Why a Stress-Test Number Is Not a Forecast
Bank stress tests are often read as if they were weather forecasts for finance. They are not. This series examines the Federal Reserve’s 2026 supervisory stress test as a concrete example of what stress testing actually does: it places large banks inside a deliberately adverse economic scenario, runs that scenario through selected models, and estimates what would happen to losses, income, and capital. The practical judgment it supports is not “What will happen next?” but “Where would this banking system be vulnerable if these conditions occurred?”
That distinction sounds modest, but it changes how almost every number should be read. A stress-test report may say that aggregate common equity tier 1, or CET1, capital falls from 12.8 percent to a projected minimum of 11.2 percent before recovering to 12.7 percent over nine quarters. It may report nearly $708 billion in losses. Those figures can be important evidence about resilience under the exercise’s assumptions. They are not a prediction that the economy will follow the scenario, that banks will realize exactly those losses, or that capital will trace that path in ordinary time.
The difference matters because banks are not isolated businesses. They take deposits, extend credit, make markets, process payments, and help households and companies finance activity. The Federal Reserve says its stress tests are intended to help ensure that large banks have enough capital to keep lending through a severe recession. A reader who mistakes the exercise for a forecast can therefore make two opposite errors. They may treat a comfortable result as proof that no serious risk exists, or treat a large projected loss as a prediction of imminent failure. Both conclusions discard the conditional structure that gives the test its meaning.
The question a stress test answers
A bank stress test begins with a bank as it is—not with an imaginary average institution. The Federal Reserve uses information about participating banks’ financial conditions and risk characteristics, together with a set of scenarios. Its models then project losses, revenue, expenses, and capital ratios over a defined horizon. The result is a conditional path: if the bank began with these exposures, and if the specified economic and financial variables followed the specified paths, how would its capital position respond under the model?
That is different from asking how the bank will perform in the most likely future. A forecast normally tries to identify an expected path, often using a central estimate and a probability judgment. A stress scenario serves another purpose. It selects a demanding but coherent set of conditions so that vulnerabilities become visible before those conditions necessarily occur. The scenario may combine a recession with falling asset prices, wider corporate spreads, and market stress. Its value lies in the pressure it applies to the system, not in its claim to be the modal future.
The Federal Reserve’s 2026 severely adverse scenario makes the distinction concrete. It included unemployment rising 5.5 percentage points to 10 percent, real GDP falling 4.6 percent from peak to trough, house prices falling 30 percent, domestic commercial-real-estate prices falling 39 percent, the BBB corporate-bond spread rising to 5.7 percent, and equity prices falling 58 percent. The results report describes the resulting projections as a hypothetical economic outcome that is more adverse than expected. That wording is not a footnote to the exercise. It is part of the definition of the output.
The scenario is therefore best understood as an input to a controlled counterfactual. The question is not whether every variable represents a separately likely event, or whether the exact combination is a forecast issued for public planning. The question is what the combination would expose. A bank heavily concentrated in credit cards may respond differently from one with more mortgages or commercial real estate. A bank with large trading positions may have a different vulnerability from a bank whose main business is traditional lending. The scenario creates common pressure; the starting balance sheet determines how that pressure is received.
This is why a stress-test number can be decision-relevant without being predictive in the ordinary sense. It can show that a bank’s capital cushion is sensitive to a particular combination of unemployment, asset prices, interest rates, and credit losses. It can show how much of the pressure is absorbed by earnings before capital is depleted. It can help supervisors and bank managers consider capital planning, lending capacity, and areas where risk deserves closer attention. It cannot, by itself, tell a reader when a recession will begin, how probable the scenario is, or what the bank’s realized earnings will be next year.
The same principle applies when the analysis comes from a research model rather than a supervisory exercise. An IMF working paper on euro-area banks describes a top-down stress test that projects capital under baseline and adverse scenarios. Its purpose is also to examine resilience within specified conditions. The framework is useful for understanding how profitability and credit risk can transmit a macroeconomic shock into capital, but it does not turn its adverse scenario into a forecast of the euro-area economy.
A severe scenario is a test world
Calling the scenario a “world” does not make it fictional in the dismissive sense. It means that the exercise holds a set of assumptions together long enough to ask a disciplined question. In the Federal Reserve’s case, the severely adverse path is designed around a severe global recession and heightened stress in commercial and residential real estate and corporate debt markets. The variables interact. Falling economic activity can weaken borrowers; falling property prices can reduce collateral values; wider spreads can raise market losses; and interest-rate movements can affect both income and the value of securities.
But the scenario still does not describe the one future that policymakers privately expect. The report’s use of “likely” must be read in context. It asks how banks are likely to perform under the hypothetical conditions—not how likely those conditions are to occur. The adjective describes the projected response conditional on the scenario. Removing that condition changes the claim.
