Demand planning: turning a forecast into a usable number

Start from a baseline that can be argued with

The statistical baseline should be generated from cleansed history, with known one-off events stripped out and stockout periods flagged, because unconstrained demand is what needs forecasting, not what was shipped. Where sales were lost to unavailability, the history understates true demand and will keep understating it every cycle if left uncorrected. A baseline nobody can inspect is a baseline nobody will challenge honestly.

Enrichment: the intelligence a model cannot hold

Commercial teams add what the series does not contain. Planned promotions and their expected uplift, new listings and their phasing, price changes, competitor activity, seasonal shifts in customer behaviour and the retirement of an ageing line. The discipline is to record each adjustment separately with its reason and its owner, so that after the event you can see whether the promotion assumption, not the baseline, was what went wrong.

Aggregation level decides what the plan is good for

Forecasts are more accurate the higher they are aggregated and more useful the lower they are disaggregated. Family-level plans support capacity and sourcing decisions; item-location plans drive replenishment. Planning at one level and executing at another is legitimate, but the split between them has to be explicit, and disaggregation rules need reviewing when the mix inside a family shifts.

Consensus without averaging away the disagreement

The sign-off meeting exists to reconcile the commercial view with the statistical one, not to split the difference. Where the two differ materially, the useful output is a named assumption and a named owner rather than a compromise number that reflects nobody's belief. Recording the gap also gives you the raw material for the accuracy review, which is where planners learn which sources of adjustment actually add value.

Feeding uncertainty forward, not just the mean

Downstream decisions need the spread as much as the central case. Inventory policy is sized on how wrong the plan tends to be, so a demand process that hands supply only a single line has withheld half the information. Publishing the expected error alongside the plan, and keeping it item-segmented, is what allows buffers to be set deliberately rather than by habit.

Frequently asked questions

Should demand planning sit in commercial or in supply chain?

Ownership matters less than separation of duties. Whoever owns the number must not also be judged on hitting a sales target with it, otherwise the plan inherits the incentive. Many firms place the process in supply chain and give commercial teams a formal, recorded input into it.

How far ahead should a demand plan run?

Far enough to cover the longest lead time you have to commit against, plus the time needed to change that commitment. A plan that stops inside the procurement horizon forces buyers to invent their own numbers, which is how two competing forecasts start circulating.

What should be done with new products that have no history?

Model them on an analogue product's shape rather than a flat line, state the assumption openly, and review it against early actuals on a short cycle. Launch forecasts are expected to be wrong; what damages the business is leaving them unrevised while stock builds against them.