Stress testing and VaR answer different questions

Translate the labels into complete statements

A VaR number needs a confidence level, holding period, loss convention, and estimation method. Without those fields, the number is incomplete. A stress result needs a scenario, horizon, portfolio boundary, and valuation approach. Read each as a conditional statement about a model, not a standalone fact about the future. Missing metadata can make two apparently comparable currency amounts answer entirely different questions.

CFA Institute's Measuring and Managing Market Risk reading distinguishes distributional risk measures, sensitivities, and scenario measures. For report interpretation, the important discipline is to preserve those categories. A sensitivity describes response near specified inputs; a scenario evaluates selected conditions; a quantile summarizes a modeled distribution. None becomes a complete account of portfolio risk merely because it is displayed as one convenient number.

Do not read a quantile as a maximum

Under a continuous loss distribution, a 99% VaR threshold leaves 1% probability above it within that model. It does not bound the size of losses beyond the threshold. In discrete distributions, the exact exceedance probability depends on mass at the quantile and the convention used. Record the convention rather than treating confidence-level terminology as a guarantee about observed exceedances in a small sample.

A stress scenario usually does not provide a probability for its outcome. Describing it as severe does not assign it a particular quantile. If a report maps scenario severity to a return period, ask how that mapping was estimated and what uncertainty surrounds it. Do not infer probability by dividing the scenario loss by VaR or by comparing the scenario with a small set of past observations.

Hypothetical calculation: inspect a discrete tail

Consider an invented distribution with ten equally probable loss outcomes, in thousands of dollars: 0, 1, 2, 3, 4, 5, 6, 7, 8, and 20. Using the smallest loss whose cumulative probability reaches 90%, the 90% VaR is USD 8,000. The worst 10% consists of the USD 20,000 outcome, so the corresponding expected shortfall is USD 20,000 under this deliberately simple distribution.

Now separately specify a hypothetical scenario that produces a USD 35,000 loss under the same portfolio valuation boundary. That scenario is outside the support of the toy distribution, illustrating an omitted possibility rather than a probability estimate. It is not an observed historical loss. The comparison asks whether the distribution's coverage is adequate; it does not establish how likely the scenario is.

Understand what expected shortfall adds

Expected shortfall addresses tail severity that a quantile alone leaves unspecified, but it still depends on the modeled distribution and data. For general discrete distributions, defining the worst probability slice may require including only part of the probability mass at the cutoff. The simple example avoids that complication because every outcome has equal probability and the tail contains exactly one outcome.

The Basel Committee's market-risk framework uses expected shortfall within an internal-models capital approach. That official use demonstrates an institutional role for tail measurement; it does not make an illustrative portfolio calculation a compliant capital model. A report should not borrow regulatory terminology as a substitute for the required scope, calibration, validation, and implementation rules of a particular supervisory framework.

Align scope and horizon before comparing

A one-day VaR on liquid trading positions and a multimonth scenario covering the entire portfolio cannot be compared without explaining both differences. Match holdings, currency, valuation date, and treatment of costs wherever the question permits. If coverage differs, show the common subset separately. A larger stress loss may result simply from a wider portfolio boundary rather than a more severe treatment of the same risk.

Avoid mechanically scaling time horizons with a square-root rule unless its assumptions are appropriate. Independent, similarly distributed increments and stable exposures are strong simplifications. Changing volatility, serial dependence, options, defaults, and forced rebalancing can invalidate the shortcut. A report that cannot support a horizon conversion should present the measures side by side with their original horizons instead of fabricating comparability through an unexplained adjustment.

Test each method on its own promises

VaR backtesting examines realized outcomes relative to forecasts under a defined profit-and-loss series and timing convention. It does not prove that every severe scenario is represented. Stress testing instead needs checks on scenario coherence, factor coverage, valuation response, and limitations. Software tests can verify arithmetic for both methods, but passing those tests is not independent evidence that the underlying economic assumptions are reliable.

Compare methods for complementary blind spots. A historically calibrated distribution can omit an unfamiliar combination of shocks, while a small scenario set can omit common exposures that matter frequently. Examine concentration, liquidity, and nonlinear response explicitly. Agreement between two measures is weak reassurance if both use the same stale prices, incomplete holdings, or favorable assumptions about offsets during stressed conditions.

Write a comparison that preserves uncertainty

A good comparison says what each measure reveals and what remains outside it. Report the VaR confidence level and horizon, the expected-shortfall convention if used, and the narrative behind the scenario. Explain material differences in coverage or valuation. Do not combine the measures into an unlabeled average, because the resulting number generally has neither a clear probabilistic meaning nor a clear scenario interpretation.

The decision question is whether the collection of measures exposes important vulnerabilities, not which single measure wins. Request better data, an additional coherent scenario, or a targeted valuation check when a gap matters. These educational comparisons do not predict a maximum future loss, recommend a portfolio allocation, or establish that any illustrative tool meets the requirements of a regulated bank risk engine.

Review checkpoints

• VaR is a quantile, not a maximum loss.
• A scenario does not acquire a probability merely by exceeding VaR.
• Expected shortfall adds tail severity but retains model dependence.

Source notes

1. CFA Institute: Measuring and Managing Market Risk
2. Basel Committee: Minimum capital requirements for market risk, January 2019

Continue the analysis