Feature Engineering Techniques to Enhance Credit Scoring Models

This review article explores and evaluates the use of feature engineering techniques for clients' credit risk assessment through the development of methodologies involving machine learning algorithms. We use publicly available credit scoring datasets in both the development of our methods and for benchmarking. After describing the development of our methods, we present a discussion about the comparison of performances of our techniques and a number of other credit scoring methods described in the literature, including traditional statistical methods and advanced soft computing-based methods. One of the existing methods analyzed involves the use of evolutionary computing algorithms for feature selection and transformation. The development of an approach for credit scoring that integrates fuzzy neural networks with granular computing to handle data uncertainty and complexity. The main conclusion of this review is that feature engineering methods generally produced better results than traditional models, improving predictability and reducing variability of the outcomes. The performance of a combination of these techniques was highly satisfactory; the vagueness present in the information was effectively managed by our method, yielding improved results. From a practical perspective, although our methods have not proved to be substantially better than some market-available options, we demonstrated that it is possible for companies to develop credit scoring mechanisms internally using past data and advanced feature engineering techniques. Under certain circumstances, companies may find this option preferable to paying high fees to large credit scoring agencies for their proprietary systems.