Why Financial Crises Are Invisible Until They’re Unavoidable

Mar 28, 2026

On March 8, 2023, Silicon Valley Bank announced it needed to raise capital. Forty-eight hours later, it was dead — the second-largest bank failure in American history.

Here is what was known at year-end 2022, three months before the collapse. SVB held $91.3 billion in held-to-maturity securities on its books at amortized cost. Their actual fair value was $76.2 billion. The gap — $15.1 billion in unrealised losses — exceeded the bank’s entire book equity of $16.3 billion. The bank was, by any honest measure, within spitting distance of insolvency.

And yet nobody acted. Not the market. Not the regulators. Not the credit rating agencies. The standard post-mortem explanations are familiar: insufficient data, regulatory capture, analytic failure, groupthink.

A recent working paper argues that all of these explanations are wrong — or rather, that they are describing symptoms of something more fundamental. The paper, “Finite-Time Certification of Latent Economic Regime Change,” makes a claim that is both simple and disturbing: the invisibility of SVB’s insolvency was not an accident of bad regulation or lazy analysis. It was a mathematical property of the reporting system itself. No test, no model, no amount of data could have certified that SVB had crossed into insolvency — because the accounting rules were designed, at a structural level, to make exactly that determination impossible.

The mechanism is geometric. And once you see it, you can’t unsee it.

The problem with magnitude-only reporting

Every financial reporting system faces a design choice: what do you show the public?

The answer, for most banking assets, is straightforward. If you hold a bond and interest rates move, the bond’s market value changes. Under mark-to-market accounting (the rules for “available-for-sale” securities), that change flows through to the financial statements. Investors can see whether you’re up or down.

But there’s a second category: held-to-maturity. Under HTM accounting, you report the bond at its original amortized cost — the price you paid, adjusted for the contractual amortization schedule. The market value might have cratered, but your primary financial statements show the original number. The logic is that you intend to hold the bond to maturity, so interim fluctuations don’t matter.

This creates an information structure with a specific geometric property. The public can see that a bank holds a large portfolio of fixed-income securities. They can see the aggregate size. What they cannot see, from the primary financial statements, is which direction that portfolio is moving — whether unrealised losses are accumulating toward insolvency or receding toward safety.

The paper formalises this as a fold map. Think of it this way. Define a bank’s hidden fragility as a number F: positive when losses are growing toward the danger zone, negative when they’re shrinking away from it. The sign of F tells you whether you should be worried.

Now suppose the reporting system shows you not F, but F². The square of any number is positive regardless of whether the original was positive or negative. A bank with F = +0.15 (recovering) and a bank with F = −0.15 (deteriorating) produce the same public signal: F² = 0.0225.

One bank is getting safer. The other is approaching insolvency. They look identical.

That is the fold. The observation map destroys the sign of the fragility coordinate. It “folds” the positive and negative sides of the fragility spectrum onto each other, so that the public sees magnitude without direction. And the paper proves — not argues, not suggests, but proves — that when the observation map has this structure, no statistical test can certify that a bank has crossed into insolvency. The worst-case detection error is exactly one-half. A coin flip.

This is not an approximation. It is not a finite-sample problem that goes away with more data. It is not a computational limitation that better algorithms could overcome. It is a theorem about the geometry of the information available to any observer, however sophisticated, over any time horizon, with any amount of data.

The paper calls this the fold bound.

Why “just get more data” doesn’t help

The natural response to any information problem is: collect more data. Run the numbers harder. Build a better model.

The fold bound says this doesn’t work, and it says so with mathematical precision.

The core argument uses a technique from statistical decision theory called Le Cam’s two-point method. The idea is simple: if you want to show that no test can reliably distinguish between two hypotheses, construct two specific scenarios that produce nearly identical data but have opposite truth values.

The paper constructs exactly this. For any observation horizon T and any detection threshold, it builds two model economies. In one, the bank is solvent. In the other, it is insolvent. The two economies produce observation paths whose probability distributions are arbitrarily close in total variation — a formal measure of statistical distinguishability. You can make the gap as small as you want by choosing the economies to be close enough to the fold.

The construction works because the fold symmetry creates exact mirror pairs. Take any trajectory where the hidden fragility coordinate F is slightly positive (safe) and mirror it to slightly negative (unsafe). The public observations are identical — F² is the same in both cases. The only thing that differs is the sign, which the observation map has destroyed.

Add to this what the paper calls attenuation: as a bank approaches distress, its reporting becomes less informative, not more. Distressed institutions have the strongest incentive to delay, obscure, and smooth their disclosures. Accounting rules contribute: the HTM classification itself discourages asset sales that would trigger fair-value recognition. Near the boundary where you most need information, the signal gets weaker.

