Out-of-stock detection is software that reads a retailer's till data to find products that should be selling but are not. A product that suddenly stops selling is often missing from the shelf, even when the store's stock records say that it is there. The software also finds the waste and the stock errors behind such gaps. It is also called on-shelf availability analytics.
Here is the problem it solves. A supermarket's back office keeps a stock record for every product. The record can be wrong. A pack may be stolen, damaged, or scanned at the till as a different product. Then the record shows stock that does not exist. This is called phantom stock. The store's ordering system trusts the record, so it does not reorder. The shelf stays empty, sometimes for weeks, and nobody notices.
The software looks at sales rather than at stock. It uses demand forecasting to predict how many units of each product each store should sell in each hour. Then it uses anomaly detection, which finds numbers that differ sharply from the usual pattern. If a product that usually sells steadily has sold nothing for several hours, the software flags it.
Imagine a supermarket in Kolkata. A 400 g cup of curd usually sells about 12 units a day there, mostly in the morning. On Tuesday, by 2 pm, it has sold none. The retail POS back office still records 30 cups in stock. The software sends an alert to the staff app: check the curd. An employee finds the shelf empty. The 30 cups are in the cold room at the back, because nobody brought them out that morning. She fills the shelf, and the next cup is sold 20 minutes later.
Each alert goes to the store, and the patterns go to the head office. If one store's stock record is wrong again and again, the cause may be theft or wrong deliveries. If fresh products are often thrown away in one store, that store is ordering too much. The head office then corrects the store's orders, and its merchandise planning.
Out-of-stock detection works from data the store already has, so it needs no cameras. That is the main difference from shelf image recognition, which reads photos of the shelf. The data method covers every product in every store, every day. But it cannot see whether a product is placed correctly. It can only tell that the product is not selling. So some retailers use both.