Cost-aware Credit-scoring Framework Based on Resampling and Feature Selection

Credit loans are fundamental to the financial industry, and effectively managing their risks is essential. Financial companies may face two challenges when performing credit scoring to control such risks. First, datasets are often imbalanced with far more non-default cases than default ones, where oversampling methods are usually applied. Few methods, however, have considered further enhancing the quality of a training dataset by addressing the critical samples that may confuse the final classifiers while maintaining the interpretability of the final model. Second, common model evaluation indicators may not accurately reflect the financial loss associated with incorrect predictions or the costs involved in collecting features. To address these challenges, we propose Cost AwarE CRediT ScorIng Framework Based on ResamplIng and FeaturESelection (CERTIFIES). In this framework, we develop a pre-learning resampling approach that employs multiple machine learning methods as assistant classifiers to detect critical data samples after oversampling. This approach further enhances the overall performance of the chief classifier, logistic regression, without compromising its interpretability. Additionally, during the model evaluation step, we design a cost-aware evaluation indicator that accounts for the actual loss due to incorrect predictions and the cost of collecting various features. This provides an approach to perform feature selection based on financial costs. To demonstrate the effectiveness of the proposed method, we apply it to our credit scoring dataset collected by local financial companies, as well as to two public datasets.