Forecast distributor demand without treating stockouts as low demand

A distributor that ships 40 units after receiving requests for 90 units has not demonstrated demand of 40. Its shipment history is constrained by available supply. A demand forecasting baseline should use a defined measure of customer requirements, preserve the timing of those requirements, and identify where the record is incomplete.

That does not mean replacing every low-sales day with a larger number. A quiet day with stock available can contain useful evidence of low demand. A quiet day with unavailable stock contains a different uncertainty. The practical first step is to distinguish the two before choosing a forecasting method.

Follow the request through the shortage

Consider an illustrative depot selling one item in individual units. All requests in this example are validated customer orders, with duplicate submissions removed. Customers cannot substitute another item in this small example.

Day 1's 90 units become 40 shipped, 30 awaiting supply and 20 confirmed lost. On Day 2, a replenishment allows the depot to ship all 70 new units and the 30-unit backlog.

Across both days, new requested demand is 160 units and shipments are 140 units. The remaining 20 units are known lost demand. There is no backlog at the end of Day 2. The ledger balances without counting any customer requirement twice.

A shipment-only series reads 40, then 100. The requirement series reads 90, then 70. Averaging shipments gives 70 units per day; averaging recorded requirements gives 80. These are descriptions of this two-day example, not sufficient history for a forecast or evidence that demand always needs a fixed percentage uplift.

Adding Day 1's entire 50-unit shortfall to the two-day shipment total would produce 190. That would count the 30 units delivered from backlog twice. Preserve the original request identifier, required date, subsequent shipment and cancellation outcome so that each unit enters the demand series once.

Choose the series for the planning decision

A depot replenishment forecast usually needs item, location, quantity and the date on which supply was required. An invoice date may suit a revenue report while shifting the apparent timing of demand. A shipment date may suit transport planning while hiding a stockout.

Agree a definition with sales and planning before assembling history:

• Use requested delivery dates consistently, distinguishing them from order creation dates.
• Resolve duplicate orders and amendments against the original order line.
• Record substitution links so that original and replacement items are not both treated as independent fulfilled demand.
• Keep genuine customer returns separate from new demand; a return does not retrospectively erase the original requirement.
• Distinguish cancellations caused by unavailable stock from duplicate orders, customer changes and credit decisions.

Promotional forward buying, pack-size changes and sales transfers between depots need explicit treatment. Converting a case into units is helpful only when the conversion was valid for that item and date. A changed pack definition can otherwise look like a demand jump.

Mark what the depot could not observe

The worked example has an unusually complete order log. Many distributors only see accepted orders. A salesperson may stop recording requests after learning that the depot is empty, or a customer may buy elsewhere without contacting anyone.

Those units are unobserved. They should not be described as confirmed lost sales. A stockout flag can identify censored periods, but it cannot reveal the missing quantity by itself.

Create separate fields for observed requested demand, observed lost demand and any reconstructed estimate. For an estimate, retain the method, version and uncertainty: comparable in-stock weekdays, a nearby location adjusted for a documented difference, or another defensible baseline. An estimate built from a promotion at a different depot should not silently become an actual order.

Where the history is too weak, show a range or exclude the period from a particular model fit with a visible reason. Also keep a separate evaluation view of those excluded periods. Otherwise, a method can appear accurate merely because the hardest weeks disappeared.

Evaluate at the horizon that replenishment needs

A forecast of tomorrow's shipments can perform well while failing to support a purchase decision made two weeks ahead. Choose the forecast horizon from the ordering process and usable replenishment lead time, then compare simple baselines and candidate methods on that horizon.

Hyndman and Athanasopoulos describe evaluation using observations withheld from model fitting and rolling forecast origins that use only earlier information. Their discussion also distinguishes performance across forecast horizons. These are useful safeguards against testing a model with knowledge it would not have had at the decision date. Hyndman and Athanasopoulos — Forecasting: Principles and Practice, forecaster’s toolbox

For the distributor, retain the forecast issued at each historical order date. Compare it with subsequently observed requirements at the same item-location grain. Report bias alongside an error measure so repeated underforecasting remains visible. Very low or zero demand can make percentage errors unstable; retain unit errors and explain the chosen aggregate measure.

Show results separately for periods with reliable availability, known stockouts and uncertain reconstruction. A model trained on estimated demand should not be presented as independently validated merely because it reproduces those same estimates.

Keep the forecast separate from the replenishment setting

A forecast estimates a requirement. The order decision also depends on usable stock, open purchase orders, lead-time uncertainty, service policy, pack multiples and capacity. Raising a forecast to compensate for every service problem can conceal the actual cause.

Microsoft documents safety stock as a planning setting within its replenishment behavior. That product setting does not establish the correct demand history or the right buffer for this depot. Microsoft Dynamics 365 — Safety stock fulfillment for items A late supplier, an inventory hold and an underestimated requirement call for different corrections.

Give a planner a small audit view: the original request, the availability evidence, the series value used, and the reason for any adjustment. The planner should be able to challenge a reconstructed spike without editing the source order record.

Start with one item-location history

The useful outcome is a demand series whose meaning survives a stockout and a later backlog clearance. Only then can a forecast comparison answer whether a different method improves the ordering decision.

Next actions

1. Reconcile one stockout period from unique customer requirements through shipment, cancellation and remaining backlog.
2. Label observed quantities and estimated missing demand separately, retaining required dates and pack conversions.
3. Compare forecasting baselines at the actual replenishment horizon using only information available at each forecast date.
4. Review bias and error by availability condition before changing either the forecast or the inventory buffer.