Unplanned CNC machine downtime significantly impacts manufacturing productivity. This research develops machine learning models for downtime prediction using systematic feature engineering that prioritizes variable independence. A dataset of 5,000 records with 21 variables from a reputed company in Bangladesh was analyzed. Statistical correlation analysis (p < 0.05) identified eight significant variables from nineteen candidates, reducing dimensionality by 58%. Multicollinearity assessment revealed zero high-risk pairs with 92% of variable combinations showing weak correlations. Seven machine learning algorithms were trained and evaluated: Gradient Boosted Trees, Random Forest, Logistic Regression, Naïve Bayes, Decision Tree, k-Nearest Neighbors, and Support Vector Machine. Models were compared using accuracy, recall, precision, and F1-Score metrics. Gradient Boosted Trees achieved superior performance with F1-Score of 67.25%, recall of 65.82%, and precision of 68.74%, demonstrating optimal balance for downtime detection. Random Forest achieved highest accuracy (96.84%) but lower recall (52.73%). Analysis revealed the accuracy paradox: all models achieved 94-97% accuracy despite substantial F1-Score differences (49.18%-67.25%), indicating accuracy alone is misleading for imbalanced classification. Tree-based ensemble methods substantially outperformed other algorithms. Gradient Boosted Trees is recommended for industrial deployment, enabling improved maintenance scheduling and reduced unplanned downtime in Bangladesh's manufacturing sector.