Stockout prediction retail — Regular demand is easy to forecast and rarely causes stockouts. The damage comes from the events your model has three examples of.
Stockout prediction retail: bu ne anlama geliyor
Availability reports show a service level of 96%. The category manager is unhappy. Both are accurate, and the gap between them is where the money is.
The 4% is not distributed evenly. It concentrates on new products, promoted lines, seasonal peaks and anything affected by an event the model has barely seen.
Why models are worst exactly where it matters most
A statistical model needs examples. A steady-selling line has three years of weekly data. A product launched last month has four points, and the promotion starting next week resembles two previous promotions, both under different conditions.
The model is confident where the stakes are low and uncertain where they are high.
What to do about it
Stop treating the forecast as a single number. Where uncertainty is high, the decision should change shape: order more conservatively, position stock centrally rather than pushed to stores, or hold a review.
An agent that says “I do not have enough history to be confident here” is more useful than one that produces a precise number from four data points.
Why the average hides the loss
A 96% service level across 8,000 SKUs means 320 items were unavailable at some point. If those 320 were the slowest lines in the range, the commercial impact is close to nothing. If forty of them were the promoted lines in the week of the promotion, the quarter is affected.
The report cannot tell the difference, because it weights every SKU equally and the business does not. Weighting availability by margin, or by promotional status, produces a different number and a different set of priorities — and it can be computed from data already held.
Most organisations discover on doing this that their worst availability sits exactly on the lines they most wanted available.
Uncertainty has to reach the decision
A model that emits a point estimate has thrown away the one thing that would have changed the action. Four data points and three years of history produce the same kind of number, and the planner cannot tell them apart.
Carrying the interval through — not as a chart for a review meeting, but as an input the decision reads — is what allows the action to change shape. Wide interval: hold stock centrally, order in smaller increments, schedule a review. Narrow interval: commit.
This is a plumbing problem more than a modelling one. Most forecasting systems compute an interval internally and discard it at the interface.
What to do about the cases with no history
Borrow, and record what was borrowed from. A product launched last month has no history of its own but sits in a category with attributes: price point, pack size, brand tier, shelf position. Similar products have histories, and their aggregate behaviour is a better prior than an extrapolation from four points.
The record matters as much as the estimate. A buyer who can see that the forecast for a new line was built from three named comparables can say the comparables are wrong. A buyer shown a number can only say it feels high.
The promotion case specifically
Promotions are where this failure concentrates, because each one is nominally unique and actually a member of a family: depth of discount, position in the store, whether a competitor ran something similar.
Modelling promotions as a family rather than as individual events is what makes the uplift estimate improve rather than reset every time. Organisations that keep a structured record of promotional mechanics get better at this within a year; those that record only the sales uplift do not.
The honest summary
The 96% is not wrong. It is answering a question — how often was any given item available — that nobody in the business is asking.
The question they are asking is whether the things that mattered were there when they mattered, and that requires weighting, uncertainty and a record of what the model did not know.