Deep Learning-Based CoVaR Forecasting

This paper examines the feasibility of deep learning for Conditional Value at Risk forecasting in international stock markets. Using daily data for the stock markets of China, the United States, Japan, Germany, and Brazil from 1 July 2010 to 1 July 2025, the study develops a CNN-Transformer quantile regression model for CoVaR prediction. The empirical analysis is based on log return series, representative forecast plots, and the Diebold-Mariano test against benchmark models. The results show that the predicted CoVaR series are clearly time-varying and become more negative during periods of market stress, indicating that the proposed model captures meaningful dynamics in conditional tail risk. The Diebold-Mariano test further shows that the proposed model outperforms both the CNN-QR benchmark and the Transformer-QR benchmark, while the overall test results remain positive for all market pairs. These findings suggest that combining local feature extraction with long-range dependency modeling helps improve CoVaR forecasting performance. The study provides empirical support for the application of deep learning to CoVaR prediction and contributes to the literature on tail risk forecasting in international stock markets.