Sales during a stockout do not reveal all the demand you lost

When a product sells out, recorded sales show how much the business supplied, not necessarily how much customers wanted. Feeding those sales into a demand forecast as though stock had been available throughout can understate future demand. The first repair is to identify the affected observations and preserve the difference between sales and estimated demand.

For a simple lost-sales model, observed sales are limited by available inventory. Demand above that limit may be unobserved. Research by Bu, Simchi-Levi and Wang examines pricing and demand learning under this kind of censoring and shows why the information in historical sales can be insufficient for some conclusions.

The operational steps below help finance prepare a more honest forecasting dataset. They do not claim to reconstruct every abandoned purchase or to implement the research paper's optimization method. Any estimated lost sales remain a planning assumption with an evidence trail.

A daily total may hide a product that sold out at 10 a.m. A store-level total may hide a size unavailable at one location while another size remained in stock elsewhere. Match the availability record to the product, location, channel and time interval relevant to purchasing decisions.

Capture starting sellable stock, receipts, reservations, quality holds and the time availability reached zero. On-hand quantity is not always sellable quantity. Inventory committed to existing orders cannot necessarily satisfy a new customer, and a positive system balance can coexist with missing or inaccessible physical stock.

Also record channel restrictions. A marketplace listing can be unavailable even when goods exist in a warehouse. That is a different cause from a physical stockout, but both can suppress observed sales. Preserve the cause because the corrective action may be a listing repair rather than a larger purchase.

Suppose a product starts a day with 20 sellable units and sells all 20 by noon. The sales file records 20 units. Actual demand might have been 20, 30 or 60; the sales total alone cannot distinguish those possibilities once availability prevented further purchases.

Now compare a day that starts with 100 units and sells 20 while remaining available throughout. That observation provides different information. Under a simplified model with no other sales restriction, it suggests demand of 20 for that day. The two days have identical sales but different usefulness for estimating demand.

A planning worksheet should therefore retain sales quantity and an availability flag separately. If an analyst estimates 35 units of demand for the stockout day using other evidence, the extra 15 units belong in an estimated-demand column. They must not become recorded orders, revenue or inventory movements.

Availability records can themselves be wrong. Reconcile a sample of stockout flags to warehouse events, canceled orders and listing history. A system that updates stock once each night may assign a full day of availability to an item unavailable for most of the afternoon. That measurement error can distort both the lost-demand estimate and any later evaluation of whether higher inventory improved sales.

Comparable in-stock days may help estimate the lost portion, but adjust for weekday, season, price, promotion and channel exposure. A Saturday promotion cannot be reconstructed from an ordinary Tuesday average without considering those differences. Keep the comparison rule visible and test how much the estimate changes under plausible alternatives.

Customer waitlists, backorders and failed add-to-cart events can provide additional signals, but they are not automatically completed purchases. One customer may make several attempts, switch to another product or abandon the purchase. Deduplicate the observations and define what each event means before assigning a conversion assumption.

Substitution complicates the total. A customer who buys Product B after Product A sells out may represent lost demand for A without representing lost company revenue. Track the question being answered: replenishment demand for A, revenue lost by the company, or the margin effect of the substitute. Those estimates should not be added together without checking for overlap.

If the forecast treats a stockout-limited sales total as unconstrained demand, the next purchase may again be too small. Another stockout then produces another low sales observation. Marking censoring helps the reviewer see this pattern before the purchasing rule repeatedly learns from its own supply restriction.

The opposite error is also possible. An optimistic lost-sales estimate can justify too much inventory, particularly when interest signals are weak or customers readily substitute. Compare the financial consequences of underbuying and overbuying, including surplus recovery, storage and the timing of cash commitments.

Where evidence is thin, use a bounded replenishment decision with an explicit review point. A modest additional purchase, a permitted preorder test or improved availability tracking may produce better evidence. Describe the operational constraints and obtain the relevant commercial approval before relying on a customer-facing test.

When reviewing the replenishment decision later, retain the estimate originally used. Replacing it with a revised lost-sales figure after observing the outcome makes the purchase look better informed than it was. Compare the original uncertainty range with the subsequent evidence, including surplus stock, so learning covers both underestimation and overestimation.

Lost-sales estimates can support inventory and capacity planning, but they are not revenue earned. Keep them outside the accounting sales ledger. A management report can show observed sales, estimated constrained demand and the assumptions connecting them without presenting the estimate as an actual transaction.

After a stock replenishment, compare actual sales with the prior estimate under similar conditions. Record whether price, marketing or availability changed at the same time. An apparent forecast success is difficult to interpret if several other drivers changed without being documented.

Archive the original availability data and the estimation method. The review should identify which observations were treated as constrained, what additional evidence was used, and how sensitive the purchase decision was to the estimate. That gives finance a defensible explanation for the stock commitment even when the unobserved demand can never be known exactly.