Tornado diagrams: reading a sensitivity analysis at a glance

A tornado diagram is a horizontal bar chart that ranks variables by how far each one moves a single outcome. One bar per variable, the widest at the top, every bar measured against a baseline running down the middle. The shape gives the chart its name, and the top bar is the answer most people came for.

One bar per variable, widest bar on top

Everything on the chart comes out of a sensitivity analysis, which is the arithmetic of recomputing an outcome while one input at a time moves across its plausible range. The tornado is the ranked picture of what that arithmetic returned, so the diagram is a presentation rather than a technique in its own right, and it can be no better than the analysis sitting behind it.

Three things are on the chart and nothing else. A vertical line for the base case, which is the outcome you get when every input sits at its expected value. A horizontal bar for each variable, drawn from the outcome produced by that variable’s low value to the outcome produced by its high value. And an ordering, widest bar at the top, narrowest at the bottom, which is the only reason the picture is worth drawing rather than tabulating.

The outcome has to be chosen before anything else, because every bar is denominated in it. Pick project cost and the axis is money. Pick forecast completion and the axis is weeks. Pick net present value, contract margin or defect rate and the axis is those. One chart, one outcome, and a tornado that mixes two of them is not a tornado, it is two charts sharing an axis.

What the bar length means is worth saying plainly, because it is the part people misread. Length is the distance between two outcomes, not a probability and not a score. A bar nine weeks long says that this variable, on its own, can put the finish date anywhere inside a nine week window. Whether the far end of that window is likely is a question the chart does not answer and does not pretend to.

Ranging each input while the rest hold still

The construction is mechanical, and its one rule carries the whole method: move a single input, hold everything else still, record what happened, put it back. NASA’s cost estimating handbook sets out five steps in its appendix on analyses for decision support, which are to compute the point estimate, select the elements for analysis, determine the range of values for each element, determine the impact, and graph or table the results. Its worked example is explicit about the fourth step: the high and low values are “applied to one component at a time”.

That is what makes the bars comparable. Each one is a measurement of a single variable’s influence taken under identical conditions, which is the only way a nine week bar and a three week bar mean anything set next to each other. It is also, in one sentence, the technique’s largest limitation, because projects do not move one variable at a time.

The soft part is step three. Ranges do not come out of the model, they come out of the people who know the work, and a variable given a generous range will outrank one given a cautious range regardless of which actually matters more. Two disciplines keep that honest. Take low and high from the same kind of evidence for every variable, whether that is historical outturn, supplier commitment or expert judgement, rather than mixing a contractual worst case for one input with somebody’s instinct for another. And write down where each range came from, next to the chart, so that a bar can be challenged on its evidence rather than on its length.

The same numbers support more than one picture, and the choice is not neutral.

Presentation	Reads best for	Loses
Tornado diagram	Ranking. Which few variables dominate, at a glance.	Detail between the ends. Only low, base and high survive.
Spider diagram	Response shape. Which variables react steeply and which are flat.	Legibility past four or five lines.
Sensitivity table	Audit. Exact values, ranges and their sources.	The ranking. Nobody sees it without reading every row.

Reading the chart: the top bar is where the delay lives

Asked which task has the greatest potential to cause delay, a schedule tornado answers directly: the one on the top row. On the chart above that is the third party data feed, whose nine week range is wider than the two bars beneath it and more than four times the range of the bottom one. Nothing else on the chart can move the finish date as far.

Read length first and direction second. A bar sitting mostly to the right of the baseline can hurt far more than it can help, which is the normal shape for schedule variables and the reason the tornado usually leans one way. Integration rework in the example has no left segment at all, because rework that goes well does not happen and cannot pull the date in; its entire range is exposure. A bar that is genuinely symmetrical is telling you something different, that the variable is an estimate with real spread rather than a hazard, and those two things want different responses.

Then read where the bars stop mattering. Below the third or fourth row the ranges usually collapse into the noise of the estimate itself, and a variable that can move the outcome by two weeks on a forty week project is not worth a monitoring regime. The value of the ordering is as much in what it dismisses as in what it promotes.

Total float is the check that stops the reading going wrong. A schedule tornado is built by moving durations and re-running the schedule, so its bars already reflect the network, but they reflect it as it stands today. An activity with three weeks of float and a six week range crosses onto the critical path partway through its own bar, and one with float to spare never appears on the chart at all despite being the thing everyone worries about. The tornado ranks potential to move the date. The schedule tells you the point at which potential becomes actual.

The last reading is the one that turns the chart into a decision. HM Treasury’s Green Book asks appraisers to “conduct sensitivity analysis to examine how changes in key assumptions, like costs, might affect the summary metrics of the appraisal”, and then to go one step further and calculate switching values, “the values that a key assumption would need to change to, in order to make an option no longer value for money”. A tornado bar shows how far a variable could move. A switching value shows how far it would have to move before the answer changes. Put the two together and a nine week bar becomes a specific question: does the delay this variable can cause cross the date the business case depends on, and if so, at what point.

Which bars earn a line in the risk register

A tornado diagram ranks variables. A risk register tracks risks, with owners, responses and dates. The two are not the same list, and copying the top three bars straight into the register produces entries nobody can act on, because “data migration complexity” is a subject rather than an event.

The translation is worth doing deliberately. Take the top bars, and for each one ask what specific event would push the variable to its high value, who would see it first, and what could be done about it. The third party data feed’s nine weeks might resolve into two register entries, a risk that the supplier’s extract format is not what the interface spec assumed, and a dependency on a data sharing agreement that has not been signed. Both are actionable. The bar they came from is not.

Scoring those entries is where the analysis pays off a second time, because a bar gives you an evidenced impact figure rather than a number somebody chose to feel proportionate. That figure is the inherent impact, before any response, and the gap between it and what remains afterwards is what the response actually bought. The distance between inherent and residual scores is the same measurement in a different notation.

What a tornado diagram cannot tell you

Four limits, and the first is the one that gets a chart quoted in a steering committee it should not have reached.

A tornado carries no probabilities. Bar width is the size of a possible movement, never its likelihood, so a wide bar attached to a remote scenario outranks a narrow bar attached to a near certainty every time. That is a defensible way to prioritise attention and an indefensible way to size a contingency, and the two get confused whenever the chart is read as a forecast. When the question is how much reserve to hold rather than what to watch, the tornado is the wrong instrument and a simulation over the joint distribution is the right one.

It also assumes the variables are independent. Each bar was measured with everything else held still, so nothing on the chart can show a correlation, and correlated variables are exactly how projects actually go wrong. A supplier delay that also drives up specialist costs while pushing testing into a holiday period appears on the chart as three modest bars in different rows.

Which is the third limit, and the reason the range column deserves as much review as the chart. A variable left out entirely has no bar, and a chart with no bar for it looks exactly like a chart where that variable does not matter. The ordering is a claim about the analyst’s judgement as much as about the project.

And no tornado tells anyone what to do. The chart is a prioritisation instrument, and its whole output is a shortlist and an order. What the shortlist is worth depends on what happens next, which is whether the top two or three bars turn into responses with owners, or into a slide that gets shown once and filed. That part has never been a modelling problem.