Using data to prevent food waste

Kickstart AI x Ahold Delhaize - Food Waste reduction Challenge 1.0

The Ahold Delhaize Food Waste reduction Challenge was a joint project between Ahold Delhaize and KickstartAI. The goal was to reduce  food waste at Ahold Delhaize’s local Belgium brand Delhaize by improving a complicated customer demand forecasting model with data science and machine learning. Seeing if we could balance product availability and waste.

Predicting customer demand with machine learning

To predict customer demand as accurately as possible, the KickstartAI team took a variety of factors and data into account. Such as product discounts, the time of day or week, and even the weather. The algorithm then looked at the relationship between all these inputs and resulting customer demand. As the algorithm processes more data, from a larger number of inputs, it can more accurately understand the impact of future conditions on customer demand.

Goals

1. Develop an improved machine learning model

   To forecast demand, outperforming the existing baseline model.

2. Quantify the impact of improved forecasting on key business KPIs

   Such as product shrinkage and product availability.

3. Create a simulation tool 

   To translate model performance into tangible business outcomes.

4. Demonstrate a measurable reduction in product shrinkage 

   While maintaining or improving product availability.

5. Share learnings and tools 

   With other Ahold Delhaize brands and KickstartAI partner companies.

Approach

We came up with a collaborative, sprint-based approach, using the expertise of data scientists from four leading Dutch companies (Ahold Delhaize, ING, KLM, and NS). This multi-disciplinary team brought a combination of skills in machine learning, data analysis, and supply chain management. We could pull historical sales data from Delhaize Belgium, use their powerful computational infrastructure, and a wide range of data science tools and libraries.

With a large, distributed team across multiple organizations, flexibility proved to be key to the process. The team started with a basic model and improved it along the way. Over three sprint weeks of development, we built and tested a simulation model to assess its impact, followed by months of in-store testing in real-world conditions at Delhaize Belgium. Rigorous data preprocessing and feature engineering ensured we could handle data complexity. With a proven MVP and validated simulation model, we successfully advanced to implementation.

• Sprint 1: initial baseline model development and data exploration.

• Sprint 2: model iterations based on data analysis and feature engineering.

• Sprint 3: translating model performance into business KPIs using a simulation tool.

After the sprints, we landed on some strong indicators to predict customer demand:

Strong cyclic patterns 
Strong weekly and yearly trends, with holidays mapping to specific peaks.

Effect of promotions 
The value of the promotion (€) is the primary driver of sales. Customers typically order 1-3 days in advance.

Product availability 
A significant predictor of sales, suggesting an opportunity to interpolate demand when supply is limited.

Price elasticity
Due to limited price changes, calculating elasticity is difficult, a potential area for business improvement.

Machine learning and data science cuts food waste by 21%

The Ahold Delhaize Food Waste Reduction Challenge demonstrated how data science and machine learning can reduce food waste. An improved demand forecasting model led to a 21% reduction in food waste at Delhaize Belgium while maintaining product availability. The project also delivered a reusable simulation tool and key business insights.

These tools and lessons are now being applied across Ahold Delhaize brands and KickstartAI companies, driving better demand forecasting and waste reduction. The project also highlighted the value of centralized forecasting to reduce the bullwhip effect and optimize supply chains industry-wide.

Results

• Simulations indicate a potential 21% reduction in product shrinkage (expired stock)

• Improved customer demand forecasting by 26%

• Our simulation tool that translates model performance into business KPIs

• Maintained similar product availability

• Outperformed the baseline across all 21 product categories