Quick answer: a bank capital stress test takes a starting balance sheet and pushes it through a deliberately adverse economic path. The model projects revenues, funding costs, credit losses, provisions, market and operational losses, changes in assets and risk-weighted assets, taxes and capital distributions. It then recalculates capital ratios quarter by quarter. The important output is not simply “the bank loses S$X.” It is the path: when losses arrive, whether earnings absorb them, whether the balance sheet grows or shrinks, when capital reaches its minimum, and whether the bank can continue meeting obligations and supporting customers without falling below required or internally chosen buffers.

A stress test is not a forecast of the most likely future. It is a controlled attempt to find out which future would break the balance sheet.

Why this belongs in mathematics

Capital stress testing combines scenario design, time-series forecasting, probability, transition models, accounting identities and recursive balance-sheet equations. It is also an excellent example of systems mathematics: a recession does not hit “capital” directly. It changes unemployment, house prices, GDP, rates, spreads and market values; those changes affect borrowers, revenues, provisions and trading positions; those effects change retained earnings and risk-weighted assets; only then does the capital ratio move.

The Federal Reserve describes its annual supervisory stress test as an assessment of whether large banks remain sufficiently capitalised to absorb losses under stressful hypothetical conditions while meeting obligations and continuing to lend. The 2026 framework again uses published supervisory scenarios and model documentation. See Dodd-Frank Act Stress Tests 2026 and 2026 Stress Test Scenarios.

1. A stress scenario is a path, not one shock

A simple sensitivity test might ask what happens if unemployment rises by two percentage points. An enterprise capital stress test is richer. It specifies a sequence of macroeconomic and financial variables across many quarters.

Represent the scenario at quarter t as a vector:

S_t = [GDP growth, unemployment, house prices, commercial-property prices, policy rates, long-term yields, equity prices, credit spreads, volatility, exchange rates, …].

The stress engine feeds S₁, then S₂, then S₃ into the bank’s revenue and loss models. The order matters. A sharp unemployment increase followed by a slow recovery can produce different defaults and provisions from a short recession with the same peak unemployment rate.

2. Begin from a reconciled starting balance sheet

Before stressing anything, the bank needs a clean starting state:

loans by product, borrower and credit quality;
securities and trading positions;
deposits and wholesale funding;
derivatives and off-balance-sheet commitments;
allowances and provisions;
tax assets and liabilities;
Common Equity Tier 1 (CET1), other regulatory capital and deductions;
risk-weighted assets (RWA) and leverage exposures.

If the starting books do not reconcile, the stress model can create an exquisitely precise projection of the wrong bank.

3. PPNR: project what the bank earns before provisions

The Federal Reserve defines pre-provision net revenue (PPNR) broadly as net interest income plus noninterest income minus noninterest expense, with specified adjustments in its supervisory methodology. PPNR matters because earnings can absorb losses before capital is consumed.

A simplified quarterly identity is:

PPNR = interest income − interest expense + fee/noninterest income − noninterest expense.

Each term responds differently to stress. Loan income may fall as balances shrink or nonaccrual loans rise. Deposit and wholesale funding costs move with rates and confidence. Trading or investment-banking fees can fall with activity. Expenses may remain sticky even when revenue drops.

The Federal Reserve’s supervisory methodology projects numerous PPNR components using firm characteristics and scenario variables. See the descriptions of supervisory models; the 2026 stress-test publication page contains the current exercise’s methodology and data.

4. Credit losses enter through transitions, not one fixed loss rate

A recession changes the probability that loans move from current to delinquent, impaired, defaulted, paid off or cured. Different products respond to different variables.

For example, the Federal Reserve’s first-lien mortgage stress models use borrower and loan characteristics together with macroeconomic conditions such as unemployment, house prices and rates. Loans move between payment states over the projection horizon. That is a dynamic system rather than a one-time haircut.

A simplified portfolio loss in quarter t can be written:

Credit Loss_t = Σ new defaults_i,t × EAD_i,t × LGD_i,t.

PD, EAD and LGD can all worsen together. Borrowers default more frequently, revolving facilities may be drawn further before default, and collateral recoveries can fall when asset prices are depressed.

5. Provisions connect expected loss to accounting capital

Credit loss does not always hit capital in exactly the same quarter that a cash recovery finally fails. Accounting provisions and allowances are forward-looking. Under applicable accounting rules, the bank updates its allowance for credit losses as expected losses change.

