How Scenario Analysis Works When Each Case Has a Named Cause
Scenario analysis works when each case starts from one named cause that moves several numbers together. How the Fed and Shell do it, and what founders can copy.
Scenario analysis tests a plan against a few alternative futures, and it works when each future starts from a named cause that moves several numbers at once. Flexing one input is sensitivity analysis; setting every input to its worst builds a world that cannot happen. The Federal Reserve, Shell and the research on forecasting point to the same practice.
Scenario analysis in three cases, labelled best, base and worst, is a format that the people who built the method gave up. Pierre Wack ran Shell’s business-environment planning from 1971 to 1981. He wrote that most scenarios “merely quantify alternative outcomes of obvious uncertainties”, such as oil at $20 or $40 a barrel, and called them first-generation scenarios, “not helpful to decision makers” (Wack, 1985). From 1989, Shell’s sets settled at two (Wilkinson & Kupers, 2013). The Fed’s 2026 stress test also publishes two, and builds the severe one from a single cause.
Exhibit 1. Scenario analysis is one of three ways to vary a financial model, and each answers a different question
| Sensitivity analysis | Stacked case | Scenario analysis | |
|---|---|---|---|
| What moves | One input; the rest stay at plan | Every input, each set to its own extreme | Several inputs, moved together by one named cause |
| How values are chosen | A range around plan | Each input’s worst (or best) value, chosen separately | Calibrated to a past episode, then linked through stated relationships |
| Question it answers | How much does the result depend on this input? | What if everything breaks at once? | What happens if this cause occurs, and what will we do? |
| Typical failure | Misses inputs that move together | A world that is improbable and often incoherent | Feels more likely than it is |
| Public example | One-variable flexes in most models | Shell’s 1972 “three-miracles” case | The Fed’s 2026 severely adverse scenario |
Source: CX Cash synthesis of Wack (1985), the Federal Reserve’s scenario design policy (12 CFR 252, Appendix A) and Tversky and Kahneman (1983).
Scenario analysis starts from a cause, as the Fed’s stress test does
The clearest public model of a coherent scenario is the one the Federal Reserve writes each year for the largest US banks. Its 2026 severely adverse scenario is “a severe global recession triggered by an abrupt decline in risk appetite”, and that one cause sets the paths of 28 variables, quarter by quarter, from early 2026 to early 2029 (Federal Reserve, 2026).
The Fed does not pick those numbers one at a time. It chooses the unemployment path, from 4.5% to a peak of 10% in the third quarter of 2027. Much of the rest is derived from it: GDP through a version of Okun’s law, inflation through a Phillips curve and the 3-month Treasury rate through a Taylor rule. Its scenario-design policy asks that the paths, taken together, present “a coherent and plausible outlook”, and that the other variables be “specified to be consistent with the increase in the unemployment rate” (12 CFR 252, Appendix A).
Exhibit 2. One cause, many numbers: the Fed’s 2026 severely adverse scenario
| Variable | Fourth quarter of 2025 | In the scenario | How the path is set |
|---|---|---|---|
| Unemployment rate | 4.5% | Peaks at 10% (Q3 2027) | Chosen: a rise of 3 to 5 points, to at least 10% |
| Real GDP | Level at jump-off | Falls 4.6% to its trough (Q2 2027) | Derived from unemployment (Okun’s law) |
| CPI inflation | 2.8% | 1% (Q4 2026) | Derived (Phillips curve) |
| 3-month Treasury rate | 3.7% | 0.1% by Q2 2026 | Derived (Taylor rule) |
| Equity prices | Level at jump-off | Fall about 58% by Q3 2026 | Guide: larger falls after larger rises |
| House prices | Level at jump-off | Fall about 30% | Guide, calibrated on past housing recessions |
| Commercial real estate | Level at jump-off | Falls 39% | Guide |
Source: Board of Governors of the Federal Reserve System, 2026 Stress Test Scenarios (February 2026); 12 CFR 252, Appendix A. The Fed states that these scenarios are not forecasts.
Two features of the design carry over to a startup, although the Fed writes for bank balance sheets rather than operating companies. First, some numbers move in a direction a worst-case spreadsheet would never choose: in the Fed’s recession, inflation falls to 1% and the short rate to 0.1%. A company holding $20M in Treasury bills earns about $740k a year at 3.7% and $20k at 0.1% (our arithmetic), so the cause that slows sales also removes most of the interest income.
