Banking has one question, and it is an embarrassment: why do banks exist at all? In an Arrow-Debreu world they should not. Every asset has a price, every contingency a contract, every saver meets every borrower in a frictionless market, and the institution that stands between them, holding illiquid loans against deposits repayable on demand, is a wheel that turns nothing. The field’s history is a sequence of answers to that embarrassment, and each answer carries a shadow. Propose a reason for banks to exist and you have specified, whether you meant to or not, the way they break: if banks manufacture liquidity, the demandable claim that does the manufacturing is what runs; if banks are delegated monitors, the assets only they can value are opaque, and opacity is what panics; if banks are deposit franchises, the sleepy depositors who fund the franchise are the ones who can wake up. Existence and fragility come as a pair, and this essay narrates the pairs.

Two things recur. The first is a rule written before any of the theory: lend in a panic, at a very high rate, on all good banking securities, as largely as the public asks. Bagehot set it down in 1873 as codified practice, without an equilibrium model of the bank anywhere beneath it, and the rule outlived every model built since; its reappearance, essentially intact, at the discount window in 2008 is one of the strangest facts in the history of economic thought, and this essay tracks the rule’s absences as carefully as its returns. The second recurrence is the crisis as examiner. Each era’s theory got tested by an event it did not choose: the Depression tested the monetary account of bank failure, the S&L collapse tested deposit insurance, 2008 tested the claim that banking had migrated safely to markets, and Silicon Valley Bank tested the deposit franchise in the spring of 2023. This is a history of a question, not a catalog of results; papers enter as evidence for claims about how the question changed, and as ammunition in the fights the question provoked. The fights are the organizing units: whether money or credit destroyed the 1930s, whether fragility is a defect or a design, whether deposit insurance stabilizes or corrupts, whether 2008 was illiquidity or insolvency, whether the deposit franchise hedges interest rate risk or manufactures it. None of them is fully settled, and the essay keeps score honestly rather than declaring winners the record has not declared.

Prologue: Practical Wisdom (1873-1959)

For a century, central banking had a manual and no theory. The manual is Bagehot’s Lombard Street (1873), and the first thing to say about it is what it is not: it is not a model, not an equilibrium, not even an argument from first principles. It is a description of “concrete realities,” reasoned inductively from the Bank of England’s published balance-sheet returns and from the panics of 1825, 1837, 1847, 1857, and 1866 (Bagehot, 1873). Bagehot’s subject is a system, not a bank: London held roughly £120 million of known deposits against Paris’s £13 million and New York’s £40 million, the greatest concentration of borrowable money the world had seen, and by what he regarded as an accident of privilege and history, the whole structure kept its single ultimate cash reserve at one institution (Bagehot, 1873, Ch. I). An average reserve near £10 million stood behind some £27 million of liabilities (Bagehot, 1873, Ch. VII). He saw plainly what the reserve arrangement implied. The public’s money is held on short notice or on demand, and in a panic some of them ask for it; against a general panic, he wrote, quoting Ricardo, banks have no security on any system (Bagehot, 1873, Ch. I). That is the existence-fragility pairing of this entire history, observed directly, sixty years before anyone modeled the deposit contract that produces it.

The panics came on a rough ten-year cadence, 1825, 1837, 1847, 1857, 1866, and the Bank of England met each with less doctrine than the last crisis had taught it; Bagehot’s complaint is that it performed the duty “hesitatingly, reluctantly, and with misgiving,” without ever acknowledging the duty existed (Bagehot, 1873). His project was to force the acknowledgment: to convert a practice the Bank performed under duress into a rule a Governor could state in advance, because a panic met by a lender known in advance to lend is a panic half-prevented. He was, by his own insistence, no alarmist; the system, though curious and peculiar, could be worked safely, if only its custodians would study it and say aloud what they would do (Bagehot, 1873).

What Bagehot supplied was the response. His formulation is worth quoting precisely, because the version everyone knows is a later compression he never wrote. There are two rules, he says in Chapter VII: first, that panic loans “should only be made at a very high rate of interest,” a heavy fine on unreasonable timidity; second, that at this rate the advances be made “on all good banking securities, and as largely as the public ask” (Bagehot, 1873, Ch. VII). The freely-lend half is argued at length in Chapter II: a panic is a species of neuralgia, and you must not starve it; very large loans at very high rates are the best remedy (Bagehot, 1873, Ch. II). The slogan “lend freely, at a penalty rate, against good collateral” is faithful in substance and false in attribution; “penalty rate” and “collateral” are not his words, and the difference matters to anyone tracking how practical wisdom gets transmitted. His live example was Overend, Gurney and Company, the most trusted private firm in England, whose 1866 suspension set off a panic the Bank answered by advancing £13 million of new money within days, running its own reserve nearly to nothing (Bagehot, 1873, Ch. VII). And the ambivalence that citation practice has stripped from him should be restored: Bagehot did not like the system his rule serves. He thought the natural arrangement was many banks of roughly equal size, each keeping its own reserve, and defended the lender of last resort only as the workable fix for a one-reserve system too entrenched to abolish; proposing to return to the natural system, he wrote, would be childish (Bagehot, 1873, Ch. II). The founding document of central banking doctrine is a second-best argument by a man who preferred a world in which the doctrine would be unnecessary.

Between the manual and the theory stands one hinge, and this history must be honest about how it was read. Gurley and Shaw’s Money in a Theory of Finance (1960) placed money inside a general theory of finance, distinguished inside money (claims backed by private debt) from outside money (claims backed by assets outside the private sector), and treated the money-creating bank as one intermediary among many, its specialness a matter of degree rather than kind (Gurley and Shaw, 1960). A book that wanted to give money a richer place in finance was received as the first step toward emptying banks of specialness altogether. The next era finished the job, and it did so, in part, by citing Gurley-Shaw not as an ally but as an example of the error still to be cleared away.

Era I: Banks as Veils (1960-1982)

The era’s answer to the existence question, stated at full strength so a reader can fight it: banks have no reason to exist as anything economically special. In a frictionless competitive world their portfolio activities fall under Modigliani-Miller and are a passive veil over portfolio choice; the only thing that makes real banks look special is regulation; and the correct null fragility of a bank is zero. This is the shortest era in this history and the one doing the most structural work, because everything after it is a specification of which friction breaks it.

Fama’s “Banking in the Theory of Finance” (1980) is the canonical statement. A bank does two things, provides an accounting system of exchange and manages portfolios, and under competition the second is a pure financing decision covered by Modigliani-Miller: perturbations by one bank are exactly offset by others, so no bank alters the portfolio opportunities investors face, and the banking sector is at most a passive force in the determination of prices and real activity (Fama, 1980, pp. 45-46). An accounting system of exchange, ledger entries clearing payments, needs no reserve requirements and no control; it is bookkeeping, and bookkeeping does not need a quantity theory (Fama, 1980). What makes actual banks look special are two named regulations: reserve requirements, which operate as a tax that differentiates bank deposits from other securities, and deposit-rate ceilings, which force banks to rebate returns as under-priced transaction services (Fama, 1980, pp. 47-49). Black had run the same experiment a decade earlier from the monetary side, in a paper the mainstream would not host; it appeared in the practitioner Journal of Bank Research in 1970 after a March presentation in Modigliani’s MIT seminar, the outsider’s provocation delivered inside the temple and published outside it. In a world of uncontrolled banking, Black argued, it would not be possible to give any reasonable definition of the quantity of money; the law of reflux cancels any excess of bank liabilities automatically, since money created beyond what the public wants is simply used to pay off loans (Black, 1970). Two routes, one destination: Fama by portfolio irrelevance, Black by reflux neutrality, and the independence of the routes is what made the destination credible. Both locate the entire specialness of banks in the same two regulations, and neither model contains a run, a panic, or a failure that matters. Fama’s deeper housekeeping is easy to miss: he separated two problems the literature had run together, rendering the monetary system determinate and giving content to a nominal unit of account, and argued that controlling outside currency alone secures the price level while the banking system, freed of both named regulations, does pure portfolio management on the side (Fama, 1980, pp. 54-55). Banking policy and monetary policy come apart; the central bank’s business is the numeraire, not the banks.

