Predictive modeling of parcel pick-up behavior and analysis of its impact on logistics profitability: case of the Sky Nine Cargo freight agency in Lubumbashi

Abstract

The freight and logistics sector is experiencing significant growth in Lubumbashi, with an increase in imports from China and other countries. The agency Sky Nine Cargo, specialized in freight transport and consolidation, faces a recurring problem: many clients do not collect their packages within the allotted 14-day period, resulting in an accumulation of goods, higher storage costs, tied-up capital, and slower inventory turnover. This work proposes a solution based on predictive modeling to anticipate package collection behaviors. By leveraging the agency’s historical data (receipt dates, client type, package value, geographic distance, etc.); we built and compared three machine learning algorithms: Random Forest, XGBoost, and LightGBM. The CRISP-DM methodology guided the entire process, from understanding the problem to model deployment. After data preparation (cleaning, handling missing values, encoding categorical variables, data normalization, and class balancing using the SMOTE technique), the LightGBM model was selected as the most effective. It achieves an AUC of 0.796, a recall of 99.2%, and a precision of 69.4%. These results make it possible to reliably identify clients at high risk of delay in collecting their packages. An operational strategy of personalized reminders via WhatsApp is proposed, with a decision threshold set at 0.45. The model thus enables the detection of 99.2% of potential delays, helping reduce unforeseen storage costs, improve warehouse space management, and optimize the company’s logistics performance. The completion of this study required the use of several tools and hardware resources. The software tools used include Python as the programming language, Jupyter Notebook as the development environment, the Pandas and NumPy libraries for data manipulation, Scikit-learn for model building and evaluation, XGBoost and LightGBM for machine learning, Imbalanced-learn for applying the SMOTE technique, as well as Matplotlib and Seaborn for visualizing results.