Fragile by Construction

A bank takes deposits that can be withdrawn at any moment and lends them out for years. This is not a flaw in how banks are run. It is what a bank is.

Douglas Diamond and Philip Dybvig's 1983 model explains why the arrangement is valuable and why it is inherently unstable, and it won them a Nobel in 2022: awarded, with some irony, in a year with a banking panic in it.

The value is maturity transformation. Savers want their money available; borrowers need it committed for years. A bank bridges the gap by pooling: not everyone withdraws at once, so it can hold a fraction in reserve and lend the rest. Both sides get what they need, and the economy gets long-term investment funded by short-term saving.

The instability is that this arrangement has two equilibria.

In the good one, everyone believes the bank is sound, so nobody withdraws beyond their ordinary needs, so the reserve suffices, so the bank is sound. In the bad one, everyone believes there will be a run. Since the bank cannot pay everyone, the money goes to whoever arrives first, so joining the queue is individually rational even if you think the bank's loans are fine. The queue forms, the reserve is exhausted, and the bank fails.

Both equilibria are self-fulfilling, and the second requires no bad lending whatsoever. That is the model's real content: a perfectly sound bank can be destroyed by the expectation that it will be destroyed.

Which is why deposit insurance works the way it does. If your deposit is guaranteed, there is no advantage to being early, so the queue does not form. The guarantee typically costs nothing, because it operates by making the event it insures against not happen.

The price is moral hazard. An insured depositor has no reason to ask what the bank is doing with the money, so the market discipline that would otherwise constrain risk-taking disappears. This is why deposit insurance arrives bundled with capital requirements, liquidity rules and supervision: the insurance removes the depositors' incentive to monitor, so somebody else has to.

The other half of the apparatus was written down in 1873. Walter Bagehot's rule for a central bank facing a panic: lend freely, at a penalty rate, against good collateral.

Freely, because a liquidity panic feeds on hesitation and half-measures confirm the fear. At a penalty rate, so that banks return to private funding when the panic passes rather than settling in. Against good collateral, so the central bank rescues institutions that are illiquid rather than insolvent.

That last clause is easy to state and nearly impossible to apply in the moment. In a crisis, the market price of everything is collapsing, so an asset that is fundamentally sound looks worthless and an asset that is genuinely worthless looks the same. Distinguishing illiquidity from insolvency requires knowing what things are worth at a time when nobody knows what anything is worth.

Two modern developments make the old model sharper rather than obsolete.

The first is speed. Silicon Valley Bank failed in 2023 in roughly two days. It had a real problem (it had bought long-dated bonds whose value fell when rates rose) and what killed it was a coordinated withdrawal by a concentrated depositor base who talked to each other on group chats and could move money instantly on a phone. The queue no longer has to form physically outside a branch, and the bad equilibrium is reachable in an afternoon.

The second is that a great deal of maturity transformation now happens outside banks. Money market funds, repo markets and structured investment vehicles all funded long-term assets with short-term borrowing while sitting outside the regulatory perimeter: no deposit insurance, no capital requirements, no supervision.

They were therefore runnable, and in 2007 and 2008 they were run on. The crisis was not primarily a run on retail deposits; it was a run on wholesale funding, and Gary Gorton's description of it as a run on repo is the clearest available.

The general lesson is that regulation defined by institutional form rather than by economic function will always leave this gap. If an entity borrows short and lends long, it can be run on, whatever it is called, and the perimeter has to be drawn around the activity rather than around the license.