Second, the Fed sets severity from history. The Fed considered a probabilistic tail and chose a recession narrative, because it “guarantees a recession of some specified severity”. People also think about bad outcomes in terms of how the numbers move, it argued, not “how far away they lie in the probability space”. Its guide for unemployment, a rise of 3 to 5 points over 6 to 8 quarters, comes from post-war recessions. The floor of 10% exists because risks built up in good times are “easily obscured by the buoyant environment”. The 2026 document adds that severity is generally set higher after expansions and rising asset prices. A founder who flexes only bookings misses every other line the same cause reaches; in the hypothetical example below, that understates the runway lost by two-thirds.
But a worst case assembled cell by cell describes a world that cannot happen
The alternative to a caused scenario is the stacked case: each input set to its own worst value, then all of them run together. Shell met the same construction from the hopeful side in 1972. Its B3 scenario let oil supply keep pace with demand, and Wack’s team called it the “three-miracles” scenario because it needed three unlikely things at once. Shell’s exploration staff gave each oil province a 30% chance of holding the reserves required, “but only a very small chance that these high reserves would be found in all areas simultaneously”. Meeting 1985 demand under B3 needed 24 million barrels a day from Saudi Arabia, 13 million from Africa and 6 million from Alaska and Canada (Wack, 1985).
The mathematics behind the joke is the conjunction rule, which Amos Tversky and Daniel Kahneman described as perhaps the simplest law of probability: “A conjunction cannot be more probable than one of its constituents” (Tversky & Kahneman, 1983). Call it the miracle count: the number of separate things that must break the same way for a case to happen. A stacked worst case with five inputs, each at a value it reaches one year in ten, has a miracle count of five. If the inputs were independent, it would describe a 1-in-100,000 world (our arithmetic).
Inputs are not independent, and that is the argument for causes. Bookings, churn and collections can weaken together when one cause drives them, so a caused scenario is both more likely than the stack and more severe than any single flex. Rising costs and collapsing interest rates do not sit together in the Fed’s recession, which takes the short rate to 0.1% and inflation down to 1%. Adding detail does not repair this. In 2012 and 2013 the Fed gave trading banks more than 35,000 specific risk-factor shocks. Its policy statement records the benefit as “at least partly offset by the potential difficulty in creating shocks that are coherent and internally consistent”. A board shown a 1-in-100,000 world is likely to discount it, and with it the decision the case was meant to force.
Three cases invite the middle, so Shell settled on two and named them
Shell tried most formats before it settled. In its early years the scenario team wrote sets of six or seven. By the mid-1970s three were common, which tempted managers to treat the middle one as a best guess. From 1989 two became the norm, according to a former Shell scenario planner and a former Shell executive (Wilkinson & Kupers, 2013). They also explain that Shell avoided good and bad futures, since neither taught managers anything they could use. Wack and his colleague Ted Newland steered clear of probabilistic forecasts and insisted on plausibility, and Shell used numbers to keep each story internally consistent.
Wack had reached a similar view by 1985. Of Shell’s six scenarios of 1972 he wrote that “Six are far too many; they had no proper names to convey the essence of what drives each scenario”.
Best, base and worst fail as scenario analysis on both counts: the labels name outcomes rather than causes, and three of them invite the middle. A sounder set starts with a surprise-free baseline, which Wack called essential in the package although it “rarely comes to pass”. Beside it go one or two scenarios named for what drives them: buyers freeze, a platform changes its terms, the largest customer consolidates vendors. The Fed’s 2026 test has the same shape, with a baseline that follows the average of a survey of economic forecasters and one severe scenario named for its cause.
Therefore scenario analysis counts only if it changes a decision
Coherence is necessary but not sufficient. Shell’s first exploratory set of four scenarios “covered a wide span of possible futures, and each scenario was internally consistent”, yet top management found that “no strategic thinking or action could be taken from considering this material” (Wack, 1985). Wack’s diagnosis was that unless the decision makers’ picture of the world changed, the scenarios would be “like water on a stone”. The sets that followed in September 1972 and in 1973 were built around a supply disruption, which the first of them timed for the 1975 renegotiation of the Teheran price agreement. The disruption came in the autumn of 1973, and Wack claims preparation for the event without claiming foresight of its date.
Two of Shell’s devices transfer to a startup with little change. The first is the phantom scenario. Shell wrote versions in which the disruption arrived 5 and 15 years late, the time needed to bring a new oil facility into service and the time to amortize it. Their purpose was to measure the “regret” of preparing for a shock that did not come. The second is timing: Shell ran the shock before, during and after a cyclical downturn, and the exercise led it to prepare for a deeper decline.
