External Validation of Ensemble Learning Models on Unseen Data for Production Output Prediction: A Comparative Study Against Garment Industry Theoretical Targets

The garment industry requires accurate production output predictions to minimize operational inefficiencies and support adaptive production planning. However, static, manual production targets often fail to reflect the real-time dynamics of the production floor, typically resulting in overestimations. This study evaluates the accuracy, stability, and generalization capability of ensemble learning models via external validation on unseen operational data, while comparing their performance against traditional theoretical targets. Employing a quantitative experimental approach, the study utilizes 700 historical data points for model training and 60 recent observations as an external test dataset, incorporating manpower and the Standard Minute Value (SMV) gap as key variables. Model performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination ( ). The results demonstrate that both models exhibit robust performance, achieving  values above 0.90 on unseen data. Comparatively, Random Forest achieved an MAE of 6.008, RMSE of 8.56, MAPE of 23.29%, and an  of 0.908, whereas Gradient Boosting yielded an MAE of 5.892, RMSE of 8.05, MAPE of 25.60%, and an  of 0.919. Although Gradient Boosting outperformed Random Forest in absolute error metrics and , Random Forest demonstrated superior relative prediction stability, winning 51.67% of daily prediction comparisons in a "Battle of Models" framework. Visual analysis using scatter plots and box plots confirmed that Random Forest maintains a more consistent error distribution and captures actual output fluctuations more realistically than the company's manual targets. These findings indicate that integrating the SMV gap variable effectively captures hidden losses within the production process. Consequently, Random Forest is recommended as the foundation for developing a data-driven Decision Support System (DSS) to facilitate more adaptive, realistic, and efficient production target setting in the garment industry.