Modelling a manufacturer

It’s the second week of the month, and four things have happened.

Second week of the month Four things, one question

1. Intake Raw milk intake is settling four percent above the contracted band.
2. Energy Diesel is up eleven percent, which will reach cold-chain distribution about six weeks from now.
3. Capacity Line 3 at the Ostrów plant is coming down for a fortnight in August for a rebuild that was scheduled for October.
4. Demand The French retailer has confirmed the new listing: six codes, first delivery 1 April, with a promotional grid attached and a rebate tier that steps at €4m.

The CFO would like to know what Q3 gross margin does. By country. By channel. Before Thursday.

Not one word of that question appears in a three-statement model.

That is not a criticism of three-statement models. The P&L, the balance sheet and the cash flow are where the answer lands: they are the report card, and a good one. They are not where the answer is computed. The computation lives a floor or two below, in a structure that knows what a litre of raw milk becomes, which line it becomes it on, how much of it is lost on the way, what that line costs per hour, and which of six hundred thousand finished-goods codes touch any of it.

This is the honest reason manufacturing is worth writing about. Not because manufacturers are more sophisticated than anyone else, but because the arithmetic that produces their P&L is genuinely deep, genuinely wide, and genuinely changes every month. You can fake it at demo scale. You cannot fake it at real scale, and the failure is not dramatic. The model simply stops being able to answer questions before Thursday, and people quietly go back to asking the plant controller for a spreadsheet.

We’ll use a process manufacturer throughout, a dairy and chilled-foods group, because process manufacturing exercises the parts that break first.

One number, all the way down

Take a single number: gross margin on a 400g pot of strawberry yoghurt, sold to one French retailer, in March.

Start at the top. Net revenue is not the invoice. It’s list price less on-invoice discount, less the off-invoice rebate accrual, less the promotional allowance for the weeks the product was on deal, less the amortised listing fee. And the rebate is volume-tiered, which means this month’s accrual is a function of the full-year forecast for that customer. A March number that depends on a December expectation.

Figure 01 · What the retailer pays, and what you keep One code, one customer, March

  List price

− On-invoice discount

− Off-invoice rebate accrual Volume-tiered, so a March accrual keyed to the full-year forecast

− Promotional allowance Only for the weeks the product was on deal

− Amortised listing fee

= Net revenue

Now go down.

Cost of goods sold is standard cost × volume, plus variances. And the standard cost of that pot is itself a small model:

Figure 02 · Standard cost, exploded Seven levels

400g strawberry yoghurt pot finished good

• Recipewhat goes in
  • Yoghurt basesemi-finished, own recipe
    • Fermented milksemi-finished
      • Standardised milksemi-finished
        • Creamjoint product
        • Skimjoint product
          • Raw milk intakepurchased, priced on composition
    • Culturepurchased
  • Fruit preparationpurchased, own price curve
  • Sugarpurchased, commodity
  • Stabiliser (contains palm)purchased, commodity-linked
• Packagingwhat it travels in
  • Pot, lid, sleeve, case, palleteach with its own waste factor
• Conversionwhat it costs to make
  • Filling line minutes× cost-centre rate per minute
  • Changeover minutesallocated across the batch
  • Start-up and flush lossesmaterial that never reaches a pot

semi-finished purchased joint product loss and waste

Then look at the last branch of that structure, because it hides the awkward part. The cost-centre rate is the fixed and direct cost of the centre divided by its planned activity volume. Which means the rate depends on the plan, and the plan depends on the rate, because cost drives price drives volume.

Figure 03 · The rate that depends on itself One deliberate second pass

Cost-centre rate → Standard cost → Price → Volume → Planned activity ↶ and planned activity is the denominator of the rate

So before we have reached anything unusual, the model already has seven levels of depth and at least one deliberate second pass over its own output. This is what “deep” means, and it is why a manufacturer’s model does not look like a revenue model with more rows.

Where the arithmetic runs backwards

Go one level further down and you meet something a discrete-assembly mindset has no slot for.

