Retail MMM: account for stockouts and distribution changes

Observed sales are not unconstrained demand

A retail model sees recorded sales, not every purchase that a customer wanted to make. When an advertised product is unavailable, the response can be capped by inventory. When distribution expands, sales can rise without a change in advertising effectiveness. Both mechanisms matter before a model recommends moving budget.

Choose the unit of analysis deliberately: national weekly sales, region-week sales or a more detailed panel. Record store openings, closures, assortment changes and meaningful availability measures on that same calendar. The data requirements guide explains why consistent geographic and temporal definitions come first.

Work through a distribution example

A fictional product sells 10 units per store per week in 100 stores, giving 1,000 units. Distribution grows to 150 stores and total sales reach 1,500 units while media spend also rises. Total growth is 50%, but units per store have not changed. This does not prove media did nothing; it shows why the total alone cannot separate the explanations.

Build a distribution series that reflects where the product is actually sellable. A simple store count may be insufficient if the new stores differ greatly in size. Document any weighted measure and avoid treating a partially available store as equivalent to a fully stocked one.

Treat availability as a causal question

Stock can fall because demand increased, which means contemporaneous stockouts may be downstream of successful advertising. Adding a stockout variable blindly could remove part of the effect. Conversely, ignoring a supply disruption can make media look ineffective. The correct treatment depends on whether the decision concerns realised sales under constraints or latent demand under adequate supply.

Record supply shocks, use pre-planned availability where appropriate, and compare sensitivity analyses with clearly stated estimands. The confounding guide helps distinguish a common cause from a mediator. There is no universal control-variable checklist that resolves every retail setting.

Constrain recommendations to what stores can deliver

If an allocator predicts extra demand, translate that forecast into inventory and distribution requirements. A £10,000 media increase with a predicted 2,000 additional units is not executable if only 500 units are available in the relevant stores. Put capacity in the scenario rather than explaining the miss afterwards.

Review model performance separately around distribution changes and supply disruptions. Ask whether the response curve is being extrapolated from normal availability into a very different operating state. A practical report should show sales, distribution, availability flags and media together, with a short explanation of excluded or revised periods.

Start the next budget test where availability is measurable and operational teams can maintain the plan. The goal is to learn a useful response under specified conditions, not to claim that inventory controls have made an observational estimate causal.

Sources and further practice

This article’s worked example is illustrative, not a measured client result. Use the linked methodology to check the assumptions before applying it.

• Vexpower: baseline sales
• Google Meridian: causal assumptions