Read the omission closely, because it is documented rather than inferred. In Fama, some deposits are literally claims against an open-end mutual fund: there is no fixed nominal demandable claim, so there is nothing to run (Fama, 1980, p. 41). In Black, reflux means excess claims cancel before they can move anything (Black, 1970). The veil model does not answer Bagehot; it deletes him. The rule’s two clauses presuppose a demandable claim and a panic, and the era’s models contain neither, so for the first and only time in this history, the lender of last resort has nothing to do. This is also why the era cannot be dismissed as a wrong turn. Its zero-fragility null is the benchmark that gives every later friction its meaning, exactly the role irrelevance theorems play elsewhere in economics; the field decided, in reception, that Fama’s deliberately overstated frictionless limit was most useful as the scaffold for proving banks special, an inversion of the paper’s own polemical intent (Fama, 1980). The reversal pattern runs through all three era papers, and each deserves its line. Gurley and Shaw wanted to give money a richer place in a general theory of finance; they were received as the first step in draining banks of specialness. Black wanted to show mainstream monetary theory rested on an institutional artifact; his provocation became foundational to the mainstream that had declined to publish it, and his risk-splitting sketch was reread, much later, as an anticipation of derivatives and shadow banking. Fama wanted to clear the mischief introduced by the concept of money out of the theory of finance; his paper became most useful as the scaffold on which the field proved banks special after all. Each author intended a corrective or negative claim; the field kept the negative result and inverted its use, turning each into a benchmark whose value is the distance later frictions travel from it.

The era’s timing is its own indictment, and the indictment is the loose thread. Fama published the zero-fragility benchmark in 1980. Three years later Diamond and Dybvig published the run. If banks are veils, why do they keep failing in ways that take the economy with them? The veil view has an answer, and the answer is that they do not: observed failures should be regulatory artifacts or benign portfolio revaluations. Every subsequent era of this history is a documented counterexample, and the interesting question became, and remained, which friction the counterexamples select.

Era II: Banks as Liquidity (1980-1990)

The era’s answer, as one contestable claim: banks exist because pooling many depositors’ privately observed liquidity needs lets a bank sell a demandable claim worth more to an early withdrawer than the liquidation value of the illiquid asset behind it; the demand deposit is an insurance contract against being the one who needs cash early, written by no market because the need is unverifiable; and the same first-come-first-served claim that manufactures the liquidity is what makes the bank runnable. Existence and fragility in a single instrument. A reader can attack every joint of that sentence, and the era’s own papers supply the attacks.

Diamond and Dybvig (1983) is the paper this history treats the way histories of other fields treat their founding theorems, so its machinery deserves the space. Three dates; a technology that pays R greater than one at date two but only returns the initial investment if interrupted at date one; consumers who learn privately at date one whether they are type 1, who must consume now, or type 2, who can wait (Diamond and Dybvig, 1983, p. 405). Autarky wastes the pooling opportunity: ex ante, everyone wants insurance against being type 1, and no market can provide it because type is private. A bank can. It offers a demand deposit paying early withdrawers more than liquidation value, funded by the pooled investment, and the contract implements full-information optimal risk sharing as a Nash equilibrium (Diamond and Dybvig, 1983). Then the theorem turns around and bites. Under the sequential service constraint, withdrawers are paid first-come, first-served, and the bank cannot condition early payouts on the total volume of withdrawals (Diamond and Dybvig, 1983, p. 408). So if a type 2 depositor expects the others to withdraw, the assets will be liquidated at a loss and there will be nothing left at date two; the best response is to join the line. A second Nash equilibrium, the run, sits beside the good one, self-fulfilling, requiring no bad news about the assets, no information at all beyond the expectation of the run itself (Diamond and Dybvig, 1983). The instrument that creates the liquidity creates the fragility; neither exists without the other. The remedies are then ranked with the same precision. Suspension of convertibility works when the fraction of type 1 consumers is known, and fails when it is stochastic, because a fixed suspension threshold strands genuine early consumers. Government deposit insurance financed by an optimal tax does better: Proposition 2 makes the good outcome the unique dominant-strategy equilibrium, because the state’s power to tax after the fact detaches asset liquidation from the withdrawal queue in a way no private bank bound by sequential service can manage (Diamond and Dybvig, 1983, p. 414). Priority deserves its sentence: Bryant (1980) built a run model first, with deposits backed by risky assets and an information-triggered run in which agents who learn bad news are indistinguishable from genuine early consumers; Diamond-Dybvig stripped the information away and showed the run survives as pure coordination failure (Bryant, 1980; Diamond and Dybvig, 1983). Bryant himself treated the run as a by-product, a step in a monetary-theory exercise about reserves and insurance, and disclaimed his early-consumers device as a convenience; the field made the by-product his claim to founding, a reversal performed on a paper that had asked for the opposite emphasis (Bryant, 1980). The two runs are different mechanisms, one triggered by news and one by expectation, and the field would eventually need both.

What Diamond and Dybvig thought they were doing and what the field did with it diverge in the way this history has learned to expect. They presented a framework, the first explicit equilibrium account of the demand for liquidity and the transformation service, built to analyze suspension of convertibility and deposit insurance (Diamond and Dybvig, 1983). The field made the framework the definition of the object: “Diamond-Dybvig” became what banking is, shorthand for run-prone maturity transformation, and the model’s conclusion that fragility is intrinsic to the demandable claim, rather than a symptom of bad assets, became the profession’s default reading of every panic since, relocated to wholesale funding in 2008 and to uninsured deposits in 2023 (Diamond and Dybvig, 1983). A framework offered for analyzing two policy instruments became the lens through which the field sees the institution itself.

The era’s first dispute was already on the table before the model existed, which is the era’s oddest chronological fact. Kareken and Wallace (1978) analyzed deposit insurance five years before Diamond-Dybvig gave insurance its rationale, and their result is the shadow over everything that followed: under a flat insurance premium independent of portfolio risk, the value-maximizing bank chooses the riskiest feasible portfolio, setting every bankruptcy-state payoff to its minimum (Kareken and Wallace, 1978, pp. 24-25). A flat guarantee works exactly like uninformed creditors; both let the bank raise risk for free. Their conclusion, that regulation is not an alternative to deposit insurance but a necessary complement, reads today like a prophecy, and it is worth recording what the prophecy was embedded in: the paper’s own headline was the opposite, a complete-information argument that uninsured bank liabilities are already safe and insurance unnecessary, veil-era logic, with the maximal-risk result offered as a secondary if-you-insist analysis (Kareken and Wallace, 1978). The field discarded the headline and canonized the footnote, a reversal so complete that the paper is now cited almost exclusively for the result its authors ranked second, and the era’s own dispute owes its terms to a reading the authors did not choose. Diamond and Dybvig, for their part, conceded the point in advance on their page 417: with risky assets and unobserved portfolio choice, insurance creates a moral-hazard trade-off requiring regulation alongside it (Diamond and Dybvig, 1983, p. 417). The two sides of the deposit-insurance dispute are thus not strangers; each names the other’s result inside its own paper. They disagree about which margin binds.

