Why Some Banks Are Run—and Others Fail
This chapter examines the difference between a bank’s risk of being run and the risk that a run will end in failure. The historical evidence makes the distinction measurable: balance-sheet weakness helps explain who is exposed to withdrawals, while public news and local contagion help explain why runs can spread beyond the weakest institutions. The practical judgment is conditional. A run report is a warning, but its meaning depends on what the bank looked like before depositors moved and on what the surrounding signal says about its assets.
What “fundamentals” means in the evidence
The word fundamentals can sound like a vague synonym for “good bank.” In the New York Fed study, it is built from observable balance-sheet measures that proxy different forms of resilience.
One measure is surplus relative to equity. It captures accumulated profitability and capitalization: a bank with more capital has a larger buffer against losses before its liabilities exceed its assets. Another is noncore funding relative to assets. This includes funding other than deposits and equity, such as wholesale obligations, bills payable, and rediscounts. That measure matters for two reasons. It can indicate a funding structure that is more expensive or sensitive to risk, and it can also be a signal that a bank has already had trouble attracting ordinary deposits.
The study also looks at liquid assets relative to total assets. More liquid holdings can make it easier to meet withdrawals without immediately selling loans or other less-traded assets. Deposits relative to assets capture reliance on deposit funding and, in the study’s interpretation, a form of leverage: a bank financed more heavily by deposits has less equity beneath those claims. Recent asset growth adds another possible signal because an abrupt contraction can precede distress.
These measures are not independent labels. A bank with low capitalization may also have limited liquidity. A bank dependent on noncore funding may be both more exposed to funding risk and more visibly troubled. The study therefore uses the variables together, including an interaction between solvency and noncore funding, rather than pretending that one ratio is a complete health score.
The basic diagnostic logic is straightforward. A bank with less capital has less room to absorb losses. A bank with fewer liquid assets has less room to accommodate withdrawals. A bank with more leverage has more claims to satisfy relative to its equity cushion. A bank relying on risk-sensitive noncore funding may have less stable financing and may already be sending a distress signal. None of these conditions guarantees a run. Together, they make the bank less able to withstand one and more likely to attract attention when the environment deteriorates.
Why no single ratio settles the diagnosis
These measures describe different margins of vulnerability, so a reader should resist turning them into a single ranking without asking what each margin does. Capital is a loss-absorption buffer. Liquidity is immediate payment capacity. Funding structure describes how the bank replaces money that leaves. Leverage describes how much of the asset base is financed by claims that sit ahead of equity. A bank can look relatively sound on one margin and exposed on another.
That distinction matters especially when a run is already under way. A bank with substantial capital may still face a severe cash problem if its assets cannot be sold quickly. A bank with liquid assets may have time to meet withdrawals while its capital position is deteriorating. A bank that depends on noncore funding may face a second pressure when risk-sensitive creditors become reluctant to roll their claims, even if ordinary depositors have not yet moved. The measures are therefore not competing explanations so much as different points in the same transmission path.
The interaction between solvency and noncore funding makes this explicit. Noncore funding is not simply a bad label attached to a bank. It can be a way to finance assets, a sign that ordinary deposits are insufficient, or a signal that other funders already perceive risk. Its meaning depends on the bank’s capital position and on the environment in which funding must be renewed. The New York Fed study treats that ambiguity as part of the measurement problem rather than pretending the variable has one interpretation.
The same caution applies to deposits relative to assets. A high share can indicate heavy dependence on deposit funding and more claims relative to equity, but the ratio alone does not reveal how stable those deposits are, who controls them, or how quickly they can be moved. Those questions become especially important in modern cases, where a small number of large depositors can dominate the funding base. The historical balance sheet can identify exposure; it cannot by itself reveal every depositor’s information, relationship, or speed of response.
This is why a fundamentals-based diagnosis should be expressed as a conditional statement: given the same withdrawal shock, the bank with less capital, less liquidity, or more fragile funding has fewer ways to absorb it. It should not be expressed as a deterministic statement that a particular ratio predicts a particular outcome. The evidence supports a change in probability, not a mechanical rule.
Why weakness predicts failure more clearly than it predicts a run
The evidence separates two questions that are often collapsed. First: which banks are more likely to experience a run? Second: once a run occurs, which banks are more likely to fail?
Bank fundamentals answer the second question more cleanly. They also matter for the first, but less decisively. In the study’s reported classification results, fundamentals alone produce an area under the receiver-operating-characteristic curve—an AUC—of 0.69 for any run. An AUC of 0.5 would indicate no useful discrimination in this comparison, while 1.0 would be perfect classification. An AUC of 0.69 therefore indicates meaningful but incomplete separation between banks that were and were not run.
The same fundamentals perform differently when the outcome is narrowed. The AUC is 0.62 for runs without failure, 0.84 for runs with failure, and 0.87 for failure itself. The contrast is the important result. Bank condition has much more information about the severe branch of the event than about the initial appearance of a withdrawal wave.
Why would that be? A run without failure can be driven by broad news, confusion, or fear that reaches a strong bank. The bank’s balance sheet may be largely unchanged, so its pre-run ratios do not fully explain why depositors moved that day. A run with failure is different. It is more likely to occur where the institution already has weak capital, limited liquidity, or a fragile funding structure. The same withdrawal pressure then has fewer routes to resolution.
