KAN LEVERANSRISK FÖRUTSES I KOMPLEXA ORDERSTRUKTURER? : En kombinerad maskininlärnings- och nätverksanalys

Sammanfattning

Cytiva manufactures and distributes products for the bioprocessing industry, where on-time delivery is critical for customer operations. Delivery precision is measured at order level, but inventory and service levels are managed at item level. When several items are bundled into a single shipment, the multiplicative nature of order-level success means that even items with high individual service levels can result in late orders. This thesis addresses three questions: whether delivery risk in customer orders can be predicted, which factors are associated with late deliveries, and how such predictions can support proactive decision-making. A binary classification problem is formulated using order-level data from 2024 and 2025. Logistic regression is used as an interpretable benchmark model and compared with XGBoost, a tree-based machine learning method. Both models are extended with network-based variables derived from a co-occurrence graph of items. In this graph, two items are linked when they are often ordered together. In a separate descriptive network analysis based on the same idea of co-occurrence, the graph is further analysed. The analysis includes clusters, cliques, edge density and risk pairs to identify item combinations associated with delivery risk. XGBoost achieves the highest predictive performance with a ROC AUC of 0.76 and a top 1% precision of 81% on test data. This means that 81% of the orders ranked as highest-risk by the model are actually late. The analysis of which variables contribute the most to the model's predictions shows that the time margin between the customer's requested date and the internal target date and the item's rolling historical delay rate contribute the most to the model's risk predictions. The network analysis identifies item groups that are associated with increased order-level risk even when the items' own row-level delay rates are lower. Combining the predictive model with the separate network analysis provides a basis for targeted decision support and leads to concrete, actionable proposals.