A Financial Model Is an Argument, Not a Spreadsheet

Assumptions

The chain has to start somewhere, and where it starts is the assumption. An assumption is a quantity you cannot derive and must instead decide: the growth rate, the gross margin, the collection period, the capex per unit of capacity, the cost of debt. Everything else in a competent model is calculated from these. That is not a stylistic preference; it is the definition of the thing. A typed number sitting in the middle of a calculation is not an assumption, it is a hardcode, and hardcodes are where models go to die, because nobody remembers that they are there.

Good assumptions have three properties. They are few, because a model with four hundred inputs is not a model but a second set of books. They are sourced, meaning each one can be traced to a contract, a historical average, a management decision or a market benchmark, and the source is written down next to the number. And they are owned: someone in the business has agreed that this is what they believe, so that when the world diverges from the model, there is a person to ask why.

The most common failure we see is not a wrong assumption. It is an unexamined one. A financing mix keyed in as sixty percent debt when the promoters had agreed sixty percent equity. One capacity figure in the input sheet and a different one in the revenue sheet. The same capex booked in two entities. None of these are errors of judgement. They are errors of attention, and they survive because nobody asked the assumptions to justify themselves.

How to build the right one

A model is right when it answers the question it was built for and no other. Before a single sheet is opened, the builder should be able to write one sentence: this model exists so that a particular person can make a particular decision by a particular date. A model for a lender needs a debt schedule, cover ratios and a downside case; it does not need a customer cohort analysis. A model for a founder deciding whether to open a second plant needs plant-level contribution and a phased capex schedule; it does not need a weighted average cost of capital to two decimal places. Building for a question nobody has asked is how a model becomes forty tabs of anxiety.

The architecture then follows a discipline older than spreadsheets: inputs in one place, calculations in another, outputs in a third, flowing in one direction. Time runs left to right on every sheet, with the same columns meaning the same periods everywhere. Historical actuals sit beside the forecast in the same layout, so the join is visible. The three statements are linked, and the balance sheet balances because the logic is correct, not because a plug makes it so. Each entity or business line is built on the same template, so that consolidation is addition rather than archaeology.

We once inherited a group model in which the operating entities were built two years out of phase with the consolidation sheet that summed them, with hundreds of broken references in the largest tab and dozens of links to workbooks that no longer existed. The numbers on the front page looked plausible. That is the danger: a broken model does not announce itself. Only structure, and the tedious habit of checking that every total agrees with every other route to the same total, keeps a model honest.

Is a model the same as a valuation?

No, and the confusion costs people money. A valuation is one use of a model. You take the forecast cash flows the model produces, apply a discount rate and a terminal assumption, and arrive at a range of what the business might be worth to a given buyer at a given time. The model is the engine; the valuation is one of the places it can be driven.

The distinction matters because the two are judged differently. A model is judged on whether its logic is sound and its assumptions defensible. A valuation is judged, additionally, on whether the discount rate reflects the actual risk of the actual cash flows, on whether the terminal value is doing eighty percent of the work (it usually is, and that should worry you), and on whether comparable transactions corroborate the answer. A beautifully built model can produce an indefensible valuation if the cost of capital is borrowed from a textbook rather than reasoned from the business. Conversely, the valuation figure is often the least interesting output of the model. The sensitivity table, showing which two assumptions the answer actually depends on, is worth more than the headline.

Where it is used

The same engine drives a surprising number of vehicles. Founders use it to decide whether a business is worth starting and what it would take to fund. Lenders use it to size a facility and test whether cash flows cover service under stress. Investors use it to price an entry and imagine an exit. Boards use it to set budgets and hold management to them. Acquirers use it to test whether the seller's story survives contact with arithmetic. Governments and grant bodies use it to judge whether an incentive is earned. And, most neglected of all, operators use it every month to see whether the business is doing what everyone said it would do, which brings us to the point most practitioners miss.

The model as MIS

The moment a forecast is approved it becomes a budget, and the moment actuals arrive the model becomes a management information system, whether or not anyone calls it that. The value of a model does not end at the fundraise or the board meeting. It begins there. Each month, the actual revenue, margin, collection period and headcount are laid against the assumptions, and the variance is the most useful sentence the business will read that month, because it tells management which of their beliefs about the business was wrong and by how much.

A model built for this purpose looks different from one built for a pitch. It carries the historical columns with the same care as the forecast. It defines its metrics precisely enough that the finance team can produce them from the ledger without interpretation. It rolls forward: the forecast for the remaining months is re-based on what has actually happened, so that the year-end view is always current. Most models never make this transition, because they were built as a one-time artefact and abandoned once the money landed. The ones that do become the operating rhythm of the company.

The value of a good model

The value of a good model is that it converts an argument about opinions into an argument about assumptions. Two people who disagree about whether a business is fundable can argue for hours. Two people looking at the same model can locate their disagreement in twenty minutes: one believes collections will take sixty days and the other believes ninety, and the model shows exactly what that difference costs in working capital. That is a smaller, more tractable, more honest disagreement, and it is the one worth having.

A good model also compounds. Built once with discipline, it serves the fundraise, then the lender, then the board, then the monthly review, then the next fundraise, each time with the historical record growing and the assumptions refined by evidence. A bad model is rebuilt from scratch at each of those moments by whoever is unlucky enough to inherit it, and each rebuild loses the memory of why the last one said what it said.

The human in the loop

Human in the loop is usually described as a safeguard, something bolted onto an automated process to catch its errors. That understates it. In financial modelling the human is not the safeguard around the work; the human is where the work is. There are four points in the life of a model where a person is not optional, and a machine, however capable, cannot stand in.

Framing. A model answers a question, and the question comes from a person who understands what decision hangs on it and who will be reading. A machine asked to “build a model for this business” will build a competent generic one. That is the forty-tab problem, now produced in minutes.

Owning the assumptions. An assumption is a belief, and a machine cannot hold one. It can propose a growth rate from the historical trend; only a person can say that this is what the business intends to do and put their name beside it. When the world diverges from the model, the board asks that person, not the software.

Verifying. Every figure that leaves our firm is traced back to a source document by a person before it does, and every total is agreed by a second route. This is not distrust of the machine; it is the same standard we applied to analysts, and to ourselves, before the machine arrived. The difference is that the machine produces more output, faster, so the verification step is now the bottleneck, and it is the right bottleneck to have.

Signing. Someone carries the consequence. A lender who was misled by a model does not accept that the model was drafted by software, and neither should a client.

The machine has made the draft cheap. It has made the reviewer expensive. Firms that understand this will spend the saved hours on judgement; the others will spend them producing more drafts.

What we build

Dissent builds financial models as a practice, not as a by-product of other work. The engagements fall into six kinds, and each is built on the architecture described above so that a model made for one purpose can be carried forward to the next.

The work so far has been concentrated in logistics and supply chain, metals and recycling, consumer brands, real estate, healthcare and wellness, and early-stage ventures across sectors, in India, Singapore and the Gulf. The sectors change; the discipline does not.

Simplicity, here as elsewhere, is not the absence of rigour. It is what rigour looks like once it has finished.