On this page
When a store runs out of a fast seller, the reflex is to raise the order quantity. It works — for that SKU, for a few weeks. Then working capital tightens, markdowns rise somewhere else, and the same conversation happens about a different product.
Many stockouts reflect demand, lead-time or phasing errors that ordering rules are then asked to absorb. Treating the symptom alone can move excess stock elsewhere without resolving the underlying signal.
Where availability is actually lost
Useful diagnoses start by separating several common mechanisms:
- Demand was under-forecast at the location level. Chain-level demand was close; store-level demand was not. Aggregate accuracy hides local misses in both directions, and they do not cancel out on the shelf.
- Lead time variability was ignored. Safety stock was sized for average lead time, leaving too little protection when delivery is late.
- The forecast was right, the phasing was wrong. Monthly volume was accurate but the demand arrived in week one.
- Nobody modelled the substitution. The out-of-stock SKU's demand moved to a neighbour, which then also stocked out, and the history now records both as "low demand".
Some replenishment reports obscure these differences, especially when forecast error is measured only at an aggregate level.
Why ordering more hides the problem
Increasing cover is a blunt instrument. It works because it buys slack against all error sources at once, and it costs because it buys that slack for every SKU, including the ones that were forecast well.
The result can be a split inventory position: overstocked in aggregate and understocked in specific locations or variants.
The four numbers worth watching
Instead of tracking fill rate alone, track these together:
| Metric | Definition | Why it matters |
|---|---|---|
| Forecast bias by location | Mean signed error over the horizon | Persistent negative bias is a structural stockout generator |
| Lead-time variance | Standard deviation of actual vs. planned lead time | Sets the floor on how much safety stock you actually need |
| Lost-sales estimate | Demand during out-of-stock periods, reconstructed | Turns availability into a revenue number executives can act on |
| Cover distribution | Days-of-cover spread across SKU-locations | Reveals split positions that averages conceal |
Bias deserves separate attention. For example, a model with respectable aggregate error and a persistent −6% bias will repeatedly understate the replenishment signal; safety-stock tuning alone does not remove that systematic error.
Reconstructing lost sales
You cannot measure what you did not sell, but you can estimate it. The workable approach:
- Identify out-of-stock windows from inventory snapshots, not from sales gaps.
- Estimate expected demand across the window from comparable periods and comparable locations.
- Subtract observed sales. The remainder is your lost-sales estimate.
- Censor those periods in training data so the model does not learn that a stocked-out week was a low-demand week.
Step 4 matters as much as the reporting. Uncensored stockout periods can teach a forecasting model that constrained sales represented unconstrained demand, allowing the error to repeat.
What good looks like
A healthy availability programme predicts demand at the grain of the decision (SKU × location × period), sizes buffers from measured lead-time variability rather than a single planning number, measures bias separately from accuracy, and feeds censored history back into training.
Order quantity then becomes one output of the diagnosis rather than the only lever.
Sources and further reading
These references support the technical concepts discussed above. Examples and recommendations in the article remain editorial interpretation.