AUROC기반의 부도예측 앙상블 모형

This paper proposes an AUCBoost model to address the class imbalance problem inherent in bankruptcy prediction. AdaBoost algorithm has the advantage of providing reinforced learning to misclassified samples. However, since the set of weights for combining the predictions of the base classifier is determined by the concept of prediction accuracy, it has a limitation in resolving class imbalance problem. AUCBoost algorithm is a boosting technique that replaces AdaBoost's predictive accuracy with AUROC. The goal of AUCBoost is to maximize the AUROC of the final classifier through iterative search for a set of weights to combine the base classifiers using quadratic programming. For empirical experiments, a sample consisting of 10,000 non-bankruptcy instances and 500 bankruptcy instances is collected and divided into 5 subsamples with different imbalance ratio values. As explanatory variables, the seven financial ratios with the highest AUROC are selected from the seven financial ratio groups. 30 cross-validations are performed for each subsample. The main findings are as follows. First, the class imbalance problem has a negative effect on the performance. Second, AUCBoost makes a significant contribution to performance improvements of AdaBoost ensembles. Finally, the effect of AUCBoost on the performance improvement is more magnified as the imbalance ratio value increases.