The 1980s answered by experiment, and the experiment is the era’s crisis-character: the savings and loan collapse. The thrifts were insured depositories whose asset side carried exactly the risks the Diamond-Dybvig model had set aside, duration mismatch above all, and whose regulators responded to insolvency with forbearance rather than closure. Kareken and Wallace’s mechanism, mispriced insurance plus weak asset-side discipline, played out item for item: deregulation without added examination resources, reduced capital requirements, delays in closing insolvent institutions (Curry and Shibut, 2000, p. 27). The bill, in the definitive FDIC accounting: 1,043 failed thrifts over 1986-1995, $152.9 billion in total resolution costs, of which taxpayers bore $123.8 billion, 81 percent, and the industry $29.1 billion (Curry and Shibut, 2000, Table 4). The crisis’s anatomy bears retelling because every piece was in the theory before it was in the newspapers. The thrifts’ original wound was interest-rate risk, long fixed-rate mortgages funded by short deposits, inflicted when rates rose at the decade’s turn; the policy response, deregulating the asset side while insurance stayed flat-rate and examination budgets stayed flat, handed insolvent institutions exactly the gamble-for-resurrection option Kareken and Wallace had formalized; and forbearance, the delay in closing institutions already dead, let the losses compound at the insurer’s expense (Curry and Shibut, 2000, p. 27). Note what the S&L collapse was not: it was not a sequential-service run. Insured depositors do not line up. It was the other fragility, the one manufactured by the run’s own cure, and the cross-country evidence later generalized the lesson: in a 61-country panel over 1980-97, 898 country-years containing forty systemic crises, explicit deposit insurance raises the probability of a banking crisis, the effect monotone in coverage, sharpest where interest rates had been liberalized, offset only by strong institutions, and robust to a two-stage design addressing the worry that crisis-prone countries adopt insurance rather than the reverse (Demirgüç-Kunt and Detragiache, 2002). The era’s scoreboard on its own remedy is therefore double-edged, and honestly so: the run-prevention benefit is real and proved; the moral hazard is real and priced at $124 billion; and the action turned out to live exactly where Diamond-Dybvig’s page 417 said it would.

The era also staged the argument about whether any of this apparatus is necessary, and the free-banking dispute deserves both its sides. White (1984) is the strongest case that the fragility is an artifact of regulation: Scottish banking from roughly 1716 to 1845 ran with no central bank, no monetary policy, and near-open note issue, and it was stable. The average annual bank-failure rate over 1809-30 was 18.1 per thousand in England against 7.8 in Scotland; total public losses from all Scottish failures to 1841 came to about £32,000; when the reckless Ayr Bank over-issued and collapsed in 1772, the note exchange returned its paper promptly, its 241 unlimited-liability shareholders absorbed the entire loss, and note holders were paid in full (White, 1984). The stabilizing machinery was specific: the note-exchange system returned an over-issuer’s paper within days and drained its reserves, an automatic and private discipline; unlimited shareholder liability put the owners’ entire fortunes behind the notes; nationwide branching diversified each bank; and every bank held its own reserve, each tub on its own bottom, so no single reserve-holder existed for a panic to besiege (White, 1984). White reads Diamond-Dybvig run-proneness explicitly as a product of English and American restrictions, not of fractional reserves. The record requires the caveats: White himself revised his failure-rate figures upward across editions after critics showed the original comparison biased in Scotland’s favor, and Scotland was not literally free (the option clause and small notes were banned in 1765) (White, 1984). Gorton (1996) is usually cited as the other side, and what it actually shows is subtler than the wildcat caricature. In the American Free Banking Era, notes of 1,673 new banks traded at an excess discount that was positive (mean 0.0258, t = 9.56), declined as markets learned about issuers (from 0.0697 before 1845 to 0.0080 after 1850), and was a fair lemons premium: portfolios of new notes earned returns indistinguishable from seasoned ones (Gorton, 1996). The market disciplined private money; the discipline device was redemption itself, a run aimed at any doubted issuer. What free banking never delivered was par: private notes traded at discounts that varied by issuer and distance, and a monetary instrument that requires a price sheet is failing at being money (Gorton, 1996). Discipline without par is what the American record supports; run-proofness is what the Scottish record claims; and the two are answers about different countries to different questions, which is why partisans of each can cite their evidence indefinitely without contact. A Tennessee bank’s note quoted at a twenty percent discount in Philadelphia is not a system in chaos; it is also not money, if money means the thing you do not have to price before accepting. The dispute stands unresolved on the facts, and its durable residue is the observation that private money can be disciplined without a central bank and still fail the test that matters, trading at par with no questions asked. That residue returns, verbatim in spirit, in the last era of this history.

Where is Bagehot in the era that finally modeled his panic? At the edges, being argued past. Diamond and Dybvig list the lender of last resort among the remedies their framework can analyze, then rank tax-backed deposit insurance above it. White turns the rule against its author: the lender-of-last-resort case was addressed to the Bank of England as sole reserve-holder in a one-reserve system, so under free banking the rationale vanishes, the need being created by the very legislation that created the central bank (White, 1984). Bagehot, who preferred the many-reserve system and defended the rule as a second best, would have recognized the argument; it is his own, aimed back at him. The era closes with its founding abstraction exposed as the loose thread. The Diamond-Dybvig technology is riskless: illiquidity is the only friction, and there is no loan anybody needs skill to value (Diamond and Dybvig, 1983, p. 405). Real banks hold loans nobody else can value, and the era’s own papers kept snagging on the difference; Bryant had built his run on risky assets from the start, Kareken-Wallace’s moral hazard lives entirely on the asset side, and the S&L crisis was an asset-side event. A theory of banking whose bank could be replaced by a bond fund with a queue was not yet a theory of banking. The gap was not pedantry. Everything distinctive about a real bank’s balance sheet, why its assets cannot simply be sold to meet withdrawals, why outsiders cannot price them, why the run is destructive rather than merely redistributive, lives in the asset the model omitted. The next era supplied it.

Era III: Banks as Monitors (1984-2002)

The era’s answer, stated so it can be fought: banks exist because monitoring borrowers is costly, and delegating the monitoring to a single diversified intermediary economizes on it; and precisely because the intermediary alone holds the borrower-specific information, its assets are opaque by construction, so its own liabilities must be information-insensitive debt, which is exactly the claim that runs when the opaque assets are doubted. The existence answer and the fragility are welded even more tightly than in the liquidity era: the monitor’s product is private information, and private information is what a panic is about.

