Research on the impact of U.S. monetary policy uncertainty on international oil price forecasting: A deep learning approach based on frequency domain decomposition

U.S. monetary policy uncertainty (MPU) has markedly increased the volatility of international oil prices, posing new challenges for global economic stability and policy formulation. Accurately forecasting international oil prices and clarifying the influence of MPU on such forecasts have thus become critical issues. The powerful nonlinear modeling capacity of deep learning offers novel solutions to these challenges. In this context, this study draws on the theoretical foundation of the New Keynesian Dynamic Stochastic General Equilibrium (DSGE) framework to systematically analyze the economic mechanisms through which MPU transmits to oil prices, including interest rates, exchange rates, financing costs, and market expectations. Building on this foundation, we propose an innovative deep learning architecture—FPN-Net (Fourier Patch and Normalization Network). This model leverages Fourier transforms to decompose input variables in the frequency domain into high- and low-frequency components: the high-frequency segment, capturing short-term shocks such as fluctuations in interest and exchange rates, is modeled by an enhanced transformer with blockwise segmentation; the low-frequency segment, which reflects longer-term factors like financing costs and market expectations, is processed by a standardized shared linear unit. The fusion of both components yields the final oil price prediction. Empirical results demonstrate that FPN-Net consistently outperforms advanced benchmarks including LSTM, Informer, Crossformer, and DLinear in both predictive accuracy and stability across different forecasting horizons and during major event windows. Further analysis reveals that incorporating MPU as a key predictor enhances forecast accuracy in most models; however, the degree of improvement varies by model: high-performing models see only marginal gains, whereas models with weaker baseline performance benefit more substantially.