A MIDAS multinomial logit model with applications for bond ratings

Abstract

In multinomial classification analysis, the issue of mixed-frequency data is becoming increasingly common. To solve the multiclassification problem in a mixed-frequency data environment with a large frequency ratio, we combine the restricted mixed data sampling (MIDAS) method with a multinomial logit (MLogit) model to construct a MIDAS-MLogit model. The MIDAS-MLogit model imposes functional constraints on high-frequency variable parameters and fully uses high-frequency information to complete multiclassification tasks in a mixed-frequency data environment. Through extensive Monte Carlo simulations, we prove that this novel MIDAS-MLogit model has better classification performance than conventional MLogit and U-MIDAS-MLogit models for solving multiclassification problems with a large frequency ratio. In addition, we use the MIDAS-MLogit model to predict the ratings of corporate bonds issued by Chinese-listed companies from 2008 to 2021 and find that its prediction accuracy outperforms several competitive models. We further verify the validity of the MIDAS-MLogit model for different periods, including the global financial crisis, the post-financial crisis, and the pandemic. The results show that all three models perform well during the global financial crisis, but the MIDAS-MLogit model predominates the post-financial crisis and pandemic periods. The empirical results also show that covariates from bond, corporate, and financial environment levels have varying effects on corporate bond ratings.