A PSO-Optimized Stacking Ensemble with Hybrid SMOTE-NC and Tomek Links for Bid-Based Winning Prediction in Procurement Projects

Abstract

This research aims to establish a classification model for the prediction of procurement winning outcomes based on bid value and owner cost estimation data. The main challenge in procurement analysis lies in severe class imbalance and complex non-linear relationships among pricing and procurement attributes. The research object utilizes procurement tender data from PT Pos Indonesia, including project owner, owner cost estimation, bid value, and procurement method. The proposed approach integrates hybrid SMOTE and Tomek links for class balancing, regulation-driven feature engineering, and a stacking ensemble model optimized using particle swarm optimization. The stacking framework combines Random Forest, Extra Trees, and Gradient Boosting as base learners. The experimental evaluation demonstrates that the proposed approach delivers the strongest performance, achieving an AUC of 0.92, an accuracy of 0.89, and an F1-Macro of 0.81, thereby surpassing all individual classifiers and homogeneous ensemble methods considered in this study. This study concludes that the hybrid optimization-based ensemble approach is effective for improving procurement winning prediction accuracy and provides a reliable decision-support tool for data-driven and regulation-compliant procurement processes.