Answers are grounded in this paper and cite their sources.
We use the mortgage data provided by Freddie Mac in the U.S. The predictive power of transformer models was empirically verified while simultaneously exploring the potential for enhancing explanatory power through the following three research questions. First, is there a statistically significant difference in predictive power between the transformer and the panel logistic regression? Second, is there a difference in predictive performance between the variables at the time of loan origination and the current ones? Finally, does altering the input order of the independent variables affect the predictive performance? Our findings are as follows. First, compared with the traditional panel logistic regression model, the transformer demonstrated overwhelmingly superior predictive performance. Second, the absolute importance of updated current month variables in default prediction was statistically demonstrated. Last, the transformer model exhibited a dual nature: practical applicability in overall predictive performance with respect to the F1 score to changes in the order of input variables and the variability of marginal effects for individual variables. These findings contribute to strengthening risk management capabilities in the housing finance sector and increasing the reliability of machine learning model utilisation.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: