The swift advancement of cross-border e-commerce highlights the growing necessity for predictive models capable of effectively managing heterogeneous market environments, particularly the challenges associated with data sparsity in emerging markets. This dissertation proposes an innovative hybrid forecasting model that combines Long Short-Term Memory (LSTM) networks with the Prophet algorithm, enhanced by a multi-resolution temporal feature extraction module and an adaptive, context-aware dynamic weighting fusion mechanism. The designed architecture capitalizes on dense user behavior data from mature markets while concurrently mitigating the intrinsic data insufficiencies of emerging ones. Comprehensive experiments conducted on real-world transaction data from the AliExpress platform reveal that the proposed hybrid model consistently surpasses conventional standalone approaches—including both deep learning models (LSTM) and classical statistical models (such as ARIMA)—across multiple evaluation metrics, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The findings demonstrate that the integrated framework not only attains higher predictive accuracy but also ensures scalability and robustness across diverse market segments. This study advances the development of hybrid modeling methodologies in e-commerce analytics and provides practical implications for optimizing global recommendation systems within the strategic framework of initiatives such as China’s Belt and Road.