A complete view of borrower risk. Predictions are more accurate, demographic bias is actively engineered out, and every decision is explained line by line to the examiners who have to sign off on it.
The challenge
Stringent regulatory requirements for model explainability (ECB and EBA guidelines). Algorithmic fairness across protected demographic groups. Integration with the bank's central systems at real-time response speeds. On top of that, an organisational shift from traditional, rules-based risk to AI-driven processes.
The approach
A constrained-optimisation framework that jointly maximises prediction accuracy and minimises demographic-parity gaps, so bias is engineered out during training rather than audited after the fact. Gradient-boosted models paired with SHAP explanations meet the explainability mandate decision by decision, while a live dashboard tracks performance, fairness and drift.
What carried the system into production at a regulated bank was auditability, not raw accuracy.
The results
Gini coefficients improved 30% over legacy models, which meant better risk discrimination and lower default losses.
- 30%
- Better risk prediction
- +18%
- Underserved approvals
- EUR 12M
- Annual savings
Approval rates for qualified underserved populations rose 18% while portfolio quality held, and the bank estimates EUR 12M in annual savings from reduced credit losses and improved efficiency.
