe
There’s more detail in a piece (here) the Times did earlier in the year (July 19, 2024) written by Danielle Kaye. And for a lot more background see this piece in Propublica (here, October 15, 2022). One wonders why the NYT waited so long until there was government prosecution to report on the issue?
Before I get to my interest in this story, the NYT coverage highlights two important political angles: (i) this may be a contributing cause to rent inflation because the software allows landlords to collude; (ii) there is an interesting issue how this fits under existing anti-trust law because it seems this software allows (the fairly large companies that use the software) to avoid competing on price. The company itself brags they have “purposely built” their platform “to be legally compliant.” I leave the first issue to economists and the second to lawyers.
There is another important angle that I haven’t seen mentioned yet. It’s not so long that we had a Great Recession caused by a financial meltdown. One of the root causes of that crisis were unexpected correlations in assets that risk models assumed to be uncorrelated. In particular, a key assumption was, for example, that local real estate markets are uncorrelated with the stock market and that local housing markets are relative uncorrelated. This is still a quite common conventional wisdom, a fairly random search, immediately found this explanation of uncorrelated asset classes at a website directed at the public at Kubera: “That coupled with the long-term nature of leases and mortgages make the value of real estate assets less reactive to economic news than traditional assets.” (I have no relationship at all to Kubera and don’t use their platform!)
It’s pretty clear that YieldStar’s algorithm will induce (stronger) correlations where there had been weaker ones or almost none before. If the government is right this will be most evidently so within local markets (which is bad for smaller and regional banks, which will make their portfolios even more sources of concentrated risk). As the Bank for International Settlements (BIS) puts it (December 2023), when it reflects on macroprudential risk of mortgages to banks “correlated defaults can result in large losses.” (p. 2) I like to call BIS the ‘central bankers’ bank.’
But the real systemic risk here is that the algorithm will also (accidentally) induce correlations among different housing (and so mortgage) markets. Undoubtedly, that’s hard to prove and may well involve tricky issues in the details of the algorithm. But the crucial issue is that the algorithm itself has become a common cause to all outputs it generates in individual markets. It becomes a systematic, structural background factor in all the markets where its users have a dominant market share. So, for example, while it will spit out different rent-levels in different housing markets, it may reduce the variance of rent-levels in each housing market. And that’s a kind of thing that modern financial markets pick up on and start arbitraging or hedging.
Financial regulators have not been very worried about Yieldstar (and products like it) presumably because it actually reduces likelihood of default of big landlords and so makes mortgages more secure than expected, and even safer going forward. (The central bankers’ bankers are, in fact, much more worried about price competition among mortgage lenders “Excessive price competition among lenders for market share can result in underpriced risks, especially as mortgages are relatively homogeneous products.” (2))
But this is why I am a little bit worried. Industry leading algorithms in all kinds of businesses will also induce correlations within and among asset classes that the historical data — and the macroprudential risk models (and not to mention the regulations for capital holdings of banks) — had suggested were uncorrelated.
This algorithmically induced macroprudential risk is already undoubtedly a fact of life. But the nerd in me hopes this case isn’t settled out of court. It would be good to have a public record on how Yieldstar’s algorithm really works. For, armed with that and all the price data of housing markets (which are relatively well understood by applied economists) it would allow us (well smarter people than me) to begin to estimate how much of a common cause it really is.*
*Research on this post was sponsored by the Dutch Research Council’s Grant 406.18.FT.014