This study introduces an enhanced ensemble framework for predicting student performance early, using only pre-midterm behavioral indicators like study habits and attendance instead of exam scores. Although the baseline models such as Random Forest initially showed high accuracy, they were less validated for early predictions. To overcome this, a pipelined approach was developed that combines K-Means clustering with an ensemble of Random Forest, Gradient Boosting, and XGBoost models. By aggregating ordinal probability boosting to capture class label more accurately, the final model achieved 84.60% accuracy and a 0.87 recall rate in labeling at-risk students. These results showcase that behavioral data alone can support effective early-warning systems, offering educators an adaptable and intuitive methodology for early-stage performance improvement intervention.