Perspective on Risk - May 15, 2026 (Grabbag)

Model Risk; Private Credit; Credit Conditions; Stablecoins; Confidence; and the Coming Revolution

Note: Over on Demographics & Capital flows I have two new posts: The EU Surplus Will Persist. The Within-EU Imbalance Will Widen and Is the US Fiscal Deficit a Symptom or a Cause? Yes. Feel free to take a read.

Chuck Prince Quote of the Week

“Does anyone really care if the Strait of Hormuz is open?” one high-powered banker posited. — ‘Blissful ignorance’: Milken elite bask in glow of roaring markets (FT)

More on … Model Risk

More Supervisory Guidance

New Guidance on Model Risk – What’s In and What’s Out (UponFurtherAnalysis) has a good writeup on the change to the model risk guidance; more detailed than my previous post. Similar conclussions; different emphasis.

My post is arguably more balanced and operational: some changes are “positive,” especially materiality-based tiering, narrowing the model inventory, and moving away from rote fixed-cycle validation. Neal’s post is more supervisory-skeptical: he sees the same changes as a retreat from post-crisis lessons, driven more by industry feedback than supervisory experience.

More on … the JPM Whale … and Model Risk

Il trader che era diventato il mercato: JPMorgan CIO e la balena che nessuno voleva vedere (Mr. Wolff) - click the translate button, a very good read. Goes into detail on the CDX position gamma, etc. He documents well five failure models and six missed signals.

But for the purpose of this post, it is how he documents the “model risk” involved in the Whale debacle.

There’s a moment in March 2012 when JPMorgan Chase’s Chief Investment Office changed its VaR model. It wasn’t because the old model was flawed—the numbers it produced were correct. It changed it because the old model was producing numbers no one wanted to see. The new model, approved internally within a few weeks, produced a VaR roughly half the previous one for the exact same positions. Management was satisfied. The positions continued to increase. Three months later, JPMorgan announced a $2 billion loss—a figure that would rise to $6.2 billion over the course of the year, the largest trading loss in the bank’s history.

the model change … is the most technically important failure mode of the case

In the London Whale, the model didn’t fail passively. It was actively manipulated to produce outputs that allowed it to continue doing what it wanted to do.

… a new internal VaR model, developed by the CIO’s own team, which for the exact same positions produced a VaR of $195 million — virtually half. The change was approved internally very quickly, without a formal independent validation process. The reported VaR dropped by 45% in just one month. The positions continued to increase, formally remaining within limits.

The fundamental principle of separation between model development, model validation, and model use — codified in the Fed (SR 11-7, 2011) and EBA guidelines—requires that those validating a model have no incentives tied to that model’s output. In JPMorgan’s CIO, this principle was systematically violated: the team using the model had direct influence on the choice of which model to use.

Note: the revised model risk guidance no longer formally mandates the 11-7 seperation of function and conceivably allows just the type of pseudo-validation that was seen here, with the front-office developing the risk model and only a cursory ERM sign-off.

More on … the Risk of AI and the Limitations of Model Risk

Michael Tsu has been killing it with his AI posts. I’m going to highlight two for you that are worth a read.

In 9 seconds, he discusses a catastrophic failure at PocketOS where an agentic system deleted their production database and all volume-level backups in a single API call to Railway, our infrastructure provider.

He uses this as a jump-off point to develop a prototype Agent Accident Report based on the NTSB’s approach to reviewing aviation accidents. I’ll excerpt his conclussion here:

The PocketOS incident was a nightmare for the CEO and the rental car companies that relied on it. But it also provided the public with a view into how an agentic AI deployment can go wrong.

It has given us all something to learn from. This post has focused on three lessons in particular:

1. The need for NTSB-style accident reports for agentic AI incidents.

2. Humility when orienting systems to face agents, instead of people.

3. The benefits of a tiered autonomy approach to agentic AI in the context of third party risk management expectations for banks.

The PocketOS incident revealed that agentic systems can expose latent weaknesses at machine speed. To build trust in agents there needs to be credible accident reporting, robust classification of actual autonomy, and well-designed controls for systems that act faster than people can supervise. That is the work AI builders, banks, and regulators should focus on now.

A second, extremely thoughtful piece he has written is Trust engineering, arguing, fairly pursuasively, that traditional model risk management (MRM) can’t oversee agents effectively.

