One Demand Signal, Four Currencies

A unit forecast by item and location is not a revenue forecast with different assumptions. It becomes revenue only by passing through price realization, trade terms, mix and shipment timing — and each of those is an assumption someone holds privately. Consumption at the shelf is not shipment either; the gap between them is inventory in the channel, and it moves. These are not four readings of one thing. They are four things that share a word.

So the spread between them is not evidence that somebody is wrong. It is the size of a conversion nobody has written down. Reconciling the four is arithmetic that has no owner, which is why the meeting settles it by seniority instead.

What actually happened

Polymath was asked to look at forecasting in a consumer packaged goods business in the United States that was running four of them.

Demand planning produced a unit forecast by item and location, on a weekly horizon, because that is what a production schedule and a distribution network consume. Sales produced a revenue forecast by customer, because that is what quotas, trade commitments and account plans are built from. Finance produced a revenue and margin forecast by period and segment, because that is what the board had already been told and what the next report would be measured against. Marketing produced a consumption forecast at the shelf, because a promotion moves consumption rather than shipments, and the two are not the same curve.

Four forecasts, four horizons, four units of measure, four owners. Every one of them was competently built. Not one of them was wrong.

Then the promotional calendar came up for approval. Deciding a single promotion required four things at once: how much consumption the offer would generate, how much inventory had to be positioned to serve it, how much revenue it would book and when, and what it would cost in margin. Those four things lived in four forecasts that did not reconcile, on four different horizons, in four different units.

The instinct in the room was to treat this as a disagreement and look for the right number. It was not a disagreement. Asked to state what it would take to turn the consumption forecast into the shipment forecast, nobody could produce the list — not because the factors did not exist, but because they lived in four different heads and no two of them would have written the same ones down.

So the number that carried was the one whose owner had the strongest position that quarter. Sometimes that was finance, because the guidance was the binding constraint. Sometimes it was supply chain, because a service level had been missed recently and the memory was fresh. The decision was real, the money was real, and the basis on which it was taken changed from meeting to meeting. The consequences were downstream and separated in time, which is why nobody traced them back. Inventory positioned against one forecast, revenue booked against another, a promotion assessed after the fact against a third — and no way to tell whether a disappointing result was a bad promotion, a bad forecast or the wrong forecast.

What we proposed was not a single forecast. Asking four teams to converge on one number takes away the frame each of them needs to do its job and produces a compromise that serves none of them. We proposed the opposite: one demand signal, built on a shared data foundation, and a written, owned conversion layer that expresses that signal in each of the four currencies. Each team’s plan then becomes a named scenario — the signal, converted into their units, plus their own adjustments listed one by one with their name against each.

Why it happens

The thing usually blamed is coordination — four teams who would produce one number if they only met more often. They meet constantly. The four numbers survive the meetings, because coordination was never the constraint.

A forecast is built to drive a decision, and the decision determines the unit. A plant needs cases by week. A sales leader needs revenue by customer. A board needs margin by segment. None of those is a rounding of another, and any attempt to declare one of them primary breaks whichever function loses. This is why the single-number instinct fails on contact: it treats an architectural problem as a behavioural one.

The conversion between them is real work. Turning cases into revenue means knowing list price, the promotional depth actually funded, the trade spend accrual, the returns rate and the mix inside the category. Turning shipments into consumption means knowing what is sitting in the channel and how fast it is drawing down. Every one of those factors exists somewhere. None of them is in a system that all four teams read, and none of them has a named owner who maintains it. Because the conversion has no owner, it gets performed privately and differently every time. Each team, needing a view of a quantity it does not own, applies its own factors quietly inside its own model. The four forecasts therefore embed four different implicit conversions, which is why they cannot be reconciled after the fact even by people acting in complete good faith.

That is why the usual remedies underperform. A consensus meeting negotiates finished figures, so it can move a number but cannot locate the factor that separated them. A single planning platform gives four forecasts one home and four private conversion sets. A cross-functional planning role has the visibility to see all four and no mandate over the arithmetic between them.

It also explains why this is now the binding constraint on automating any of it. Every proposal to put a model on planning assumes there is a computable path from a demand signal to a decision. A model can sweep a thousand promotional scenarios faster than a team can discuss one — but only if the conversions are written down. It cannot infer a trade spend accrual convention or a channel drawdown rate from a spreadsheet, and it certainly cannot arbitrate four numbers whose disagreement lives in factors nobody has recorded. The missing conversion layer is the reason the planning meeting is slow and the reason the planning agent is impossible, which is one problem and not two.

The strongest objection

The strongest case against this: we run sales and operations planning, this is exactly what it is for, and a conversion layer is a modelling project bolted onto a process that already works.

That is fair, and S&OP is the right instrument. Where it works, it works because it forces four owners into one room on a fixed cadence with a decision at the end, which is more than most planning processes manage. An organization without it should build that before anything described here. The difficulty is what arrives at the meeting. S&OP as practised brings four finished forecasts and argues them to a consensus. The conversions stay inside the models that produced them. So the consensus is a negotiated position rather than a derived one, and when it proves wrong there is no way to say which factor was wrong, because no factor was ever separated from a figure. None of this replaces that meeting. It changes what walks into it.

The second objection is the better one: these conversion factors are estimates, they move, and writing them down gives a false precision that a good planner’s judgment handles better. The first half is true — a promotional lift factor is an estimate and it does move. But a factor that is written down and wrong can be measured against outcome and corrected. A factor held privately and wrong is indistinguishable from a forecasting difference of opinion, and it survives indefinitely. The claim here is not that the conversions can be made precise. It is that they can be made visible, versioned and owned, which is the only condition under which they improve.

