In Bangladesh's vital ready-made garment (RMG) sector, unplanned machine downtime is a major cause of operational inefficiencies and financial loss. A solution is provided by Predictive Maintenance (PdM), a crucial element of Industry 4.0, which uses sensor data to predict machine faults before they happen. Using the benchmark AI4I 2020 Predictive Maintenance dataset, this paper conducts a rigorous comparative analysis of three baseline models—Logistic Regression, Random Forest, and XGBoost—for machine failure prediction. To ensure a robust and reliable evaluation for this high-imbalance problem, models were assessed using 5-fold stratified cross-validation, focusing on their average F1-Score, Precision, and Recall. The findings unequivocally demonstrate that the XGBoost classifier (Avg. F1-Score: 0.731) provides the most balanced and reliable performance, significantly outperforming Random Forest (Avg. F1-Score: 0.571) and Logistic Regression (Avg. F1-Score: 0.235). Furthermore, a Business Impact Analysis (BIA) translates this performance into a practical metric, estimating the XGBoost model could reduce downtime-related costs by 70.7%. The model's primary predictors are logical physical elements, with torque and tool wear being the most important, according to an explainability (SHAP) study. Rather than proposing a new algorithm, this paper builds a methodological bridge from benchmark datasets and industrial application. We provide a practical, end-to-end framework that: (1) identifies the most reliable baseline model (XGBoost) using robust 5-fold cross-validation, (2) translates its performance into a clear financial case using a Business Impact Analysis (BIA), and (3) validates its logic using SHAP explainability. This provides a direct, data-driven recommendation for RMG factories.