Essays on LGD models for residential mortgage loan : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Finance at Massey University, Albany, New Zealand

Abstract

This thesis introduces innovative models for predicting loss (recovery) rates of defaulted and uncured residential mortgages through three research topics: (1) decomposing loss (recovery) rates into three stages: prior to collateral disposition (Stage 1), collateral disposition (Stage 2), and post-collateral disposition (Stage 3); (2) further breaking down loss (recovery) rates by resolution types; and (3) examining loss (recovery) rates across periods that include the 2008 global financial crisis (GFC). All analyses utilize account-level US single-family prime mortgage data from Freddie Mac spanning 1999-2019, extended to 2021 for the third topic.

For the first research topic, we demonstrate that splitting uncured recovery rates into three meaningful stages and improving accuracy for each stage enhances overall accuracy in specific cases. Each recovery stage exhibits distinct loss (recovery) rate distributions, historical trends, and driving factors with varying magnitudes and directions. This challenges the traditional approach of predicting loss (recovery) as a single proportion of exposure at default (EAD), which overlooks valuable information. This contribution enables recovery rates to be modelled in three distinct, banker-relevant modules that can be independently improved. Several models for each stage outperformed nominated benchmarks in RMSE and r-square metrics when applied to the Freddie Mac dataset.

The second research topic extends the first by developing separate three-stage models for each resolution type, as evidence suggests different outcomes yield varying recovery levels and distributions across stages. These models are combined with estimated probabilities of occurrence for each resolution type.

The third research topic investigates new external predictors relevant to periods of economic stress, incorporating them into models from the second topic to enhance prediction accuracy.

Collectively, these research topics contribute to more accurate mortgage recovery rate predictions, thereby improving collection strategy effectiveness and the accuracy of capital and accounting provisions. This thesis advances the literature by introducing new perspectives and uncovering valuable information in loss given default (LGD) modelling, ultimately enhancing credit risk quantification accuracy.