Sensitivity Analysis

Thinking Tools5.3 Validating Impact and ExecutionAlso known as Tornado diagram

From operations research and financial modeling, with no single originator: the post-optimality analysis of Abraham Charnes and William W. Cooper (1961), the Monte Carlo method of Nicholas Metropolis and Stanislaw Ulam (1949), and the capital-budgeting simulation of David B. Hertz (1964).

Varies each input of a model across its plausible range to show which few assumptions move the answer, where the recommendation flips, and how wide the real range is.

A new plant's value as each assumption swings across its range, widest swing on top

Each bar is the plant's net present value when one input moves from its low to its high case while the others stay at plan; at the opening settings the bars run widest swing first, the shape that gives the chart its name of tornado. Set how far each input could plausibly be off. Plan marks the value at plan. A bar that crosses zero is an assumption on which the go decision hangs, and the break-even price shows how little room price has. Illustrative.

10 percent
10 percent
20 percent
20 percent
Net present value at plan, ten years at 8 percent
$2.1M
Value with price at its low case
−$4.6M
Break-even price: the plant is worth nothing below it
$48.45
How far price can fall from plan before the decision flips
3.1 percent
Selling priceUnit costVolumeBuilding cost−$5M$0M$5M$10MZeroPlanSelling priceUnit costVolumeBuilding cost−$5M$0M$5M$10MZeroPlan

An example

A packaging manufacturer's board was asked to approve a $10 million plant for a new line of molded containers. The business case showed a net present value of $2.1 million over ten years and recommended building.

The finance director asked for each input's plausible range and reran the case one input at a time. The building cost, the input the project team had spent the most time on, barely mattered: 20 percent either way left the plant worth building. Volume and unit cost mattered more. Price mattered most by far. A 10 percent swing in the selling price moved the value from minus $4.6 million to plus $8.8 million, and the case broke even at $48.45 a unit, only 3 percent under the $50 assumed.

The decision hung on one untested number. The board deferred approval by six weeks while sales went back to the three largest buyers for priced letters of intent. Two came in at $51 and one at $49, and the plant was approved with a price floor written into the first contracts.

Why it works

A business case gives one answer, built on dozens of assumptions, and some of those assumptions matter far more than others. Varying each one across its plausible range, with the rest held at plan, shows how much the answer moves for each. Sorted by swing, the results almost always show the same shape: two or three inputs do most of the moving, and the rest hardly register.

That changes where effort goes. The few inputs that move the answer deserve research, testing, and contract protection; the many that do not can be left at a reasonable guess. Finding the value of an input at which the recommendation flips is the sharpest form of the test. If that break-even value sits close to the estimate, the decision rests on a guess that needs firming up before anyone commits.

The thinking behind it

David B. Hertz David B. Hertz, an operations researcher, showed managers in Risk Analysis in Capital Investment (1964), in the Harvard Business Review, how to replace the single best estimate in a capital budgeting decision with ranges for each uncertain input. Simulated many times, the ranges give a distribution of returns with its odds in place of one number. The article brought Monte Carlo simulation into corporate finance, and the one-at-a-time swings on this page are its simplest relative.

Abraham Charnes Abraham Charnes, a mathematician and operations researcher, wrote with William W. Cooper Management Models and Industrial Applications of Linear Programming (1961), a two-volume work that set out how to apply linear programming to management problems, including how far a solution holds as its inputs change. That question, how much an input can move before the best answer changes, is the break-even test at the center of this tool.

Stanislaw Ulam Stanislaw Ulam, a mathematician, conceived the Monte Carlo method at Los Alamos in 1946 while estimating the odds of winning at solitaire by playing it many times, and saw that electronic computers made such sampling practical. Nicholas Metropolis named it, and their paper The Monte Carlo Method (1949), in the Journal of the American Statistical Association, was the first public account. Sensitivity analysis itself has no single originator; it grew out of this work and out of operations research.

How to run it

  1. Build the model on one page. Write the answer as a formula of its inputs, so each input can be changed alone.
  2. Set a range for every input. Give each a low and high case that the people who know it would call plausible, not extreme.
  3. Swing one input at a time. Recompute the answer at each input's low and high, with the others at plan.
  4. Sort by swing. Rank the inputs by how far they move the answer and draw them as bars, widest first.
  5. Find the break-even values. For the top inputs, find the value at which the recommendation flips, and compare it with the estimate.
  6. Firm up what matters. Spend the research, testing, or contract terms on the two or three inputs at the top.
  7. Check combinations. Run the top inputs together, or simulate them, since they may move at the same time.

Where it goes wrong

The ranges are too narrow. Teams set each input's range by what they hope, so every bar is short and the case looks safe; spot it when no range is wider than 10 percent. Ask the people closest to each number for honest low and high cases.

Inputs that move together are swung alone. A recession lowers volume and price at once, and one-at-a-time swings miss the combined hit. Run the top inputs together, or simulate them with their links.

The analysis is done and nothing changes. A tornado chart in an appendix does not firm up the price. Tie each top input to an action: a test, a contract term, or a decision to wait for better data.

When to use something else

  • Monte Carlo Simulation. When several inputs are uncertain at once and the full distribution of outcomes is needed.
  • Wind Tunneling. When the inputs move together in coherent stories that are better tested as whole scenarios.
  • Decision Tree Analysis. When the uncertainty comes as a few discrete events with probabilities, not continuous ranges.

Where it sits in SOLVE eX

Go to the source

  • Management Models and Industrial Applications of Linear Programming, Wiley, 1961
  • The Monte Carlo Method, Journal of the American Statistical Association 44, 1949
  • Risk Analysis in Capital Investment, Harvard Business Review 42, 1964
  • Global Sensitivity Analysis: The Primer, Wiley (Andrea Saltelli and colleagues), 2008

Back to the catalog