Field evidence points the same way. In three field experiments, Shardul Phadnis and colleagues found that scenarios did not raise or lower experts’ overall confidence. Judgments moved with how each investment fared in each scenario, and the experts came to prefer more flexible options (Phadnis et al., 2015). Sequoia’s 2022 deck for its founders ties scenarios to decisions in startup terms. It starts from the input outside the company’s control, revenue, with a few discrete cases: down 0%, 25% and 50%, with a warning that these may not be “the right numbers for you”. Each case maps to a cost plan, and spending is released at “defined trigger points” (Sequoia, 2022). The deck also names the cause that links the lines: “Their spending is your revenue. Talent is your cost, funding is your need for profitability.”
Exhibit 3 runs these cases for an invented Series A software business with $20.0M in Treasury bills, $6.0M of ARR, a 75% gross margin and $900k of monthly operating costs, which rise by $20k each month as it hires. New ARR comes from the existing team ($150k a month) and from each month of hiring ($6k a month of extra capacity after a three-month ramp); gross churn is 1.0% of ARR a month. With no new funding, its surprise-free baseline runs out of cash in month 30.
Exhibit 3. One flex, one stack and one caused scenario for a hypothetical company
| Case | What changes | Miracle count | Cash runs out | ARR at month 24 |
|---|---|---|---|---|
| A. Surprise-free baseline | Nothing | 0 | Month 30 | $9.11M |
| B. Sensitivity | New bookings −30% | 1 | Month 29 | $7.79M |
| C. Stacked worst case | Bookings −50%, churn 2.0% a month, collections +30 days, short rate to 0.1%, costs +10% | 5 | Month 23 | $5.69M |
| D. Scenario: buyers freeze | One cause: bookings −30%, churn 1.5% a month, collections +15 days, short rate 3.7% to 0.1% | 1 | Month 27 | $7.11M |
| E. Buyers freeze, hiring stops after month 3 | D plus the decision | 1 | Month 36 | $6.51M |
| F. Phantom: hiring stops, no freeze | A plus the decision | 0 | Month 47 | $8.24M |
Hypothetical company; monthly model over 48 months with no new funding. Short-rate paths approximate the Fed’s 2026 baseline (3.7% falling to 3.1%) and severely adverse (3.7% falling to 0.1%) scenarios. All other inputs are assumptions, and the results come from our monthly model.
The one-input flex costs a month of runway, which suggests nothing needs deciding. The caused scenario costs three, because the same freeze slows bookings, raises churn, stretches collections and cuts the monthly interest on $20M from about $62k to under $2k. The stack runs out in month 23, but it needs five separate breaks at once and pairs rising costs with near-zero rates. Scenario D is the case that should change a decision, since stopping hiring after month 3 moves the cash-out from month 27 to month 36.
Case F prices the regret. If the company stops hiring and the freeze never comes, ARR at month 24 is $0.87M lower than plan, about 10%. That argues for preparing the cut in advance and pulling it on a signpost, such as two consecutive months of bookings below a set level, instead of cutting on the forecast.
Do vivid scenarios mislead the people who build them?
Scenario analysis has its sharpest critic in the psychology that makes it work. In July 1982, Tversky and Kahneman asked professional analysts at an international forecasting congress in Istanbul to rate outcomes for 1983. One group rated a complete suspension of diplomatic relations between the US and the Soviet Union; another rated a Russian invasion of Poland followed by that suspension. The second outcome is necessarily less likely, yet its geometric-mean estimate was 0.47% against 0.14%, about 3.4 times higher (our arithmetic). The authors’ explanation names the risk in every scenario deck: “The representativeness heuristic generally favors outcomes that make good stories or good hypotheses” (Tversky & Kahneman, 1983).
The evidence that scenario planning improves results is also thin. A 2023 review of reviews in Futures points to a “scarcity of evidence on its effectiveness”. It finds evidence of impact only for specific techniques (Cordova-Pozo & Rouwette, 2023); one exploratory study it cites found better financial performance, possibly at the cost of customer service. Shell’s record comes mostly from insiders, and the business historian Jenny Andersson describes the 1980s management literature on it as “considerable hyperbole” (Andersson, 2020).
