Profit-based uncertainty estimation with application to credit scoring

Credit scoring is pivotal in financial risk management and has attracted significant research interest. While existing studies primarily concentrate on enhancing model predictive power and economic value, they often overlook the crucial aspect of predictive uncertainty, especially in the context of deep neural networks applied to credit scoring. This study addresses uncertainty estimation in credit scoring and evaluates three widely used uncertainty methods across various credit datasets. Additionally, guided by the maximum profit criterion, we propose two profit-based uncertainty metrics to assess profit uncertainties stemming from predictive uncertainty, specifically targeting class-dependent and instance-dependent cost scenarios. Subsequently, we develop a classification system with a rejection mechanism based on these metrics. Our approach aims to improve model profitability and reduce predictive uncertainty, specifically regarding model profit. Empirical results across several benchmark credit datasets indicate that our proposed framework outperforms existing methods in terms of increasing model profit in different credit-scoring scenarios. Furthermore, sensitivity analyses of varying cost parameter settings highlight the robustness of our framework.