In short: The account plan and the item forecast are slices of the same volume taken on axes that cross rather than nest, so nothing forces them to agree and the reconciliation becomes a negotiation. The customer dimension earns its place by carrying decisions that are knowable in advance, being range changes, promotional calendars and inventory policy moves, which a statistical model can only see months after they happen. One account at thirty percent of volume roughly triples the coefficient of variation of your total demand against an evenly spread base, and the buffer moves with it. The Herfindahl index borrowed from the merger guidelines calls that customer base unconcentrated, so measure stranded inventory and dedicated capacity instead.
The quarterly review runs in two halves. In the first, the account manager for the largest grocery customer presents nine percent growth, built from three new listings, a fourteen event promotional calendar and a category ambition both sides signed in September. In the second, the demand planner shows the statistical forecast for the same items over the same weeks, and it says four.
Both numbers go into the pack, on different pages, in different units. The consensus that comes out is six, and nobody can name which component of the nine was cut to get there. Supply builds against six. In February the listings arrive at full store count and the plant is short.
Nothing in that sequence connects the two numbers arithmetically, so whichever one wins is settled by seniority.
The plan and the forecast are slices of different cubes
Demand planning holds item by ship-to by week, in cases. Key account planning holds customer by category by quarter, in net revenue with trade spend and margin attached. Both are defensible views of the same business and they intersect at exactly one grain: customer by item by period, in cases. Almost nobody stores it in the planning system, even though it sits in every invoice line in the ERP.
The structural point underneath is that the product hierarchy and the customer hierarchy cross rather than nest. Any item can go to any customer, so you cannot roll one up inside the other the way you roll stores into regions. Hyndman and Athanasopoulos call this a grouped structure rather than a hierarchical one in Forecasting: Principles and Practice, and the practical consequence is that the number of internally consistent ways to aggregate is much larger than in a single tree. Reconciling forecasts across that kind of structure is D8's subject.
What matters here is cheaper than the method. Without that intersection grain there is no shared floor forcing the two numbers to add up, which is why both survive contact with each other for years.
Building the floor is a data exercise rather than a modelling one. Four thousand items against six hundred ship-to points gives 2.4 million cells, of which three to six percent usually carry volume in a year, so you land on seventy to a hundred and forty thousand live series. You do not forecast at that grain. You reconcile and measure at it.
What the customer dimension actually adds
The largest single-period movements in your demand are decisions taken by a handful of customers, and most of those decisions are knowable weeks before they take effect. A range review outcome. A depot closing. A change in the retailer's own weeks of cover. A promotional grid that was agreed in a meeting you attended.
A statistical model at item level sees the sum of these and calls it noise, because at item level that is what it looks like.
Take an account at twenty-two percent of an item's volume moving from three facings to two across four hundred and eighty stores. Say the step down is four hundred cases a week and it was decided in a range review fourteen weeks before it lands. Simple exponential smoothing with a smoothing constant of 0.2 carries a residual error of the step size times 0.8 to the power of the number of periods elapsed. Five weeks after the change it has recovered about two thirds of the shift. Fourteen weeks after, it has recovered ninety-six percent. Sum the whole transition and the cumulative over-forecast is the step size divided by the smoothing constant, which is five weeks of the lost volume, or two thousand cases.
Those two thousand cases get built, and they sit somewhere until they are marked down, on the back of a decision someone in the building knew about a quarter earlier. The customer dimension delivers an earlier input rather than better statistics, which is worth being clear about because it usually gets sold as the second thing.
The joint forecast and the unit it has to be agreed in
The reason to build a forecast jointly with an account is that each side holds something the other cannot compute. They have point of sale, loyalty data, store level inventory and their own promotional intent at a granularity you will never be given. You have supply constraints, launch dates, allocation logic and a view across the category that you cannot share with them and should not try to.
Two mechanical failures kill most joint forecasts before the modelling starts.
The first is the unit of agreement. The number gets agreed in value, at total account level, for a quarter. Supply needs cases, by item, by week, by ship-to. Converting the first into the second requires a price assumption, a mix assumption and a phasing assumption, all invented afterwards by whoever is left holding the spreadsheet, and the invented part is usually larger than the agreed part. Agree the unit at the point of agreement or accept that the number is decoration.
The second is accountability. The customer's forecast is an intention, and nobody at the customer is measured on it. So measure it yourself. Treat their submitted number as a stage in a forecast value added ladder, score it against what they actually ordered, and compare it with your own statistical baseline at the horizon your supply chain needs. D7 covers how that ladder is constructed.
The pattern worth testing is that the customer forecast wins at two to four weeks, where they are reading their own order book, and loses at twelve to sixteen, where they are restating a budget. If that holds, take their number inside the short horizon and your model outside it, which is a per-horizon blend rather than a decision to trust or distrust the account.
One large customer breaks the arithmetic underneath aggregation
Aggregate forecasts beat item forecasts because many buyers behave roughly independently, so variances add while means add faster. Standard deviation grows with the square root of the account count, volume grows linearly, and the coefficient of variation falls.
