Context & objectives

A Belgian company found that over 5% of their ultra-fresh products were unsold and discarded, resulting in hundreds of thousands of euros in annual losses. The products could only be kept for up to three days across several stores before spoiling.

The company faced significant forecasting challenges due to:

• a large variety of products with high sales volumes,

• frequent catalog changes,

• and high product substitutability,

which led to inconsistent sales data over time.

The biggest challenge was the inefficiency of manual demand prediction methods, which were time-consuming and error-prone. Our objective was to automate and improve demand predictions using AI. We identified two key opportunities for food stock optimization:

• Reducing overstock: the client needed flexible production capabilities and the ability to hire staff on short notice to meet demand fluctuations.

• Reducing stockout: the client maintained low stock levels to minimize waste. However, this approach risked lost sales when demand exceeded supply.

By developing a predictive algorithm, we aimed to provide precise demand forecasts considering historical sales, weather, and marketing events.

Approach

Our solution began with a thorough analysis of the context and close collaboration with the client to ensure alignment with their expectations and maximize project impact.

We shadowed a key contact to align the solution with high-impact areas, understand their challenges, and build trust in the final result.

3 layers of prediction

1. First, we forecasted demand per product several days ahead using a global forecasting model. With abundant sales data available, we built this model using the complete dataset. Most importantly, we split the sales data into "normal" and "celebration" days, as forecasts were more reliable for normal days.

   
   

2. Second, we optimized production orders using key business rules. We determined production quantities based on expected stock one day ahead and forecasted demand.

   
   

3. Third, we forecasted demand per product and shop one day ahead using the rolling median of sales for each product.

Model architecture

We accurately forecasted demand for each product by splitting our food stock optimization solution into two parts.

Model architecture

1. The baseline model

The baseline model was not AI-trained: it was simply calculated as the rolling median of sales per product for the last 4 weeks. It gave us a simple and robust computation of base sales that we could later incorporate with an AI-trained model. This layer acts as a normalisation step, enabling the forecasting of each product using a single model.

2. A multiplier

The multiplier, on the other hand, is an AI-trained model. This took the form of a global forecasting model that we trained on the last 14 months of historical data. We then applied the multiplier to the baseline to obtain the final forecasted demand.

This model allowed us to incorporate various additional factors that could influence the demand for sales. For example, we included weather, product details, holidays (national and school), and weekend versus weekday timing. Like all AI-based models, this approach requires sufficient data to generate accurate predictions; since our client regularly introduced and rotated a large variety of products, this posed a real challenge that we were ultimately able to tackle.

Another important success metric for this food stock optimization project was the estimated losses due to stockout. Many retail companies struggle to compute this metric because stockouts are difficult to detect and can result in significant lost sales. We developed a unique method for detecting when a product was missing from a store and calculating the resulting volume of missed sales.

Results

Our solution demonstrated very promising results.

The test phase of the model resulted in one of the best performing months so far for our client in terms of overstock and stockout.

Before and after model introduction: impact on stockout and overstock

Since both overstock and stockout could be reliably converted into a monetary value, our client observed tangible improvements to their profits.

Segmenting the data between "normal" and "celebration" days allowed the company to direct production efforts more efficiently. Products with sufficient historical data in the "normal" day segment showed the highest prediction confidence, with a mean error of less than 20%.

The client saw positive results across all key project objectives, and the solution became central to their supply chain planning. They even noticed that, when they overrode the algorithm's demand predictions, performance declined.

Conclusion

We focused on delivering a practical, working solution; not just a proof of concept. This required a robust yet simple approach, developed through close client collaboration and rapid iteration.

The result: a solution tailored to the client's pain points, deployed quickly and delivering maximum impact.