What a Bank Stress Test Actually Computes
The Federal Reserve’s 2026 stress test starts with a simple-looking question: if a large bank faced a severe recession and financial-market shock, what would happen to its capital? Answering it requires more than applying a percentage decline to today’s balance sheet. This chapter follows the calculation from bank data and scenario variables through projection models to the capital ratio and its minimum. The goal is to make the reported number traceable: a reader should be able to ask which input, transformation, or assumption produced it.
The key distinction from the opening chapter remains in force. A stress test does not first predict the economy and then report the bank’s likely response. It defines a hypothetical path, applies that path to a particular starting institution through a particular model system, and reports a conditional result. The result is therefore best understood as a chain, not a score.
From starting bank to projected capital path
The first input is the bank itself. The Federal Reserve’s models use information about participating banks’ financial conditions and risk characteristics. That information matters because a scenario cannot affect an abstract “bank.” It affects loans, securities, trading positions, counterparties, funding structures, and sources of income held by a particular institution at a particular starting point.
The second input is the scenario. In 2026, the severely adverse scenario supplied paths for unemployment, GDP, house prices, commercial-real-estate prices, interest rates, corporate-bond spreads, equity prices, and other economic and financial variables. These paths are not yet losses or capital ratios. They are drivers that the models translate into bank-level consequences.
The third element is the model system. Models connect macro-financial variables to specific outcomes: loan losses, provisions, trading and counterparty losses, securities effects, net interest income, noninterest income, noninterest expenses, and other gains or losses. The model is the mediator between the scenario and the bank’s reported result. It determines how a change in unemployment becomes a change in credit quality, or how a change in interest rates becomes a change in income or the value of securities.
The final element is the capital calculation. The projected components are combined with capital-action assumptions and tracked over the projection horizon. The Federal Reserve’s 2026 report covers nine quarters, from the first quarter of 2026 through the first quarter of 2028. It reports both the path and the minimum capital ratio reached during that period. A bank can begin with a substantial capital cushion, absorb losses and income effects, reach its lowest point before the endpoint, and then recover within the same hypothetical path.
The calculation is repeated across time rather than performed as one large subtraction at the end. A scenario variable has a path: unemployment can rise, interest rates can move, asset prices can fall, and credit conditions can deteriorate at different points in the horizon. The models apply those conditions to the bank’s relevant exposures and business characteristics, producing projections for each period. Those projections then affect the capital calculation, which becomes the starting point for understanding the next period. This does not mean that the model predicts a sequence of realized quarterly events. It means that the counterfactual has internal timing, so the model can show when pressure accumulates and when it eases.
That timing also explains why the same input can have more than one effect. A higher interest-rate path may support some interest income while reducing the value of certain securities or increasing credit pressure. A decline in property prices can affect borrowers, collateral, and portfolios with different sensitivities. The scenario supplies the common movement; the bank data and model relationships determine which channels receive it and how strongly. The published ratio is the endpoint of those linked transformations, not a direct quotation from any single scenario variable.
That sequence can be stated plainly:
The diagram is not a claim that the Federal Reserve runs one simple model in a single straight line. It is a map of dependencies. Bank data and scenario variables enter together; models transform them into financial components; the capital calculation combines those components across time; and the published result is meaningful only with its assumptions attached.
The 2026 aggregate result illustrates the chain. The 32 participating banks began with an aggregate CET1 ratio of 12.8 percent at the end of 2025. Under the severely adverse scenario, the projected minimum was 11.2 percent, followed by a projected 12.7 percent at the end of the nine-quarter horizon. The 11.2 percent is not “the bank’s capital” in the abstract. It is the lowest point generated by this starting population, this scenario, these models, these capital-action assumptions, and this time window.
That qualification makes the number more informative, not less. It tells the reader what kind of question the number answers. A current capital ratio describes a condition at a point in time. A stress-test minimum describes the result of applying a defined pressure path to that condition through a defined analytical system.
How financial effects become a capital ratio
A stress-test result becomes easier to interpret when the projected effects are separated before they are recombined. Losses reduce income or capital. Revenue provides an offset. Expenses consume part of that revenue. Other effects, such as changes in the value of securities or trading and counterparty positions, can move capital through additional routes. The capital ratio is the combined outcome of these channels over time.
In the Federal Reserve’s 2026 results, aggregate pre-provision net revenue was projected at $719 billion over the horizon. That figure is not the same as profit and does not mean the banks would be comfortably profitable under stress. Against it were $634.7 billion in provisions for loan and lease losses, $6.5 billion in credit losses on investment securities, $37.1 billion in trading and counterparty losses, and $39.5 billion in other losses or gains. Aggregate cumulative pre-tax net income ended at only $1.2 billion.
This arithmetic shows why a reader should not jump from one component to a conclusion about resilience. Revenue can absorb losses without leaving much net income. A large loss category can coexist with a capital ratio that remains above a regulatory minimum. A positive aggregate result can conceal individual banks with negative pre-tax income. The calculation is not designed to produce one intuitive label such as “safe” or “unsafe.” It exposes how much pressure different channels place on capital under the specified conditions.
