<p>Convolutional Neural Networks (CNNs) excel in feature extraction and pattern recognition in areas like image classification and speech processing. However, their application in the financial sector has been limited due to the complexity, high dimensionality, and temporal nature of financial data, as well as the need for model interpretability. This study, based on CNN technology, proposes a Reordering-Enhanced Grad-CAM algorithm to improve model interpretability and reliability, offering transparent and dependable tools for financial decision-making. The innovation of this study lies in two key aspects: firstly, it replaces the traditional manual variable selection approach with automatic feature extraction and fusion using CNNs, demonstrating the effectiveness of deep learning in handling large-scale financial data. Secondly, we propose a novel reordering-based iterative algorithm that adapts Grad-CAM, originally designed for image classification, to multi-source financial time series data, treating sliding window data segments as pseudo-images to improve interpretability and identify critical features. Using data from Shanghai and Shenzhen A-shares (1990–2020), the Reordering-Enhanced Grad-CAM technique generated heatmaps that identified key predictive indicators, leading to improved model performance. Robustness analysis demonstrated that over 70% of important variables were consistently identified, with some models reaching up to 100%, confirming the reliability and stability of our method in financial distress prediction.</p>