An IC pack arrives with a tornado chart on page nine and a table titled "downside scenario" on page ten. The tornado ranks the model's response to plus or minus 200bps on WACC, plus or minus five per cent on revenue growth, plus or minus 100bps on terminal margin. The "scenario" table flexes the same WACC by plus 200bps and reports a lower equity value. The two pages are presenting the same arithmetic in different costumes. Sensitivities answer a question the committee did not ask, and answer it precisely. This note sharpens the distinction between the two techniques, names what each one actually computes, and offers four diagnostic questions for the IC reader. It extends the earlier argument in Layered scenario architecture vs single-variable [1], which set out why a layered scenario beats a single-variable flex. This piece sits one level above that argument, at the conceptual line between the two methods.
What does each method actually compute?
A sensitivity analysis perturbs one input while every other input is held at its baseline. The output response is recorded and, in the one-at-a-time (OAT) form universal to IC packs, ranked on a tornado chart [2]. The mathematics is a local partial derivative. The implicit assumption is that inputs vary independently, that the chosen perturbation has no probability weight, and that the response is interpretable in isolation from any other move in the model.
A scenario analysis is a different object. It specifies a set of inputs that move together in a manner consistent with a named state of the world. The variables share a joint distribution rather than being flexed independently. BCBS d450 frames this requirement at supervisory scale, asking banks to design scenarios coherent with bank-specific vulnerabilities, periodically reviewed, with governance over scenario approval as distinct from sensitivity analyses [3]. NGFS Phase V demonstrates what the joint specification looks like in practice, with macroeconomic, financial and physical variables specified across sectors and countries within named narrative pathways such as orderly, disorderly, hot house world, and too little too late [4].
The two methods therefore answer two different questions. Sensitivity analysis answers a local question: if one input moves and nothing else does, how much does the output change. Scenario analysis answers a structural question: if the world enters a particular jointly coherent state, what happens to the asset, the cash flows and the financing. The questions have different mathematics, different governance footprints, and different decision uses. They are not interchangeable.
Why does the conflation persist?
OAT outputs are cheap to produce. A tornado chart is a half-day of model work. The reader sees a clear ranking, the rows are sorted by absolute impact, the visual reads as discipline. Joint scenario construction is harder. It requires a named narrative, a consistent set of macro and operational moves, and a willingness to admit that the chosen scenarios are a hypothesis about how the world might co-move rather than an algorithmic exercise in input ranges.
There is also a defensive economy at work. "We stressed it" is a defensible phrase. It implies the modeller did something deliberate; it satisfies the committee minute. A single sensitivity table reads as governance even when the underlying exercise is mechanical. The conflation is not malicious. It is the natural equilibrium when production cost favours one method and presentation convention does not penalise calling the cheaper output by the more expensive name.
The ICAEW working-modeller guidance draws the line plainly: a sensitivity perturbs one input, a scenario changes a coherent set of inputs in tandem, and a model should be built as an engine capable of running either [5]. The distinction is not a contested matter of taste in the modelling community. It is the standard working definition. The pack conflates the two because the pack writer is rewarded for producing pages, not for labelling them precisely.
What does a properly constructed scenario look like?
A recession scenario is not "revenue down ten per cent." A recession scenario specifies revenue, working capital, default rates, financing spreads and FX, with each move chosen to be consistent with the others under a named narrative. Revenue contracts because demand contracts; working capital extends because customers pay later; default rates rise; financing spreads widen because credit conditions tighten; FX moves in the direction implied by the relative monetary stance. Each line is a derivation from the narrative, not an arithmetic convenience. The IFoA has flagged that scenario analysis is frequently performed in isolation, missing correlated risk factors and dependencies, with the result that the exercise misleads when used as the basis for a strategic decision [6]. The fix is not more variables. The fix is joint specification.
PRA SS3/18 places scenario specification inside the model risk management perimeter, subject to validation and independent review under a four-principle framework covering model inventory, governance, development and implementation, and validation [7]. This is the architecture the firm's financial modelling work draws on for scenario construction. The discipline is not bureaucratic. It is the mechanism by which the committee can interrogate whether the scenario set covers the firm's real exposure rather than the variables that happen to be easy to flex.
What can sensitivities still tell us?
Sensitivities are not wrong. They are correct answers to a narrower question: local elasticity. They identify which inputs the model is most responsive to, which is useful for two purposes. The first is model-risk triage: if the output is highly sensitive to an input that has been weakly evidenced, the pack should say so. The second is scoping: a sensitivity scan tells the scenario designer which variables to specify carefully and which to fix at baseline without losing much.