The empirical case arrived before the theory, which is the era’s first oddity. Bernanke (1983) is a Depression paper written against the era’s inherited account. Friedman and Schwartz (1963) had established the money view with the most influential decomposition in monetary history: from August 1929 to March 1933 the money stock fell by more than a third, more than triple the largest prior decline in their near-century of data, and the fall was driven almost entirely by the two deposit ratios, the public fleeing deposits for currency and banks hoarding reserves against runs, while high-powered money actually moved the other way (Friedman and Schwartz, 1963, p. 334). The panics came in waves, four of them by their count, each converting more depositors into currency-holders and more banks into reserve-hoarders, and on their account the panics mattered because of what that conversion did to the money stock, not because of anything intrinsic to the failing banks. The collapse was not the inescapable consequence of other forces; it was a largely independent contraction the Federal Reserve had the tools to stop and did not, and the cure was the one Bagehot had written sixty years earlier, a lender of last resort supplying reserves freely, which the System conspicuously failed to be (Friedman and Schwartz, 1963). The money view’s implicit model of the bank, note, is liquidity transformation, deposits and reserves; it sits in this era not by its own theory but because the question it fought over, what exactly do bank failures do to an economy, is the question the monitoring era answered (Friedman and Schwartz, 1963). Bernanke accepted the channel and showed it was too small. In dynamic simulation, money-only equations capture at most half of the 1930-33 output decline; adding the real deposits of failing banks and the real liabilities of failing firms cuts the unexplained error by roughly half, and his yield-differential proxy for intermediation cost cuts it by ninety to ninety-five percent, results that survive downweighting the March 1933 bank holiday and adding free dummies for 1931-32 (Bernanke, 1983). His mechanism is the one the era would spend two decades formalizing: banks specialize in lending to households and small firms whose claims trade in no public market, and a failed bank’s accumulated knowledge of those borrowers dies with it. The failures raised what Bernanke named the “cost of credit intermediation,” rationing exactly the borrowers with no other channel, and the damage was informational, slow to repair because borrower-specific capital cannot be rebuilt by printing reserves (Bernanke, 1983). The concept had, by his own note, no exact antecedent in the formal literature; the theory that would rationalize it arrived a year later (Bernanke, 1983). The dispute between the money view and the credit view was framed by its own younger side as complementary rather than rival, and it remains unresolved in that form. Its resolution came, if anywhere, outside the literature: the author of the credit view spent 2008 running the Federal Reserve, executing at scale the policy his 1983 paper implied. A Depression paper became the intellectual license for a rescue, written by the person who performed it.

Diamond (1984) turned Bernanke’s prose concept into a primitive one year later. The problem: lending requires monitoring that costs K per borrower, and m small lenders either duplicate it, paying mK, or free-ride and nobody monitors. Delegating to one intermediary economizes, but who monitors the monitor? Diamond’s answer is the diversification theorem, the era’s central result: as the intermediary adds independent monitored projects, the cost of disciplining the delegate, per project, goes to zero, because a diversified portfolio makes the intermediary’s aggregate performance informative about its diligence, and the discipline can be automated by a debt contract with a nonpecuniary penalty in default (Diamond, 1984, Prop. 2). Three things follow that organized twenty years of work. The intermediary is highly levered, financed by debt, because debt is what makes delegation incentive-compatible. Its loans are illiquid and “special” because only it holds the borrower information. And, the step everyone now reads into the paper, its debt must be information-insensitive. Read the paper’s own scope honestly: the model is one-period and risk-neutral, there is no sequential service and no run, and Diamond himself flagged demandable deposits and runs as a separate matter handled elsewhere (Diamond, 1984). Runnability is the field’s extension, not the theorem’s content; the paper the field cites for why bank debt runs contains no run. Ramakrishnan and Thakor (1984), in the same volume of the same journal, built the mirror-image machine, ex-ante certification rather than ex-post verification, and reached the same limit by the same engine, diversification within the coalition driving incentive costs to zero as N grows (Ramakrishnan and Thakor, 1984). One engine, two masks; a paper about rating agencies and underwriters was reread as a paper about banks, and the pair became the twin foundations of “banks as informed agents.”

The era now had two complete existence theories standing side by side, and their rivalry deserves to be stated as the field experienced it. The liquidity answer needs no credit risk: Diamond-Dybvig’s asset is riskless, and the bank exists to insure withdrawal timing. The monitoring answer needs no liquidity: Diamond’s model has no early consumers, no sequential service, and the bank exists to economize on verification. Each theory generates a bank the other does not require, and each pairs its bank with a different fragility, the run against the opacity. For two decades the field carried both, citing whichever fit the application, and the honest scoreboard is that no reduction of one to the other was ever produced; what was eventually produced, at the era’s boundary, was a reason one institution should do both jobs at once, which is a different and humbler thing.

Holmström and Tirole (1997) closed the hole the delegation step had left. Diamond’s diversification disciplines the monitor statistically; Holmström-Tirole make the monitor post a bond. In their double-decker moral-hazard model, monitoring is unverifiable, so it is credible only if the intermediary injects its own capital into each firm it monitors, and intermediary capital becomes an aggregate state variable in its own right (Holmstrom and Tirole, 1997). Firms sort into three regimes by net worth: the well-capitalized borrow directly from markets with no monitor; a middle tier borrows through intermediaries, paying for monitoring because their own capital cannot carry a direct claim; the poorest cannot borrow at all. The comparative statics that made the paper canonical after 2008 fall out immediately: a shock to firm collateral squeezes the bottom of the distribution, and a shock to intermediary capital squeezes the bank-dependent middle, two different crises with two different footprints, distinguishable in the data by who loses credit (Holmstrom and Tirole, 1997). The paper called itself a simple model of credit and matched itself to Scandinavian banking distress and the American 1990-91 credit crunch; the field made bank net worth a macroeconomic state variable and the model the standing microfoundation for the argument against procyclical capital rules (Holmstrom and Tirole, 1997). The distance between a simple model of credit and the analytical core of post-2008 macro-finance is the era’s reversal pattern in one example. The existence answer now had teeth: banks are monitors whose own capital makes the monitoring credible, and the implied fragility migrates from the depositor’s psychology to the intermediary’s balance sheet.

The evidence that markets price all this exists, though the claim is narrow. James (1987) found that announcements of bank loans move borrowers’ stock prices upward, where announcements of public debt and private placements move them not at all or down: the market pays for the bank’s approval specifically, treating the credit decision of the informed lender as news in a way an arm’s-length investor’s money is not, which is the uniqueness of bank loans the era’s theory predicted (James, 1987). Petersen and Rajan (1994) found that lending relationships raise credit availability more than they lower its price, with multiple-lender borrowing raising price and cutting availability (Petersen and Rajan, 1994). The shared point is narrow but important: the market and the lending contract both behave as if the bank’s private information and relationship capital are priced objects, not institutional decoration.

The era’s sharpest turn is the dispute over what fragility is for. Era II treated the run as the price of liquidity insurance, a pathology to be engineered away by deposit insurance. Two papers inverted the sign. Calomiris and Kahn (1991) set liquidity demand to zero, and demandable debt is still optimal: the withdrawal option is a monitoring device, and sequential service, the very feature Diamond-Dybvig identified as the run’s enabling constraint, is repurposed as a reward schedule. Depositors who pay the cost of monitoring learn the bank’s condition first, withdraw first, and collect in full; the uninformed, later in line, bear the shortfall; and the banker, knowing informed money watches the door, does not abscond (Calomiris and Kahn, 1991). A withdrawal is a vote of no confidence, and the run is the election. The authors were explicit that this is a theory of normal-times discipline and not of systemic panics, a scope restriction the field acknowledged and then ignored (Calomiris and Kahn, 1991). Diamond and Rajan (2001) pushed the inversion to its strongest form. Their relationship banker has a collection skill nobody else has, so her loans fetch only a fraction of their value in any sale; the fragile deposit-funded capital structure is what lets her commit that skill, because any attempt to renegotiate rents out of depositors triggers a run that destroys her, and it is exactly this hair-trigger that raises the pledgeable value of the loans to their full worth (Diamond and Rajan, 2001). Financial fragility allows liquidity creation; insulate the bank, by full insurance or by heavy capital, and you do not make it safer, you make it unable to commit, no better than an all-equity fund; narrow banking, they argue against Simons’s tradition by name, would kill liquidity creation (Diamond and Rajan, 2001). The honest scoreboard is that this branch supplies the design-feature side of the dispute with unusual force. Its direct descendant is the capital-requirement war two decades later: if fragility is the price of liquidity creation, then capital is not free, and that “if” became the most expensive conditional in bank regulation.