The combination — sign destruction plus near-boundary signal collapse — makes the impossibility exact and complete. Every detector, from a simple ratio test to a machine learning classifier with unlimited computational power, achieves worst-case error no better than random guessing.

Three reforms, all necessary

The paper doesn’t just prove an impossibility. It proves a sharp converse: three specific structural reforms, implemented together, restore the ability to certify insolvency. Each is individually necessary — drop any one and the impossibility returns.

Signed disclosure. The observation map must reveal the sign of the fragility coordinate, not just its magnitude. In practice, this means mark-to-market accounting for stress-relevant portfolios. When a bank’s fixed-income holdings are losing value, the financial statements must show that they are losing value — not just that a position of a certain size exists.

Non-collapsing transmission. Signal strength must not vanish as the bank approaches distress. In practice, this means inverse disclosure triggers: more reporting as capital deteriorates, not less. Current practice often works in the opposite direction — banks in trouble reduce disclosure frequency, reclassify assets to avoid mark-to-market, and exploit reporting lags.

Transverse crossing. Boundary crossing must be detectable as a discrete event, not a gradual graze. In practice, this means hard capital thresholds — automatic regulatory intervention at specific, pre-committed levels — rather than the discretionary forbearance that allows banks to drift along the insolvency boundary without triggering action.

Drop signed disclosure, and the fold symmetry persists. Drop non-collapsing transmission, and the signal vanishes near the boundary even with full sign information. Drop transversality, and the bank can approach the boundary tangentially, making the crossing indistinguishable from a near-miss.

The converse theorem’s conditions are tight. Each is proved individually necessary via explicit counterexamples (Propositions S13.1–S13.2 in the supplementary appendix). The reform prescription is not a policy preference — it is a mathematical consequence of the information geometry.

What the data say

The theory makes a cross-sectional prediction: banks with more of their securities hidden behind the fold map (larger HTM share) should be harder to certify from public data, and their hidden losses should predict failure more strongly than visible losses.

Using FFIEC Call Report data on 9,356 US banks from 2005 to 2023, the paper tests this. For each bank, it computes two variables:

Fold severity (κ): the fraction of securities that are not in the HTM category. A bank with κ = 0 has everything hidden behind the fold. A bank with κ = 1 has everything marked to market. Higher κ means more sign disclosure.
Signed fragility (F): unrealised HTM losses divided by book equity. A bank with F = −0.50 has unrealised losses equal to half its equity. This variable is computable from footnote disclosures, but it does not appear on the primary financial statements — it is precisely the information the fold map hides.

The results are striking. At year-end 2022, SVB had κ = 0.22 (78% of its securities were HTM — deep in the fold) and F = −0.98 (unrealised losses had consumed 98% of equity). First Republic had κ = 0.11 and F = −0.27. Both sat squarely in the “hard region” — the parameter space where the theorem predicts certification is impossible.

The most telling result is the placebo test. Among banks with HTM exposure, unrealised HTM losses strongly predict failure (β = −18.5, p < 0.001). Among banks without HTM exposure, unrealised losses on available-for-sale securities — which are marked to market — have the opposite sign (β = +4.9, p < 0.001). Larger visible losses are associated with lower failure probability.

Read that again. Hidden losses predict failure. Visible losses predict survival.

The fold bound explains why: mark-to-market disclosure forces corrective action. When losses are visible, management acts, regulators intervene, and the market reprices. When losses are hidden behind the fold map, none of these mechanisms trigger. The bank drifts toward insolvency in silence.

A model trained on 2008 crisis data — a completely different crisis, driven by loan losses rather than interest-rate losses — ranks SVB and First Republic in the top 0.6% of predicted failures when applied to 2022 data, 14 years out of sample. The fold variables identify vulnerable banks across crises because the underlying geometry — signed fragility hidden behind magnitude-only reporting — is the same regardless of what drives the losses.

The fold is an equilibrium, not an accident

One might ask: if the fold map is so destructive, why does it exist? Why did regulators permit HTM accounting in the first place?

The paper’s answer is that the fold is not a regulatory accident. It is the equilibrium outcome of a game between three players: the institution (which benefits from opacity), the standard-setter (which faces political costs of mandating transparency), and the market (which prices what it can observe).

When the perceived cost of undetected insolvency is low — which it is during booms, when nobody believes a crisis is coming — the political cost of mandating signed disclosure exceeds its perceived benefit. Banks lobby against mark-to-market. Regulators accept the argument that “these are held to maturity, interim fluctuations don’t matter.” The resulting reporting rule is exactly the fold map.

