Two-Step Classification for Solving Data Imbalance and Anomalies in an Altman Z-Score-based Bankruptcy Prediction Model

Differences in bankruptcy regulations with varying value parameters cause data anomalies when implemented in the Altman Z-Score model. Another common problem in bankruptcy predictions is imbalanced data; the number of companies that fall into the bankruptcy category is much smaller than those that do not. Therefore, a novel method was proposed to address data imbalance and anomalies in an Altman Z-Score-based bankruptcy prediction model. The proposed method employs a two-step classification controlled with data binning. Assumption values were used to set the proportion of distress and non-distress classes. Quartile calculation-based data binning is then used to ordinally rank the non-distress category into three classes. Furthermore, a two-step classification was performed using the Long-Short Term Memory (LSTM) method, followed by a rule-based classification method. The LSTM method predicts output in the form of one class representing the distress zone and three classes representing non-distress zone subcategories. The results are then processed using a rule-based classification to summarize the output into a two-class classification, where all data not in the distress zone class is part of the non-distress zone. The performance evaluation shows promising results, with outcomes closely matching the source bankruptcy data. These findings strengthen the evidence that the Altman Z-Score is a powerful tool for bankruptcy prediction and demonstrate that the proposed method can improve the Altman Z-Score model in handling differences in data value parameters.