<p>International trade risk prediction has become increasingly challenging due to globalization, digitalization, and the rapid expansion of heterogeneous data generated from logistics platforms, customs systems, and Internet of Things (IoT) enabled supply chains. Traditional trade risk assessment approaches mainly depend on statistical indicators and rule-based evaluation methods, which limit the ability to perform intelligent and proactive risk prediction. Moreover, existing analytical models face difficulties in handling high-dimensional trade datasets and identifying the most significant risk indicators effectively. To overcome these challenges, a proposal is made for an optimized deep learning framework known as the Stellar Oscillation Optimizer-Based Convolutional Attention Network (SOO-CAN) for international trade risk prediction and analysis. The dataset used in this research is the Cross-Border Trade and Customs Delay Dataset obtained from Kaggle, containing more than 10,000 shipment records with 22 attributes related to shipment characteristics, compliance indicators, trade routes, and customs inspection information. Data preprocessing includes data cleaning and Z-score normalization to ensure consistency and remove noise from the dataset. Principal Component Analysis (PCA) is applied for feature extraction and dimensionality reduction, retaining the most significant trade risk attributes while eliminating redundant information. The model implementation is performed using Python. The CAN extracts hierarchical feature patterns from multidimensional trade data and highlights important indicators using an attention mechanism. The SOO optimizes model hyperparameters to enhance convergence and prediction stability. Experimental results demonstrate that the proposed SOO-CAN framework achieves 96% prediction accuracy, confirming its effectiveness for reliable international trade risk prediction and analytical decision support.</p> Graphical Abstract <p></p>