Reference · Cross-cutting
Sensitivity and Scenario Analysis
Load when stress-testing an existing model — sensitivity, tornado, scenarios, breakeven, margin of safety.
Part of the Financial Analysis skill · loaded on demand from SKILL.md
A post-processor, not a standalone analysis. It attaches to a model that already exists — a DCF, an LBO, an operating model, a deal — and answers what it costs to be wrong. Run the underlying model first; this playbook consumes its output.
The point is not to vary every assumption. It is to find the few that decide the answer and quantify the exposure to each. A sensitivity section that walks twenty variables by ±10% has told the reader nothing they can act on.
When: "sensitivity", "scenarios", "what if", "what breaks this", "how wrong can we be", "downside case", "tornado", "margin of safety", "stress the model".
Phases
-
Assumption hierarchy — do this before any table. List every material assumption in the model and rank by impact if wrong × probability of being wrong. Output the five that drive most of the outcome uncertainty, and justify the ranking with the arithmetic rather than asserting it. Assumptions with large impact but genuine certainty rank low; assumptions that are shaky but barely move the output rank low. Only the intersection matters.
-
One-way sensitivity. For each of the top five: base value, bear value, bull value, and the resulting output at each. The range from bear output to bull output is that assumption's sensitivity. Present ranked widest-first — a tornado — so the dominant driver is visible without reading the numbers.
-
Two-way tables. At least two, crossing the top drivers against each other, with the key output in the cells. Mark each cell against the hurdle: clears or fails. A table the reader has to compare cell by cell against a threshold stated elsewhere is a table they will misread.
-
Scenario analysis. Scenarios are coherent states of the world, each with a cause. Give the narrative first — what has to happen for this case to occur — then the driver values that follow from it, then the output.
| Bear | Base | Bull | |
|---|---|---|---|
| Narrative — what causes this | |||
| Revenue growth | |||
| Margin | |||
| Other key driver | |||
| Key output | |||
| Probability weight | |||
| Probability-weighted output |
Two warnings that belong in the output, not in a footnote. The probability weights are subjective — label them as a judgement and say whose. The weighted average assumes the drivers are independent, and they are not. In a real downside, volumes fall while input costs rise and credit tightens together, so the true bear case sits below the one the table produces. State which drivers you expect to move together.
-
Distribution framing. Where a full simulation is in scope, define for each top-five assumption a distribution type — normal, triangular, uniform — with its parameters, plus the correlation structure between them. Report P10, P50 and P90. Where simulation is out of scope, describe the inputs and the expected shape of the outcome distribution and say that no simulation was run. Do not present a distribution figure that was not computed. A Monte Carlo run on correlated inputs without a correlation matrix produces a confidently narrow answer and is worse than no simulation.
-
Breakeven and margin of safety. For the investment to clear its hurdle, state the minimum level of each key driver — revenue growth, margin, exit multiple. Then report the gap between each breakeven and the bear case. That gap is the margin of safety, and it is the number the analysis exists to produce. Where a breakeven sits above the bear case, say plainly that the downside does not clear the hurdle.
-
Mitigation. For the top three exposures, what hedges, contract structures, covenants or operational actions reduce them, and what residual exposure remains after each.
Required tables: ranked tornado; at least two two-way tables with hurdle marking; the scenario table; breakeven against bear case.
Challenge test: which single assumption, at its bear value alone with all others held at base, is sufficient to fail the hurdle. If more than one qualifies, the base case is thinner than it looks.
Inputs to gather: the underlying model type and its output; the key output metric and the hurdle it must clear; the assumption set already made in that model; historical volatility of the key drivers if known; management's own scenario thinking if any, kept separate from yours.