The method

Six steps. One and two are diagnosis, three and four build the layer, and five and six are what make it hold. We do not reorder them.

1. We convert the four forecasts into one another and record what breaks. We take the same period and we push units to shipments to revenue to margin, and consumption back to shipments, and we write down every factor each conversion demanded of us. The trade-off we accept: the first pass fails in places, and we publish where it failed rather than papering over it. Skip it and the conversation stays about whose forecast is better, which is a conversation with no end.
2. We list the missing factors and find who is already deciding them. Each one is being set by somebody today, inside a model, without knowing they own it — a promotional lift, a returns rate, a channel drawdown, a trade spend accrual convention. We go and find that person, and we ask them how they arrived at it. The trade-off we accept: some factors turn out to be set by nobody, and we mark those as open rather than assigning them to whoever is nearest. Skip it and you build a conversion layer on assumptions with no custodian, which decays within two cycles.
3. We build one demand signal on the shared foundation. We use the same history, the same item and customer masters, the same hierarchy and the same calendar, and we produce one signal that every conversion starts from. The trade-off we accept: agreeing the masters is unglamorous work that surfaces disputes about hierarchy that have been deferred for years. Skip it and each conversion runs off a different history, which puts you back where you started with more machinery.
4. We write the conversion layer down and give every factor an owner. We give each factor a value, a source, a review cadence and a name. This is what we hand over, and it is a short document rather than a system. The trade-off we accept: people are markedly more comfortable owning a forecast than owning a single factor inside somebody else’s forecast. That discomfort is the mechanism working. Skip it and the conversions go back inside four private models.
5. We express each team’s plan as the signal plus a named stack. We make demand planning’s service-level buffer a line. We make marketing’s incrementality assumption a line. We make sales’ commitment coverage a line. Each team still gets its own forecast in its own units — it is now legible to the other three. The trade-off we accept: an assumption on a page can be argued with by someone who does not own it, which is uncomfortable and is the point. Skip it and the biases go back into hiding.
6. We run one category through one full cycle before extending anywhere. We take it through baseline, conversion, four scenarios, an approval that names the scenario it was granted against, and a post-event read against that same scenario, and we stay in the room for all of it. The trade-off we accept: the first cycle is slower than what it replaced, because people are learning to argue about factors in public. Skip it and this becomes a methodology document instead of a working cadence.

What carries the weight is step four. Everything else in the library that compounds does so because a definition stopped being re-derived; here it is the conversion between quantities that stops being re-derived, privately, four times a cycle. That is worth doing for the planning meeting alone. That it is also the precise artifact an automated planning capability needs in order to exist is a consequence rather than a justification.

Evidence and measures

The four quantities, and the conversion each one hides:

Five measures worth tracking over ninety days. None of them requires a new system:

— Number of conversion factors that are written down, valued and owned, against the number the four conversions actually require.

— Proportion of the spread between the highest and lowest forecast that resolves into a named factor rather than remaining unexplained. — Proportion of spend approvals that record which scenario, and which conversion, they were granted against.

— Elapsed time from a factor changing to every plan that depends on it reflecting the change.

— Number of separate places the same conversion is being performed. If this is still four, nothing structural has moved yet.

Move those five and the arithmetic of the business has an owner. Move only the number of forecasts in the pack and you have edited an agenda.

How Polymath solves this

We do not begin by building a forecast, and we do not ask anyone to give one up. We begin by converting the four you already have into one another and recording every factor the arithmetic demanded — which produces, usually within the first cycle, a short list of the numbers your company plans on and nobody maintains. That list is the specification for everything after it. The discipline is AI Enablement, because a planning capability that a machine can operate is a planning capability whose conversions are written down.

The work happens against the cycle in front of us. The demand signal is built from the masters and history you already hold; the conversion layer is written with the demand planner, the financial planning lead and the revenue growth manager who are setting these factors privately today. We test it on periods you have already closed before anyone plans against it, then take one category through a live cycle and hand the next one to the people who own it.

What accumulates is the conversion layer, not the forecast. Once the factors are written, valued and owned, every plan that touches them inherits rather than re-derives: a new category starts from a known conversion instead of a workshop, a post-event read has something specific to be read against, and a model asked to evaluate a season of promotions has explicit levers to vary rather than four defended positions to referee. The forecasts stay four, which is correct. The arithmetic between them stops being performed four times in private.

What it costs to do nothing

The visible cost is the meeting: senior people spending the first part of every planning review reconciling rather than deciding. Of the three costs here, that is the cheap one. The second is spend approved through the wrong instrument. A promotion sized on consumption, served from inventory built on units, booked against a revenue figure from a third model — the mismatch does not announce itself. It surfaces as excess in one place, a service miss in another and a margin result nobody can attribute, and every one of those was correct according to the forecast that produced it. That is precisely why the error is so hard to find afterwards: nothing failed, the conversions between the things that succeeded were simply never performed by anyone accountable.

The third accumulates, and it decides what is available to you later. Each cycle run this way teaches the organization that a forecast is a position to be defended rather than a conversion to be examined, and it leaves the factors that connect the quantities undocumented for another quarter. A company that writes its conversion layer this year can put a model on its promotional calendar next year. A company that does not will be arbitrating four numbers with better tooling and the same meeting.

None of this needs a planning platform decision. It needs the arithmetic between your four forecasts written down once and given owners — which is work for the cycle already in front of you, not for a programme after it.