Kashyap, Rajan, and Stein (2002) closed the era by asking why one institution does both of the things the two eras had separately explained, and the answer is a synergy, not a synthesis. Deposits and loan commitments are both promises of liquidity on demand, one to households and one to firms, and both require holding a costly buffer of liquid assets against the day the promise is called; a bank doing both holds a smaller combined buffer than two specialists would hold separately, provided deposit withdrawals and commitment takedowns are imperfectly correlated, so the institution that puzzled two eras, lender and deposit-taker under one roof, is an exercise in buffer-sharing (Kashyap et al., 2002). The institutional fact motivating the paper is that roughly seventy percent of banks’ small-business lending runs through lines of credit, promises of future liquidity rather than spot loans, so the lending side is itself a liquidity business (Kashyap et al., 2002). The empirical anchor: across banks, the commitment ratio rises with the transactions-deposit ratio, coefficient 0.116 with a t-statistic of 20.21 in the full sample, positive in every size class (Kashyap et al., 2002, Table IV). Be precise about what the paper claims, because the field has been careless: it explains the coexistence of the two products; it does not reduce monitoring to liquidity or liquidity to monitoring, and it concedes it models one synergy among possibly many (Kashyap et al., 2002). The condition in the proviso is not decoration. The synergy holds while the two liquidity demands stay imperfectly correlated, and the paper names the state of the world where they do not, both sides drawing at once, which a later generation would meet in March 2020 and call the dash for cash.

Every era of this history reverses its papers, but Era III’s pattern is the cleanest: every author wrote a bounded, scope-limited contribution, a missing step, a complement to the money view, an explicitly not-systemic theory of normal times, a simple model of credit, and the field promoted each into a pillar of a general theory of banking fragility none of them set out to build. Diamond’s run-free model became the microfoundation of runnable debt; Calomiris-Kahn’s normal-times discipline became a theory of panics; Bernanke’s complement became a license. And the era’s loose thread is the most ironic in this history. Its crowning objects, the relationship lender whose value is her untransferable skill, the deposit-commitment synergy bank, were perfected between 1997 and 2002, exactly as the system they described was ceasing to be the system. KRS’s own pages concede that in Diamond’s model it is hard to see why the lender must be a deposit-funded bank rather than a finance company rolling short-term paper, and report that finance companies already did half of small-business lending (Kashyap et al., 2002). Holmström-Tirole’s own sorting sends every well-capitalized firm straight to the markets, no bank required (Holmstrom and Tirole, 1997). The theory perfected the deposit-funded relationship lender just as the actual system stopped being one. Where the functions went is the next era, and Bagehot’s rule, which in this era appears only once, as Friedman and Schwartz’s indictment of the Fed for not using it, would have to wait there too, because the balance sheets the functions moved to had no discount window to lend at any rate.

Era IV: Banks Escape the Model (1990-2007)

The era never states its existence answer, which is the point of calling it one: by the 1990s the field’s implicit answer to “why do banks exist” had become “because they hold charters and sit inside a safety net.” The productive functions, liquidity manufacture, safe-asset production, monitoring, risk sharing, were understood to be migrating to markets, so what distinguished a bank was a legal wrapper. The implied fragility follows with a logic nobody drew at the time: a system defined by charter rather than function cannot see risk that has left the charter, so the safety net keeps protecting institutions that matter less each year while the run-prone activity accumulates where no insurance reaches and no discount window opens. A reader can contest this from inside the era’s own record, on both fronts, and should; the era is unusual in this history for having its correction and its complacency published side by side, by the same journals, in the same years.

The first contest is about whether the migration even happened. Boyd and Gertler (1994) took apart the “decline of banking” statistics and found a measurement artifact: count off-balance-sheet credit equivalents and underreported offshore lending, and banks’ share of intermediation was roughly stable, averaging a bit above 40 percent across four decades (Boyd and Gertler, 1994). Noninterest income more than doubled, from 0.70 percent of assets in 1979 to 1.87 percent in 1992, precisely as the on-balance-sheet share fell from 45 to 34 percent; off-balance-sheet credit equivalents reached roughly 30 percent of on-balance-sheet loans by 1990-91; properly counted foreign-bank activity, with some $175 billion of unreported offshore commercial loans, took the adjusted foreign share to 21 percent (Boyd and Gertler, 1994). A backup line to a commercial-paper issuer, they argued, is intermediation not fundamentally different from making the loan (Boyd and Gertler, 1994). Banking was transforming, not shrinking. Both halves of that sentence proved load-bearing. The share was stable, and the comfort drawn from stability was inverted by events: what Boyd-Gertler measured as reorganization was the leading edge of the migration of run-prone claims to uninsured venues, and their own passing observation that steep-yield-curve spread income was building interest-rate exposure reads, three decades later, like a footnote addressed to Silicon Valley Bank.

The second contest is theoretical, and it had been written before the era began. Gorton and Pennacchi (1990) is the era’s deepest paper and its most reversed. Its stated subject is security design. Uninformed agents systematically lose to informed traders in any market where the security’s value depends on private information; the fix is not to inform them, which is impossible, but to give them a claim whose value does not depend on what anyone knows: split the cash flows into an information-insensitive debt claim for the uninformed, safe so long as asset returns cover the promised payment, and an equity residual that concentrates the information sensitivity on those equipped to bear it (Gorton and Pennacchi, 1990). Trading losses to insiders are a tax on transacting; the debt claim is the exemption; and the institution that manufactures the exemption at scale is a bank, which is why deposits look the way they do (Gorton and Pennacchi, 1990). Deposits are the archetype, and the authors note in passing that money market fund shares backed by commercial paper could serve as well. Read in 1990, this was a theory of why intermediaries exist. Read after 2008, it is a description of the shadow money system, written fifteen years before the machinery it describes was assembled, by authors who explicitly did not model a run; the mechanism they built, safety that survives only while no one asks questions, is the run, waiting for a questioner (Gorton and Pennacchi, 1990). The paper thought it was microfounding deposit insurance. It was drafting the repo market’s obituary in advance.

The era’s public dispute was about system architecture, and its verdict was a null. Allen and Gale’s Comparing Financial Systems (2000) argued that bank-based and market-based systems excel at different functions, that neither dominates, and, in the book’s signature mechanism, that long-lived intermediaries can smooth nondiversifiable risk across generations in a way no cross-sectional market can, so disintermediation is not costless: the smoothing franchise unravels the moment a lucky generation can defect to spot-market returns (Allen and Gale, 2000). Levine (2002) supplied the scoreboard: across 48 countries over 1980-95, no measure of financial structure, bank-based or market-based, relates significantly to growth at even the ten percent level, and the null survives every robustness check the paper runs, instrumenting with legal origin, comparing the extreme bank-heavy against the extreme market-heavy tails, re-measuring structure every way available (Levine, 2002). Overall financial development, by contrast, enters strongly, a coefficient implying that raising Peru’s financial activity to Chile’s level would add on the order of a full percentage point of annual per-capita growth, and the legally determined component of development passes the overidentification tests, favoring the view that what finance does, and what law lets it do, matters, while the institutional form doing it does not (Levine, 2002). Structure is the wrong axis; development and law are the levers. The verdict retired the bank-versus-market question for growth, and it was correct, and its correctness is the era’s blind spot in miniature: a structure-neutral growth regression is built so as not to see the fragility that lives in the form of the claim. Whether a service is produced by an insured deposit or an uninsured repo does not move a growth regression; it decides whether there is a queue outside the building in a panic. The growth question got answered while the fragility question went unasked. Allen and Gale’s other 2000 paper, “Financial Contagion,” belongs intellectually to the liquidity era: banks in different regions hold deposits in one another as insurance against imperfectly correlated liquidity demands, and the insurance network doubles as the contagion channel, because cross-holdings can redistribute liquidity but cannot create it, so a small aggregate shortage travels the network and a complete network absorbs what a ring amplifies (Allen and Gale, 2000). The first-best works until the state arrives that the interbank market cannot insure, at which point the connections that were mutual insurance become the transmission lines. It appears in this era because of what its timing says. The field’s most polished model of fragility in the chartered interbank system was published in 2000, at the exact moment the actual fragility was leaving the chartered system.

