Explainable Machine Learning for Inventory Shrinkage and Stockout Prediction in U.S. Retail Supply Chains: A Quantitative Study using Gradient Boosting and SHAP

Authors

  • Md Saddam Hosain Business, Data & Research Analyst; Financial, Supply Chain & Operational Analytics; California, USA; Author

DOI:

https://doi.org/10.63125/6hzy9c10

Keywords:

Inventory Shrinkage, Stockout Prediction, Explainable Machine Learning;, XGBoost, SHAP

Abstract

Inventory shrinkage and stockout occurrence undermine U.S. retail supply chains through inaccurate stock records, uncertain deliveries, and ineffective replenishment, while opaque prediction systems complicate managerial adoption. This study investigates the relationships between inventory information quality, operational uncertainty, inventory risks, and acceptance of explainable analytics. Guided by Organizational Information Processing Theory, it employs a quantitative, cross-sectional, embedded multiple-case framework covering six U.S. retail organizations, 60 stores, 24,000 store-stock-keeping-unit records, and 270 employee questionnaires from 300 invitations (90.0% response). Inventory record accuracy, inventory control effectiveness, supplier lead-time variability, demand forecast error, replenishment performance, and explanation usefulness constitute principal predictors; shrinkage, stockouts, and AI acceptance form the outcomes. The analytical plan combines five-point Likert measures, descriptive statistics, Cronbach’s alpha, Pearson correlations, store-clustered logistic regression, held-out XGBoost classification, logistic regression benchmarks, ROC-AUC, precision-recall metrics, and SHAP explanations. The reported analytical results show shrinkage and stockout incidence of 8.08% and 16.35%, respectively. Higher record accuracy (OR = 0.440) and control effectiveness (OR = 0.479) are associated with lower shrinkage odds, whereas greater supplier variability (OR = 2.140) and forecast error (OR = 2.052) correspond to higher stockout odds. XGBoost achieves ROC-AUC values of 0.877 for shrinkage and 0.850 for stockouts, compared with 0.874 and 0.847 for logistic regression, without statistically significant superiority. SHAP highlights record accuracy and audit age for shrinkage, and replenishment performance and supplier variability for stockouts. Explanation usefulness correlates with AI acceptance (r = 0.596), with adjusted R² = 0.358. Questionnaire reliability ranges from 0.765 to 0.827, and perceived SHAP usefulness predicts acceptance (B = 0.643, p < .001), emphasizing human interpretability alongside predictive discrimination for practical deployment. These patterns support prioritizing reliable inventory records, supplier monitoring, calibrated model comparisons, and interpretable alerts. The framework offers a basis for testing whether explainable prediction improves inventory decisions and operational performance across retail contexts.

Author Biography

  • Md Saddam Hosain, Business, Data & Research Analyst; Financial, Supply Chain & Operational Analytics; California, USA;

    MBA in Business Management, International American University, Los Angeles, California, USA;

    Bachelor’s Degree in Accounting & Finance, SEGi University, Malaysia;

Downloads

Published

2026-10-03

How to Cite

Md Saddam Hosain. (2026). Explainable Machine Learning for Inventory Shrinkage and Stockout Prediction in U.S. Retail Supply Chains: A Quantitative Study using Gradient Boosting and SHAP. American Journal of Advanced Technology and Engineering Solutions, 16(10), 01-29. https://doi.org/10.63125/6hzy9c10

Cited By: