This thesis was conducted on behalf of a company in the automotive industry, where material shortages recurrently affect production and deliveries. The classification of material shortages is currently performed manually based on employees’ experience and judgement, which may lead to inconsistent prioritisation and limited transparency in the decision-making process. At the same time, the company operates a well-developed enterprise resource planning system that generates large amounts of data related to, among other things, material flows, lead times, and delivery status. This information is currently used only to a limited extent for systematic analysis.The purpose of this study was to investigate whether machine learning can be used to model historical prioritisation patterns for material shortages. The modelling was performed using a dataset consisting of 74 357 observations distributed across four classes: red, yellow, green, and blue. Random Forest and XGBoost were used as classification models and evaluated together with different strategies for handling class imbalance. The results show that XGBoost with automatic class weighting achieved the best performance and that weighting-based strategies consistently outperformed undersampling. Variables related to time deviations and material structure had the greatest influence on the models’ predictions. The findings indicate that machine learning may provide support the classification of material shortages, but that model performance is affected by dependent observations and the risk of data leakage.
Keywords: Machine learning, classification, material shortages, Random Forest, XGBoost,class imbalance, SHAP, feature importance, predictive modelling.