A supermarket buying fresh bread chooses between two kinds of loss every morning. Order too much and it goes in the bin at closing time. Order too little and the shelf is empty by four. That is the newsvendor problem, and with fresh produce it returns every single day, per product and per store.
What makes it hard: the demand is invisible
With a one-day shelf life, real demand cannot be read from the sales figures. On a day the shelf sells out, you do not know how many customers came afterwards. And those days are common: across the measured series, between forty and eighty-five per cent of selling days ended in a sell-out.
There is also little history. Per product per store, roughly twenty to forty days of usable sales data remained. Season, weekly pattern and promotional effects differ per store and per item as well. A standard demand forecaster has too little to go on.
What we researched
A method to forecast demand per product per store and set the ordering parameters accordingly. We researched it on a limited set of fresh products; it did not become a full replenishment system under our management.
The client, a retailer, prefers not to be named. A reference is available on request.
At its core is a mathematical model that ties demand, sales and replenishment settings together rather than estimating them separately. That way a sell-out day still yields information: the model knows sales ran up against the order quantity there, and therefore that demand was higher than what crossed the counter.
Why Bayesian, and not a standard model
We estimate using Bayesian statistics. It suits this problem because it reckons explicitly with uncertainty instead of quoting a single number: with twenty observations, the range is the answer, not the average. Sequential Monte Carlo methods, particle filters, keep the computation fast enough to take each new selling day into account without recalculating everything.
Why we test it against business results
A demand forecast you cannot check is an assumption with decimals. So we translate the forecasts back into the outcomes the store does observe, waste, lost sales and margin, and set those against what actually happened. That makes the estimate verifiable rather than plausible.
What it delivered
A study of one fresh product across ten randomly selected supermarkets showed that optimising the ordering parameters alone can raise profit by 4 to 7 per cent, without touching recipe, price or assortment.
That figure applies to that product and those ten stores. It is not a promise for an entire assortment, but it was measured rather than estimated. How the approach looks as a standalone service is on the page about inventory management.
About this article. Twentynext is responsible for the content, but cannot guarantee that it is accurate or complete. See our disclaimer.
