2025-04-23 –, Zeiss Plenary (Spectrum)
Ever craved your favorite dish, only to find its key ingredient missing from the store? You're not alone - stock outs can have significant consequences for businesses, resulting in frustrated customers and lost sales. On the other hand, overstocking can lead to wasted storage costs and potential write-offs. The replenishment system is responsible for striking the right balance between these opposing risks.
The key to successful replenishment is making accurate predictions about future demand.
This presentation takes a deep dive into the intricate world of demand forecasting, at Europe's largest retailer. We will demonstrate how enhancing simple machine learning methods with domain knowledge allows to generate hundreds of millions of high-quality forecasts every day.
This talk will provide an in-depth look at the forecasting engine, the heart of Lidl's replenishment system.
Each day at Lidl, hundreds of millions of various products journey from suppliers to warehouses before reaching the shelves. Our so-called forecasting engine helps to automate the supply chain at every step along the way.
Even with the vast amount of data at our disposal, the problem is still extraordinarily intricate. Each item, store or warehouse has unique demand patterns influenced heavily by a wide range of factors, such as holidays. While most of the effects are quantifiable, others remain unavailable and a certain degree of stochasticity is inherent to the process. The objective of our demand prediction may also vary based on their usages. Accuracy on the day level typically matters for short-term predictions, while it doesn't for long-term predictions.
We'll present our pragmatic modeling methodology on a simplified version of the problem at hand: The warehouse forecasting of single items.
We explain the rationale for training separate models for each item-warehouse combination and go into the reasons why we opted for using a LGBM model and why we believe it is best suited for our application. In addition to outlining our high-level modeling approach, we demonstrate how business and domain expertise are integrated into the modeling process through the use of sample and feature weighting and examine the impact of this integration on prediction quality. Following the base model, extensions are introduced that enable the incorporation of higher-level information at the finest level of granularity. This is achieved through decomposition and recomposition of the time-series at hand. In detail, we will present uplift decomposition for different use-cases, which include handling of promotions and holidays.
To conclude, we will give an overview of how all the presented methods synergize in delivering reliable forecasts for happy customers, so that you will hopefully never find yourself in front of an empty shelf!
Novice
Expected audience expertise: Python:Novice
Machine Learning Engineer at Schwarz IT, Germany, where I'm passionate about harnessing the power of AI to revolutionize the retail industry
Versatile data scientist with 3+ years of experience building AI-products at the service of the industry. I believe that the key for success revolves around embracing shared best practices, upholding high quality standards for code development and having a team composed of complementary skill sets.