The demand a sales record
cannot show.

A replenishment system can record every completed sale accurately and still underestimate what customers wanted. When availability limits purchases, the sales history partly records an earlier stocking decision. Passing that history upstream without its availability context can turn a shortage into an apparently weak market.

Start with what was observable

Consider a deliberately simple lost-sales model. A shop begins each day with ten units, receives no replenishment during the day, and records completed purchases. Customers who cannot buy leave without registering an order. There are no substitutions, returns or reserved units.

If demand is D, recorded sales are the smaller of D and ten. Selling seven units identifies demand as seven under these assumptions. Selling all ten establishes only that demand was at least ten. It does not distinguish ten willing purchases from twenty.

For illustration, suppose the underlying demand over four days is 12, 18, 7 and 20. The sales record contains 10, 10, 7 and 10. Average recorded sales are 9.25 units; average demand is 14.25. Twenty desired purchases disappear from the completed-sales record.

Those demand values are supplied by the example, not recovered from its sales. A retailer observing only the four sales totals cannot calculate the missing twenty. Treating the totals as complete demand observations conceals precisely the uncertainty that matters.

When the shelf emptied

Aditya Jain, Nils Rudi and Tong Wang examine this observation problem in Demand Estimation and Ordering Under Censoring: Stock-Out Timing Is (Almost) All You Need. Their multiperiod inventory model uses the timing of sales to update demand beliefs. The public abstract identifies particular demand models in which stockout timing captures the relevant timing information. That is a model-specific result, rather than a general promise that one timestamp reconstructs every lost customer.

A practical implication follows. Ten units sold before lunch and ten units sold just before closing produce identical daily sales totals but different records of exposure to potential demand. Preserving the availability interval gives an analyst information that the total alone discards.

Timing still needs interpretation. Opening hours, promotions, weather and customer arrival patterns may differ. A midday stockout does not justify doubling sales mechanically. An estimation method must explain how observed purchasing opportunities relate to the unobserved period.

Orders introduce another transformation

Demand censoring is also different from the bullwhip effect. Hau L. Lee, V. Padmanabhan and Seungjin Whang’s Information Distortion in a Supply Chain: The Bullwhip Effect examines how upstream orders can become more variable than downstream sales. Its public abstract identifies demand-signal processing, rationing incentives, batching and price variation as sources of distortion.

The distinction matters for diagnosis. A manufacturer may receive a large order because a retailer is replenishing several days together, responding to a discount, or requesting extra stock under an allocation rule. Meanwhile, that retailer’s sales may understate customer demand because its shelves were empty. Both records can be accurate accounts of the transactions they contain.

Sharing more records helps only if their meanings survive the transfer. An upstream order, a completed retail sale and an unfulfilled customer request measure different events.

Preserve the observation boundary

A useful operational record should therefore retain more than quantity and date. For a proposed evaluation, I would distinguish opening availability, receipts, sale timestamps, stockout intervals, cancellations and any separately observed unmet requests. Substitutions and backorders should remain explicit rather than being silently combined with lost sales.

Inventory records also need reconciliation. A computer reporting available stock while the shelf is empty creates a different observation failure from a correctly recorded stockout. Neither should be converted automatically into evidence of low demand.

The decision rule should be tested against these distinctions. Compare a sales-only policy with an availability-aware policy using the same service objective and cost assumptions. Report holding costs, spoilage where relevant, service failures and uncertainty, rather than rewarding forecast accuracy alone.

Historical replay has a limit: the demand hidden by the original stocking policy remains hidden. Simulation can expose the consequences under stated assumptions, but cannot establish realised savings. A prospective evaluation needs a defensible observation design and bounds on the additional inventory exposure.

A better record preserves what remains unknown

Supply-chain resilience depends partly on knowing whether a low number describes weak demand or restricted opportunity to buy. The same restraint applies upstream: a large order is evidence of an order, not an unqualified measurement of final consumption.

The useful improvement is to preserve the boundary between what happened, what could have happened and what the model must infer.