An Explainable, Calibrated, and Operationally Interpretable Framework for Early Identification of Production-Line Backorder Risk in Garment Manufacturing

Backorder risk on garment production lines undermines on-time delivery and buyer confidence, yet few studies combine rigorous feature validation, a transparent probabilistic alternative to opaque classifiers, and deployment-oriented evaluation. This study develops and evaluates such a framework using 1,000 production-line observations from Bangladeshi ready-made-garment (RMG) factories, with 27.40% backorder prevalence. A three-stage preprocessing pipeline and multi-criteria statistical validation, integrating hypothesis testing, Lasso regularization, and composite Justification Scoring, reduced twelve candidate predictors to a stable, low-collinearity ten-feature subset (maximum VIF = 1.66). Logistic Regression, Random Forest, XGBoost, and a Discrete Bayesian Network were benchmarked using repeated stratified 10-fold cross-validation (50 fits for machine-learning models and 10 for the Bayesian Network), with Precision-Recall AUC (PR-AUC) as the imbalance-appropriate metric. Contrary to a single-split evaluation favoring XGBoost, cross-validation identified Logistic Regression as the best discriminator (mean PR-AUC = 0.661, versus 0.652 for Random Forest and 0.622 for XGBoost), a ranking that persisted after matching XGBoost’s imbalance handling to that of the other models. Although the Bayesian Network had the lowest discrimination (AUC-ROC = 0.750), it achieved the best calibration (Expected Calibration Error = 0.048). Sensitivity analyses confirmed Logistic Regression’s robustness to decision threshold, regularization strength, and moderate input noise, while ablation analyses showed that five to six features, including Daily Production Quantity, Overtime Hours, Machine Failures, Planned Production Target, and Number of Operators, captured all achievable predictive value. SHAP attributions and the Bayesian Network structure converged on these drivers, providing four-way cross-method validation of a parsimonious, calibrated, and operationally interpretable backorder early-warning framework.