One paper pointed at the migration and called it dangerous. Rajan (2005), at Jackson Hole, argued that financial development had made the world riskier, and the argument was incentive-theoretic rather than nostalgic. Investment managers paid on performance relative to peers carry a convex claim on outcomes: writing deep out-of-the-money insurance, hidden tail risk, books steady fees against a loss that arrives rarely and all at once, and relative evaluation makes managers herd into whatever their benchmark holds (Rajan, 2005). Securitization moves credit risk to these markets, but the originating banks keep the most complicated and volatile slice of what they created, and the whole arrangement is tied back to the banking system through committed liquidity lines that will be drawn at precisely the moment markets seize, so bank balance sheets “may not be any safer than in the past” (Rajan, 2005). The liquidity backstop evaporates exactly when it is needed, because the backstopping banks’ own funding is suspect in the same state of the world. The reception is part of the era’s record: the discussion was dismissive, with Summers reported to have called the paper’s premise slightly Luddite and largely misguided (Rajan, 2005). Received as alarmism in 2005; canonized after 2008 as the warning that named the mechanism. The era-level reversal pattern is uniform, and Rajan is its exception that defines it: each of the era’s papers correctly described a static object, a safe claim, a stable share, a structure-neutral growth relation, and was received as settling a question, and each settled question was the wrong one, because the object was migrating and the migration carried the fragility the models were not built to see.

The era’s papers reversed with a uniformity that suggests the reversal was structural rather than accidental. Gorton-Pennacchi microfounded deposit insurance and were reread as the theory of the uninsured system. Boyd-Gertler corrected a statistic and were reread as having documented, reassuringly, the buildup they could not price. Levine retired a dichotomy for growth and was complicated by a crisis his regression was not built to see. Each paper described its static object correctly; each settled question was the wrong one; and the field’s capacity for polished answers grew in exact proportion to the questions’ irrelevance. When a literature’s craft improves while its subject escapes, the craft becomes the evidence: the models of this era are the best-documented record of where the field was looking, which is how we know where it was not.

So the loose thread, stated as the era should have stated it: liquidity transformation had not disappeared, it had moved to balance sheets with no deposit insurance and no discount window. Bagehot’s rule does not appear in this era, and for a reason more damning than neglect: the rule presupposes a lender of last resort standing behind the demandable claim, and the demandable claims were accumulating in venues the lender could not reach. The era did not violate the rule. It built a system the rule could not see. The run, when it returned, would have to be met by improvisation.

Era V: The Run Returns (2008-present)

The era does not offer one existence answer; it offers three, and its arc is that each was proposed as a resolution and each turned out to name a new way to run. Stated as one claim a reader can refuse: banks exist to manufacture money-like safe claims from a stable, inattentive funding base, and the same inattention that makes the claim money-like is the fragility, so every version of the answer manufactures its own run. In 2008 the funding base was wholesale and the inattention was lenders not questioning collateral until they did. After the crisis the answer relocated to the insured deposit base and the sleepy depositor. In March 2023 the second answer produced its own run, on schedule. The reader’s available refusal is that fragility is not intrinsic to the answer, and the era’s live disputes are exactly about that refusal.

The crisis as Diamond-Dybvig at institutional scale. Gorton and Metrick (2012) is the era’s empirical keystone, and its central object is a number series nobody had thought to treat as a monetary aggregate: the repo haircut. In “securitized banking,” firms funded portfolios of securitized bonds with repurchase agreements, collateralized, overnight, repriced daily, the institutional-scale equivalent of a demand deposit, held by depositors, money funds, corporate treasurers, other dealers, for whom insured accounts were never an option because insurance caps out six orders of magnitude below their balances. The run took the form of haircuts: an equally weighted index across collateral classes rose from approximately zero in early 2007 to 45 percent at the panic’s peak, with some collateral classes at 100 percent, meaning the market for them had simply closed (Gorton and Metrick, 2012). A haircut is a withdrawal. Raising it from zero to 45 percent on a dealer funding itself overnight removes nearly half its funding in a day, and no depositor need appear at any window. The identification points at counterparty fear rather than fundamentals: spreads on securitized assets far from housing move with LIB-OIS, the interbank counterparty-risk gauge, and not with the ABX subprime index (Gorton and Metrick, 2012). Brunnermeier (2009) had mapped the amplification in real time, cataloguing the channels while they operated: the loss spiral, where falling asset prices erode capital and force sales; the margin spiral, where rising haircuts do the same through funding; runs on the off-balance-sheet vehicles that had warehoused securitized assets; and the counterparty fear that made every dealer hoard against every other. Through those channels a shock of a few hundred billion dollars in subprime losses became roughly eight trillion in lost equity value (Brunnermeier, 2009). Adrian and Shin (2010) supplied the balance-sheet mechanics from data assembled as the system failed; their paper began as a staff report written, in their words, as a contemporaneous record of the last months of the once illustrious Wall Street investment banks. The finding: broker-dealer leverage is procyclical, rising as balance sheets expand, the adjustment running through repo, so that marked-to-market balance sheets turn price moves into funding moves and back into price moves (Adrian and Shin, 2010). The dispute the era could not close is whether this was a liquidity run on a solvent system or a solvency crisis wearing a run’s clothes, and the honest answer starts from the lead paper’s own text: Gorton-Metrick’s abstract calls the banking system of 2008 “effectively insolvent,” describes its evidence as correlation rather than causation, and its authors later conceded that the bilateral repo run they documented was smaller in aggregate than the tri-party and money-fund channels their critics emphasized, critics whose objections became part of the later reception (Gorton and Metrick, 2012). The record supports a run mechanism operating on a system that was also plausibly insolvent, and no cleaner verdict; the two readings were not separable in 2008 and are not separable now. What is beyond dispute is what the run’s return did to a rule written in 1873. Runs on traditional banks had ended, Gorton-Metrick note, with the discount window and deposit insurance; the 2008 run happened where neither reached, and the response, lending freely against collateral at facility after improvised facility, was Bagehot’s rule executed at institutional scale by a central bank rediscovering, at the securitized-banking perimeter, the practical wisdom the theory eras had argued past (Gorton and Metrick, 2012). The rule returned intact. The perimeter had moved.

What banks turned out to be. The post-crisis decade rebuilt the existence answer from the ground up, and the rebuilding began with the demand side. Krishnamurthy and Vissing-Jorgensen (2012) priced what the shadow system had been manufacturing: Treasuries carry a convenience yield averaging 73 basis points a year, decomposable into a liquidity attribute and a safety attribute, and it behaves like a demand curve, the premium widening when Treasury supply is scarce and compressing when it is abundant (Krishnamurthy and Vissing-Jorgensen, 2012). Money-like safety is a priced good with a downward-sloping demand, which means a supply gap is an arbitrage: whenever the public sector under-issues safety, a spread stands open for anyone who can manufacture a claim the market will treat as safe. A priced premium is an invitation, and private issuers accepted it; the securitization machine was the market’s answer to a safe-asset shortage that had been widening for a decade, emerging-market reserve accumulation and regulation and demography all raising the demand for safety faster than sovereigns supplied it (Caballero et al., 2017). The tranches the machine produced were, in the phrase the record supplies, micro-AAA rather than macro-AAA, engineered against idiosyncratic risk and helpless against the tail, so the private safety unravelled exactly when safety was most demanded, absent a public backstop, and the shortage the machine had answered reopened wider than before (Caballero et al., 2017). The authors of the demand curve thought they were writing asset pricing, not a theory of financial institutions; the field decided they had priced the product banks exist to make, which makes the paper a theory of financial institutions whether it wanted to be or not (Krishnamurthy and Vissing-Jorgensen, 2012).

