International Stock Forecasting Using Ensemble Deep Graph Models and Complex Network Analysis

International stock forecasting faces challenges as global stock markets become increasingly synchronized. This study develops a prediction model for 46 global stock prices by examining the complex interconnectedness of the global stock network. We propose a multilevel fusion approach that integrates technology and knowledge for accurate international stock predictions. First, for technical convergence, we introduce a hybrid model combining graph convolutional networks (GCNs) and long short‐term memory (LSTM) to capture both temporal dynamics and spatial dependencies. Additionally, we construct an ensemble of multiple GCN‐LSTM models via AdaBoost, which reduces mean absolute error (MAE) by an average of 36.7% compared with baseline models. Second, for knowledge fusion, we construct stock centrality features that represent influence on the entire network market for each country and use this feature as input to models, further decreasing MAE by 23.2%. In profit tests, trading based on our predictions significantly improved cumulative profit and the Sharpe ratio by 2.06× and 2.83×, respectively, compared with a simple long‐holding strategy. Finally, using Shapley Additive Explanations (SHAP), we derived the insight that our stock network feature has a stronger impact on future stock prices during high‐volatility periods, such as when the financial system experiences shocks. These findings suggest that the network index becomes particularly valuable in enhancing predictive performance under turbulent market conditions.