Food industry demand forecasting: baseline, promotions, use-by dates

Demand forecasting in the food industry has a feature other industries do not: the product goes off. With a use-by date only a few days away, there is no way to absorb a forecasting error with stock. What is not sold is thrown away, what is missing is a lost sale. Yet many SMEs in the sector still forecast with an inherited Excel file, adjusted every week by one person who keeps it all in their head.

Structuring the forecast does not require a complicated model. It means separating the regular from the exceptional, measuring seasonality, projecting stock with the use-by date in mind, and measuring the error. This article describes that step-by-step approach, as we have applied it in the food industry, and the mistakes we come across most often.

Why food industry demand forecasting is unlike any other

Four constraints add up:

• Short use-by dates. The use-by date caps the stock cover you can hold. A product with a seven-day use-by date, where the customer demands five days of remaining shelf life on delivery, leaves two days of margin.
• Retailer promotions. A catalogue promotion can triple an item's volumes over two weeks, then drop back. Mixed into the history, it distorts every average.
• Seasonality and weather. Barbecue season, Christmas and New Year, school holidays, public holidays: demand shifts from one week to the next, and even from one day to the next.
• Production lines. Line rates, recipe changeovers and cleaning mean the load has to be smoothed; a daily forecast serves that purpose first.

Separating the baseline from promotions

The baseline refers to an item's regular sales: what the customer orders week after week, outside any promotion. Promotions are one-off volumes, negotiated per customer, with a start and an end. The two are not forecast the same way, and mixing them is the leading cause of wrong forecasts in the sector.

The baseline: a daily reference week

The first step is to define, for each item, a standard week with a quantity per day, which applies by default to all upcoming weeks. It is built from the history excluding promotions and is corrected by an increase or reduction coefficient decided by the planner: start-of-month uplift, school holidays, public holiday, weather forecast. The coefficient can apply to a period, a product family or an item.

Promotions: managed per customer and item, with statuses

Each promotion is recorded as soon as it is proposed to a customer, and follows a cycle:

1. Simulation: a feasibility check, i.e. the added load against the available capacity of the lines.
2. Proposal: once approved internally, the promotion is sent to the customer.
3. Accepted: the customer confirms it and it enters the production plan.

Once accepted, the quantity must be spread over the duration of the promotion, with three levels of smoothing:

• simple smoothing: even spread over the period;
• automatic smoothing: coefficients calculated from the same customer's past promotions (for example 40% in the first week, 60% in the second);
• advanced smoothing: readjustment as actual orders come in, with the balance spread over the remaining days.

Seasonality, public holidays, weather: explicit coefficients

Rather than adjusting quantities by hand, it is better to set readable coefficients that can be explained and corrected:

• Weekly seasonal coefficient = the week's sales excluding promotions ÷ the year's weekly average. Averaged over two or three years, it smooths out the accidents of any single year.
• Day-of-week coefficient = the day's sales ÷ the week's daily average. It shapes the reference week.
• Event coefficient, decided by the planner for a public holiday or a bridge day, documented with a comment.

Example: a reference week of 1,000 units, a seasonal coefficient of 1.25 for the week of 14 July (Bastille Day, a public holiday in France) and an event coefficient of 0.8 on the holiday itself give 1,250 units for the week, with the 14th reduced by 20%. Every figure can be explained, and the rule will be reused next year.

Projected stock and cover when the use-by date is short

For a fresh product, the projected stock formula does not change, but how you read it does:

Projected stock on day D = closing stock D−1 + planned production D − sales forecast D

Cover (in days) = projected stock ÷ next day's forecast

Cover has a ceiling: the use-by date minus the remaining shelf life demanded by the customer. Beyond that, the stock will end up as waste. A fictitious example: an item with a seven-day use-by date and five days of remaining shelf life required, i.e. a maximum cover of two days, produced at a fixed rate of 1,200 units a day, with 500 units forecast for Saturday:

At a fixed rate, stock drifts from mid-week because demand falls. The answer is not to produce “just enough” on Friday, but to cut production from Wednesday or switch the line to another item. A daily forecast lets you decide two days before rather than throw away two days after.

Retrieving and classifying the history

For the coefficients to be right, every past sale must be classified as baseline or promotion, by matching order lines against the recorded promotions, customer by customer. Without this classification, the history contains the promotional peaks and the reference week is overestimated.

You also need to measure the gap between forecast and actual by item and by week, again separating baseline and promotions: a recurring gap on an item signals a reference week to review; a gap on a promotion, a smoothing rule to correct for that customer.

The classic mistakes

1. Mixing promotions and baseline in the same history, then being surprised that the baseline rises every year.
2. Forecasting by the week when the line is planned by the day and the use-by date is counted in days.
3. Correcting quantities by hand with no coefficient or comment: the following year, nobody knows why week 28 was at +30%.
4. Letting sales change the forecast without a trace, after the cut-off, and discovering the gap on the shop floor.
5. Waiting for the perfect tool before structuring the process: process first, then the tool to keep it running.

What we see in food forecasting projects

We have run this type of project in a food group, with significant gains in accuracy and efficiency (see our case study on the digitalisation of an Excel forecasting file). Three observations stand out.

First, the process is simple on paper, but it relies on rigorous data structuring: items, customers, promotions with their statuses, classified history. A tool is essential to sustain it over time: standard forecast and promotions visible side by side, adjustable coefficients, promotion statuses and entry into the production plan, automatic updating of the history and the smoothing coefficients.

Second, the Excel file that does all this often already exists and contains the business rules. Its problem is not the logic, it is the volume, the fragility and the lack of history.

Finally, the question of granularity always comes back: a forecast that is accurate at site level is not accurate item by item, and the plant works at item level (see our article on centralising sales forecasts).

Conclusion: four building blocks for fresh food forecasting

Structuring demand forecasting in the food industry comes down to four building blocks: a daily reference week for the baseline, promotions managed per customer and item with statuses and smoothing, explicit coefficients for seasonality and events, and projected stock read with the use-by date as a ceiling. All of it measured every week, one block at a time.

Planiprev follows this structure: baseline kept separate from promotions, item-by-item forecasting, coefficients and comments, projected stock, cover and anticipated stock-outs, change history. It reads your ERP or your Excel workbooks. See how Planiprev handles baseline and promotions or request a demo.