Loss given default decomposition using mixture distributions of in-default events

Modeling loss in the case of default is a crucial task for financial institutions to support the decision making process in the risk management framework. It has become an inevitable part of modern debt collection strategies to keep promising loans on the banking book and to write off those that are not expected to be recovered at a satisfactory level. Research tends to model Loss Given Default directly or to decompose it based on the dependent variable distribution. Such an approach neglects the patterns which exist beneath the recovery process and are mainly driven by the activities made by collectors in the event of default. To overcome this problem, we propose a decomposition of the LGD model that integrates cures, partial recoveries, and write-offs into one equation, defined based on common collection strategies. Furthermore, various levels of data aggregation are applied to each component to reflect the domain that influences each stage of the default process. To assess the robustness of our approach, we propose a comparison with two benchmark models on two different datasets. We assess the goodness of fit on out-of-sample data and show that the proposed decomposition is more effective than state-of-the-art methods, maintaining a strong level of interpretability.