Algorithmic auditing of AI credit scoring: integrating IFRS 9 model governance with ethical AI assessment

Abstract

This study aims to propose applying the governance principles associated with the implementation of IFRS 9 to the algorithmic audit of artificial intelligence (AI) models used for credit rating. It examines how the principles of validation, documentation, traceability and forward-looking risk assessment used in the implementation of IFRS 9 can be adapted to the audit of AI systems. The framework is validated on the Kaggle Credit Risk data set using four machine learning algorithms. It combines the model governance principles associated with the implementation of IFRS 9 with criteria for evaluating responsible AI. The assessment framework integrates predictive performance metrics alongside fairness evaluation measures, such as Disparate Impact and Demographic Parity. In addition, model interpretability is examined to identify the influence of sensitive attributes. The results indicate that XGBoost achieves the highest predictive performance. However, all models exhibit varying degrees of dependence on sensitive features, particularly age and property ownership status, raising concerns regarding algorithmic bias in credit decision-making. The proposed framework can help financial institutions improve the transparency and auditability of AI-based credit scoring systems. It provides auditors and regulators with a structured approach to assessing the performance, fairness and explainability of AI models in accordance with model governance principles. This research establishes a link between the governance principles of models associated with the implementation of IFRS 9 and the ethical auditing of AI, two fields that are rarely associated with one another. It contributes to the development of a multidimensional framework that integrates predictive performance, algorithmic fairness and explainability for the auditing of AI-based credit scoring models.