Raw milk goes into separation. Cream and skim come out. Both are products. Neither is a by-product of the other. And there is no honest way to say what the cream cost, because the cost was incurred on the milk, before cream and skim existed as separate things.

IAS 2 handles this in a single sentence that is easy to read past: when the costs of conversion of joint products are not separately identifiable, they are allocated between the products on a rational and consistent basis, for example in proportion to relative sales value at the point the products become separately identifiable.

Read that as a modeller and it is startling.

Figure 04 · The one place the arithmetic reverses Separation

Everywhere else

Children sum into the parent

A bill of materials is a tree that flows one way. Costs roll up, and every number above is the total of the numbers below it.

At the separator

The parent splits into children

You take a parent cost and divide it down, on a basis that is itself a forecast. Change the butter price and the cost of skim milk powder moves. Nothing physical changed. The split moved.

Two more process realities that a parts-list mental model doesn’t carry:

Scale: the number everyone quotes, and the one that matters

Now put the whole company on the table.

Figure 05 · The company on the table Before scenarios

20Plants

100Lines

600kCodes

50Countries

200Cost centres

× scenarios  ·  × versions  ·  × twelve months  ·  × the rolling forecast

The naive reading is that this is an enormous grid: multiply it out and you’re in the trillions of cells. That reading is wrong in two directions, and both matter.

Multiply that by every code that contains the ingredient, and you have the number that actually decides whether the model is usable.

This is why “it handles ten million rows” tells you nothing. Ten million rows of a flat table with no dependencies is a report; it’s a big read. A hundred thousand cells with a seven-level dependency graph, a joint-cost split, and a second pass for under-absorption is a model, and it is a different order of problem.

And note which direction the finance question runs.

Flexibility: the model changes as often as the plant does

Here is the part nobody budgets for. A manufacturing model is not built once and then operated. Its structure changes, routinely, as ordinary business:

Structure, moving Ordinary business, not exceptions

1. Reformulation A recipe is changed mid-year, palm to rapeseed, or a sugar reduction, and the old recipe is still needed for the months before the change.
2. Capacity A line is commissioned. Another is mothballed.
3. Rationalisation Six hundred thousand codes become four hundred and eighty thousand, and the deleted ones still have history that must report.
4. Acquisition A plant is acquired, and runs its own costing method for two more years because that was the deal.
5. Reorganisation The cost centre structure is redrawn because the organisation changed, not because the model wanted it to.
6. New dimension A new packaging format arrives that needs a dimension nobody had.

The question is not whether a tool can express these. Everything can express them eventually. The question is how long each one takes, and who does it. If the answer is “raise a change request, six to eight weeks”, the model has stopped tracking the business it describes.

Finance goes back to spreadsheets, not out of nostalgia, but because a spreadsheet can be changed on Tuesday.

And structure is only half of it. The other half is exceptions, and every real manufacturing model is one general rule plus several hundred local ones:

The local rules Roughly fifteen percent of the P&L

1. Tolling The arrangement where you never own the material. You’re paid a conversion fee, so material cost must be out of that plant’s margin and in the group’s.
2. Excise The country with a duty sitting between gross and net.
3. Rebate year The customer whose rebate year starts in April.
4. Scrap booking The plant that books scrap to a different cost centre because it always has, and the auditors have signed it for a decade.
5. Co-packing The co-packer making three of your codes, so those carry a purchase price rather than a standard cost, and still have to appear in the same margin report as everything else.

These are not untidiness to be normalised away. Taken together they can be fifteen percent of the P&L, and they are disproportionately the part people argue about. So the design question is precise: can an exception be local? Can a rule apply to one plant, one customer, one product family, one period, without forking the model, and without spawning a workbook beside it?

When it can’t, one of two things happens. The exception is ignored, and the model is wrong in a way somebody knows about and nobody has written down. Or it’s handled outside the model, and now there are two models, and they begin to drift the day they’re created.

Events: the things that happen once

Models are built to express recurring relationships: volume times rate, quantity times price. But a large part of a manufacturing year is not a relationship at all. It is a dated event.

