In the ultra fresh food retail industry, stock optimization has the potential to deliver the greatest impact on profits, while saving precious resources.

A Belgian company found that, on average, 6% of their final products were not sold fast enough and had to be thrown away. This resulted in hundreds of thousands of euros in losses annually.

The biggest challenge our client faced was the inefficiency in predicting their demand accurately. They relied on manual methods, which were time-consuming and prone to errors.

Our objective was to automate and improve demand predictions for our client using AI. Most importantly, we needed to reduce excess inventory while avoiding stockouts.

Context & Objectives

Our client produced ultra fresh products. This meant that the products could only be kept for up to three days in one of the 20+ stores before they had to be thrown out.

There was a large variety of products with a high volume of sales, frequent changes in the daily catalog of products on sale, and a high degree of substitutability of the products. This leads to inconsistent sales data over time and makes forecasting demand challenging.

Using AI, our solution had the greatest potential to benefit our client in two ways:

1. There is the potential for improved efficiency by reducing overstock. Our client needs to be very flexible in their production capabilities and hire on very short notice to adapt to the demand.

2. There is the potential to reduce stockout. Our client aims to keep very low stock levels because they do not want to throw out excess. However, this could also lead to lower sales if they do not make enough of a product to meet demand.

Therefore, our solution needed to find a good balance in the amount produced to minimize overstock and stockouts.

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

Approach

Our solution began with a deep dive into the context and close collaboration with the client. We needed to make sure that the proposed solution aligned with what the client was expecting and that the project could deliver the highest impact.

To do this, we closely shadowed a key point of contact with the client so that the solution was clearly aligned with where the highest impact could be obtained. We also had to make sure we understood their pains and the client knew they could trust the final result.

We used three layers of prediction

First, we forecasted demand per product for 6 days ahead using a global forecasting model. There was a lot of available sales data so we could this model based on all the available data.

Most importantly, we decided to split the sales data between "normal" and "celebration" days. This is because there was a higher degree of confidence in the forecast for days that were not celebration days.

Second, we optimized production orders using key business rules. We determined the quantity that the client should produce based on the current expected stock one day ahead and the forecasted demand.

Third, we forecasted the demand per product and shop for 1 day ahead using the rolling median of sales of every product.

Model architecture

We were able to accurately forecast demand of each product by splitting our solution into two parts:

• The baseline model

• A multiplier

Model architecture

The baseline model was not AI-trained. It gave us a simple and robust computation of base sales that we could later incorporate with an AI-trained model. It acts as a normalisation step, enabling the forecasting of each product using a single model.

So, how did we define the baseline model? It was simply calculated as the rolling median of sales per product for the last 4 weeks.

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 information, whether it was a national or school holiday, or even whether the sales occurred on the weekend or weekday.

However, as is the case with all AI-based models, the only limitation is that we require sufficient data to obtain predictions. Since our client regularly introduced and rotated a large variety of products, this could pose a real challenge.

In addition, an important metric determining the success of the project was the estimated losses due to stockout. Many companies in the retail sector can find it difficult to compute this metric. This is because it is often difficult to detect and can result in large losses in sales.

We developed a unique way of detecting when a product was missing from a store and the volume of sales that was missed as a result.

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

Both overstock and stockout could be reliably converted into a monetary value. Thus, our client was able to observe tangible improvements to their profits.

We found that segmenting the data between "normal" and "celebration" days enabled our client to efficiently direct its production efforts the most. Products that had sufficient historical data available and in the "normal" day segment had the highest degree of confidence in the predictions. We observed a mean error of only 17% in this category.

Our client realized a positive impact across all the key objectives of the project. In addition, the solution became a central part of their supply chain planning. Not only that, our client even noted that when they overwrote the algorithm's demand suggestions, they observed worse results.

Conclusion

Throughout the development of this project, we were committed to delivering a solution that our client could work with right away, rather than just a theoretical proof of concept.

To do this, we had to develop a solution that was both robust and simple. This required us to work very closely with the client so that we could iterate and improve our model rapidly.

Our approach allowed us to develop a working solution very quickly. We tailored it directly to the pain points that the client faced, leading to a solution that delivered the highest impact possible.

—

Written by Joleen Bothma