The Impact of Machine Learning-Based Credit Scoring on Borrower Risk Assessment in European Banking

Sammanfattning

The EU AI Act classifies credit scoring as a high-risk application but does not define how transparency and oversight should be measured or evaluated. This creates a gap between regulatory obligations and the technical tools available to financial institutions. This thesis develops and empirically tests a compliance-oriented evaluation framework that translates the Act’s requirements into four measurable explainability dimensions: stability, fidelity, interpretability, and human oversight. These dimensions are applied across multiple credit scoring datasets and modelling approaches. The findings show that no single approach satisfies all four dimensions simultaneously. Different models and explanation methods perform well on different dimensions, indicating that explainability is not a single property but a profile of separate strengths and weaknesses. A trade-off between explainability and predictive accuracy appears only for some dimensions and is not universal. The study provides European banks with an evidence-based decision tool for model selection under the EU AI Act. It demonstrates that compliance is a strategic choice about which transparency risks to accept rather than a binary outcome.