Client Order Forecasting in a Semi-Digital Meal Delivery System Using Machine Learning

Abstract

This master’s thesis was conducted during an internship at Nubios on behalf of Alice Traiteur, a company based in Liège that delivers daily meals to households. The main goal of the project was to design a forecasting system capable of predicting what each client is likely to order and when, based on historical data, and to integrate it in an already existing application. The solution aimed to improve operational planning, reduce manual errors, and improve customer satisfaction. Over one million historical orders were analyzed. An Exploratory Data Analysis (EDA) provided insights into customer behavior, ordering frequency, dish combinations, and time gaps between orders. The fore- casting problem was approached using supervised learning, with a focus on time-aware validation. Multiple models were tested, including Random Forest, XGBoost, neural networks (MLP, RNN), and Transformers (TFT). Among these, the customer-based Random Forest model achieved the best overall performance. The system was designed to be non-intrusive. Predictions are delivered through a dedicated dash- board separate from the main application. Several features were implemented: future order forecasts, forgotten order detection, revenue predictions, and client-specific follow-up alerts. The machine learning pipeline, deployed on Amazon SageMaker, automatically retrains the models monthly and stores results in a dedicated back-end. Feedback from the client was very positive. The system provided accurate forecasts of daily orders and expected sales. The predictions closely matched real-world values. One suggested improvement was the addition of an automatic carry-over feature for recurring orders. Overall, the project demonstrated the real-world value of AI in supporting data-driven decisions without disrupting existing workflows.