The Impact of Data Aggregation and Risk Attributes on Stress Testing Models of Mortgage Default

Abstract

Stress testing models have been developed at various levels of data aggregation with or without risk attributes, but there is limited research on the joint impact of these modeling choices. In this paper, we investigate how data aggregation and risk attributes affect the development and performance of stress testing models by studying residential mortgage loan defaults. We develop mortgage default models at various data aggregation levels including loan-level, segment-level, and top-down. We also compare the models with and without risk attributes as control variables. We assess model performance for goodness-of-fit, prediction accuracy, and projection sensitivity for stress testing purposes. We find that the loan-level models do not always win among models with various data aggregation levels, and including risk attributes greatly improves goodness-of-fit and projection accuracy for models of all data aggregation levels. The findings suggest that it is important to consider data aggregation and risk attributes in developing stress testing models.