MRM can be effective for models that function as single-purpose calculators, where the inputs and outputs are of a specific type and form, e.g., for the calculation of credit scores, valuations, the Greeks, probabilities of default, or stress loss estimates. In such cases, outputs can be statistically validated by backtesting against historical data.

Large language models (LLMs) are not single-purpose calculators and they can’t be statistically validated. Their inputs and outputs are open ended and user-AI interactions can be dynamic. Users’ prompts can take a nearly infinite range of possibilities, as can LLMs’ outputs. MRM, which is heavily grounded in documentation and validation, was never designed to handle that.

He goes on to espouse the “harness engineering” approach

The harness engineering paradigm does not try to understand why the model produced a given output. It does not require a human to sign off before each action is taken. Instead, it builds an environment where the range of possible agent behaviors is constrained by design, deviations are detected automatically, and the system improves through use rather than through periodic reviews.

Adopting a harness engineering approach to overseeing agents requires a shift in the role of humans.

In the MRM paradigm, the human is an expert gatekeeper who reviews documentation and approves or rejects models.

In the harness engineering paradigm, the human is a trust engineer who designs the environment, defines the constraints, specifies the success criteria, monitors the system’s performance, and provides continuous feedback over time. The human’s job is not to grant or withhold permission every time an agent asks. The human’s job is to build a world in which the agent’s autonomy is bounded by design, where verification is continuous, automated, and grounded in outcomes, and feedback loops frequently.

More on … the AI Risks to Cybersecurity

New cybersecurity industry coalition aims to lead US critical infrastructure protection signals how serious the threat is. Industry is no longer waiting to follow the government’s lead.

In February, a coalition that includes corporate titans JPMorgan Chase, Mastercard, AT&T and Berkshire Hathaway Energy launched the Alliance for Critical Infrastructure (ACI), vowing to take the lead in helping infrastructure sectors work more closely together to understand and mitigate the shared cybersecurity risks they face. Reading between the lines, the message was clear: The critical infrastructure community, increasingly alarmed at the Trump administration’s retreat from decades-long partnerships, is trying to fill the growing void of coordination and leadership.

More on … Private Credit

More on … Obscuring Deteriorating Asset Quality

The Vanishing Footnote (Mispriced Assets) tells an absolutely facinating tale of financial engineering and the increasingly opaque world of private credit. As detailed by Nick Nemeth, the story traces how questionable debt, specifically loans tied to companies like Medallia, is quietly shuffled through rated collateralized loan obligations (CLOs) and ultimately offloaded into retail funds.

The vanishing footnote itself represents the subtle, almost imperceptible ways in which private credit managers can obscure deteriorating asset quality, effectively using retail investment as a quiet dumping ground for underperforming loans while gatekeeping firms look the other way.As private credit managers aggressively push their products to everyday investors through wealth advisors, the true risk profile of the underlying collateral is masked by a severe lack of transparency. When a crucial disclosure can simply disappear from an annual report without triggering widespread alarms, it highlights how the sheer complexity of these structures outpaces the average investor’s ability to properly underwrite the credit risk they are absorbing.

Ultimately, this tale serves as a stark warning about the shifting of systemic risk. The illusion of steady, high-yielding private credit relies heavily on valuation marks that are rarely stress-tested by transparent, public markets. If the sector continues to rely on financial sleight-of-hand to sustain its massive inflows, the eventual reckoning will not just penalize institutional players, but will disproportionately harm the retail investors left holding the bag when hidden defaults finally materialize.

It’s a great-read if you are into financial chicainery or market-microstructure.

More on … Bank Concerns About Private Credit

Goldman is out with Cracks in Private Credit that suggests that software exposure is the new “CRE office” for private credit. They lay out a causal chain:

AI → software borrower impairment → BDC redemption pressure → valuation/liquidity confidence problem.

Goldman is, of course, talking its book, but intelligently. It argues its software exposure is lower than peers. The bottom line insight is that AI disruption may be the first genuinely novel credit-cycle trigger for private credit, not just a macro/higher-rates story.

Meanwhile, Banks are pretending not to notice a big problem with private credit (American Banker) points the finger at Wells Fargo.

The question about private credit that many people are asking right now is whether this is 2007 again. That’s the wrong framing. The real question is what happens when the funds that banks have quietly financed for years face stress — and when the collateral banks think they’re holding is worth less than quarterly net asset value statements imply.