For analysts who own the model, the technical frontier moves past OAT. The Morris method, introduced in 1991, computes the mean and standard deviation of elementary effects across a designed sample, distinguishing factors with strong main effects from factors whose influence runs through interactions with other inputs [8]. Sobol indices go further. The first-order index Si captures the share of output variance attributable to a single input acting alone; the total-effect index STi captures its contribution including all interactions, computed by Monte Carlo over the full joint input space [9]. Most IC packs do not need either. The point of naming them is to mark where the methodological frontier sits for analysts who do own the model, and to make clear that the OAT tornado is not the only tool available.
Four diagnostic questions for the IC reader
Four questions are enough to separate sensitivity from scenario in any pack the committee receives.
- Which variables co-move in this scenario, and how have the co-movements been specified.
- What real-world state is this scenario meant to represent, in one sentence.
- Who approved the severity calibration, and against what risk appetite.
- Where in this pack are the OAT sensitivities, and are they labelled as sensitivities.
The four questions are not a checklist for the modeller. They are diagnostic prompts for the reader. A pack that cannot answer the first three is presenting sensitivities and labelling them as scenarios. A pack that cannot answer the fourth has presented scenarios without the parameter triage that should sit underneath them. The same logic applies to discount-rate stress in IRR revisited for investment committees and to capital-recovery framing in NPV vs payback for infrastructure committees; the diagnostic generalises across the IC-architecture cluster.
What does the institutional standard look like?
The institutional standard is public. BCBS d450 is the international supervisory baseline for stress testing in banks; SS3/18 is the PRA's model risk management framework as applied to stress testing; NGFS Phase V is the operational exemplar for joint scenario specification at supervisory scale. The CFA Institute Standard V(A) sets the conduct expectation for analysts: a duty to test models across a broad range of assumptions, including scenarios outside the observable historical database, and to evaluate model components as an integrated whole [10]. The standard exists. The IC member can ask for any of these documents by name and will not be told they are obscure.
The institutional grounding sits in the firm's financial services sector work, where regulators learned, expensively, that a portfolio of one-at-a-time stresses tells the firm almost nothing about its real exposure. The principle applies outside banking with equal force. The category error is not a banking artefact. It is the natural state of any IC pack whose production economics reward tornado charts and whose minute culture rewards the phrase "we stressed it."
The standard for the pack is not "include both sensitivities and scenarios." It is: label each correctly, do not present one in the costume of the other, and require any scenario set to specify which variables move, how they co-move, and what real-world state the combination is meant to represent. Where a scenario cannot pass that test, it is a sensitivity, and the paper should say so.
Notes
1. GIVE Analytics, Layered scenario architecture vs single-variable, 2026-03-31. /insights/layered-scenario-architecture-vs-single-variable
2. ICAEW, Intro to Financial Modelling, Part 18: Sensitivities and Scenarios, Excel Community, 2021. https://www.icaew.com/technical/technology/excel-community/excel-community-articles/2021/intro-to-financial-modelling-part-18
3. Basel Committee on Banking Supervision, Stress testing principles (BCBS d450), October 2018. https://www.bis.org/bcbs/publ/d450.htm
4. Network for Greening the Financial System, NGFS Climate Scenarios for central banks and supervisors, Phase V, 2024. https://www.ngfs.net/en/publications-and-statistics/publications/ngfs-climate-scenarios-central-banks-and-supervisors-phase-v
5. ICAEW, Intro to Financial Modelling, Part 18 (as above).
6. Institute and Faculty of Actuaries, Risk Alert on climate-related scenario analysis, 2022. https://actuaries.org.uk/media/ue4hdq3l/risk-alert-climate-change-scenario-analysis.pdf
7. Prudential Regulation Authority, Bank of England, SS3/18: Model risk management principles for stress testing, April 2018, effective 1 June 2018. https://www.bankofengland.co.uk/prudential-regulation/publication/2018/model-risk-management-principles-for-stress-testing-ss
8. Morris, M.D., Factorial sampling plans for preliminary computational experiments, Technometrics 33(2): 161-174, 1991.
9. Saltelli, A. et al., Global Sensitivity Analysis: The Primer, Wiley, 2008; Sobol, I.M., original Russian 1990, English translation 1993.
10. CFA Institute, Standards of Practice, Standard V(A) Diligence and Reasonable Basis. https://www.cfainstitute.org/standards/professionals/code-ethics-standards/standards-of-practice-v-a