On the chartered side, the rebuilt answer came in two layers. Hanson, Shleifer, Stein, and Vishny (2015) described banks as patient fixed-income investors: funded by stable, sleepy, insured deposits, banks can hold illiquid assets with low fundamental risk to maturity and ignore transitory repricing that would force a market-funded holder to sell; the bank’s two sides are complements, stable funding purchasing the right to be patient (Hanson et al., 2015). The model’s comparative statics say where such a bank will live: in assets more illiquid than a market fund would hold and less fundamentally risky, agency mortgage securities rather than either Treasuries or credit card receivables, because patience is valuable precisely where prices swing more than values (Hanson et al., 2015). Its authors presented the empirics as a synthesis of known facts rather than a test, a modesty the field overrode by making the paper the foundational statement of the modern view, which means the era’s load-bearing existence answer rests, by its own authors’ description, on a calibration (Hanson et al., 2015). Drechsler, Savov, and Schnabl then gave the funding side a name and a price. The deposit franchise: banks hold market power over sticky depositors, and when the policy rate rises they widen the deposit spread rather than pass the rate through, a spread beta of 0.54 on average. The identification is the paper’s quiet triumph: comparing branches of the same bank across counties of different concentration, so that everything about the bank differences out, the high-concentration branches raise spreads by more and shed more deposits when the Fed hikes, market power measured at the branch level (Drechsler et al., 2017). The paper’s own subject was monetary policy: when the Fed hikes, banks widen spreads, depositors slowly leave, and deposits, the cheapest funding there is, drain out of the system, contracting lending without any reserve requirement in the mechanism, a lending channel rebuilt on market power (Drechsler et al., 2017). The field took the mechanism and enlarged the object: the franchise is the modern answer to what a bank is, and it comes with the era’s new existence-fragility pairing, which should be stated with the same weight as liquidity-implies-runs: the deposit franchise implies interest rate risk. The benign reading came first in logic though not in print (the two papers were developed together, the hedge circulating before the channel paper appeared): the franchise is a negative-duration asset, deposit-spread income rising with rates, and it hedges the bank’s long-duration assets so well that a measured duration mismatch of three and a half years produces equity losses of roughly 4 percent, not the textbook 34, when rates rise a point; income and expense betas match nearly one for one, and net interest margins have held a band of 2.2 to 3.8 percent since 1955 against short rates that ranged from zero to sixteen (Drechsler et al., 2021). Maturity transformation without interest rate risk. It is the most reassuring sentence the field produced in a century, and its own authors flagged the condition: the hedge is the franchise, the franchise is depositor stickiness plus market power, and the Regulation Q deregulation had already once unhedged an industry and left the thrifts insolvent (Drechsler et al., 2021).

Silicon Valley Bank tested the condition in March 2023, and the test ran at a speed no earlier run had rehearsed: the deposits left by wire and by phone, tens of billions in a day, no queue outside any building, the sequential-service constraint executed in software. The era’s live dispute is the test’s interpretation, staged quantitatively in the record. Jiang, Matvos, Piskorski, and Seru (2024) marked the banking system’s assets to market after the 2022 tightening: a decline of roughly 10 percent on average, $2.2 trillion below book value, about 96 percent of aggregate pre-tightening bank capital (Jiang et al., 2024). On their arithmetic the run exposure was broad: a run by half the uninsured depositors would impair 186 banks, and SVB itself was unexceptional on the asset side, some 500 banks had larger unrecognized losses, but sat at the 99th percentile of uninsured leverage, 78 percent of assets funded by uninsured deposits, with enough liquid assets that a classic illiquidity run is ruled out; what failed was a self-fulfilling solvency run on the liability structure (Jiang et al., 2024). Drechsler, Savov, Schnabl, and Wang (2023) answer from inside the franchise theory: the franchise is itself the runnable asset. Its value rises with rates but exists only if the bank survives to collect the spread, so uninsured depositors who might flee can destroy the very asset that would have justified staying; the run equilibrium requires no fire sale and no illiquid loan, and their model produces it for a bank holding nothing but fully liquid high-quality assets, which no earlier run theory could do (Drechsler et al., 2023). Runs become likelier as rates rise, as the uninsured share rises, and as the uninsured deposits’ beta falls, each movement raising the value of the thing the run destroys (Drechsler et al., 2023). On their capitalization, which marks the liability franchise symmetrically with the assets, fewer than one percent of banks sat in the run region, essentially the ones that failed, against the aggregate franchise having risen from about zero to 11 percent of assets during the tightening (Drechsler et al., 2023). Roughly a third of the system at risk on one accounting, as the franchise paper itself states the contrast, or less than one percent on the other: the gap is not an empirical disagreement but a modeling choice about whether to mark both sides of the balance sheet, and the reconciliation the record itself points to is precise. The hazard is not uninsured deposits per se; it is low-beta uninsured deposits, funding that is simultaneously flighty in a crisis and cheap in the calm. SVB’s deposits were uninsured business balances paying almost nothing; Citigroup held comparably uninsured balances at near-market rates and was not exposed (Drechsler et al., 2023). The sleepy depositor, the load-bearing assumption of the entire post-crisis answer, had been priced as an asset, and SVB is what the price looks like when the sleeper wakes. Between the two poles stands the structural bridge: Egan, Hortaçsu, and Matvos (2017) estimated the run on modern data before 2023 supplied one. In their structural demand system, uninsured depositors respond to bank distress, demand falling with default risk with an elasticity near minus 0.6, while insured depositors do not move at all, and the asymmetry is enough to generate multiplicity: on identical March 2008 fundamentals, their model supports Wachovia at a 3.3 percent default probability and at 52 percent, two self-consistent equilibria for one bank, the depositors’ beliefs selecting between them (Egan et al., 2017). Their capital counterfactuals belong in the capital-wars ledger too: the model’s robust conclusion is that capital below roughly 18 percent leaves room for equilibria with large welfare losses, a floor far above Basel’s (Egan et al., 2017). Diamond-Dybvig’s multiplicity, estimated rather than assumed, forty years on.