Let’s be direct: U.S. banks have extended nearly $300 billion in credit to private credit funds, business development companies, or BDCs, and collateralized loan obligations, or CLOs, per Moody’s data as of October 2025. Wells Fargo alone sits at $59.7 billion — nearly double the next‑largest lender. Nondepository financial institution loans are now 10.4% of total U.S. bank lending, up from 3.6% a decade ago.

More on … Regulatory Interest in Private Credit

Meanwhile, the FSB has published Report on Vulnerabilities in Private Credit. The paper is important because it translates a fast-growing, opaque market into a set of monitorable risk channels. It does not argue that private credit is inherently destabilizing; it argues that its opacity, interconnections, leverage, valuation discretion, and limited downturn experience make it a financial-stability issue and a firm-level risk-management priority.

They lay out a surveillance framework for private credit. The FSB notes that direct bank lending to private credit funds appears relatively small, but uncertainty is large, and risks may arise through indirect channels such as NAV lending, subscription lines, asset-based lending, CLOs, SRTs, revolving credit lines to the same borrowers, and strategic bank–asset manager partnerships. The paper repeatedly emphasizes that definitions and datasets are inconsistent, and its heatmap shows uneven availability of basic data across jurisdictions. That gives policymakers a concrete basis for harmonizing reporting templates. Expect this to come down the road.

More on … Credit Conditions

More on … Lending to NDFIs

From the most recent SLOOS:

More on … Commercial Real Estate

Legacy office loans deteiorating; data-center construction booming (but with weak underwriting standards)

Office

A Fire Sale Has U.S. Office Buildings Going for 90% Off (WSJ)

Even higher-quality properties on average have dropped about 35% in value from their peak, according to analytics firm Green Street.

Lenders to Commercial Real Estate Owners: Pay Up Now (WSJ)

Lenders to commercial real-estate owners are reaching the breaking point, calling in tens of billions of dollars of troubled loans.

Refinancing property debt has become difficult since interest rates started to soar in 2022. Many lenders initially extended maturing loans they made when borrowing costs were far lower, hoping that either interest rates would fall or that cash flows would grow. It is a strategy known as “extend and pretend.”

Now, many lenders have stopped pretending, and the default rate is surging. The delinquency rate for office loans in commercial mortgage-backed securities climbed to a record 12.34% in January, the highest level since Trepp began tracking in 2000.

Construction

If you’ve followed this substack for a while, you know that I like to look at the SLOOS data on a detrended and normalized basis. May be a bit hard to read: Green lines are demand; red lines are standards. Demand up; standards down. From 2014 to current day.

More on … Stablecoins

More on … Stablecoins vs Tokenized Deposits

Dan Heller, the former head of the secretariat of the Committee on Payments and Market Infrastructures of the BIS, writes in The stablecoin stumbling block (FT). You should listen to Dan. He writes:

… today’s stablecoin arrangements are not safe enough for them to function as large-scale settlement assets, at least in their current form.

In the wake of the 2007-09 financial crisis, leading central banks and securities regulators agreed on a comprehensive set of 24 principles that FMIs must observe, which have since been embedded in national legislation around the world. Particularly important is the standard on the quality of the money used to settle financial transactions. It requires FMIs to settle in the highest-quality form of money available — ideally central bank money or reserves, which carry neither credit nor liquidity risk.

… the balance sheets and operating models of today’s leading stablecoins are grossly misaligned with those of regulated FMIs. They hold hundreds of millions in uninsured bank deposits and typically take one to three business days to honour fiat redemptions of stablecoins. The securities they hold may have relatively short maturities, but would still need to be sold in the event of large-scale redemptions, with the potential to trigger dislocations in securities markets. As a result, current stablecoin arrangements offer neither the credit quality nor the immediacy of liquidity required of wholesale settlement assets.

So that’s a “no thank you, dawg” for the use of existing stablecoins in intra-bank settlement.

He further writes:

Tokenised central bank money would clearly satisfy regulatory expectations, and several central banks are exploring “tokenised central bank money” or “wholesale central bank digital currency”

Perhaps there is a world where consumers use stablecoins, but inter-bank settlement happens through tokenized deposits.

More on … Confidence (or the Lack Thereof)

More on … the Coming Revolution