GRADIENT BOOSTING MACHINE LEARNING FOR PREDICTING QUALITY FAILURE COSTS IN NIGERIAN TABLE WATER MANUFACTURING SECTOR

Abstract

This study developed an XGBoost gradient boosting framework to predict quality failure costs from Total
Quality Management practices in Nigeria's table water industry. The research addresses limitations of
conventional regression by capturing nonlinear relationships, threshold effects, and synergistic
interactions. Drawing on survey data from 397 managers across 223 firms and cost records from 50 firms,
the study augmented the dataset through simulation of 5,000 intervention scenarios. The XGBoost model
was implemented with SHAP analysis for interpretability and compared against Random Forest and linear
regression benchmarks using five-fold cross-validation. The findings revealed that XGBoost achieved 86%
directional accuracy and 12.4% mean absolute percentage error, significantly outperforming linear
regression. Feature importance analysis identified complaint resolution speed as the dominant predictor
accounting for 48.2% of predictive power, followed by equipment maintenance adequacy at 31.8% and
customer feedback system presence at 20.0%. A critical threshold effect was identified whereby firms
scoring below 3.0 on complaint resolution incurred predicted quality failure costs of 22% to 25% of sales,
whereas achieving 4.0 or above reduced predicted costs to 10% to 12%, representing annual savings of ₦3
million to ₦5 million for a medium-sized producer. The study recommends that table water producers adopt
this XGBoost framework as an interpretable decision support tool for predicting quality failure costs,
prioritizing improvement actions, and building data-driven business cases for TQM investment.