Tornado Chart

Why it matters

A tornado chart is the picture of a sensitivity analysis. You have a model with one number you care about — a project’s net present value, a launch’s projected revenue, a trial’s expected effectiveness — and a handful of uncertain inputs feeding it. The chart draws one horizontal bar per input, showing how far that output number swings when you move that single input from its low plausible value to its high one, with everything else held at its base case. Sort the bars longest-at-top, and the silhouette tapers like a funnel — wide at the top, narrow at the bottom — which is where the name comes from. Its whole job is to answer one question: of all the things I’m unsure about, which few actually move the answer?

For example: an investor is valuing a deal off a discounted-cash-flow model with four assumptions — growth rate, discount rate, terminal multiple, and churn. All four feel uncertain, and it’s tempting to research all four equally. The tornado swings each one across its plausible range and draws the result: the discount-rate and terminal-multiple bars are long, the growth and churn bars are short. The picture says, plainly, that the valuation lives or dies on two assumptions — so tighten those before committing capital, and stop agonizing over churn, which barely moves the number whatever value it takes.

  • What it shows. How sensitive one output is to each uncertain input, one input at a time — bars sorted so the variables that drive the answer sit at the top and the ones that don’t sink to the bottom.
  • When to reach for it. You’ve built a model or estimate, several inputs are uncertain, and you need to know which ones to measure more carefully, hedge against, or worry about — and which to set aside.
  • How to read it. Top-down. The long top bars are the variables the answer is most sensitive to; the short bottom bars barely matter. A bar that crosses the value at which your decision flips is a variable the decision actually turns on.
  • What you’d miss without it. The disproportion. A flat list of “things we’re unsure about” treats every assumption as equally worth chasing; the tornado reveals that usually two or three dominate and the rest are noise.
  • Where it misleads. It swings each input in isolation, so it misses interactions — two inputs with short bars that jointly move the output a lot won’t show up; and an arbitrary choice of low-high range produces an arbitrary tornado, so the ranges must be honest and stated.

How to read it

Picture a horizontal axis showing the output you care about, with a vertical line through it marking the base case — the answer when every input sits at its best estimate. Each uncertain input gets one horizontal bar straddling that line. The bar’s left end is the output when you set that input to its low plausible value; the right end is the output at its high value. A long bar means swinging that one input alone shoves the answer a long way — the output is sensitive to it. A short bar means the output barely budges no matter where that input lands within its range.

The bars are stacked longest-at-top, shortest-at-bottom, and that sorting is the whole point: it ranks your uncertain assumptions by how much they actually drive the result. The eye runs down the funnel from the variables worth fighting over to the ones safe to ignore. So you read it top-down. The top few bars are where the answer is decided — these are the inputs worth measuring more precisely, modelling more carefully, or hedging against. The short tail at the bottom is the set of worries you can drop: their value doesn’t change the conclusion enough to matter.

Two things make the picture trustworthy, both about the ranges. First, the low and high values must be deliberate — typically a stated percentile band (the 10th and 90th, say, or the 5th and 95th for a more cautious read), drawn from data, expert judgment, or a clearly-labelled assumption. A bar is only as meaningful as the range that produced it: a mechanically influential input with a narrow plausible range will rightly show a short bar, because in practice it can’t move much. Second, watch for the decision threshold — the output value at which your choice would change. Any bar that crosses it flags a variable the decision genuinely hinges on, which is exactly where to spend effort before committing.

When to use it

The tornado chart belongs to the DECISION/STATISTICAL family — the diagrams that help you reason about choices and uncertainty quantitatively. Within that family it is the sensitivity-ranking tool: you reach for it after you’ve built a model or estimate, when you want to know which of its inputs the answer is most sensitive to. That places it among three relatives, and the boundaries are how you pick the right one:

  • A decision tree maps the structure of the decision itself — the sequence of choices and chance events and their payoffs. The tornado doesn’t model the decision; it interrogates the inputs to whatever model produced your numbers.
  • A distribution plot shows the full spread of one uncertain variable — its shape, range, and likely values. The tornado collapses each variable to a single low-high swing and instead compares many variables against each other.
  • A Monte Carlo histogram shows the output’s full distribution once all the input uncertainties are turned loose together — the joint result. The tornado gives up that joint picture in exchange for something the histogram can’t: a per-variable ranking of which input drives the output most.

Reach for a tornado when you have a quantitative output, several uncertain inputs, and the immediate need is to triage them — to find the vital few worth more analysis and dismiss the trivial many. Skip it when the question is the decision’s shape (use a decision tree), when you care about one variable’s full spread (use a distribution plot), or when you need the combined effect of all uncertainties at once (run a Monte Carlo and read its histogram). The tornado ranks sensitivities; it doesn’t replace the model or the simulation.

How Ora builds it

Ora produces a tornado from a semantic spec — a structured description naming the output metric and its base-case value, then, for each input variable, its low value, its base value, its high value, and the resulting low and high outputs (with a note on whether the input-output relationship runs direct or inverse). That spec is rendered to a horizontal-bar chart: bars centred on the base case, sorted by length descending, input names down the y-axis and the output value along the x-axis. The accessibility layer generates a layered description — the output and its baseline, the most-sensitive inputs and their ranges, the least-sensitive ones, and any inverse-relationship bars that read backwards — plus keyboard navigation across the bars in sensitivity-rank order. The caption is load-bearing: a tornado without an explicit input-range and percentile interpretation isn’t interpretable, so Ora won’t emit one that omits it.

The diagram is the visual face of Ora’s Decision Under Uncertainty mode (its risk-analysis work): when you ask “which assumptions does this answer actually depend on — draw the tornado,” that mode swings each input across its plausible range, computes the swings, and this artifact is how it shows which few drive the result and where buying better information pays off.

The technique comes from the sensitivity-analysis tradition in decision analysis founded by Ronald Howard at Stanford, where the tornado became the standard companion to an influence-diagram model: build the diagram, fill in the probabilities and utilities, compute the recommendation, then run a tornado to see which inputs the recommendation is sensitive to — and whether it flips across any input’s range. It is documented in the canonical decision-analysis texts, including Robert Clemen’s Making Hard Decisions and Howard and Abbas’s Foundations of Decision Analysis. The underlying move — separating the variables the answer depends on from the ones it doesn’t — is Pareto’s vital-few-and-trivial-many applied to uncertainty.

  • Decision Tree — the DECISION-family member that maps the structure of the choice itself (sequential decisions, chance events, payoffs), where the tornado interrogates the inputs to the model behind the numbers.
  • Influence Diagram — the decision model the tornado is the classic companion to: draw the diagram, then tornado its inputs to find which the recommendation is sensitive to.
  • Distribution Plot — the STATISTICAL-family member for the full spread of a single variable; the tornado instead reduces each variable to one swing and ranks many against each other.
  • Decision Under Uncertainty (mode) — the analytical operation this diagram renders: the risk analysis that swings each assumption to find the few that drive the answer and where information is worth buying.