Tversky and Kahneman’s result does mean that a scenario is not a forecast, and that a vivid one should never carry a probability set by feel. The IPCC defines a scenario as a plausible description of the future built on “a coherent and internally consistent set of assumptions” and adds that scenarios “are neither predictions nor forecasts” (IPCC, 2021). As a reason to stop using scenarios, it proves less. Paul Schoemaker’s experiments found that scenarios widened people’s confidence ranges, and he described the benefit as exploiting one bias, the conjunction fallacy, “to counteract another (such as overconfidence)” (Schoemaker, 1993). The single plan is a story too, only an unlabelled one. Wack made the point in 1985 about forecasts, which are often reasonably accurate: “And that is what makes them so dangerous.”
A baseline, one caused scenario and a signpost
A founder’s set needs a surprise-free baseline and one or two scenarios, each named for its cause rather than its outcome. For each scenario, choose the driver the cause hits first and set its path from a past episode, such as the 2022 funding reset or the loss of a top customer. Then derive the other lines through stated links: pipeline to bookings, bookings to cash, market rates to interest income. Count the miracles, because a case that needs more than one independent break is a stack to split or drop. Tie each scenario to a decision, a signpost and a date, and price the regret with a phantom case before acting.
After a good year, write the harsh scenario harder; the Fed does so after expansions because risks built up in good times are easy to miss. The 3-statement model template and scenario planner below has a column for each case.
A board should ask what causes each case, rather than how far it sits from plan. It should also ask which variables the cause moves and which links connect them, since a case in which only revenue moves is a sensitivity test, and what decision each scenario triggers and what a false alarm would cost. A probability attached to a scenario deserves suspicion unless it comes from data.
Decisions written before the world changes
Scenario analysis earns its place when it changes what a company does before the world changes. Cases built cell by cell either understate the damage, because they move one number, or overstate it, because they move every number in ways no single story allows. The Fed and Shell reached the same design from different directions: one named cause, numbers derived from it and calibrated to history, a plain baseline beside it, and decisions written in advance.
Sources and notes
- Andersson, J. (2020). Ghost in a Shell: the scenario tool and the world making of Royal Dutch Shell. Business History Review 94 (Winter 2020), 729–751.
- Board of Governors of the Federal Reserve System (2026). 2026 Stress Test Scenarios, February.
- Board of Governors of the Federal Reserve System. Policy Statement on the Scenario Design Framework for Stress Testing, 12 CFR Part 252, Appendix A (as amended 2019).
- Cordova-Pozo, K. and Rouwette, E. A. J. A. (2023). Types of scenario planning and their effectiveness: a review of reviews. Futures 149, 103153.
- Grady, P. and Gupta, R. (2022). Forecasting & Scenario Planning. Sequoia Capital presentation to founders.
- IPCC (2021). Annex VII: Glossary (Matthews, J. B. R. et al., eds.). In Climate Change 2021: The Physical Science Basis, Working Group I contribution to the Sixth Assessment Report.
- Phadnis, S., Caplice, C., Sheffi, Y. and Singh, M. (2015). Effect of scenario planning on field experts’ judgment of long-range investment decisions. Strategic Management Journal 36(9), 1401–1411 (abstract).
- Schoemaker, P. J. H. (1993). Multiple scenario development: its conceptual and behavioral foundation. Strategic Management Journal 14(3), 193–213 (abstract).
- Tversky, A. and Kahneman, D. (1983). Extensional versus intuitive reasoning: the conjunction fallacy in probability judgment. Psychological Review 90(4), 293–315.
- Wack, P. (1985). Scenarios: uncharted waters ahead. Harvard Business Review, September–October.
- Wilkinson, A. and Kupers, R. (2013). Living in the futures. Harvard Business Review, May.
The findings of Phadnis et al. and Schoemaker come from their published abstracts, and the exploratory study by Phelps, Chan and Kapsalis (2001) is cited as the 2023 review reports it. Interest amounts, runway effects and ratios that carry no citation come from CX Cash’s own calculation on the cited figures or on the monthly model behind Exhibit 3. The Series A company in that exhibit is hypothetical, and its assumptions are stated above it.
More in Financial Modeling, Scenario Planning & Risk
6 Financial Modeling Best Practices for the Second Reader
Financial modeling best practices exist so someone else can check the model. Builders misjudge their own errors, so size the review to the stakes.
Financial Risk Management for a Small Business, Line by Line
Financial risk management for a small business is mostly concentration: customers, banks, currency, rates, payment authority. How to measure and price each.
Sensitivity Analysis: Stress-Test a Startup Plan to Failure
Sensitivity analysis is only as good as its ranges. Set them from evidence, then test the plan to failure: the likeliest break is a mix of moderate misses.