Three forms show sensitivity and they answer three different questions. Choosing the wrong one is the commonest quantitative sin in a research deck after the dual axis.
A two-way grid moves two drivers against each other and answers what the value is at each combination. A tornado moves many drivers one at a time and answers which driver matters. A scenario set moves many drivers together coherently and answers what the world looks like if the story is wrong. A grid can't rank drivers, a tornado can't show interaction and neither can describe a coherent world.
What goes wrong with a two-way sensitivity grid?
The grid has a specific failure that's easy to show. Cross Cascade's discount rate against terminal growth, five values by five values, and the grid spans $21.28 to $41.29 per share. The maximum is 1.94 times the minimum, and the range is 70.4% of the base case. A grid that wide tells the reader the model can produce nearly any answer. That's the opposite of what the exhibit was meant to prove.
Advertisement
Trim it to the middle three by three, with the discount rate from 8.9% to 9.9% and terminal growth from 1.9% to 2.9%, and you get $24.41 to $33.79, a range of 33.0% of the base case. Now you have a usable exhibit. It keeps 46.9% of the full grid's range while removing 16 of the 25 cells.
Trimming is fair to the reader on one condition. State in the footnote what you removed and why, and make the reason about the driver. A discount rate of 8.4% is below the range any reasonable cost-of-equity build produces for this company, so the column goes. A terminal growth rate of 3.4% approaches the long-run nominal growth of the economy, so that column goes too. Trimming until the answers look tight isn't acceptable. That's choosing the exhibit's message from the exhibit's output.
How does a tornado chart rank the drivers?
The tornado is sorted by swing, which is just the ordinary sort-by-value rule, and it's the one sensitivity exhibit that ranks. For Cascade, over a plausible range on each driver:
| Driver, and the range tested | Low ($) | High ($) | Swing ($) | Share of total swing |
|---|---|---|---|---|
| Discount rate, ±80 bp | 24.92 | 32.85 | 7.94 | 32.3% |
| Capex intensity, ±80 bp of revenue | 25.78 | 31.08 | 5.30 | 21.6% |
| EBITDA margin, ±100 bp | 25.95 | 30.92 | 4.97 | 20.2% |
| Terminal growth, ±50 bp | 26.43 | 30.74 | 4.32 | 17.6% |
| Revenue growth, ±2.0 points a year | 27.41 | 29.47 | 2.05 | 8.4% |
The top two drivers carry 53.9% of the total swing, and that's the sentence the action title should make. It also tells you where the next two hours of work belong. Cutting the discount rate by 20 basis points adds $1.0078 a share, while adding a full point to annual revenue growth adds $0.5153. So the cost-of-capital build is worth nearly twice as much attention per unit of revision.
What can't a tornado chart show?
One limitation has to be stated on the exhibit itself. Don't assume the reader knows it. A tornado moves one driver at a time, so it can't show the interaction that the independent valuation bridge also missed. Where the drivers are correlated, and growth and margin usually are, the true joint swing can be wider than the widest bar.