This distinction also explains why two stress-test scenarios can produce different results without either being a correction of the other. The Federal Reserve notes that the 2026 scenario had more severe paths for some variables than the 2025 scenario, including corporate-bond spreads and commercial-real-estate prices, but smaller projected declines in interest rates. At the same time, loan balances had grown by around 10 percent, concentrated in credit cards and wholesale loans. A change in the reported capital decline can therefore reflect a changed stress world, a changed starting portfolio, or both. It is not automatically evidence of a new trend in underlying bank health.
A stress test also has a time dimension. The Federal Reserve’s 2026 result is not just a starting ratio and an ending ratio. The aggregate CET1 ratio begins at 12.8 percent in the fourth quarter of 2025, reaches a projected minimum of 11.2 percent during the nine-quarter horizon, and rises to 12.7 percent by the first quarter of 2028. The minimum matters because stress can arrive before the end of the projection period. The recovery matters too, but only as a recovery within the same hypothetical world and under the same assumptions.
The path is easier to understand when the report’s headline is decomposed. Across the 32 participating banks, projected losses were nearly $708 billion. Loan losses accounted for about $625 billion; trading and counterparty losses accounted for $37 billion; securities losses for $7 billion; and other losses or gains for $39.5 billion. At the same time, projected pre-provision net revenue was $719 billion. After provisions and other losses, aggregate cumulative pre-tax net income was only $1.2 billion.
That near-offset is not a paradox. It is the point of tracing the channels separately. Revenue can absorb some losses, while provisions, trading losses, securities effects, and expenses consume the rest. The capital ratio records the combined result, not a single measure of “how bad the economy is.” A reader who sees only the $708 billion figure misses the earnings cushion. A reader who sees only the projected capital minimum misses the different forms of pressure that produced it.
Read the number as a conditional chain
The most reliable way to read a stress-test result is to keep its chain intact: starting bank data and exposures, hypothetical scenario, chosen models, projected losses and income, capital assumptions, and the resulting path. Break that chain, and the number starts to look more certain than it is. Keep it intact, and the number becomes a map of conditional vulnerability.
The starting conditions matter because the same scenario does not act on a blank balance sheet. The models matter because they determine how an economic variable becomes a loss, a revenue effect, a risk-weighted-asset change, or another capital adjustment. The capital assumptions matter because the projected ratio is not simply a neutral observation of what management would do in a crisis. The horizon matters because the minimum may occur well before the endpoint. The scenario matters because the entire result is conditional on the path chosen.
The Federal Reserve’s 2026 methodology makes model dependence visible. While the Board reviewed public feedback on proposed changes to stress-test transparency and accountability, it generally used the models implemented in 2025, with targeted updates. Those updates included a Fair Value Option model assumption of a constant 50 percent loss given default and a 65 percent loan-equivalent factor for corporate loans; changes to how the largest counterparty default model selected sovereign entities; and a broader treatment of domestic credit-card revenue and loss-sharing agreements. A compensation-model adjustment was also fully implemented.
These details do not mean that the reported result is arbitrary. They show that a model is part of the measurement instrument. A targeted change can make the instrument more representative of a risk profile, more consistent with historical outcomes, or more transparent. It can also change the result. The source documents do not isolate a percentage-point contribution for each adjustment, so a careful reader should not claim that one named change caused a particular share of the aggregate decline. The defensible conclusion is narrower: the model version and its assumptions belong inside the interpretation of the number.
The Federal Reserve also warns that the 2026 aggregate results are not fully comparable with 2025 because 32 banks participated in 2026, compared with 22 in 2025. Category IV banks generally participate every other year. A year-over-year chart can still be informative, but it is not a clean experiment in which the sample, exposures, scenario, and models are held constant. Before calling a smaller decline an improvement, a reader must ask what changed in the test itself.
The central lesson is therefore not that stress tests are “only hypothetical” and can be ignored. Their conditional nature is precisely why they are useful. They force a bank, supervisor, or reader to connect a particular vulnerability to a particular set of conditions and a particular capital response. The right conclusion from the 2026 exercise is not that banks will experience an 11.2 percent CET1 ratio, or that they will lose $708 billion. It is that, under the stated severely adverse world and the Federal Reserve’s models, the tested banks retained capital above regulatory minimums while absorbing a large, uneven set of projected losses.
That sentence carries more information than either a forecast or a slogan. It identifies the scenario, the model, the population, the output, and the boundary. The next step is to open the calculation itself: how bank data and scenario variables pass through the models to produce the capital ratio that readers see in the report.