Put numbers on it. A hundred accounts, a hundred cases a week each, each with a coefficient of variation of 0.4. Total demand is ten thousand a week, total standard deviation is the square root of a hundred times forty, which is four hundred, and the coefficient of variation of the total is four percent.
Now move one account to thirty percent. It takes three thousand a week with the same 0.4 variability, so a standard deviation of twelve hundred. The remaining ninety-nine share seven thousand, seventy-one cases each, standard deviation twenty-eight each, combining to two hundred and eighty-one. Total standard deviation is the root of twelve hundred squared plus two hundred and eighty-one squared, which is one thousand two hundred and thirty-two. The coefficient of variation of the total is twelve percent.
Same volume, same per-account variability, three times the aggregate volatility. Safety stock scales with the standard deviation, so the buffer required to hold the same service level roughly triples, and it triples for a structural reason no forecasting improvement addresses.
There is a second consequence. That large series records ordering decisions rather than consumption: their reorder policy, their batching, their period end buying, their response to your own price moves. Lee, Padmanabhan and Whang set out the four causes of that distortion in Management Science in 1997, being demand signal processing, order batching, price fluctuation and rationing behaviour, and every one of them lives at customer level rather than at item level. Which means the customer dimension is where you can see it and also where you can negotiate it down: a fixed order calendar, an agreed weekly slot, a cap on period end loading.
Measuring the concentration so that it changes something
IFRS 8 paragraph 34 and the equivalent requirement in FASB ASC 280 both call for disclosure of any single external customer at ten percent or more of revenue. That line was drawn for investors and tells an operator very little.
The Herfindahl-Hirschman index is the more useful instrument. Square each customer's percentage share and sum. It moves when you win or lose an account, it is comparable across your markets, and it takes one query. Its published cut-offs do not transfer. The 2010 Horizontal Merger Guidelines from the US agencies called a market unconcentrated below 1,500 and highly concentrated above 2,500, and the 2023 Merger Guidelines set the structural line at 1,800. Run the hundred-account base above through the formula and you get roughly 950, with a thirty percent customer sitting inside it. By the guidelines it is unconcentrated. Borrow the index and set your own thresholds.
The measure that actually changes a decision is stranded exposure. For each item, compute the largest customer's share of the last twelve months. For each of your top accounts, total the value of on-hand and on-order inventory for the items where that account exceeds seventy percent, then add the packaging and components specific to them, plus any tooling or qualified capacity that serves nobody else.
Two businesses can each have a thirty percent top customer and be in completely different positions. One ships standard items from the shared distribution centre, so losing the account means selling the stock elsewhere at some markdown. The other has forty bespoke items, a dedicated artwork library and a line qualified for that customer alone, so losing the account means a write-off and idle capacity until requalification. Identical Herfindahl index. Exposure different by an order of magnitude. The contribution the account was covering is a third question, and W4 handles it.
What belongs in the account plan
Decompose the delta into named components, each with an owner and a date. Base run rate. Distribution change. Promotional volume. Pipeline fill on new listings. One-offs.
Pipeline fill is the component that gets double counted every year, because it is real volume that is not repeatable. Six new listings across four hundred and eighty stores at two cases of initial fill is five thousand seven hundred and sixty cases in the launch month, none of it consumption. Leave it in the baseline and it becomes next year's comparative, so next year's plan starts nearly six thousand cases too high and the account manager spends the year explaining a decline that never happened.
Every assumption carries an expiry date rather than living on as a permanent uplift, and B3 covers the mechanism for that. The supply implication of the range and promotional commitments gets stated when the plan is agreed rather than when the order arrives, which is a sequencing question K3 works through. What does not belong in the plan is a growth percentage with nothing underneath it.
The limit
Customer-level planning pays only where the account behaves differently from the aggregate and the difference is knowable ahead of time. Split too far down and you are estimating many short, sparse, intermittent series, and the sum of their forecasts will be worse than a forecast of their sum. There is no universal cut-off, so run the test on your own history: forecast each account separately, add them up, and compare that total against forecasting the aggregate directly over the same weeks. Where the direct aggregate wins, the customer dimension belongs in the plan as an input and not in the model as a level.
The concentration work has a harder limit, which is that it sizes an exposure without suggesting a remedy. Diversifying a customer base takes years. The things that can change inside a quarter are contractual: notice periods, inventory liability on bespoke items, a commitment against components ordered on their forecast. Those clauses are worth more than the analysis that identified the need for them, and they are negotiated by people who do not sit in planning.
One political constraint is worth naming. Reconciling the account plan to the demand forecast makes the account manager's number falsifiable, and that number is usually also their target. Where plan and target are one document, authority settles the reconciliation every time, and separating them is an operating model decision.
Pull last year's invoice lines at customer by item level and, for each of your ten largest accounts, total the inventory value of the items where that account took more than seventy percent of the volume.