A simplified provision relationship is:

Provision_t ≈ charge-offs_t + desired ending allowance_t − beginning allowance_t.

If the stress scenario suddenly raises lifetime expected losses, provisions can rise before those loans actually default. That reduces net income and capital earlier in the path.

For the underlying expected-loss mechanics, see How Banks Estimate Expected Credit Loss.

6. Trading, counterparty and operational losses need separate engines

Loan losses are only one channel. Large trading banks can face a global market shock, counterparty losses and stressed trading/credit valuation effects. Operational-risk losses can also reduce earnings.

The Federal Reserve’s annual stress-test design includes additional market-shock components for firms with substantial trading activity and, where applicable, counterparty-default components. Those losses are distinct from ordinary loan PD/LGD models because their mechanism is different.

This is why enterprise stress testing should not create one giant generic “loss percentage.” A credit default, a market-price shock, an operational event and a counterparty failure require different data and different falsifiers.

7. Balance-sheet projection changes the denominator as well as the numerator

Capital ratios can deteriorate because CET1 falls, because RWA rises, or both. A stress model therefore needs a balance-sheet path rather than holding every exposure fixed without reason.

During recession:

loan demand may fall;
credit lines may be drawn;
weak borrowers may migrate to higher risk weights;
securities values may change;
deposits may grow or run depending on confidence;
the bank may deliberately reduce risk or preserve lending capacity.

If the bank’s credit quality deteriorates, RWA can increase even while total assets shrink. The denominator can therefore work against the bank at the same time retained earnings fall.

8. The capital recursion

A teaching version of quarter-by-quarter CET1 evolution is:

CET1_t+1 = CET1_t + net income_t − distributions_t ± regulatory adjustments_t.

Then:

CET1 ratio_t = CET1_t / RWA_t.

The stress engine repeats the calculation every quarter. The minimum ratio can occur well after the worst GDP quarter because provisions, defaults and RWA migration have lags.

9. A worked miniature path

Suppose a teaching bank begins with S$12 billion of CET1 and S$100 billion of RWA: a 12% CET1 ratio.

In the first stressed year:

PPNR contributes +S$2.2bn;
credit provisions reduce income by S$3.0bn;
market/operational losses reduce income by S$0.6bn;
tax effects add back S$0.2bn;
distributions are S$0.3bn;
RWA rise to S$106bn.

Simplified ending CET1 becomes:

12 + 2.2 − 3.0 − 0.6 + 0.2 − 0.3 = S$10.5bn.

Ending CET1 ratio ≈ 10.5 / 106 = 9.9%.

The capital decline came from both numerator loss and denominator expansion. A model that froze RWA at S$100bn would report 10.5%, a materially different conclusion.

10. Stress Capital Buffer: a US supervisory use of the result

In the United States, supervisory stress-test results feed the stress capital buffer (SCB) framework for large banks. At a high level, the stress test’s decline in the CET1 ratio over the supervisory horizon helps determine a forward-looking capital buffer, subject to the governing rules and minimums.

This should not be confused with a bank’s entire internal capital-planning process. A bank can and should run additional stresses relevant to its own business model, concentrations and emerging risks even when those stresses are not part of the supervisory scenario.

See the Federal Reserve’s Stress Tests and Capital Planning page for the current supervisory framework.

11. Supervisory stress and internal stress answer different questions

Test	Primary job
Supervisory common scenario	Comparable, standardised view across large banks
Bank-specific internal scenario	Test concentrations and business-model weaknesses unique to the bank
Sensitivity test	Move one or a few variables to understand local dependence
Reverse stress test	Start with failure or buffer breach and solve backwards for the conditions that cause it

Basel’s current Stress Testing Principles deliberately cover a range from sensitivity analysis through complex scenario and reverse stress testing. The important requirement is that the method match the decision.

12. Reverse stress testing: find the combination that breaks the buffer

Instead of asking “What does our chosen recession do?”, reverse stress testing asks:

What combination of credit loss, revenue decline, RWA growth, market loss and distributions would push CET1 below our critical threshold?

This is useful because standard scenarios can become cognitively comfortable. Reverse stress starts from an unacceptable state and forces the model to discover routes that ordinary scenario design may never generate.

13. Model interaction is often the hidden weak point

Stress testing is a system of models: macro scenario → credit model → revenue model → balance-sheet model → provision model → RWA model → capital calculation.

