Combining Sampling Methods, Cost-Sensitive Learning, and Ensemble Techniques for Highly Class-Imbalanced Financial Distress Prediction

Abstract

In highly class-imbalanced financial distress prediction, standard classifiers often fail as they assume balanced class distributions and equal misclassification costs. We propose a heterogeneous voting ensemble method that addresses extreme class imbalance by synergizing sampling, cost-sensitive learning, and ensemble techniques. First, coordinating sampling and cost-sensitive learning reduces model bias in identifying distressed firms. Sampling rebalances class distributions, which reduces representation bias and provides an effective data foundation for cost-sensitive learning. Cost-sensitive learning assigns higher costs to distressed firms, adjusting the optimization objective to minimize total misclassification costs. Second, by aggregating diverse base classifiers, the heterogeneous voting ensemble reduces potential additional variance introduced by sampling and cost-sensitive learning, thereby improving generalization. Using data from China, Poland, and the United States, we show that the proposed method significantly outperforms benchmark models in detecting distressed firms. Moreover, the improved predictions translate into reduced misclassification costs, particularly when the cost difference between distressed and healthy firms is greater. Ablation experiments validate the effectiveness of each component, where sampling contributes most to detection accuracy for distressed firms, followed by cost-sensitive learning, while the ensemble ensures model stability. These results highlight that combining sampling, cost-sensitive learning, and ensemble techniques is crucial for stakeholders (e.g., investors, creditors, and regulators) to detect rare events such as financial distress.