Practical framework / 5 MIN READ
Do not teach the forecast that unavailable products were unwanted
Mark inventory constraints before interpreting low sales as low demand.
AI-assisted research and analysis for the SNS Creative Fieldbook
Sources checked 2026-11-06
Sales show what customers bought under the available conditions. When an item is unavailable, the observed total can hide unmet demand, substitution or abandoned purchases, so a forecast trained on sales alone may reproduce the inventory constraint instead of estimating customer interest.
What you will leave with: Build an availability-aware demand brief with censored periods, alternative signals and estimation limits.
Use this when
- A product sells out quickly and later appears to have weak demand.
- A campaign’s apparent decline coincides with unavailable variants or locations.
This is demand interpretation, not an automatic lost-sales estimator or an instruction to change inventory.
Define availability at the customer’s point of choice
Record when and where the relevant item or variant could actually be purchased and fulfilled. Separate warehouse stock from saleable stock visible to the customer. Include channel, location, size or configuration where those determine the buying opportunity.
Check inventory record reliability and timing. A system may show positive stock while the shelf is empty, or zero stock while an approved alternative is available. Ask the operations owner how stockouts are detected and logged. Mark uncertain availability explicitly rather than creating a false precise history from an unreliable end-of-day snapshot.
Separate observed sales from unconstrained demand
Mersereau’s research explains that inventory-limited sales censor underlying demand and that inaccurate inventory records make estimation harder. The practical implication is to flag constrained observations before asking a model to learn demand from them. The paper does not supply a universal correction for every retailer.
Keep three quantities distinct: fulfilled sales, observed attempts or enquiries, and estimated unmet demand. A back-in-stock signup or product-page visit can provide context, but it is not automatically an intended purchase. Deduplicate permitted signals and avoid summing several actions by the same customer into multiple lost sales.
Build a defensible estimation comparison
Ask a qualified analyst which evidence can support estimation: comparable in-stock periods, similar locations, alternative products or an appropriate censored-demand model. Review seasonality, promotion changes and substitution so that the comparison does not simply import another bias.
Do not fill every stockout day with the prior day’s sales and call the result true demand. Show assumptions and a plausible range. Where availability itself responds to forecast demand, explain the feedback loop. More stock may reveal demand previously hidden, while a reduced assortment can change customer behavior rather than merely censor it.
Use the limitation in campaign and inventory decisions
Report campaign outcomes with availability coverage. A weak conversion rate during widespread stockouts does not cleanly diagnose weak creative or audience quality. Coordinate the interpretation with commercial and operations owners before cutting a product or expanding spend.
Preserve actual sales separately from any modeled demand adjustment and label revisions. Evaluate future forecasts on periods where availability is understood, while recognizing that unconstrained demand is still not directly observed in every setting. Use the analysis to guide a bounded planning decision, not to promise that replenishment will recover every estimated lost unit.
Make the reasoning concrete.
A hypothetical item sells 20 units by noon and then becomes unavailable for the rest of the day. The next day’s sales are zero because it remains unavailable. Those observations do not show that demand fell from 20 to zero, and they do not reveal exactly how many additional units would have sold.
Train on sales without context
Opening: Use 20 and zero as ordinary demand observations.
Hypothesis: The input table is simple and complete.
Risk: The model may learn the stock constraint as low interest.
Mark constrained periods
Opening: Preserve sales and availability, then estimate unmet demand with justified assumptions.
Hypothesis: The planning team can distinguish observed fulfillment from latent demand.
Risk: The estimate remains uncertain and may depend on substitution.
Flag the stockout periods before interpreting the trend. Do not replace actual sales with an invented demand total or promise that all estimated demand can be recovered.
Take the questions into your next brief.
Use these prompts in your own notes. Nothing is submitted from this page.
Availability
Could the customer actually buy the specific item?
Evidence
How reliable and time-specific are inventory records?
Signals
Which observations show interest without proving purchases?
Estimate
What comparison and assumptions support unmet demand?
Decision
How does the constraint change campaign or assortment interpretation?
Watch for the wrong conclusion.
- Treating stockout zeros as ordinary demand zeros.
- Assuming website visits equal lost purchases.
- Using a fixed imputation without validation.
- Blaming creative performance for an unavailable offer.
Your next action
Annotate one sales series with verified availability and separate constrained observations before the next demand review.
Sources & scope
Frameworks, worksheets and hypothetical examples are original editorial synthesis. Linked primary sources support only the stated concepts. Source checking is not an independent audit of effectiveness. No named human author or reviewer is claimed.
- Adam Mersereau: Demand Estimation from Censored Observations with Inventory Record Inaccuracy →Primary research abstract · Supports: The primary paper abstract explains why sales constrained by inventory censor demand and why inaccurate inventory records complicate estimation. No paper-specific correction or measured benefit is generalized.