This does not mean fundamentals are irrelevant to surviving runs. A weak bank may lose deposits and lending even when it manages to remain open. Nor does it mean that a strong bank is immune. The evidence says something narrower and more useful: fundamentals are relatively weak predictors of the broad category “run,” but much stronger predictors of “run followed by failure.”
That distinction also guards against a common error in interpreting predictive statistics. A high AUC for failure does not show that depositors consciously calculated the same ratios, and it does not prove that the ratios caused the outcome. It shows that the variables contain information about the classification of events in this historical sample. The mechanism still has to be reasoned through: capital absorbs losses, liquidity meets withdrawals, and fragile funding can make pressure harder to manage.
Public news changes the risk before the balance sheet does
A bank’s own condition is only part of the information available to depositors. The New York Fed study includes macroeconomic and local signals that are not specific to the bank under observation. Runs are more likely after weak stock-market returns, elevated business failures, and a run on another bank in the same city.
These signals can change the perceived risk of deposits before the target bank’s published balance sheet changes. A falling stock market may suggest a worsening economic outlook. A rise in business failures may suggest that borrowers are under pressure. A neighboring bank’s run may make depositors wonder whether the problem is isolated or systemic. The bank has entered a new information environment even if its own reported ratios are unchanged.
That is the basis for information contagion. Depositors do not need to believe that every bank is equally weak. They only need to believe that identifying the safe bank quickly is difficult, or that waiting carries a cost. A visible run becomes evidence about what other people may know. Fear can therefore move from one institution to another through observation, not through an immediate transfer of losses on the target bank’s books.
But a run at another bank is not the same kind of signal as a rise in business failures. The study finds that a run following another run in the same city is associated with a 14-percentage-point lower probability of passing through to failure than other runs. That pattern is consistent with a jittery response that draws some healthy banks into the episode. It does not prove that every target bank was sound; the reported failure probability for these runs remained material.
The distinction becomes sharper when the signal points to borrowers. A rise in nonfinancial business failures can mean that bank assets are deteriorating because borrowers are failing. In the study, a one-percentage-point increase in the state-level business-failure rate—roughly a one-standard-deviation increase—is associated with about a seven-percentage-point increase in the pass-through from a run to failure. The signal is not merely “people are nervous.” It is evidence consistent with losses that threaten solvency.
The diagnostic contrast can be summarized like this:
The arrows should not be read as a mechanical forecasting formula. They separate pathways that can coexist. A local run can be both a contagion signal and a response to a common economic shock. A business-failure rate can carry information about bank assets while also reflecting the same downturn that changes depositor behavior. The value of the map is diagnostic: it forces us to ask what the signal is revealing rather than treating all adverse news as equivalent.
What the time dimension adds
The study’s classification results show that run risk is not only a cross-section of weak and strong banks. It also varies over time. For any run, adding year effects raises the AUC from 0.69 with fundamentals alone to 0.817. Adding macroeconomic and local conditions produces an AUC of about 0.799 in the reported specification, close to the year-effect result. This is consistent with runs clustering in particular panic years rather than arriving as isolated, bank-specific events.
The improvement is especially important for runs without failure. Fundamentals alone produce an AUC of 0.617 for that group, while adding macro and local information raises it to 0.755. For runs with failure, the corresponding relationships are already much stronger because bank-specific condition carries more information. In the richest reported specifications, the AUC reaches 0.895 for runs with failure and 0.911 for failure itself.
The numbers do not say that time effects cause runs or that a model can identify the next failing bank with certainty. They show that the informational environment matters more for contained runs than for failures. A broad panic can make a healthy bank look exposed. A bank with deep weakness can fail in a way that is more legible from its own balance sheet.
This is also why a headline comparison can mislead. Two banks may both be run during the same crisis year, but one may have been pulled into a system-wide information shock while the other was already losing capital and liquidity. The date is shared; the diagnosis is not.
From public signal to bank failure
A useful diagnosis therefore separates symptom, cause, contributing condition, and consequence. The symptom is the run itself. Possible causes include a false rumor, a run at another institution, broad economic news, or evidence specific to the bank. Contributing conditions include weak capitalization, low liquidity, leverage, and fragile funding. The consequence may be a contained withdrawal episode, a temporary suspension, permanent failure, or a wider contraction in credit and activity.
The evidence does not let us observe every depositor’s reasoning. Nor does it turn a balance-sheet ratio into a complete measure of hidden losses. A published statement can miss problems that have not yet been recognized, and a variable such as noncore funding can capture several channels at once. The AUC results are in-sample comparisons, not out-of-sample forecasts and not causal estimates.
That uncertainty does not make diagnosis impossible. It changes the level of confidence appropriate to each claim. We can say that weak banks were more likely to be run and that fundamentals were more predictive of runs with failure. We can say that runs followed public and local signals, and that different signals had different associations with failure. We should not say that every run on a weak bank was rational, that every run on a strong bank was irrational, or that the statistical model has identified a single motive.
The most important conclusion is an asymmetry. A bank may be run because the surrounding information environment has become frightening. It is much more likely to fail when its own condition leaves little margin for error. That is why the next analytical step is not another list of warning signs. It is a pass-through test: given the trigger and the bank’s condition, what outcome should we expect, and what evidence would distinguish survival from failure?