Two families exist, and they answer different questions. Deterministic one-at-a-time analysis moves each input across a range while holding the others fixed — Excel data tables can do it — and answers “what happens if this one number changes?”. Simulation-based analysis runs a Monte Carlo simulation and ranks every uncertain input by how much of the output's variation it actually drives, with all inputs moving together as they do in reality. That ranking is what a tornado chart reports, and it is the one that tells you where to spend your next hour of analysis or your next euro of de-risking.
This page explains how to read a tornado chart properly, using a fully worked example whose every number you can check — and the three classic traps that make tornado charts lie to careful people.
A tornado chart is a ranked bar chart of how much each uncertain input drives a simulated output — widest bars on top, which gives it the funnel shape and the name. The type matters. A rank-correlation tornado scores each input from −1 to +1; it is simple, but the scale means nothing to a decision-maker and it misbehaves when inputs are correlated with each other. The conditional-mean tornado is the more informative choice: each bar spans the output's average when that input sits in its lowest values versus its highest — in the output's own units. A board reads “civil works swing the expected total by €2.3m” without any statistics briefing.
For project schedules rather than costs, the same chart ranks tasks and risks by their effect on the completion date — that is the version Tamara draws for Primavera and Microsoft Project schedules.
The model is the €12m capital project from our cost contingency guide — ten line items with three-point ranges, six of them correlated at ρ = 0.7 through a common market/productivity driver, plus two discrete risk events. It is a free download, so every bar below can be reproduced rather than believed. The conditional-mean tornado of total project cost, from a 400,000-sample run (means of the total, €000, when each input sits in its bottom versus top 20%):
Three things in this table teach more than any definition. First: the top six bars are really one bar. All six correlated items show near-identical €2.1–2.3m widths — not because each item individually swings the project by that much, but because each bar carries the whole shared market driver. Read them as a group (“the market/productivity driver dominates this project”), never individually, and never add bars together. The giveaway is temporary works: a €250k line item with the sixth-widest bar, outranking €900k of design work (bar: €234k) purely through correlation. Second: the ground-conditions risk is one clean bar of €1,277k — the largest genuinely independent driver — because it is modelled with VoseRiskEvent. Model the same risk as a probability cell times an impact cell and its sensitivity splits across two shorter bars, quietly demoting your biggest single risk. Third: the ordered mechanical package barely registers (€56k) despite being the second-largest line item — its price is contracted, its range is narrow, and the tornado correctly tells you to spend zero further effort on it. Where the effort should go instead is a question with real money attached: the contingency guide's €150k ground-survey scenario shows what resolving the top independent bar is worth.
1. The monotonic assumption. A tornado assumes each input pushes the output in one direction. An input with a U-shaped or peaked effect — a temperature with an optimum, a staffing level that hurts both under and over — averages its two tails and shows a short bar despite mattering greatly. When the relationship might not be monotonic, use a spider plot, which draws the shape of each input's effect across its full range, or a scatter plot of input against output.
2. Split risk events. As the worked example shows: a risk written as =probability × impact in separate cells appears as two weak bars instead of one strong one. Keep probability and impact inside a single VoseRiskEvent and the tornado ranks the risk as decision-makers think of it — one thing that either happens or does not.
=probability × impact
3. Correlated inputs read individually. Correlation is usually the right modelling choice — ignoring it understates the tail, as the contingency guide quantifies — but it changes how the tornado must be read: correlated bars share credit for their common driver. If you need each input's own contribution, examine the driver itself as an input, or rerun the tornado with the correlation switched off and compare (the downloadable model has exactly that switch built in).
In ModelRisk, mark each uncertain cell with VoseInput and the result cell with VoseOutput, run the simulation, and choose Insert → Tornado → Conditional Mean in the results viewer; the spider and scatter plots for trap-checking live in the same menu, and every chart exports to PowerPoint, Word or PDF for the report. The worked-example model is the fastest way to try it: download, click Start, and reproduce this page's table before pointing the same analysis at your own estimate. If the tornado then changes which input your team investigates next, our guide to the value of information puts a euro figure on whether that investigation is worth commissioning. And for where the tornado sits in the full workflow — the methods, the model structure, and the other outputs a review board expects — see our overview of cost risk analysis.
How do you run a sensitivity analysis in Excel? Two families exist. Deterministic one-at-a-time analysis moves each input across a range while holding the others fixed — Excel data tables can do it — and answers what happens if this one number changes. Simulation-based analysis runs a Monte Carlo simulation and ranks every uncertain input by how much of the output's variation it actually drives, all inputs moving together — that is what a tornado chart reports.
What is a tornado chart? A ranked bar chart of how much each uncertain input drives a simulated output, widest bars on top. The most informative type is the conditional-mean tornado: each bar spans the output's average when that input falls in its lowest values versus its highest, in the output's own units, so a decision-maker reads euros rather than an abstract correlation score.
Why do correlated inputs make tornado bars misleading? Because each correlated input's bar carries the whole shared driver, not that input's own contribution. In the worked example, six items sharing one market driver all show near-identical €2.2m bars — collectively real, individually overstated — and a €250k line item outranks a €900k one purely through correlation. Read correlated bars as a group, and never add bars together.
How should risk events appear in a tornado chart? As one bar per risk. Modelling a risk as a probability cell multiplied by an impact cell splits its sensitivity across two bars and understates it; modelling it with VoseRiskEvent keeps probability and impact together, so the ground-conditions risk in the worked example shows as a single €1.28m bar.
What does a tornado chart miss? It assumes each input pushes the output in one direction — an input with a U-shaped or peaked effect can show a short bar despite mattering greatly. When you suspect a non-monotonic relationship, use a spider plot, which shows the shape of each input's effect across its full range, or a scatter plot of input against output.
Conditional-mean tornados, spider plots and scatter plots on any Excel model — run the free 15-day trial and see which of your inputs actually drives the answer.