Abstract

Against the backdrop of rising geopolitical uncertainty, this paper uses daily data from 2013 to 2025 on the geopolitical risk index and China’s stock, bond, money, foreign exchange, commodity, gold, and real estate markets to construct a “quantile time-frequency connectedness–machine learning early warning” framework. It examines the state-dependent characteristics of risk spillovers between changes in geopolitical risk and China’s financial markets, as well as the ability to identify high-risk states. The results show that, first, financial market risk connectedness exhibits a pronounced tail amplification effect: the total connectedness index is 13.72% under normal conditions, but rises to 77.97% and 78.47% under extreme downside and extreme upside states, respectively. Second, risk connectedness is mainly concentrated in the short term, although long-term connectedness strengthens under extreme states. Third, the stock, real estate, and commodity markets generally act as net transmitters of risk, while the bond and foreign exchange markets generally act as net receivers. Fourth, machine learning models based on dynamic connectedness indicators can effectively identify future high-risk connectedness states. However, naïve benchmarks and ablation tests indicate that the total connectedness index and its distance from the rolling threshold are the main sources of information, while the contribution of machine learning models lies mainly in probability calibration, multi-horizon risk ranking, and the integration of auxiliary variables. The findings provide empirical evidence for cross-market risk monitoring and the construction of early warning indicators under geopolitical risk.

IPC Classification

Keywords