The IMF’s top-down euro-area framework offers a useful conceptual distinction inside this combined result. It describes a profitability channel and a credit-risk channel. The profitability channel affects the numerator of the capital ratio through income and retained earnings: weaker economic conditions can reduce income, increase provisions, and slow the accumulation of organic capital. The credit-risk channel affects the denominator through risk-weighted assets: worsening probabilities of default can make a given exposure require more capital. The two channels can move together, but they are not the same mechanism.
This distinction explains why “capital fell” is an incomplete diagnosis. A bank may be pressured because earnings deteriorate, because risk-weighted assets rise, or because both occur. The policy meaning differs. An earnings shock points toward profitability, expenses, credit provisions, or market income. A risk-weighted-asset shock points toward the risk characteristics of exposures and the way the capital framework maps those characteristics into required capital. The reported ratio combines the effects, while the model structure helps separate them.
Consider a simple conditional example. Two banks can face the same rise in unemployment, but the first may have more consumer loans while the second has more commercial exposures. The same macro input can therefore produce different modeled loss paths. If the first bank also earns more interest income under the scenario, that revenue may offset part of its provisions; if the second has greater market exposure, trading or securities effects may matter more. None of these differences contradicts the common scenario. They show why bank data must enter before the model can produce a bank-specific output.
The numerator-and-denominator distinction also prevents a common shortcut. A fall in income is not automatically the same event as a rise in risk-weighted assets, even though both can reduce the final ratio. One changes the capital available to absorb risk; the other changes the amount of risk that the capital measure must cover. The stress test brings them together because resilience depends on both the resources available and the risks being measured. Keeping them separate until the final calculation makes the reported result easier to question and compare.
The same logic applies to the Federal Reserve’s categories. Loan losses and provisions are related but not identical: provisions include the amount needed for the allowance for credit losses, not simply the raw projected loss on loans. Trading and counterparty losses apply to the banks subject to those components and reflect their trading positions and counterparty exposures. Other losses and gains include several items, such as changes in the fair value of loans and losses or gains on hedges. Each label represents a different route through the calculation.
The horizon adds another layer. The Federal Reserve reports the minimum over nine quarters rather than only the final quarter because stress is a process. Credit losses may build as borrowers deteriorate. Revenue may respond differently as interest rates and market prices move. Capital can reach its low point before the scenario ends. The minimum is therefore a timing judgment within the model, not merely a subtraction from the starting ratio.
A reader should also resist treating the endpoint as a promise of recovery. The projected rise from 11.2 percent to 12.7 percent occurs within the same hypothetical world that produced the decline. It does not mean that management will necessarily make the same capital distributions, that markets will follow the same path, or that the bank will recover on that schedule in reality. It shows how the modeled system behaves after its most severe point under the exercise’s assumptions.
Where the instrument gets its conditions
The model system is not a neutral window onto the future. It is part of the stress-test instrument. The Federal Reserve’s 2026 methodology says that the Board generally used the models implemented in 2025 while proposals about model transparency and public accountability were under review. That continuity supports an orderly comparison, but it does not mean the instrument was unchanged.
The methodology identifies targeted updates. The Fair Value Option model adopted a constant 50 percent loss-given-default assumption and set the loan-equivalent factor for corporate loans at 65 percent. The Largest Counterparty Default model changed the treatment of certain high-rated sovereign entities, including the United States, in selecting a firm’s largest counterparty. The Domestic Credit Card model broadened its treatment of revenue and loss-sharing agreements and tailored the adjustment to the form of the agreement. An earlier compensation-model adjustment was fully implemented as well.
These examples matter for a reason beyond their technical detail. They show that a model choice can change how a bank’s risk is represented even when the broad scenario remains recognizable. A model may be adjusted to reflect historical recovery outcomes, business practices, or a firm’s risk profile more appropriately. But the output remains conditional on the adjustment. The source documents do not isolate how many percentage points of the aggregate result came from each update, so a defensible interpretation should stop at the documented boundary: the model version and its targeted changes are part of what produced the result.
Capital-action assumptions create another condition. The Federal Reserve’s ratios are calculated using assumptions provided in the stress-testing rules. Those assumptions make banks comparable within the exercise, but they are not a complete prediction of what every bank’s management would do if the scenario occurred. A projected ratio is therefore not simply a measurement of assets after an imagined recession. It is a projection under specified rules about capital actions as well as specified economic and financial conditions.
The data date matters too. The report says that the trading and counterparty losses for the relevant banks are based on positions and exposures held on October 17, 2025. A different as-of date could have produced different losses because trading positions and counterparty exposures change. That does not make the estimate unreliable; it tells the reader what population of positions the estimate actually describes.
This is the boundary between a useful conditional measurement and a timeless claim about bank strength. The Federal Reserve can say how the tested banks performed in the model under the severely adverse scenario. It cannot, from that result alone, say what their future capital will be under an unmodeled path, a different exposure date, a different set of capital actions, or a different model specification.
Once the chain is visible, the next question becomes more precise. The scenario is not a single shock word attached to the end of a report. It is a path whose timing, severity, and combination of variables shape the outputs at every stage. The next chapter follows that hypothetical world through time and asks why the 2026 path produced the particular minimum that it did.