Class-imbalanced dynamic financial distress prediction based on Adaboost-SVM ensemble combined with SMOTE and time weighting

Abstract This paper focuses on how to effectively construct dynamic financial distress prediction models based on class-imbalanced data streams. Two class-imbalanced dynamic financial distress prediction approaches are proposed based on the synthetic minority oversampling technique (SMOTE) combined with the Adaboost support vector machine ensemble integrated with time weighting (ADASVM-TW). One is the simple integration model of SMOTE with ADASVM-TW, which uses SMOTE to make each data batch class-balanced before ADASVM-TW is applied for dynamic financial distress prediction modeling. The other is the embedding integration model of SMOTE with ADASVM-TW, which embeds SMOTE into the iteration of ADASVM-TW and creatively designs a new sample weighting mechanism. Namely, in each round of iteration, it does not only resample the majority financial normal samples by time weighting, but also generates more synthetic minority samples around new and difficult minority samples and less synthetic minority samples around old and easy minority samples to make the training dataset class-balanced. The empirical experiments are carried out based on the financial data of totally 2628 Chinese listed companies. Since there is certain degree of randomness in the proposed models, 50 times of experiments were performed under the same computational environment, so that the experimental results of model performance can be statistically compared and tested. The results indicate that both the simple integration model and the embedding integration model can greatly improve the recognition ability for the minority financial distress samples, and the embedding integration model is even more preferred because it also significantly outperforms the simple integration model.