Each component can be reasonable by itself while the combined machine is inconsistent. A credit model may assume balances run off while the PPNR model assumes the same balances remain. A deposit model may predict outflows while the balance-sheet model funds asset growth anyway. An RWA model may use a credit-quality path that disagrees with the provision model.

This is why model validation must include interfaces, not only individual models. See How Banks Validate Risk Models.

14. Creative-work lens: The Martian and capital as a survival path

The Martian is not a banking source, but it offers a useful systems lens: survival is not determined by the amount of food, oxygen or power available on Day 1. What matters is the sequence of consumption, replenishment, failures and repairs through time. A bank stress test works the same way. Starting capital matters, but losses, revenue, provisions and RWA arrive on different dates.

The creative work makes path dependence memorable. The bank’s conclusion must still come from reconciled exposures, macro scenarios, accounting rules and validated models.

15. The capital-stress algorithmic pipeline

Reconcile the starting balance sheet and capital.
Define the macroeconomic and financial scenario path.
Project loan, security and off-balance-sheet exposures.
Project PPNR and funding costs.
Model credit-state transitions, defaults and recoveries.
Update allowances and provisions.
Apply market, counterparty and operational-loss shocks where relevant.
Project taxes and other accounting effects.
Apply capital distributions/actions according to the test rules.
Project RWA and leverage exposure.
Recalculate CET1 and other capital ratios each quarter.
Identify the minimum capital point and binding constraint.
Run bank-specific and reverse stresses.
Challenge interfaces and re-run when assumptions change.

16. Failure modes

One-shock thinking. A multi-quarter recession is compressed into one instantaneous loss.
Static balance sheet by habit. Exposures are frozen without asking whether borrowers, deposits or the bank would change behaviour.
PPNR optimism. Revenue remains strong while the scenario says economic activity is collapsing.
RWA freeze. Credit deterioration reduces capital but never increases the denominator.
Model-interface inconsistency. Separate models assume different portfolio paths.
Most-likely contamination. Stress scenarios are weakened because they “do not look realistic enough” to forecasters.
Supervisory-scenario complacency. Passing the common scenario is treated as proof against bank-specific concentration risk.
Capital/liquidity confusion. A bank remains solvent in the model while running out of usable cash before the capital loss materialises.

17. Diagnostics and falsifiers

Which quarter produces the minimum CET1 ratio?
How much of the decline comes from credit loss, PPNR, market loss or RWA expansion?
Which loan class contributes most to provisions?
Does the balance-sheet model agree with the liquidity model on deposits and funding?
What happens if collateral values and LGD worsen together?
How much capital is preserved if distributions stop earlier?
Which bank-specific concentration is not represented in the common scenario?
What scenario combination pushes the bank below its internal critical threshold?

Suppose someone claims, “The bank has 12% CET1 today, so it has plenty of capital.” A falsifier is a plausible stress path in which cumulative losses reduce CET1 while RWA rises enough to push the ratio below the relevant requirement or risk appetite. Today’s ratio is an initial condition, not a survival proof.

18. Verification and update triggers

reconcile starting data to regulatory and financial statements;
backtest component models separately and together;
compare scenario output with simpler sensitivity benchmarks;
validate model interfaces and accounting identities;
re-run after material portfolio acquisitions or disposals;
update scenarios when new concentrations emerge;
review whether modelled management actions are operationally credible;
keep reverse stresses alive so the scenario library does not become conventional and self-confirming.

Connections across the finance-and-banking algorithms lane

Bank capital models — the static capital/RWA architecture this article pushes through time.
Expected credit loss — the provision engine inside the stress path.
Credit concentration — bank-specific stresses should target common-factor concentrations.
Liquidity stress testing — solvency and liquidity must be tested as distinct but interacting constraints.

Research anchors

The deeper lesson

Capital stress testing is the mathematics of a bank moving through a hostile timeline. The scenario changes the economy. The economy changes borrowers, markets and funding. Those changes move revenue, losses and provisions. The balance sheet changes the denominator. Capital absorbs what remains. A strong stress test therefore does not ask only “How much capital do we have?” It asks “What sequence of events consumes it, what else changes while it is being consumed, and what evidence would show that our scenario still missed the true weak route?”

Educational note: This article explains public banking and stress-testing mathematics. It is not investment advice, regulatory advice or a capital-planning model for any specific institution.

Small Group Tutorials

How Banks Stress-Test Capital Under Macroeconomic Scenarios: PPNR, Credit Losses, Balance-Sheet Projections, RWA and Stress Capital Buffers