Your Financial Model Is Probably Lying To You

Most quantitative answers in business are dressed in math that isn't a measurement — that's false precision: a guess with a confidence interval drawn around it.

Revenue projections, polls, attribution reports, AI benchmark tables — they all look like measurements. What they actually are is guesses dressed in mathematics: about inputs, distributions, and how variables move together, inherited from the previous version of the same guess. It's not analysis. It's dressed-up guessing.

Two decimal places. The confidence interval. The curve sweeping into the future. None of those track the underlying inputs. The inputs are guesses; the outputs are guesses multiplied together. Everyone checks the multiplication. Nobody checks the guesses. The capital allocated off those outputs doesn't know it.

A typical board-level revenue forecast at a $10M–$50M Series B SaaS company rests on three guesses: CAC, churn, sales cycle length. Nobody can name where the original assumption came from, or defend it from first principles.

(I once spent a week tracing the original source of an assumption in a Fortune 500 forecast. Nobody could name it. The math had been compounding on top of itself for three years.)

Run this the next time someone shows you a quantitative answer that's supposed to change what you do: ask them to name the three assumptions that would have to be wrong for the answer to change by thirty percent. Without irony. The ask itself is unusual. The presenter may think you're being adversarial. You're not — you're testing the math.

The thirty-percent figure is arbitrary but defensible. Anything less and the presenter can talk you into the result; anything more and the model would already have to defend itself. Thirty is where the model earns its place or gets retired.

If they can name the three assumptions, and tell you how confident they are in each, you have something. If they cannot, you have a curve. Most of the time, they can't. The guess got made in a private conversation, or pulled from a previous model, or inherited from an industry benchmark. By the time it reaches the room, the guess is a fact.

A model is not its output. It's a stack of decisions about the world. If you cannot name the three that would have to be wrong to change the answer by thirty percent, the model is just a story about a curve.

The categories that fail most are the ones used most. Pipeline forecasts. CAC payback. Hiring plans. Political polls assume a likely-voter screen and weighting by education or region — change any one and the result moves by ten to thirty points. AI benchmarks assume a specific test, prompt template, configuration, and rubric. I have watched the score on a "settled" benchmark move thirty points in an afternoon when any one of those specifics changed — and a benchmark that swings that much on a single rubric tweak isn't settled, it's a calibration problem.

The model you are about to defend probably is one, because you cannot name the three assumptions.

The model has two or more decimal places of precision on its outputs.

The presenter can't name the three assumptions that would have to be wrong for the answer to move by 30%.

The confidence interval was drawn after the guess, not before.

The math has been audited but the inputs haven't.

The model was built by inheriting assumptions from a previous version of the same model.

The output is being used to allocate capital without re-checking the inputs.

"Show me the math" produces a spreadsheet, not an explanation of the assumptions.

A model is a stack of decisions about the world. If you cannot name which three would have to be wrong, you have a wish, not a model. See our financial modeling guide for the structural cleanup that makes one auditable.

The fix is small but expensive. It costs the room its false confidence and the presenter the protection of the spreadsheet. In exchange it gives you a forecast you can defend in a year, and a meeting where the guess gets named before the capital does.