Financial modeling
A model is an argument about how the business works, expressed in arithmetic. Its value is the argument, not the output precision.
Structure
Three separated layers, always:
- Inputs — every assumption, in one place, each with a source and a date. An assumption buried inside a formula is invisible and therefore never challenged.
- Calculations — no hard-coded numbers. Ever. A constant inside a formula is an untraceable assumption.
- Outputs — the statements and the summary a decision-maker actually reads.
One row, one calculation, carried consistently across periods. Models become unauditable through inconsistent rows more than through complexity.
Build revenue from drivers
Never grow a top-line by a percentage. Build it: volume × price, or accounts × retention × expansion. Driver-based models can be argued with, and being argued with is the point — a growth rate cannot be wrong, only optimistic.
Cost structure separated into fixed, variable, and step-fixed. The step-fixed items are where plans break, because they move in jumps nobody modeled.
Sensitivities are the deliverable
A single-scenario model tells you nothing about risk. For every model, produce:
- Which two or three assumptions actually move the answer. Usually far fewer than expected.
- Breakeven on each — how wrong can this be before the decision reverses?
- Downside case — not a haircut on the base case, but a coherent story where things go badly.
If a plan only works in the base case, that is the finding.
Reviewing someone else's model
The description of a model is not evidence about the model. Check these, in this order, because each one invalidates everything after it.
- Trace one number end to end. Pick an output that matters and follow it back to inputs. If you cannot, nobody else has either, and the model has never actually been reviewed.
- Find the hard-coded constants. Search the calculation area for typed numbers. Each one is an assumption that escaped the input sheet, and they are where overrides hide.
- Check the row consistency. A formula that differs partway across a row is either a deliberate change nobody documented or an error, and the two look identical.
- Test the extremes. Set a key driver to zero and to double. Models frequently break, go negative in impossible ways, or fail to respond at all — which tells you the driver is decorative.
- Check that the statements tie. Cash flow reconciles to the balance sheet movement; the balance sheet balances in every period, not just the first.
- Ask what is missing. Working capital, hiring lag, churn, price changes, tax, and the step costs that come with growth are the omissions that flatter a plan most.
Then find the assumption doing the work. Most models rest on one or two numbers, and those are usually the least evidenced. Ask where each came from and what it is based on — the answer is frequently that it was chosen to make the case work, which is a fine thing to know before relying on it.
Presenting
Lead with the answer, then the two assumptions it rests on most heavily, then what would change it. Never present a model without stating what it is most sensitive to — the recipient will assume robustness you did not claim.
Sources
references/sources.md in this skill lists the outside authorities that settle the questions
here — what each one is authoritative for, and what you may do with it. Check them before
answering on anything they cover, and cite what you used. Most are free to read and not free
to reproduce; the use note on each is binding.
Never
- Report a number to more precision than the assumptions support. Five significant figures from a guessed growth rate is false confidence.
- Build a model whose logic you cannot explain in three sentences.
- Change an assumption to reach a desired output without labeling it as a target case.
