A1154
Title: Incomplete inference model to impute loss given default within a regulatory capital model: A South African case study
Authors: Helgard Raubenheimer - North-West University (South Africa) [presenting]
Tanja Verster - North-West University (South Africa)
Gerbrand Breed - North-West University (South Africa)
Ayla Coop - Independent researcher (South Africa)
Abstract: Accurate estimation of Loss Given Default (LGD) in regulatory capital models is complicated by accounts with incomplete LGD outcomes. A methodology is proposed to impute LGD values for incomplete defaults, drawing on principles traditionally used to infer the behaviour of rejected applicants in credit scoring. Within the LGD framework, performing and non-performing exposures are modelled separately, with LGD values calculated either at the point of Default or dynamically throughout the workout process. However, incomplete defaulted accounts lack final outcomes, creating a structural gap in model development. An LGD model is built on the complete case population to predict realised LGD outcomes. This model is then applied to incomplete accounts to generate inferred LGD values, which are subsequently incorporated into the final model development alongside the actual LGDs of complete accounts. This enables a more comprehensive representation of the defaulted population and mitigates potential bias introduced by excluding incomplete observations. The method is illustrated on a real-world dataset from a retail bank in South Africa. The findings offer practical guidance for practitioners and researchers seeking defensible approaches to integrating incomplete defaults into regulatory LGD estimation, while also identifying opportunities for further refinement.
