Research on listed companies’ credit ratings, considering classification performance and interpretability

Any credit evaluation system must be able not only to identify defaults, but also to be interpretable and provide reasons for defaults. Therefore, this study uses the correlation coefficient and F-test to select the initial features of a credit evaluation system, and then a validity index for a second selection to ensure that the feature system has the optimum ability to discriminate in determining default status. We omit one feature in each iteration by replacing each feature, calculating the changes in validity index values after deleting this feature and, finally, calculating the ratio of the change value to the sum of all change values. This ratio is then used as the feature’s weight. This study also introduces a data gravity model in predicting defaults, as predicting a validation set’s default status derives the classification threshold to maximize classification accuracy. An empirical analysis of the listed company samples reveals that the feature system selected from 610 features can distinguish between both defaults and nondefaults. Compared with eight other models, our data gravity model not only exhibits good classification performance, but also has interpretability; further, this model can provide at least five-year-ahead forecasting, and can offer a timely risk warning for banks.