Every few years, something goes badly wrong in finance. A bank fails, a market seizes up, a systemic shock reverberates through the global economy. In the aftermath, regulators invariably reach for the same tool: the stress test. Banks are subjected to hypothetical disasters—a severe recession, a housing crash, a spike in unemployment—and asked to prove they can survive. The results are scrutinized, capital requirements are adjusted, and the financial system is declared, more or less, sound.

The template was set in the spring of 2009, when the Federal Reserve published the results of its first large-scale stress test. Nineteen banks had been subjected to a hypothetical severe recession and asked to calculate their losses. Ten were told to raise more capital. Markets rallied. The exercise was widely credited with helping to restore confidence in a sector that had come close to collapse. Stress testing had arrived as a tool of financial supervision, and it has never left.

But a deceptively simple question has lurked beneath the surface ever since: are regulators actually asking the right questions? Do they choose the hypothetical scenarios most likely to reveal dangerous hidden exposures? A stress test conducted with great rigour but built around the wrong scenario is still the wrong stress test. The right answer to the wrong question is still wrong.

Whether the scenarios are right, it turns out, depends critically on what regulators plan to do with the results.

Learning versus acting

A stress test is fundamentally an information-gathering exercise. Regulators cannot directly observe how exposed each bank is to, say, a collapse in commercial real estate or a sharp rise in unemployment. They design hypothetical scenarios and ask banks to estimate their losses under each one. From those reported figures, they try to infer the true underlying risk profile of each institution.

Choosing a scenario is therefore choosing what to learn about and how much. A scenario that stresses GDP growth hard and house prices gently is a decision to acquire precise information about one exposure and relatively little about the other. This makes scenario design a consequential act—not a technical formality.

The critical insight is to separate two things that are almost always conflated: learning and intervening. A regulator learns something from a stress test, then uses that knowledge to act. Whether the information is valuable depends entirely on what kinds of actions she can take afterwards.

If the only lever available is a broad capital requirement—telling all banks to hold more equity as a buffer—then the information revealed by stress tests turns out to be worth surprisingly little. Capital requirements can only target average risk across the whole system. Imagine one bank sitting on a mountain of commercial real estate loans while another has barely any—a blanket capital requirement hands them the same instruction. It cannot differentiate. And because regulators already have reasonable beliefs about average risk levels before running any test, the stress test barely moves the needle. Calibrated to U.S. banking data and the Federal Reserve’s own stress-testing framework, the welfare gain from running an optimally designed stress test—when capital requirements are the only tool—amounts to less than a quarter of the gain from simply lowering the cost of bank capital by 10%. For this purpose, stress tests are closer to a rubber stamp than a revelation.

The picture changes dramatically when regulators can also deploy targeted interventions: loan-to-value caps on specific mortgage categories, collateral requirements against particular asset classes, or the supervisory letters—known in Federal Reserve parlance as “matters requiring attention”—that demand a named bank reduce a named exposure. These tools can be aimed precisely, and precision requires information. When targeted interventions are on the table, the welfare gains from a well-designed stress test are four to five times larger than under pure capital requirements. The same test, the same data, but a dramatically different payoff—because the information can now be put to surgical use.

What makes a scenario well designed?

The answer is not simply to make scenarios as severe as possible. Think of it like adjusting a microscope: too little magnification and you see nothing useful; too much and the image dissolves into blur. More extreme hypotheticals do make banks’ loss estimates more informative about their risk exposures—but they also make those estimates noisier, because banks’ models become unreliable when extrapolating far from historical experience. The 2011 European stress tests were widely criticized for scenarios that proved too mild; but a scenario so catastrophic that no bank can model it credibly is useless in a different way. The sweet spot lies roughly one to two standard deviations beyond the historical average of bad outcomes—severe enough to be revealing, not so extreme as to be unintelligible.

Two further principles govern which risk factors deserve the most attention. Systemic risks—those to which many banks are exposed simultaneously—should be stressed disproportionately, not merely because they are dangerous, but because they are efficient carriers of information: a single scenario that probes a correlated exposure yields intelligence about the whole sector at once. And regulators should stress hardest the factors about which they know least. The intuition is direct: a stress test that will not change the regulator’s mind is, in a meaningful sense, useless. If prior beliefs about a particular risk are so entrenched that no result could shift them, running the test produces no actionable information. The most valuable tests are those where genuine uncertainty remains—where the results could actually surprise.

Knowing what makes a good scenario, however, is only useful if it is matched to a clear sense of purpose. And that brings the analysis back to its central finding.

Figure 1. The stress-testing process has three stages: regulators design the scenarios, banks report their losses, and regulators act on what they learn. The first stage only matters because of the third. When the available action is a broad capital requirement, scenario design adds little — regulators already know enough about average exposures to set sensible buffers. When regulators can instead target specific banks and specific risks, the design of the scenarios that preceded the test determines how accurately they can act. Welfare gains are four to five times larger in the latter case. The implication is that the test and the intervention it enables should be designed as a single exercise, not as separate steps.

Implications for practice

The consequences of scenario choice are higher than is often recognised. In the United States, the outcome of the Federal Reserve’s annual stress test directly determines how much capital banks are permitted to return to shareholders through dividends and share buybacks. A scenario that is too mild allows banks to pay out capital they might need in a genuine downturn; one that is poorly calibrated to actual risk exposures may restrict payouts at the wrong institutions while leaving the genuinely vulnerable ones unchecked. And the consequences are particularly concrete when it comes to the relative assessment of institutions. Which banks pass and which fail—and by how much—depends critically on the scenario chosen. A scenario that stresses credit risk heavily may rank one set of banks as most vulnerable; a scenario that stresses interest rate risk may produce an entirely different ranking. The choice of scenario shapes who is seen as risky and who is not.

History illustrates how much this matters. European regulators were criticized for failing to include sovereign debt exposures in their 2010 and 2011 stress tests, just as the Greek debt crisis was exposing precisely those risks. More recently, the collapse of Silicon Valley Bank in 2023 raised questions about whether standard stress scenarios—typically calibrated to credit risk and recession—are well suited to reveal interest rate duration risks of the kind that proved fatal. These episodes do not reflect failures of rigor; they reflect the difficulty of knowing, in advance, which scenario to choose. Yet much of the attention in stress testing naturally falls on implementation and governance—the models, the data, the institutional machinery—rather than on the scenario itself. This research suggests that is where more of the attention should go.

The recurring debate in supervisory circles—are stress tests a capital-setting device or a genuine intelligence operation?—turns out to have a clean answer: they are both, but with very different requirements. As a capital-setting device, a single plausible adverse scenario is sufficient; elaborate scenario design adds little. As an intelligence operation, scenario design is everything—and the scenarios should be chosen not for plausibility but for informativeness, deliberately deviating from the average bad state to probe the dimensions of risk where uncertainty is highest and targeted intervention is most valuable.

The practical implication is a two-part stress-testing architecture. One scenario anchors capital requirements, as today. One or more additional exploratory scenarios are designed from scratch around the specific risks that targeted interventions will address. The two serve different masters and should be designed accordingly.

Stress testing has already proven its worth as a tool for anchoring capital requirements—that much is not in question. But the research points to an underexploited opportunity. Where regulators have the appetite and authority for targeted intervention, the design of the scenarios that precede those interventions deserves far more deliberate attention. The greatest gains come when the stress test and the actions that follow are conceived together—as two parts of a single, coherent exercise.