Integrated demand planning: intelligence, collaboration, governance
"Integrated demand forecasting" is sold as a tool. It isn't.
It's three distinct capabilities — intelligence, collaboration, and governance — usually built at different times, by different people, with different sponsors. And the process almost never fails inside one of the three. It fails at the joins.
Layer 1 — Intelligence: the baseline nobody has to type in
This is the statistical part: baseline per series, seasonality, trend, handling of events and promotions, intermittency, and the choice of aggregation level where the signal actually exists.
The point of the baseline isn't to be right. It's to be the starting point nobody has to build by hand, freeing the conversation for what the model can't know.
What kills this layer, in practice:
- Black box. If nobody can explain why a SKU's forecast moved 20% from one month to the next, the commercial team will override everything — and statistics has become decoration.
- History with a break inside it. You train the model on what sold, not on what the market wanted to buy. Where product was out of stock, history teaches the model to forecast the stockout. Handling censored demand is tedious and almost always skipped.
- Wrong level. Forecasting at SKU-customer-day when the decision happens at family-DC-week isn't more precision, it's more error.
Layer 2 — Collaboration: what the model can't know
Sales knows about the client that will open 40 stores. Marketing knows about the campaign starting in September. Trade knows the real promotional calendar, not the planned one. Supply knows the line will stop for maintenance.
None of that is in the history. That's why the collaboration layer exists.
The distinction that organises this step: collaboration means bringing information, not opinion. Information has an owner, a date, a magnitude, and a reason. Opinion has a target.
Two rules that change the process more than any feature:
- Every adjustment is a record — who changed it, by how much, and why. It's not about policing anyone. It's about being able to measure later.
- Sell-in and sell-out are different conversations. In many industries, the collaboration that matters most is with the channel, not internal. Forecasting sell-out and deriving sell-in is harder and almost always more accurate.
In an S&OP we implemented for a leading pharmaceutical company in Brazil, sales, marketing, and operations moved to a single cycle, with detailed sell-in and sell-out forecasts. The result that mattered wasn't accuracy in the report: it was stockout reduction and faster market response.
In a multinational leader in agribusiness, the problem had a different shape: demand isn't born in one area. It travels through the franchisee network and only afterwards receives input from the internal areas — commercial, marketing, supply, finance — all weighing in on the same number, in a process built on Anaplan. That's the scenario where the record rule stops being bureaucracy: when four areas contribute and nobody records who contributed what, there is no consensus, there is whoever spoke last.
And there's a kind of information no manual adjustment can carry: the effect of a launch on the rest of the portfolio. At another client — a multinational with a broad portfolio and frequent launches — we built a cannibalisation simulator across products, assessing the impact of a launch and mix changes on the whole portfolio before the decision, not after. It's the case where the right answer isn't "sales adjusts the number by hand": it's to explicitly model what the statistical baseline can't know, instead of letting the conversation turn into an opinion fight in a consensus meeting.
Layer 3 — Governance: one number, with an owner and a measure
It's the most thankless layer, and the one that holds the other two up.
- One consensus number, versioned, frozen on date, feeding S&OP and the financial plan.
- Accuracy measured (MAPE, bias) at the level where decisions are made — not at the aggregate that always looks good.
- Bias before MAPE. A consistently optimistic forecast is a correctable behaviour; random noise isn't. Bias is the metric that generates actionable conversation.
- Value added by step. The uncomfortable, necessary question: did the human adjustment improve the number over the baseline, or make it worse? Measuring that per contributor is how you discover — with data, not fights — who systematically adds and who systematically subtracts value.
That last one turns the process from "monthly meeting" into a cycle that learns. Without it, a forecast can run for three years without improving once.
Where the process breaks: at the joins
Almost every demand process that never took hold broke on one of these five points:
- Baseline delivered without explanation. The commercial team gets a number they don't understand, doesn't trust, and overrides by default.
- Adjustment without a record. Without a trail, no measurement; without measurement, the process never improves — only repeats.
- Consensus meetings that re-do numbers. Three hours arguing about SKUs instead of deciding exceptions. Consensus meetings should handle what diverged, not review what converged.
- A consensus number that binds nothing. If Finance runs its own number in parallel, there are two truths — and demand planning became theatre.
- Inherited granularity. The level was defined by what the ERP exports, not by what the decision requires.
What "integrated" actually means
Integrated isn't "the areas talk". Integrated is: the demand number is the same object feeding the supply plan, S&OP, and the financial plan.
A direct test: if a change in the demand forecast doesn't move revenue in the financial plan without human intervention, the process isn't integrated. It's coordinated — which is better than nothing and very different from what was sold.
In the agribusiness case above, that meant bringing finance inside the supply plan: the supply and production plan stopped being just volume and started to carry the financial impact in the same model. It's the difference between supply knowing how much to produce and supply knowing how much the decision it just made costs.
Where to start, if the process doesn't exist yet
Counter-intuitive: start with governance, not intelligence.
A middling baseline inside a governed cycle, with accuracy measured, is worth more than an excellent statistical model that nobody consumes. And it's governance that makes the improvement of the other two layers visible — without it, you can't prove that investment in statistics or collaboration paid off.
Order that works: cycle and calendar → one frozen number per version → accuracy and bias measurement → adjustment record → and only then serious investment in the statistical model.
When it isn't worth it
Few SKUs, stable demand, short replenishment lead time. If your lead time is shorter than the forecast horizon that matters, the answer is probably ability to react, not ability to forecast. Investing in a sophisticated forecast in that scenario is solving the wrong problem very elegantly.
We treat demand forecasting and S&OP as one auditable process — see demand planning and supply chain, or get in touch.