The paper proves this: for sufficiently low perceived systemic risk, the unique equilibrium of the standard-setting game is κ = 0 — the pure fold. The impossibility theorem then applies endogenously. The reporting system is designed to make crisis detection impossible, and it is designed that way because the players’ incentives produce exactly that outcome when nobody thinks a crisis is coming.

The welfare cost is first-order. In the fold regime, the entire certification gain is lost. The paper calibrates this at roughly $76 billion per systemically important bank, or $1.4 trillion in aggregate across the 19 largest US bank holding companies — the same order of magnitude as the total fiscal cost of the 2008 crisis. This is not a coincidence. The fold bound is the mechanism by which crisis costs accumulate undetected.

What stress tests actually do

There is one institution that bypasses the fold: the bank supervisor conducting a stress test.

When the Federal Reserve runs CCAR stress tests, it gets access to granular, signed, position-level data — exactly the information the fold map destroys for public observers. In the paper’s language, a stress test sets κ_eff = 1: the supervisor sees F directly, not F².

The paper notes that all three banks that failed in 2023 — SVB, Signature, and First Republic — were outside the CCAR stress-testing regime. They were too small to be covered. Zero CCAR-covered banks failed. This is consistent with the converse theorem, though not causal proof (CCAR banks are larger and have more diversified revenue).

The policy implication is clear: banks in the “hard region” — high fold exposure, negative fragility — cannot be certified as safe by the market. They require supervisory certification. Stress tests are not a nice-to-have supplement to market discipline. For banks behind the fold, they are the only mechanism that can work.

The honest caveats

The paper is not without weaknesses, and intellectual honesty requires naming them.

The mapping from HTM accounting to the fold map F → F² is not quite right as stated. Amortized cost is not literally F² — it is a constant determined by original purchase price. The fold structure arises from the effective information environment: the combination of accounting blackout (primary statements reveal nothing about F) and market-based signals that partially reveal magnitude but not sign. The paper’s formal machinery is correct, but its headline mapping — “amortized cost is g(F) = F²” — overstates the precision of the correspondence.

The empirical results are strong on pattern but thin on identification. The placebo test (hidden losses predict failure, visible losses do not) is the cleanest result. The out-of-sample prediction is compelling as a case study but has only three positive examples — a ranking, not a statistic. The fold-severity variable κ is insignificant in all regression specifications, which the paper honestly acknowledges but which also means the paper’s main theoretical variable has no direct predictive content in its own regression.

The game-theoretic equilibrium result relies on a maintained assumption (concavity of the certification-value function V in fold severity) that is supported by simulation but not proved analytically. And the proof that the fold is the unique equilibrium for low systemic risk establishes a local optimality condition but does not formally verify global optimality.

These are real issues. They do not invalidate the paper’s core contribution, but they constrain how much weight to put on the specific empirical claims versus the general theoretical insight.

Why this matters beyond banking

The fold bound is not specific to bank insolvency. Any system where public data report magnitude without sign has the same structure.

Interbank lending markets report trading volume but not signed net flow. High volume is consistent with healthy lending (positive net flow) and desperate rollover borrowing (negative net flow). The paper formalises this as a second application: the September 2019 repo market stress, where the SOFR rate spiked from 2.2% to 5.25% while tri-party volume barely changed, is a clean demonstration. The signed signal (rates) revealed the stress. The magnitude signal (volume) did not.

Central banks face the same structure in monetary transmission. Credit aggregates report total lending volume but not whether lending is expanding or contracting at the margin. If the credit channel has reversed sign — banks hoarding reserves rather than lending — the aggregate volume is unchanged. The fold bound implies the central bank cannot certify monetary transmission failure from standard credit aggregates.

The pattern is general. Wherever a reporting system collapses positive and negative states into a single observed magnitude, the fold bound applies. The impossibility is not about banking specifically. It is about the geometry of observation.

The bottom line

Financial crises are not invisible because we lack data, or models, or computing power. They are invisible because the reporting system is designed — in equilibrium — to destroy exactly the information needed to certify that a crisis is underway.

The fold bound formalises this. It identifies the precise geometric property (sign destruction) that makes certification impossible, the exact conditions (signed disclosure, non-collapsing transmission, transverse crossing) that would restore it, and the strategic mechanism (the standard-setting game) that produces the fold as an equilibrium outcome.

SVB’s $15.1 billion unrealised loss was not hidden because someone chose to hide it. It was hidden because the reporting rule — amortized cost for HTM securities — was the fold map, and the fold map was the equilibrium. The math says: under that map, no test could have certified insolvency. Not with better analysts. Not with more data. Not with faster computers.

The invisibility was structural. And until the structure changes, it will happen again.

Craig’s Substack

Discussion about this post

Ready for more?