Machine Learning for Waste Reduction in Food Businesses

Waste is a forecasting problem wearing an apron

A bakery chain bakes to a production plan built on last week's sales plus a buffer, because running out of croissants at 10am is visibly bad and binning twenty at close is invisible to customers. A restaurant preps for a busy Saturday that turns out rainy. A grocer orders salad for a heatwave that ends on Wednesday.

Each of those is a demand forecast that was too high, on a product that cannot be held over. The buffer feels safe because the cost of waste is spread across hundreds of small bins rather than one painful moment. Add it up across a year and it is often one of the largest controllable costs in the business.

Where waste comes from, by type of business

The table points to the main theme. For most food businesses, the first and largest lever is a better demand forecast. For manufacturers, batch losses matter too, which we cover in yield prediction on production lines.

Forecasting fresh demand per site and day

Fresh product forecasting is harder than general retail forecasting because the unit is small and the horizon is tomorrow. Sales of one pastry in one shop are noisy. That noise is manageable with the right inputs.

• Day of week and time of day, since morning and afternoon mix differ
• Weather forecast, which matters a great deal for salads, cold drinks, soups and ice cream
• Local calendar: school holidays, events, nearby office occupancy
• Promotions and menu changes
• Stockouts in the history, which hide true demand and must be corrected for

That last point catches many projects. If croissants sold out at 11am yesterday, recorded sales understate demand. A model trained on raw sales learns to under-produce, which looks like success on the waste report and failure in the shop. Recording sell-out times, even roughly, fixes this.

Turning a forecast into a production or order quantity

The forecast is a range, and the quantity to produce is a choice about balancing waste against lost sales. For perishable items this is a textbook newsvendor problem.

1. Estimate the margin lost by running out of one unit
2. Estimate the cost of one unit wasted, net of any markdown or donation value
3. Produce to the percentile of the demand forecast that balances those two

A high-margin item that is cheap to make should be produced generously. A low-margin, expensive item should be produced tight. Applying the same buffer to both is how waste and stockouts appear together on the same shelf. The underlying method is similar to the one in our post on setting safety stock with a model.

Markdown pricing for short-dated stock

For grocers and retailers, a second lever is reducing prices on stock approaching its date. Machine learning can predict how much will sell at each price over the remaining time and suggest the smallest discount that clears it.

This works best with electronic shelf labels or at least a simple label-printing workflow, and with clear rules: never discount below cost unless the alternative is disposal, never discount items within food safety limits that require removal. Redistribution partners and staff sales are also part of the answer, and a model can tell you early which items will not sell so they can be donated while still useful.

When this is not worth building

A single café with a good manager who adjusts orders by feel will often do better than any model, and should not pay for one. Forecasting at site and product level earns its keep across several sites, many products or high volumes, where no single person can adjust everything each day.

• Waste is not recorded, or only recorded as a weekly total
• Tills do not record product-level sales by time
• Staff will override suggested quantities without any feedback loop
• The real waste source is storage failure or poor rotation, which is a process problem

Food safety deserves a firm line. A model decides how much to produce and when to discount. It never decides whether something is safe to sell. Use-by dates, temperature control and HACCP stay exactly as they are.

How SpiderHunts would approach it

We would start with a waste baseline: what is thrown away, by product, site and day, joined to sales and production. Many businesses discover their waste is concentrated in a small number of products and days, which focuses everything that follows.

Then we build a forecast per product and site, a suggested production or order quantity each afternoon, and a simple screen for managers to accept or adjust it, with reasons. SpiderHunts builds the forecasting as machine learning development and connects it to the tills and ordering systems already in use. For the wider operational side of food businesses, automation for food and drink distribution is a useful read.

Frequently asked questions