Betala mig kanske: Förutsägelse av sena fakturor med tolkningsbar maskininlärning : Minska manuellt arbete genom att förutsäga betalningsrisk vid orderbekräftelse

Predicting whether business-to-business invoices will be paid on time is critical for cash-flow planning and for reducing manual effort in credit control. Yet existing rule-based or heuristic screens struggle with the sheer scale and heterogeneity of enterprise data and cannot adapt quickly to shifting economic conditions. This thesis frames invoice payment behaviour as a supervised learning task at order level and investigates whether modern ensemble methods can deliver operationally useful risk scores. Using ≈ 40 000 credit-blocked orders from two European markets of Hilti Group, we engineered features capturing recent delay history, open exposure, payment terms and regional context, then benchmarked logistic regression against three gradient-boosted decision-tree implementations. The best model, LightGBM, reached a receiver-operating-characteristic area above 0.85 and a precision–recall area above 0.96 in both markets while maintaining F-scores around 0.94 at business-defined operating points. Shapley Additive Explanations confirmed that predictions relied on intuitive drivers, such as past average delay, open credit and failed credit checks, providing the transparency required for governance and user trust. A supplementary regression showed delay-length estimation is viable in one market (mean-absolute-error ≈ 15 days) but unstable in the other, underlining the context-sensitivity of payment dynamics. These findings make three contributions. First, they demonstrate that gradient boosting can flag late-payment risk with high recall at tightly controlled false-positive rates, enabling pre-emptive actions such as automated order release or targeted reminders. Second, they show that market-specific models outperform a hypothetical global model, suggesting that local data profiling and transfer-learning strategies are necessary when expanding to new regions. Third, the work delivers a reproducible pipeline for deploying interpretable machine-learning credit screens in other accounts-receivable contexts.