딥러닝 모델을 활용한 국내 채권 신용스프레드 예측 연구

This study examines the practical applicability of time-series deep learning models for predicting the Korean corporate bond credit spread. By employing Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolution Network (TCN), Attention and Transformer Encoder, our study attempts to overcome the limitations of traditional approaches by capturing complex variable interactions and non-linearities while processing sequential data. For better model interpretability, we utilize the permutation importance technique to identify variables that impact predictions. Our results imply that the GRU model generally outperforms the others in terms of area under the curve (AUC) and accuracy. However, leveraging the predicted probabilities by filtering out intermediate outputs enables even higher predictive accuracy. Model performances differ by bond rating and prediction horizons. AA- bonds illustrate robust predictability during all test horizons, whereas BBB- bonds exhibit a decline in predictive accuracy as the test horizon extends. Although the stock market variables play a significant role in predicting spread directions of BBB- bonds, other variables are more useful for AA- bonds. These results suggest that time-series deep learning models provide a practical framework for credit spread forecasting and interpreting its key drivers, supporting data-driven decision-making in financial markets.