Do Macroeconomic and Geopolitical Risk Factors Improve FX Volatility Forecasts? A Mixed-Frequency Transformer Approach

Forecasting exchange rate volatility is essential for risk management, portfolio allocation, and policy decisions, especially in emerging markets. This study investigates the USD/CNH market and develops a Mixed-Frequency Transformer that integrates three layers of information: technical signals from high-frequency foreign exchange trading data, macroeconomic fundamentals, and geopolitical risk indices. The dataset spans August 2015 to March 2025 with Parkinson volatility as the dependent variable. The MF-Transformer architecture aligns monthly and daily predictors, fuses them with technical features through cross-attention, and models temporal dependencies using stacked Transformer layers. Forecasting performance is benchmarked against a technical-only model using statistical accuracy, regime-specific analysis, and event-sensitive analysis. Results show that including macroeconomic and GPR variables substantially improves Directional Accuracy and Spike-Hit rate, particularly during high-volatility stress regimes and event-driven periods. The findings demonstrate how mixed-frequency deep learning can capture external shocks, offering new insights for global investors, multinational firms, and policymakers seeking to better manage exchange rate risks in turbulent environments.