[Week 11 of 2026] Coins, Classifications, and Complacency

Welcome back to Price and Prejudice with a few musings from Week 11 of 2026. Now that the Capital Markets & Investments class at Columbia is in progress, the blog will also reference to some contents from class as well.

Deposit Déjà Vu

This WSJ article discusses how stablecoins are starting to catch on as a place to park cash and earn yields, with some crypto users reporting returns of 5-12% compared to near-zero on traditional savings accounts. For obvious reasons, banks are not thrilled, with industry estimates saying that $6.6 trillion in deposits could drain if yield-bearing stablecoins are allowed to flourish. Two main usual pushbacks against stablecoins still stand: stablecoins aren't FDIC-insured and they may not be as stable as you might think.

As with many things in Crypto, the playbook here is remarkably familiar. In the 1970s, money market funds offered savers higher yields than banks could pay under Regulation Q interest-rate caps, and banks made nearly identical arguments about deposit flight and systemic risk. They lost that fight. MMFs grew from nothing to trillions in assets, and Reg Q was eventually repealed. The stablecoin debate follows the same script: an upstart product offers better rates, incumbents invoke consumer protection and financial stability, and regulators are caught in between.

What might be different this time is that banks can actually fight back. In the 1970s, Regulation Q capped what banks could pay on deposits, so they had no way to compete with money market funds on yield. Today there's no such constraint. Banks pay near-zero on savings accounts because depositors are sticky and don't shop around, not because they're legally prohibited from offering more. If stablecoins are the thing that finally wakes depositors up to the opportunity cost of their cash, banks can just raise rates. And the other difference is probably infrastructure. Money market funds plugged into existing financial plumbing quickly, offering check-writing and brokerage integration that made them feel like bank accounts within a few years. Stablecoins are nowhere close. You can earn 5% on them, but you can't pay rent with them (at least for now). Banks have both the tools and the time to respond in ways they didn't fifty years ago.

When Labels Lag

This WSJ article points out something odd about the current war trade: investors aren't hiding in the usual places. Healthcare and consumer staples, the textbook defensive sectors, are both down more than the S&P 500 since the Iran conflict began, while tech has barely budged. Part of the explanation is crowding: investors had already rotated into defensives to hide from AI uncertainty in the weeks before the war, so by the time hostilities started, the lifeboat was already full. But the more interesting part is what actually predicted performance within these sectors. It wasn't the sector label, but geography: the top performers in healthcare and staples generated about 72% of revenue in North America; the laggards were closer to 59%.

This makes sense once you stop thinking of industry classification as description of risk. "Consumer staples" tells you what a company sells, not what risks it's exposed to. But a domestic food company and a global food company with heavy emerging-market revenue share a sector label and almost nothing else from a risk perspective. When the shock is geopolitical, what matters is geographic exposure which cuts across sectors entirely. In other words, SIC codes are static labels slapped on dynamic businesses, and they haven't kept up.

The better framework is to think in terms of factor models: decomposing stock's returns into exposures to underlying risks. Instead of asking "is this a defensive stock?," the more useful question is "what is this stock's loading on the risks that matter right now?" The market already does this intuitively. The only catch is that factor models are estimated from historical data, which means they work best for risks we've seen before. When a genuinely new risk shows up, like COVID in 2020, there's no historical loading to look at. You're flying blind (unless you are creative) until enough data accumulates, and by then the trade is over.

Redundant Checks

Risk management in capital markets is a broad tent which includes things like position sizing, hedging, stress testing, counterparty limits, value-at-risk models, and the compliance infrastructure that wraps around all of it. After every crisis, the tent gets bigger – for example, after the great financial crisis, banks added layers of oversight: more risk officers, more sign-offs, more automated monitoring.

In a recent episode of Bloomberg's Odd Lots podcast, former Goldman Sachs CEO Lloyd Blankfein makes a point about one dimension of risk that rarely gets discussed. He argues that when you build nine layers of checks into a system, nobody takes any single check seriously because everyone assumes the other eight will catch the problem.

This is a well-documented phenomenon in other fields. Aviation safety where people call it "automation complacency": the more reliable the autopilot, the less attentive the pilot. In finance, the pattern played out with credit ratings before 2008. Ratings were supposed to be one input among many, but issuers, investors, and regulators all treated the triple-A stamp as a substitute for their own analysis. Each participant assumed someone else had done the real diligence. Adding another layer of oversight didn't make the system safer but rather gave everyone else permission to look away.

The question is whether this gets worse as AI enters the picture. As firms layer in automated anomaly detection and algorithmic risk flags, the humans in the loop have even more reason to assume the machine will catch it. So Blankfein's worry isn't about sophisticated cyberattacks or rogue traders but more about the fat finger and the mundane error that ten systems were supposed to catch and none of them did. This tells you that in a world where risk management keeps getting more automated, the binding constraint is whether anyone is still paying attention.