The perimeter question. The era’s policy fight was over capital, and its terms were set by a theorem from another literature. Admati, DeMarzo, Hellwig, and Pfleiderer (2013) applied Modigliani-Miller to banks: requiring more equity lowers the required return on that equity, doubling capital from 10 to 20 percent roughly halves the sensitivity of returns on equity to asset returns, so the funding-cost arguments against capital confuse private costs with social ones; the tax shield, the implicit-guarantee subsidy, the debt-overhang losses are transfers, and the social cost of substantially more equity is, if it exists, very small (Admati et al., 2013). Their fire-sale arithmetic carries the fragility point: a bank at 3 percent equity must shed a third of its assets to recover from a 1 percent loss; a bank at 20 percent, about a twenty-fifth (Admati et al., 2013). And the argument connects to the era’s existence answer rather than fighting it: more equity makes the bank’s debt safer and therefore more information-insensitive, so capital serves the very money-manufacture function the fragility-as-design tradition claimed capital would kill (Admati et al., 2013). That collision, Admati’s arithmetic against Diamond-Rajan’s commitment theory, is the monitoring era’s sharpest legacy dispute restaged with a balance sheet: if fragility is the price of liquidity creation, capital is expensive; if the private costs of equity are transfers, the price was always an accounting illusion. Kashyap, Stein, and Hanson (2010) occupied the middle ground with numbers: conceding the MM logic for the long run, each percentage point of additional capital raises loan rates by 2.5 to 3.5 basis points in steady state, 25 to 45 for a ten-point increase, small; the real frictions are the transition, where requirements met by shrinking assets are contractionary, and, above all, migration (Kashyap et al., 2010). Cheap funding is often a bank’s only competitive advantage, so a capital wedge on the charter pushes the runnable activity to vehicles outside it, run-prone as ever but beyond the insurance and the window; their prescription, haircut regulation attached to the asset rather than the institution, follows from taking seriously that fragility is a property of funding structures, not of charters (Kashyap et al., 2010). That migration argument is the era’s through-line made policy. The functions had escaped once, into repo and conduits, and the era watched them escape again along both margins at once: on the asset side, private credit funds absorbing the lending that capital charges price out of the chartered bank, relationship lending reconstituted outside the deposit franchise and outside the window; on the liability side, stablecoins reperforming the oldest experiment in the book, private money issued at par by institutions that are, in the record’s phrase, essentially unregulated banks. Gorton and Zhang (2023) read stablecoins as antebellum banknotes with better typography: privately issued money that fails the no-questions-asked criterion, trades at par only until doubted, and can be redeemed to zero at run speed; their historical evidence is the free-banking record itself, notes trading at discounts that varied by issuer and distance, and their policy menu is the 1863 National Banking Act’s menu updated: of the options they canvass, only issuance from inside the insured perimeter, one-for-one backing by Treasuries or reserves, or replacement by a central bank digital currency achieves no-questions-asked money and eliminates the run; everything softer preserves the discount, and the discount is the run in slow motion (Gorton and Zhang, 2023). The innovation side’s best reply in the record is a sentence inside the same note, that on-chain transparency and composability may resolve the very opacity that made private notes information-sensitive. That possible defense is treated here as an opening rather than an evaluated counterclaim. Later run predictions belong to the paper’s reception rather than to its original evidentiary base.

Bagehot’s ledger for the era deserves its own entry before the close. In 2008 the rule was executed at the discount window and then extended beyond it, facility by facility, to the securitized perimeter where the run actually lived. In 2023 it was executed again at the franchise perimeter: the record notes that the Bank Term Funding Program and the guarantees extended over uninsured deposits likely short-circuited a broader run on the banks Jiang, Matvos, Piskorski, and Seru had marked (Jiang et al., 2024). Lending against good securities, as largely as the public asks: the 1873 formulation required only one emendation each time, an answer to the question of who counts as the public.

One more reversal completes the era’s set, because it bears on who gets believed. The paper that told the field what banks had become, patient investors funded by the sleepy, offered itself as a synthesis rather than a test; the paper that documented the run offered itself as correlation rather than cause; the paper that measured the franchise offered itself as monetary economics. The era’s load-bearing structure was assembled from materials their manufacturers had labeled provisional, and the SVB episode is what quality control looks like when it arrives late: one label, the hedge’s proviso about depositors staying put, was tested, and the structure held or failed bank by bank according to a variable, the beta of uninsured money, that no supervisor had been reporting.

The era leaves its question open because the record does not resolve it. Regulation pushed money manufacture and liquidity transformation out of the chartered bank, into repo, money funds, conduits, private credit, stablecoins. If the runnable claim can be a haircut, an uninsured deposit, or a token, then “bank” names a function rather than an institution, and the object of study is no longer the chartered balance sheet but the perimeter of the lender of last resort: who gets Bagehot’s two rules applied to them, and who gets to find out, in the panic, that they were standing outside.

Epilogue: The Balance Sheet in the Middle

Banking’s history of thought has a shape, and the shape is a circle that closes somewhere unexpected: not on either side of the bank’s balance sheet, but in the middle of it. The eras walked the perimeter first, a full lap, before the field noticed that every route home crossed the same ground.

The field began at the edges. The liquidity era was a theory of the liability side: the demandable claim, the pooling of withdrawal risk, the run as the liability’s shadow. The monitoring era was a theory of the asset side: the loan only the bank can value, the delegation of costly verification, opacity as the asset’s shadow. For two decades the two theories coexisted as rival answers to the existence question, and the field’s best effort at joining them, the shared liquidity buffer behind deposits and commitments, was a synergy conditional on the two sides not drawing at once, a proviso that named its own failure state (Kashyap et al., 2002). What the last era kept discovering, result by result, is that the fragility lives in neither edge but in the join. The patient-investor bank is a complementarity between sleepy funding and illiquid assets, valuable only as a pair (Hanson et al., 2015). The deposit-franchise hedge is a duration match, income betas against expense betas, a statement about how the two sides move together or fail to (Drechsler et al., 2021). Silicon Valley Bank failed on neither side alone; its asset losses were ordinary and its funding structure extreme, and the failure was a property of the match (Jiang et al., 2024). Even the era’s great policy dispute, capital, is an argument about the thickness of the wedge between the two sides (Admati et al., 2013). The existence answers moved from the edges to the middle, and the fragilities followed, because in this field they always travel together.

Step back further and the same convergence is visible across the neighboring fields of economics, each arriving from its own direction. Asset pricing spent decades asking whose marginal utility prices securities and ended by asking about the capital of the intermediaries who actually hold them. Corporate finance spent decades cataloguing frictions inside the borrowing firm and ended at the health of the lender in the middle. Monetary economics built transmission mechanisms of interest rates and money stocks and ended with balance sheets, bank capital and reserves and collateral, as the transmission itself. All of them, pursuing their own anomalies, converged on the same object: the levered balance sheet that stands between savers and borrowers, manufacturing safe claims on one side and holding information-dense assets on the other. Banking’s distinction is that this object was never its destination; it was the field’s home address. What the neighboring literatures reached by convergence, banking had been studying all along, badly at first, then well, and the study’s own arc, veil to liquidity to monitoring to escape to return, is the history this essay has told.

The arc has a discipline to it that a reader should not mistake for fatalism. Each era’s existence answer was true, locally and permanently: banks do manufacture liquidity, do economize on monitoring, do harvest a franchise from inattention. What failed, each time, was the inference from truth to safety, the hope that naming the function had domesticated its shadow. The pairing held instead: every function, priced, became a fragility, waiting.

Which is why the field’s oldest text reads less dated than its newest models. Bagehot had no equilibrium, no existence theorem, no run equilibrium and no franchise value; he had a description of a balance sheet in the middle, £10 million of reserve against £27 million of demandable claims, and a rule for what the holder of the ultimate reserve must do when the claims come due at once (Bagehot, 1873). Every era of theory since has either deleted his problem, formalized a piece of it, argued past his solution, or rediscovered it under a new name at a new perimeter. The rule itself returned in 2008 essentially as written, applied to institutions he could not have imagined holding claims he could not have priced, and it worked, again, at the cost of extending the reserve-holder’s reach further than any charter had contemplated. The open question the field carries forward is Bagehot’s, generalized: not why banks exist, which has been answered too many times, but where the lender of last resort’s perimeter should fall when the thing that runs no longer needs a bank to run on. A history of thought about banks ends, fittingly, unable to say precisely what a bank is, and knowing precisely what one does when it breaks.