A year of things that happen once Each one dated, scoped, shaped

1. August Line 3 down for a fortnight. Lost capacity, volume transferred to another plant at a different cost, and under-absorption at the origin.
2. September A new line commissions. Three months of a yield curve, scrap well above steady state, and a rate that only makes sense once it reaches design speed.
3. 1 April A listing win. First delivery on the first of the month, pipeline fill in March.
4. Dated A delisting. Revenue stops on a date, and packaging obsolescence hits once.
5. 1 April A price increase, notified sixty days ahead, so Q1 carries a forward-buy distortion that reverses in Q2.
6. October An energy contract resets.
7. Unplanned A recall.

Date

When it starts, and when it stops. Rarely on a month boundary.

Scope

This line, this family, this country. Not the whole model.

Shape in time

An event is rarely a step. It’s a ramp, a curve, a partial month, a fill followed by a run rate.

Spreadsheets handle events by hardcoding. Someone types a number over a formula, in one cell, in one column. It works. It is invisible. And it is the single most reliable source of a model nobody fully trusts, because the moment the event has to move by a month, or apply in the downside case but not the base, the override is orphaned and no one can find all of its siblings.

A model handles an event by making it a thing: named, dated, scoped, shaped, owned, participating in the calculation rather than overwriting it. Then it can be switched off in one scenario and on in another, moved a month without archaeology, and, most valuable of all, explained to someone who asks why August looks like that.

Simulation: one price, a million cells

Now the part that ties it together, and it comes with a real example.

In April 2026 the FAO Vegetable Oil Price Index rose 5.9 percent in a single month, to its highest level since July 2022. Palm oil rose for a fifth consecutive month. The FAO’s own explanation of why is the useful bit: demand from the biofuel sector, supported by policy incentives, and higher crude oil prices. The FAO notes plainly that higher crude oil prices increase biofuel demand and put additional pressure on vegetable oil markets.

Read that as a modeller and the shape of the problem appears.

Figure 06 · One driver, two doors Same margin line, different lags

Crude oil moves one external driver

Into the recipemonths, through a commodity market

Biofuel demand → Palm oil → Stabiliser → Standard cost

Into distributionabout six weeks, through the pump

Diesel → Cost per pallet-kilometre → Cold-chain cost

Both arrive in the same gross margin line, for the same customer, in the same month. Two unrelated entry points · different lags · different structures

Follow palm oil at plus twelve percent and count the ground it covers:

Figure 07 · How far one price has to travel Palm oil, +12%

1. 01Every recipe containing it, at every level: a where-used inversion across the whole graph.
2. 02Yield and scrap, because now the losses are on a more expensive material.
3. 03The standard cost of every semi-finished and finished code above it.
4. 04Margin by code.
5. 05Margin by customer, which moves the rebate accrual, because some tiers are margin-linked.
6. 06Margin by country, after transfer prices that were set on the old standard.
7. 07Inventory revaluation, because the standard changed mid-year and there is stock on hand.
8. 08The group P&L, in reporting currency.

And then the honest part: that is only the naive simulation. At plus twelve percent, the reformulation becomes economic. So the recipe changes, which changes the cost, and the yield, and the line speed, and the shelf life, and possibly the artwork and the label claim. The question worth asking is never “what if the price rises”. It is “what if the price rises and we respond”, which means your response has to be inside the model too, as an event with a date and a scope.

And it is never one run. It is a grid:

Thirteen runs, rebuilt every time the buyer sends a new forward curve. Which makes speed a design property, not a nicety. If a run takes overnight, nobody runs the grid. They run one case, once, and call it the scenario.

Recalculation time doesn’t just determine how fast you get an answer. It determines how many questions get asked.

And the difference between a company that asks four questions a quarter and one that asks forty is not a reporting difference. It’s a management difference.

In Novi

The reason to describe all of this in one article is that these are not six problems. They are one problem seen from six sides, and they have to be solved by one structure, otherwise the seams between the solutions become the place the numbers go wrong.

And the formulas stay legible: readable by the analyst who inherits the model in two years, and readable by the machine that has to recompute six hundred thousand codes when one price moves.

None of that is exotic. It’s just what it takes for the answer to arrive before Thursday.