A machine learning approach for engineer-to-order firms to predict supplier performance in critical supply chains

Predicting supplier performance is critical to downstream decision-making and a firm's success. Current industry practice relies on institutional knowledge, simple rating scales, and expert opinion instead of empirical data to predict performance during supplier selection. The nuclear industry is a critical supply chain that relies on engineer-to-order (ETO) manufacturing environments that cannot rely on traditional risk-mitigating activities like allocating orders to multiple suppliers or maintaining inventory like make-to-stock (MTS) environments can. Supplier performance management is important in ETO environments because of fewer qualified and capable suppliers, irregular demand, higher product costs, and longer lead times. Motivated by industry practice, this article proposes using gradient boosting (GB), combined with synthetic minority over-sampling technique and hyperparameter tuning, to predict supplier delivery, hardware quality, and documentation quality. Based on its superior prediction performance, GB was selected from three other machine learning (ML) methods - binary logistic regression, random forest, and support vector machine. A GB model was trained and tested using a firm's operational data. The resulting model was implemented to illustrate its usefulness in critical decisions, such as supplier selection. This work is the first to use actual production data from the nuclear industry to predict individual supplier performance using ML.