The breaking point
What would have to be true for this answer to be wrong
A forecast with a band around it tells you how uncertain the model is. That is not the question you have. The question you have, standing in front of a board with a decision to make, is which number this rests on, how far that number would have to be off before the recommendation reverses, whether that distance is plausible, what a delay costs, and what is inside the bad case. This answers all five, and the first of them by moving the number until the answer changes sign.
Build a twinThe analysis layer
All of it runs on every analysis, is reported next to the answer rather than buried in a drawer, and says so plainly when it could not be computed.
Sensitivity tells you which number matters. That is a different question from which number could change your mind.
The largest number in a model usually moves the forecast the most, and can never change the decision, because scaling it scales both the thing you are considering and the thing you would do instead. The numbers that can change a decision are the behavioural ones: how customers answer a price move, how fast a rival follows, how much of your base is locked under contract. Ranking by how much a number moves the level put six numbers that cannot reverse an answer at the top of the list.
How it works
Four steps to the breaking point, and none of them is a heuristic
- 01
Establish the sign of the answer
The decision is run against the same company doing nothing, on the same seeds. The difference between them is the answer, and whether that difference is positive or negative is what a breaking point reverses. If the decision makes no difference at all, that is said instead, and nothing further is computed.
- 02
Score every number that has a band around it
About sixty of them on a typical twin. Each is moved to both edges of its own band and the answer is recomputed, at reduced replications because this is a ranking rather than a measurement. What is scored is how far the number moves the answer, not how far it moves the forecast. A number that moved nothing at either edge is reported separately as having no bearing on this question.
- 03
Bisect the ones that carry it
The top candidates are moved outward, to three times their own band, to find out whether a sign change exists in that direction at all. Where one does, nine rounds of bisection close on the value at which it happens, at full replications. Where one does not, in either direction, the number is reported as unable to overturn the answer, which is a useful thing to know.
- 04
Report the value, and where it came from
Gross margin would have to fall to twenty four percent, against the thirty one percent your statement says, before this stops paying. That number came from a file you uploaded, so you can judge how likely it is. If instead it came from an industry default, the most valuable thing you can do next is send the document that settles it, and the analysis says so.
What it found first
Building this uncovered a fault in every what-if the software had ever run
Every run draws each assumption from a distribution over that assumption's band, with the value from your ledger as the peak. That is what makes a forecast a range rather than a single line.
Anything that asked what if this number were different wrote a new value into the model and left the band where it was. That asks for a distribution whose peak is outside its own range. The engine went on drawing from the original band, so a test that said gross margin was sixty two percent was answered at thirty one, and every sensitivity figure the analysis layer produced was a measurement of almost nothing.
Values now travel with their bands, through one function that every such test has to call. The draw refuses a peak outside its own range rather than returning numbers from outside it. And the engine test now runs the same question both ways and asserts the difference, so this cannot come back quietly.
What you get
Three shapes of answer, and all three are useful
A breaking point is computed from this model, and this model is wrong in ways that page on the limits sets out plainly. What it tells you is how much the recommendation depends on each number, which is true of the model whether or not the model is true of the world.
Two decisions
And whether they help each other, or fight
A question naming two decisions gets a fourth run. Neither of them, each of them alone, and both together. The answer is almost never the sum of the two.
Hiring salespeople while raising price puts more people in front of a market that has just been given a reason to leave. Cutting cost while cutting price takes the money out of the thing that was supposed to absorb the cut. Those are interactions, they are the whole reason to own a simulator rather than a spreadsheet, and a plan that counts two decisions at their full separate value is counting some of it twice.
The gap between doing both and adding the two up is reported as a number, with a verdict: they amplify each other, they fight each other, or they barely touch and you can decide them one at a time. Where doing both is worse than doing the better one alone, it says that too.
And when
Every plan has a date on it, and nothing was testing the date
- 01
The same decision, started in four different months
Now, a quarter into the window, halfway, and three quarters. Each one is measured against the same company doing nothing over the same window on the same seeds, so the only thing that differs between the four rows is the month it begins.
- 02
What a month of delay is worth
Reported as a number, in the metric the question was about. A price rise that compounds from whenever it starts loses most of its value to a six month delay. A cost cut in a business that is already shrinking may not.
- 03
Or that waiting is free
Where the answer moves by under five percent of the decision itself across the whole delay, it says so, and that settles the meeting in the other direction. An argument about timing that nobody has costed is an argument about temperament.
- 04
Or that later is better
Which happens, and is worth knowing, because it usually means something in the way has to clear first. Find out what, before committing to a delay for the wrong reason.
The bad case
A band tells you how bad. It does not tell you what went wrong
One run in ten ends below this number. That is a fact about the distribution, and it is the last thing most forecasting tools will say to you. It is not enough to sign anything off on.
So the worst tenth is taken apart. Which month cash was at its lowest, because that is the month to fund rather than the last one. What share of runs ended out of money, because a company that runs out in month seven does not get to enjoy month twelve. Which month carries the widest band, because that is where the model is least sure and the obvious place to put a tripwire. And which severe events showed up in the bad runs that did not show up in the rest.
Where no single event drove them, it says that too, and that is a different situation: a downside made of ordinary bad luck accumulating is harder to hedge against and easier to live through than one built on a named event.
And what it said last time
An answer that reverses itself has to say so
Every analysis is checked against the analyses already stored on the same twin. Same decision, opposite direction, and it is reported as a disagreement with the date of the earlier one.
A model that quietly says the opposite of what it said last month is worse than a model that does not answer, because somebody acted on the first answer. The earlier analysis stays where it is, so the two can be read next to each other rather than one of them replacing the other without anybody noticing.
Ask it something and see what would make it wrong
The free tier takes real files and runs the whole analysis.