Abstract
Macroeconomic forecasting plays a pivotal role in policy formulation and global risk management. However, dealing with nonlinear high-dimensional data in macroeconomic forecasting has long been proven to be an NP-hard problem that traditional models find extremely challenging to handle. These models often struggle to break free from local optima, thereby failing to provide accurate and reliable forecasts, which is a significant obstacle that urgently needs to be overcome. This study puts forward a hybrid model (EBA-BPNN) that combines an Enhanced Bat Algorithm (EBA) with dynamic inertia weight and adaptive frequency mechanisms to optimize Backpropagation Neural Networks (BPNN). The framework is structured in three stages. Firstly, Dynamic Time Warping (DTW) is employed to align cross-country economic cycles, and then Granger causality analysis along with mutual information is utilized to select 32 core variables. Secondly, during the EBA optimization process, the dynamic inertia weight mechanism is used to strike a balance between exploration and exploitation. Meanwhile, a Cauchy-Gaussian perturbation is introduced to enhance the population diversity, and a variance-driven pulse emission rate is adopted to regulate the local search intensity. Finally, a gradient-assisted search strategy is applied to accelerate convergence and prevent the model from falling into local optima. Experiments conducted on datasets from the World Bank and the OECD demonstrate that the EBA-BPNN model achieves remarkable results. Specifically, it reduces the Mean Absolute Error (MAE) in quarterly GDP forecasting by 29.3%, 17.8%, and 15.6% when compared to BPNN (with an MAE of 2.15), PSO-BPNN (with an MAE of 1.85), and BA-BPNN (with an MAE of 1.92), respectively. Even under extreme scenarios, the increase in MAE is limited to only 12.3%. Overall, this model offers a high-precision approach for dynamic economic modeling, providing valuable support for the formulation of forward-looking fiscal policies and cross-border investment decisions.