Default forecasting based on a novel group feature selection method for imbalanced data

The presence of several redundant features poses a significant challenge to default forecasting, but this can be effectively resolved through feature selection. However, the number of defaulting firms is much smaller than that of nondefaulting firms. If the selected features do not adequately reflect the information on defaulting firms, it is likely that the prediction model will incorrectly identify defaulting firms, misleading banks and financial institutions into granting loans to these firms, resulting in huge capital losses. This study proposes a group feature selection method combining optimal instance selection with weighted comprehensive precision (OIS-WCP) to improve the performance of prediction models, especially for defaulting firms. The empirical results for the three data sets (Chinese listed firms, and German and Chilean credit data) show that the OIS method effectively removes noisy majority samples and significantly improves the classification ability of the prediction model compared with the two popular instance selection methods. The type II error of the prediction model for imbalanced data from the OIS-WCP method is significantly lower than that of the six existing feature selection methods. This study contributes to the field of default forecasting on imbalanced data and demonstrates that integrating instance selection into the feature selection method is worth considering for improving the performance of default prediction models.