Sequence-based clustering applied to long-term credit risk assessment

Abstract

This paper studies the effectiveness of estimating credit rating transition matrices using sequence-based clustering on historical credit rating sequences. The data set used in this study consisted of monthly credit rating sequences from Korean companies from 1986 to 2018. The credit rating sequences were converted to sequence matrices and was clustered using PCA-guided K-means. Representative transition matrices of the resulting clusters were then generated to be used in the classification process. The proposed clustering model is evaluated under the 3 different long-term classification scenarios; 7 class credit rating prediction, credit rating transition direction (upgrade, stay, or downgrade) prediction, and default behaviour prediction. All three classification scenarios produced promising results suggesting that the representative transition matrix of the K clusters better describes future credit rating behaviour than a single transition matrix.