In credit risk modelling, Loss Given Default (LGD) estimation is characterised by a persistent trade-off between interpretability and predictive accuracy. Industry practice typically relies on parsimonious grid-based approaches, favouring simplicity over performance. This thesis proposes a new framework to overcome this limitation: a non-parsimonious machine learning estimation followed by hierarchical clustering to restore parsimony. Applied to a representative dataset, the methodology shows that an XGBoost-based model achieves an Accuracy Ratio (AR) of 25.97%, compared to 22.30% for a parsimonious baseline, while delivering a similar number of final grades. From a validation perspective, we demonstrate that exclusive reliance on Somers’ d(C|R) is misleading, as it tends to reward overly simplistic estimates.