The chilled forecast gave every converted store a distribution of tomorrow’s demand for every product: a most likely number and a range around it. That is not yet an order. Order the middle of the range and a store sells out before the commuter peak on about half of all days. Order the top and it throws away a fifth of what arrives. The right point differs by product, because a salad with a good margin and two days’ life is a different bet from a dessert with a thin margin and five.

And stores do not order independently. Chilled products arrive on shared multi-temperature trucks with a cage limit per store. The central kitchen can make only so many of each recipe per day. A back-room chiller holds four cages, not six. A product with one day of life left can be marked down at 17:00, and the markdown price determines whether it sells or is binned. Every store’s best order depends on every other store’s through the truck and the kitchen.

The large grocers solve this with specialist systems. When Morrisons automated ordering with machine-learning replenishment across 26,000 products in 491 stores, it reported up to 30 percent fewer shelf gaps and two to three days less stock held in store. That programme covered ambient and long-life lines first, with fresh to follow, and it was built by a specialist vendor at national-grocer scale. An 85-store convenience chain does not get that project. It gets store managers ordering from a forecast and a feeling.

Order quantities as a solved problem

The supply chain manager asked for the whole decision at once:

> Set tomorrow’s chilled ready-meal orders for all 66 converted stores to maximize expected margin after waste, keeping evening availability above 92 percent on core products, within each store’s chiller space, each delivery run’s cage limit and the kitchen’s recipe caps. Mark down anything with one day of life left at the price most likely to clear it.

The Optimization Engine’s formulation followed the business logic closely. Decide: an order quantity for every store and product, the cages on every delivery run, and one of three markdown prices for stock close to expiry. Estimate: expected sales and waste at each order quantity, from the Prediction Engine’s demand distribution, and markdown response from past markdown sales. Must hold: chiller space per store, cages per truck run, recipe caps per day, and the availability target on core products. Maximise: margin after waste and markdowns.

That is about 3,200 linked store-product decisions every night. The engine linearised the sales-and-waste curves so the whole network could be solved together, and reached a plan within 0.2 percent of the proven bound in about ninety seconds after the day’s sales closed at 21:00.

On the first weekend the kitchen cap on chicken wraps was binding in fourteen stores. The engine reported what that cap was costing: the margin an extra two hundred wraps of daily capacity would add. The supply chain manager called the kitchen supplier on Monday with that number. Capacity rose the following week.

Over ten trading weeks, waste fell from seven to about 4.6 percent of units, and sales lost to empty shelves fell by more than a third. Both improved together because the plan puts stock in the stores and on the days where the forecast says it will sell, rather than spreading caution evenly.

Store manager orders 70 wasted96 sales lost to empty shelves

Order the forecast median 52 wasted88 sales lost to empty shelves

Optimization Engine plan 46 wasted58 sales lost to empty shelves

0255075100

Units per 1,000 ordered

Both numbers fall together: the plan puts stock in the stores and hours where the forecast says it sells.

Composite scenario; figures illustrative. Lost sales estimated from forecast demand on sell-out days.

Space for what sells

The trading director’s next question was about the chilled space freed by the deli conversion. The conversion dashboard had caught the small-format planogram problem in week six. She now wanted the allocation right before the next reset:

> Allocate chilled facings across categories in each converted store to maximize weekly margin. Keep at least two facings for every core product, use our fixed module sizes, and limit small-format stores to three metres.

Facings are integers, modules are fixed, and sales respond to space with diminishing returns - each extra facing sells less than the last. The engine estimated that response from past resets, stated that the estimate was uncertain for rarely changed categories, and solved per store format. Commuter stores went deep on wraps and sandwiches. Suburban stores gained hot pots. Small-format stores dropped sushi and kept wraps deep, which is the decision the week-six dashboard had pointed to, now made before the reset rather than after it. Projected margin per chilled metre rose by about six percent.

	Commuter	Suburban	Small format
Sandwiches & wraps	14 facings	9 facings	8 facings
Salads	8 facings	7 facings	4 facings
Hot pots	4 facings	8 facings	3 facings
Sushi	4 facings	2 facings	not ranged
Desserts	2 facings	4 facings	2 facings
Chilled drinks	6 facings	6 facings	4 facings

• Small format capped at 3 msushi dropped, wraps kept deep
• Every core product keeps ≥ 2 facings

Composite scenario; figures illustrative. Space response estimated from past resets and reported with its uncertainty.

What the store manager keeps

Store managers can still override any order. Overrides are logged, and the solver re-plans the rest of the truck around them, so one manager’s local knowledge does not quietly overfill another store’s cage. The pattern from the forecasting work continued: overrides before local events were usually right, and those events became inputs.

Larsen & Holt’s trading meeting has now lost two pieces of folklore. “It drives footfall” was tested in the Analytics Engine. “Order a bit more, it’s the weekend” has been replaced by a forecast with a measured error and an order with a proven gap. Neither conversation needed a new data team. Each needed the person responsible for the decision to be able to ask for the answer directly.