Abstract

Purpose: This study aims to predict monthly prepayment rates of Mortgage-Backed Securities (MBS) issued by the Korea Housing Finance Corporation, focusing on the effects of external shocks such as the financial crisis and COVID-19. Research design: The research compares traditional fixed-effects regression models with machine learning techniques (ElasticNet, LASSO, Ridge) to determine which model best predicts MBS prepayment rates before and after external shocks. Data and methodology: The study uses monthly data from June 2004 to December 2020, analyzing MBS prepayment rates alongside various macroeconomic variables. The performance of each model is assessed using cross-validation and blocked cross-validation methods to evaluate stability under different economic conditions. Results: Machine learning models, particularly ElasticNet, consistently outperform traditional regression models. ElasticNet showed the highest predictive accuracy, with a stable performance even after the financial crisis and COVID-19, unlike traditional models that struggled to adapt to the shocks. Conclusions: The study concludes that machine learning models, especially ElasticNet, offer superior predictive performance in forecasting MBS prepayment rates, especially in volatile market conditions, and should be considered over traditional models for financial predictions.