Beyond Accuracy in Artificial Intelligence Based Credit Scoring Systems : Explainability and Sustainability in Decision Support

The use of artificial intelligence (AI) and machine learning (ML) in decision support has become vital for financial services companies to remain competitive. Appraisal of AI/ML has largely focused on prediction accuracy. We argue the need to look beyond accuracy as the sole determinant for model choice. Explainable artificial intelligence (XAI) and model transparency are vital in financial services to fulfill not only regulatory requirements, but also to foster trust in the overall system. Also, in an era where the urgency of climate action is paramount, the carbon impact of AI is emerging as an important environmental concern. More broadly, digital sustainability is increasingly being seen as a social responsibility and an economic imperative. Two modeling approaches currently receiving attention in comparative studies in the credit risk management space are deep learning (DL) and gradient boosting (GB). We compare those two algorithms in the context of credit scoring but provide new dimensions to the comparative analysis. While we consider accuracy for benchmark purposes, we add to the existing literature by considering (i) explainability and (ii) digital sustainability. We show that the GB model produces explanations consistent with extant theory and empirical evidence, while highlighting inconsistencies in explainability under the DL model. Furthermore, the GB model has the advantage of computational speed, thus lowering its environmental impact for comparable operations. We contend that this makes GB the preferred choice for credit scoring, extending existing evidence on the preference of GB over DL from an accuracy perspective