Modeling an early warning system for household debt risk in Korea: A simple deep learning approach

Abstract

We construct an early warning indicator for household debt risk by analyzing the relationship between household debt and certain important macroeconomic determinants using a simple deep learning approach. A precise and informative indicator can help inform economic policies, especially in light of the recent growth in the ratio of household debt to income. Although several studies have analyzed the determinants of the household debt crisis, very few have examined early warning indicators for household debt risk. Some studies suggest that a situation can be regarded as a crisis if the household debt ratio is greater than 50% or 85%. However, as the household debt ratio in Korea is already over this threshold, this criterion is neither informative nor useful. Accordingly, we propose a transformed index that addresses long-term memory characteristics. Moreover, five categories for the degree of household debt crisis are considered instead of the binary variable that has been frequently used in previous studies. Furthermore, we use a well-known deep learning approach to find a non-linear relationship between crisis indices and many factors. The empirical results demonstrate that the proposed early warning indicator explains the household debt crisis quite well.