Abstract

With the very unbalanced data found in financial risk prediction, this study hopes to aid in anticipating the financial risk that corporations may encounter. We can improve performance by employing oversampling and under-sampling algorithms. We were able to better understand how the performance of each classifier changed in each dataset by using a variety of classifiers across three distinctively sampled datasets. In addition, we analyzed our dataset using three different evaluation metrics: accuracy, sensitivity, and specificity, rather than being limited to just one. The results indicate that the accuracy on three separate datasets with various sampling methods differs greatly. The sensitivity and specificity of the under-sampled dataset differ from those of the original dataset and the oversampled dataset, which are fairly comparable to one another. It was discovered that gradient boosting trees produce better outcomes than other algorithms. When using oversampled data and measuring accuracy, logistic regression was found to be the most effective. However, when using under-sampled data, LightGBM Classifier had the best performance. Additionally, when considering sensitivity and specificity, CatBoost Classifier was the best choice.

Keywords: Bankruptcy, unbalanced data, classification