Custom ML Predicts Grocery Demand
Pradeep Elankumaran, CEO of Farmstead, on the future of online grocery
The real moat in online grocery is not the warehouse, it is the prediction system that decides exactly how many cartons of milk, bunches of basil, and trays of salmon to buy for each location and each day. Grocery margins are thin, so a few points of spoilage can decide whether an order makes money. Farmstead ties low waste to data collected from its own operating software, then uses per item models to balance two things at once, keeping items in stock for customers and keeping fresh inventory from expiring on the shelf.
-
This is why manual rules break down. A buyer can make broad calls like stock more milk on weekends, but perishables have different shelf lives and demand patterns. Milk may sell through over two to three weeks, while fresh salmon may need to move in two or three days, so ordering has to happen at SKU level, not by gut feel.
-
The payoff shows up directly in gross margin. Farmstead reports under 2.5% food waste in its best markets and says sub 2% is possible while still maintaining high availability. That matters because supermarkets often absorb double digit waste across stores and distribution centers, while dark store models live or die on contribution margin after labor and delivery.
-
This also explains the split between models. Instacart built scale by using existing stores and shoppers, which avoids owning inventory risk. Vertically integrated operators like Farmstead take inventory risk themselves, but in return they can improve margins through tighter purchasing, direct supplier relationships, and a narrower SKU set of roughly 1,500 items instead of supermarket scale assortments.
As online grocery matures, the winners are likely to look less like marketplaces that simply move store shelves onto an app, and more like software driven grocers that treat forecasting, replenishment, and assortment as core product. Better models let operators carry more fresh items with less waste, raise gross margin into the 30% range, and expand without repeating Webvan era inventory mistakes.