Abstract: To achieve accurate measurement of VaR and risk spillover effect ΔCoVaR, considering the typical characteristics of carbon financial markets such as volatility aggregation, thick tail and asymmetry, QRNN-GARCH-CoVaR model is constructed based on neural network quantile regression (QRNN) model and GARCH model to fit the advantages of volatility aggregation. Taking carbon trading returns from Beijing, Guangdong, Hubei and London as the research objects, the empirical results show that, first, QRNN-GARCH-CoVaR model is not only better than the traditional model in measuring VaR, but also captures the financial risk spillover effect. Second, the risk transmission direction and sensitivity of domestic markets are different. Hubei market has high stability and strong risk absorption ability. Beijing and Guangdong markets fluctuate greatly, and Beijing market is vulnerable to foreign markets.