Explainable AI-Driven Supplier Risk Assessment and Predictive Analytics for Resilient Supply Chain and Procurement Management

Abstract

Supplier disruptions are among the most frequent and costly risks that procurement organisations manage, and machine-learning models that forecast them are increasingly offered as a remedy. Two questions are seldom answered: how much such models add over the scorecards and rules that buyers already use, and whether their explanations identify the cause of a supplier's risk well enough to choose the right mitigation. This article answers both in a transparent simulation study. A synthetic base of 500 suppliers over 72 months carries four latent risk states (financial stress, operational fragility, quality-system weakness and compliance exposure), regional disruption episodes and four disruption types drawn from logistic hazards with a known ground-truth driver, observed through lagged, noisy and partly missing signals. Gradient-boosted trees forecast disruption within six months with a test AUC of 0.72 (0.71–0.75 across five bases), against 0.71 for an expert scorecard, 0.68 for a single on-time-delivery rule and a ceiling of 0.84 for the generator's true risk; an interpretable logistic model matched the boosted model. Explanation diagnostics show stable global rankings (Spearman 0.94 across refits) but moderate local stability and weak agreement between methods, and SHAP-based driver attribution identified the true driver of a realised event in 55% of cases against a 42% majority baseline, with financial and regional causes mostly attributed to operational symptoms. A procurement policy simulation with common random numbers shows that an expected-value flagging rule on the machine-learning score cut total cost by 32% against reactive management and by 4% against the scorecard (not significant), that explanation-guided driver-specific actions matched generic buffers and dual sourcing in cost while randomly targeted actions cost 21% more, and that even perfect risk probabilities added only 12%, whereas perfect timing would have halved cost. A five-year value model and sensitivities show a business case that depends on disruption cost and on what the organisation already does. All data are synthetic; the contribution is a reproducible testbed and an evaluation protocol that tests explanations against causes and decisions.