What-If Analysis in Financial Models: Structuring It So It Changes a Decision

What-if analysis is the practice of changing assumptions in a model and watching what the output does. That is the whole definition, and it is why the term covers everything from a five second cell edit to a formal multi case framework. What separates useful what-if work from a wall of numbers is not the technique. It is whether the question being asked could change a decision.

Most models accumulate what-if output that nobody reads. The fix is structural: decide what decision is open, work backwards to the question that would settle it, and only then start moving inputs.

Three shapes of what-if question

It helps to notice that what-if questions come in distinct forms, and they get answered differently.

Forward questions. “What happens to the output if this input takes that value?” You set an input, you read an output. This is the default form, and it is the least useful one, because the answer arrives without a reference point. Told that an output moves to some level, a reader still has to work out whether that matters.

Threshold questions, sometimes called goal seek or breakeven. “How far would this input have to move before the conclusion changes?” You fix the output at a decision boundary and solve backwards for the input. This is almost always the better framing, because the answer arrives as a testable statement about the world: growth would have to fall below some level, or the margin would have to compress by some amount, before the case stops holding. That is something you can then go and check against disclosures, industry data, or management commentary.

Attribution questions. “Which of these changes drove the difference between the old answer and the new one?” You move from one full set of assumptions to another and decompose the gap. This is what turns a revised model into an explanation rather than a replacement.

Most analysts run forward questions by default. Threshold and attribution questions are where the decision content lives.

Structuring the work

Start from the decision, not from the model. Write the open decision as a sentence before touching a cell. If no decision is open, the what-if run is documentation, and it should be scoped and time boxed like documentation rather than expanded indefinitely.

Define the boundary that matters. Every decision has a level at which it flips. It might be a valuation, a coverage ratio, a return on capital threshold, a portfolio weight limit, or simply a comparison against an alternative use of the money. Naming the boundary explicitly is what lets you run threshold questions instead of forward ones.

Move inputs, and know which ones are genuinely independent. As in sensitivity analysis, the one at a time approach only makes sense for inputs that can plausibly move alone. Where inputs are linked, move them as a group and be clear that you have done so, which is the territory of scenario analysis.

Version the base case, and never edit it in place. A live model where someone has been flexing assumptions is no longer the model that produced the original conclusion. Keep the base case as a stored, labelled set of inputs that can be restored exactly. Models that lose their base case are the single most common source of numbers in a memo that nobody can reproduce.

Log what you changed. Every what-if run should leave a trail: which inputs moved, from what to what, and what the output did. Without this, a set of results is uninterpretable a week later, including by the person who produced it.

How to do it well

Convert every result into a sentence about the world. A number in a grid is not a finding. “Revenue growth would have to run below the level implied by current order book disclosure for two consecutive years before the conclusion changes” is a finding, because it names something observable. If a what-if result cannot be written that way, it probably was not worth running.

Attach a tracking marker to each threshold. Once you know the boundary, decide what evidence would tell you it is approaching. That may be a quarterly disclosure, a segment level trend, a capacity utilisation number, a working capital metric, or a shift in guidance. This is what turns a one off exercise into monitoring, and it connects to the discipline of monitoring a portfolio of holdings.

Prefer few, sharp questions to many shallow ones. A model that can answer forty what-if questions will usually be asked all forty and read on none. Three well chosen threshold questions with clear answers beat a forty cell grid every time.

Keep the units and the basis consistent. A large share of what-if errors are not conceptual but mechanical: a ratio flexed as a percentage where the model expects a fraction, a consolidated assumption applied to a standalone base, a quarterly rate applied annually. Label units in the input block and check the direction of every result against intuition before reporting it.

Test the model at the extremes before trusting it in the middle. Push an input to an absurd value and check that the output moves in the right direction and does not break. Models with hardcoded values, circular references or broken links often behave correctly across the small range anyone tests and produce nonsense outside it. Finding that early is much cheaper than finding it in a committee meeting.

Separate assumptions from calculations physically. Every input a what-if run might touch should live in one clearly marked block, not buried inside formulas. This is basic modelling hygiene, and it is what makes what-if work auditable by someone other than the author.

Test the same question historically, with care. If you want to know whether a threshold rule would have been informative in the past, you have to feed it the data as it stood then. Restated financials and today’s index membership both leak information backwards, which is the problem covered in why point-in-time data matters and what is lookahead bias.

Where it breaks down

It only explores inside the model you built. What-if analysis cannot discover that the model has the wrong structure, misses a subsidiary, uses the wrong accounting basis, or ignores a real driver entirely. Every answer it gives is conditional on the model being a fair representation, and that condition is the one most likely to fail.

Wide output ranges get read as knowledge. Producing many results feels like rigour. In practice, a large grid of outputs without an accompanying judgement about which regions are plausible transfers the analytical work to the reader, who usually does not do it and instead anchors on whichever number appears first.

It carries no probabilities. Every what-if answer is conditional. Knowing what happens if an input takes a value tells you nothing about whether that value is likely. Combining what-if outputs into anything resembling an expected value requires probability judgements that the technique itself does not supply, which is the same limitation that applies to stress testing.

Linked variables get moved apart. The convenience of flexing one cell encourages combinations that cannot occur: volumes down with prices unchanged, growth up with working capital unchanged, capex flat with capacity rising. These produce answers that are arithmetically valid and commercially impossible.

It is trivially easy to steer. Because the analyst chooses which questions to run and which results to report, what-if analysis is one of the easiest techniques to use as confirmation rather than investigation. The tell is that only the questions supporting the existing view were run. The countermeasure is to write the question list before seeing any answers, and to report the full list including the results that were unhelpful.

Threshold answers can be brittle. A breakeven computed from a model with one dominant assumption inherits all that assumption’s fragility. A threshold expressed to two decimal places implies a precision the underlying model does not have, and rounding it deliberately is usually more honest than reporting it exactly.

It does not tell you whether the decision boundary is right. The technique tells you where the model flips. Whether that boundary is the correct one, whether the valuation approach is appropriate for the business, whether the benchmark is fair, are all judgements made outside the analysis. As the argument in why the P/E ratio is not enough shows, running elaborate what-if work on top of the wrong framing produces sophisticated confidence in the wrong answer.

The discipline that makes what-if analysis worth doing is narrow and unglamorous. Ask fewer questions. Frame them as thresholds. Write the answers as statements about observable things. Log what you changed. Everything else is a spreadsheet exercise.

Related reading

• Portfolio and Backtest Metrics, Explained: the hub guide to the metrics and methods behind portfolio analysis.
• Sensitivity Analysis Explained: the formal one input at a time version, and how to rank assumptions.
• Scenario Analysis Explained: moving linked assumptions together into coherent futures.
• Monte Carlo Simulation in Investing: running the same what-if question thousands of times.
• What Is Lookahead Bias?: why testing a rule on today’s data flatters it.

This article is educational. Altys Labs is not a registered research analyst or investment adviser, and nothing here is investment advice or a recommendation to buy, sell, or hold any security.