Phantom inventory screening
Flags products whose sales have stopped for longer than their demand history makes plausible while the inventory record still shows stock, and sends a store worker to check the shelf.
Discovery, not a referral queueTechnique: Anomaly Detection
Phantom inventory screening looks for products that should have sold and did not, in stores where the system believes they are in stock.
For each product and store it compares the run of days without a sale against how unlikely that run is given the product’s normal sales rate, so a fast seller is flagged after a day or two and a slow seller only after a much longer gap.
The ranked flags go to store staff as a short daily list, sized to the time they actually have, and each one is checked on the shelf and in the backroom.
A confirmed empty shelf has its record corrected, which lets the reorder point place a normal order, and stock found in the backroom goes straight to the shelf.
This is a discovery tool, not a decision tool: an unusual gap in sales only means something is worth checking.
It is also reactive, since a flag can only appear after the shelf has already been empty for a while, so it complements frequent counting rather than replacing it.
It is the wrong place to invest first in a store estate that already counts inventory frequently and accurately, since there will be little phantom stock for it to find.
Strengths for this problem
- Finds empty shelves that the replenishment system will never reorder, because its record says the stock is there.
- Needs no labelled examples, only sales history and the inventory record that are already collected.
- A confirmed flag fixes availability immediately, by correcting the record and triggering a normal reorder.
- Each confirmed or rejected flag becomes a label that can later tune the screen or train a model.
Limitations for this problem
- A run of zero sales is ambiguous for slow sellers, which can go weeks without a sale while stock is present.
- Cannot tell stock missing from the store apart from stock in the backroom that never reached the shelf, which only a store visit resolves.
- Overstated records that still show sales are invisible to it, because sales keep the product looking healthy.
Data needed
- Daily units sold per product and store, and the expected sales rate from the stockout-corrected forecast.
- Daily perpetual inventory on hand, with the date of the last count or adjustment.
- Receipts, transfers, write-offs, and count adjustments, to explain sudden changes in the record.
- Store visit outcome per flag, recorded as confirmed empty, found in backroom, or stock present.
How it is measured
- Precision of flags confirmed empty or found in backroom by store staff, per category.
- On-shelf availability in screened stores against a matched set of unscreened stores.
- Recovered sales on products in the weeks after a confirmed flag is corrected.
- Store labour minutes per confirmed flag, which sets how many flags a store can be asked to check.
Pitfalls
- Sending more flags than stores can check, so staff stop checking and the screen loses its only source of truth.
- Using one fixed run of zero-sale days for every product, which floods stores with slow sellers and misses fast sellers that stopped yesterday.
- Correcting the record automatically from the flag without a shelf check, which writes off real stock that was merely slow.
Human oversight
A store worker checks the shelf and backroom before any record is changed, and records what they found. Loss prevention, not the screen, decides whether a pattern of confirmed flags in a store warrants investigation.