Lost Sales Are Silent Demand

The hardest part of measuring a stockout is that the missing sale does not write itself anywhere.

A customer lands on a product page and leaves because the item is unavailable. A store customer sees an empty shelf and buys a substitute. A distributor loses the order to a competitor. Finance sees no invoice. The sales table records nothing. The demand forecast sees lower history. The replenishment model concludes the item did not need as much inventory after all.

That is how stockouts turn into fake efficiency.

Inventory turns look better because inventory is lower. Forecast error may even look acceptable because the system is comparing forecasts to censored sales. Margin looks worse, but not always in a way that can be traced back to the specific stockout. The business then congratulates itself for being lean while the customer silently went somewhere else.

Lost sales are not a reporting nuisance. They are a decision problem.

Sales are not demand

The first mistake is treating observed sales as true demand. This is convenient because sales data is clean, audited, and trusted by finance. It is also wrong whenever availability constrains the transaction.

Observed sales are the minimum of demand and supply availability. If the item is in stock, sales may be a decent demand signal. If the item is out of stock, sales are censored. The system only observes what it was able to fulfill.

A replenishment model that ignores this distinction learns the wrong lesson. It sees low sales during out-of-stock periods and interprets them as low demand. The next order is smaller. The next stockout is more likely. The bias compounds.

A practical planning system does not need philosophical perfection, but it does need to stop pretending that stockout weeks are normal weeks.

Frame the decision first

Before building a lost-sales model, ask what decision the estimate will influence. Are you changing reorder quantities? Safety stock? supplier priority? assortment? transfer logic? markdown timing? allocation across stores? customer promise dates? The answer matters.

If the decision is replenishment, you care about the incremental expected profit of buying more units before the next stockout. If the decision is assortment, you care about whether the product deserves shelf space or catalog visibility. If the decision is allocation, you care about where scarce units create the most value. If the decision is vendor escalation, you care about the economic cost of late supply.

A lost-sales estimate without a decision attached becomes another dashboard number that people debate and then ignore.

The useful framing is: what action would we take differently if the true lost-sales number were higher or lower?

What can be observed

Lost sales are not directly observed, but the exposure often is. The exposure is the opportunity for demand to occur while the product was unavailable.

In ecommerce, useful signals may include product-page sessions, search impressions, add-to-cart attempts, back-in-stock signups, abandoned carts, substitution clicks, notify-me requests, customer-service contacts, and paid traffic that landed on unavailable products. These signals are imperfect, but they are closer to demand than a zero in the sales table.

In physical stores, useful signals may include shelf-availability audits, point-of-sale gaps, inventory-on-hand corrections, substitution patterns, loyalty-basket changes, store traffic, planogram exposure, and replenishment cutoffs. Stores are harder because a customer who does not find the item often leaves no obvious trace. That does not mean the demand did not exist. It means the measurement problem is harder.

For B2B and distribution, useful signals may include quote requests, unfulfilled order lines, backorder cancellations, customer allocation cuts, service tickets, distributor complaints, and competitor loss notes. Many firms already have fragments of this information scattered across sales operations, customer service, and ERP workflows. The problem is not always missing data. It is that nobody has connected the data to the inventory decision.

A defensible floor

A practical lost-sales estimate should start with a defensible floor rather than a heroic point estimate.

For ecommerce, one simple floor is:

[ \text{lost units} \geq \text{out-of-stock product-page sessions} \times \text{pre-stockout conversion rate} ]

This is not perfect. Conversion rates can change. Customers may behave differently during stockouts. Some sessions would not have converted anyway. But it gives the business an observable, explainable lower bound.

A stronger version segments by traffic source, customer type, device, price, promotion, and season. A paid-search session landing directly on an unavailable product page should not be treated the same as a casual browse from an unqualified source. The goal is not to make the estimate complicated. The goal is to make it economically credible.

For stores, a similar floor may use historical sell-through during comparable in-stock periods, adjusted for store traffic, day of week, promotion, season, and local events. The key is to exclude censored periods from the baseline. Do not use stockout weeks to prove that stockouts were harmless.

Policy structure

Once lost demand is estimated, it should enter the decision system as an economic signal, not as a vanity metric.

For a replenishment policy, the decision might be:

[ x_{i,t} = \text{units to order for item } i \text{ at time } t ]

The objective should include expected margin from fulfilled demand, holding cost, purchase cost, obsolescence risk, capacity usage, and the cost of lost sales when inventory is unavailable. If demand is probabilistic, lost sales are scenario-dependent:

[ \text{lost}{i,t}^{s} = \max(0, D{i,t}^{s} - \text{available}_{i,t}^{s}) ]

This formulation forces the model to confront the real tradeoff. Buying more units is not automatically good. It reduces lost sales in some futures and creates excess inventory in others. The best decision depends on margin, demand uncertainty, lead time, MOQ, shelf life, cash, storage, and substitution.

The important point is that lost sales should affect the economics of the decision, not just appear after the fact in a service dashboard.

Uncertainty and censoring

Lost-sales estimates are uncertain. Pretending otherwise creates false precision.

The best practice is to treat lost demand as a distribution or scenario input. For each stockout event, estimate a range: conservative floor, most likely value, and high-demand scenario. Then test whether the recommended decision changes across that range.

If the decision does not change, do not over-engineer the estimate. If the decision changes dramatically, invest in better measurement. This is how you avoid turning analytics into academic theater.

Censoring should also be explicit in the forecasting model. Out-of-stock periods should be flagged. Depending on the product and channel, those periods may be excluded, adjusted, imputed, or modeled with a censored likelihood. What you should not do is feed censored sales into a forecast and pretend the resulting demand history is clean.

Metrics that matter

Inventory teams often report turns, fill rate, service level, forecast accuracy, and sales. Those metrics are useful, but they are incomplete without stockout exposure and lost-sales economics.

Track out-of-stock exposure hours, unavailable page sessions, shelf-empty duration, lost-margin estimate, censored-demand share, substitution rate, back-in-stock conversion, customer repeat impact, and time to recovery. Also track how often the replenishment policy stocked out on products where the expected value of additional inventory would have been positive.

The last metric matters because it directly evaluates the decision. A stockout is not automatically a failure. Sometimes the item was not worth carrying more deeply. The failure is a stockout where the economic case for more inventory was obvious before the decision was made.

Common failure modes

The first failure mode is counting every unavailable session as a lost sale. That overstates the problem and destroys credibility.

The second failure mode is counting nothing because the sale did not occur. That understates the problem and rewards bad availability.

The third failure mode is using post-stockout conversion as the baseline. Of course conversion is low when the item is unavailable. That is the symptom, not the counterfactual.

The fourth failure mode is ignoring substitution. If customers buy a substitute with similar margin, the lost economics may be smaller than the lost unit count. If they leave the brand, the economics may be larger than the immediate sale.

The fifth failure mode is reporting lost sales but not changing the planning policy. Measurement without decision change is just expensive guilt.

What to do in practice

Start simple. Add availability flags to demand history. Build an exposure table. Estimate a conservative lost-sales floor. Segment only where segmentation changes the decision. Feed the resulting demand signal into replenishment, allocation, and assortment choices. Track whether recommendations change and whether those changes improve economic outcomes.

Do not wait for perfect lost-sales measurement. The current default in many companies is not neutral. It is biased toward underestimating demand during stockouts. A rough, transparent estimate is often better than a clean zero.

Lost sales are silent demand. The job of a serious planning system is to make that silence economically visible.