Highly class-imbalanced dynamic financial distress prediction based on Bagging-XGBoost

Zhao, Mengru, Sun, Jie, Li, Nengfei, Yang, Xingyi
Risk Management · Springer Nature · 2026-08-28 · DOI: 10.1057/s41283-026-00242-7

Abstract

Financial distress prediction (FDP) can help enterprises prevent financial distress in advance and assist investors in reducing investment risks, and it is of great significance for improving the performance of FDP models. The paper constructs a new highly class-imbalanced dynamic Bagging-XGBoost FDP model which combines the two ensemble learning algorithms of Bagging and Boosting, and tests the model based on a dataset from Chinese listed companies during the period 2010–2020. The results indicate that the dynamic Bagging-XGBoost FDP model developed on the balanced training set constructed by random under-sampling achieves a G -value of 93.88% and a high true positive rate ( TPrate ) of 97.35% for distressed enterprises, suggesting that the combination of Bagging and Boosting can enhance the performance of FDP models. Meanwhile, in handling class imbalance, this model outperforms the models developed on the slightly class-imbalanced training sets constructed by random under-sampling (optimal TPrate = 95.94%) and on the balanced training sets constructed by the combination of SMOTE and random under-sampling (optimal TPrate = 95.97%). The dynamic models based on rolling time window mechanisms are superior to the static models ( G -value = 93.54%), and the full memory time window mechanism ( G -value = 93.88%) slightly outperforms the fixed-width time window mechanism ( G -value = 93.80%), indicating that the models must learn from new samples over time and the early sample data can still provide valuable information for FDP. In addition, the Bagging-XGBoost model outperforms other models, such as decision tree (DT), XGBoost, Adaboost-DT and Bagging-DT, in highly class-imbalanced dynamic FDP scenarios, and this conclusion remains valid through a series of robustness tests.

Read at publisher →

Similar articles