Sensitivity Analysis: Testing One Input at a Time
Sensitivity analysis changes one input at a time, holding everything else fixed, and records how much the result moves. Its purpose is ranking: which input dominates the outcome and which barely matters, so attention and checking effort go where an error is most expensive. Sensitivity analysis deliberately ignores that inputs move together; building the linked world is scenario analysis's job, not a flaw to fix here.
Scenario analysis moves whole worlds. Sensitivity analysis does the opposite: it isolates. The deliverable is not a forecast, it is a priority list, built to be read at a glance and never to be mistaken for a scenario.
What question does sensitivity analysis actually answer?
Not what will happen. Which input matters most. Any cook learning a new dish knows the method: change only the salt, taste; change only the flame, taste. Change both at once and neither change can be told apart from the other, so the tasting settles nothing. Isolation is what buys attributionBeing able to say which input caused the change in the result. Attribution requires moving one input while the others hold still., and attribution is the entire product here.
Tessora Weaves carries four forces on its exposure map: volume, rupee realisation, cotton yarn, the floating rate. The question is which one of the four moves profit before tax (PBT) most across its plausible range. Answering that shows where a checking hour is worth the most.
Why does the method insist on holding everything else still, when in reality inputs move together?
How is a fair test built for each input?
Give each input its own plausible range, never one uniform percentage. Here is why: one per cent on rupee realisation and one per cent on the floating rate are completely different events, one routine, one substantial. Swing every input by a flat 10 per cent and the ranking is rigged before it starts. The fair test moves each input across the range it could plausibly cover, stated openly: volume 10 per cent either way, realisation 5, cotton 10, the rate one point. The ranges are judgements, and stating each range openly is what makes them honest.
Before computing: one per cent on rupee realisation, against one per cent on the floating rate. Which of the two moves Tessora Weaves' PBT more, and by roughly what multiple?
How are the results ranked and read?
Run each input across its range, record the PBT swing, and stack the bars widest-first. Practitioners call the result a tornado chart, for its shape. Tessora Weaves' tornado runs like this: volume 10 per cent moves PBT Rs 2,40,00,000; realisation 5 per cent also moves Rs 2,40,00,000; cotton 10 per cent moves Rs 1,44,00,000; the rate one point moves Rs 10,00,000. Read the bars as a priority list: the top bars get the checking hours, and the bottom bar earns permission to be rough.
| Input, over its stated range | PBT swing | Rank |
|---|---|---|
| Volume, 10 per cent either way | Rs 2,40,00,000 | joint first |
| Rupee realisation, 5 per cent either way | Rs 2,40,00,000 | joint first |
| Cotton yarn, 10 per cent either way | Rs 1,44,00,000 | second |
| Floating rate, one point either way | Rs 10,00,000 | a distant last |
Predict the order of the four bars before the simulation draws them live: volume, realisation, cotton, rate, over their stated ranges.
Pick one input. Move it. Watch the tornado answer.
One input live at a time, the other three locked at base, shown locked. Pick an input, then drag it across its own plausible range.
What does the ranking change about where attention goes?
Everything about how the checking week is spent. A small error in a top-ranked input costs more than a large error in the bottom one: mis-estimate volume by 2 per cent and PBT is off Rs 48,00,000; mis-estimate the rate by half a point and it is off Rs 5,00,000. So the demand and currency assumptions get the calls, the confirmations and the second looks, and the rate assumption is allowed to be a round number. Precision is a budget; the tornado shows where to spend it.
The ranking puts the floating rate a distant last. A colleague wants to spend the week refining the rate assumption to two decimal places. What does the ranking say?
Where does sensitivity analysis stop and scenario analysis begin?
At the moment two inputs must move together, sensitivity analysis has been left behind. The objection "but a demand fall would also weaken the rupee" describes a linked world, and building it honestly is the discipline of scenario analysis. The two tools are a pair, each the opposite of the other. Sensitivity moves one input and holds the rest, and its output is a ranking. Scenario moves every linked input together, and its output is a range. Each is wrong at the other's job, and most misuse is one tool doing the other's work with the wrong label on top.
An analyst moves cotton yarn down 10 per cent, sees PBT rise Rs 1,44,00,000, and reports it as "a downside protection case". What actually happened?
How does the tornado change as the business changes?
The ranking is a photograph, not a portrait: it ages. Suppose Tessora Weaves locks next year's exchange rate with its bank for most of its invoices. The realisation bar, joint first today, shrinks toward a sliver, and suddenly cotton yarn is the second-biggest force in the business. Every structural change, a hedge, a new buyer, a renegotiated loan, redraws the tornado, and a team still checking last year's top bar is guarding a door that has been bricked up while the open one stands unwatched.
So the working routine is a loop, not a one-off: build the base, set each range with its reason, run the inputs one at a time, rank, point the checking effort at the top bars, and rerun the whole exercise whenever the base case or the business's structure changes. The tornado is cheap to redraw; a stale one is expensive to believe.
Tessora Weaves hedges its currency for the year. The old tornado had realisation joint first. Where should the checking hours move now?
How often should the tornado be redrawn?
The error that gets made, and what it costs
A report titled "downside case" contains one moved input and three frozen ones. The arithmetic inside is correct; the title is the error. A reader plans against it believing a world was tested, when only a knob was turned. The title promised a world; the method delivered a ranking entry.
The cost is a downside estimate that nobody argued as a world, carrying the credibility of one that was.
Which deliverable should each tool hand over at the end: sensitivity, and scenario?
References
| Source | Document | Where |
|---|---|---|
| National Stock Exchange of India (NSE) | Listed-company presentation material where sensitivity analysis is used | nseindia.com |
Tessora Weaves Private Limited is invented.
Educational material. Not advice on any investment, tax, budget or market position.