Bank customer churn prediction is a critical task in the banking industry, where prediction accuracy directly impacts the effectiveness of customer retention strategies and the stability of business revenue. This study proposes a lightweight and highly interpretable multi-model fusion method to enhance the accuracy and business applicability of bank customer churn prediction. The method integrates three base models with complementary strengths-XGBoost, CatBoost, and Multilayer Perceptron (MLP)-each excelling in numerical feature modeling, categorical feature encoding, and capturing nonlinear interaction relationships, respectively. During the fusion stage, a lightweight Attention-MLP module is introduced to dynamically weight the outputs of the three base models at the sample level. Additionally, a shallow MLP is employed to model the nonlinear mapping between the fused representations and churn probability, thereby constructing an end-to-end trainable and efficient architecture. The proposed fusion model was thoroughly validated through experiments on a public bank customer churn dataset. The results demonstrate that the fusion model significantly outperforms various single models and traditional fusion approaches (such as simple averaging and Stacking) across multiple key metrics, including Precision, Recall, F1-Score, and ROC-AUC. Ablation experiments were also conducted, further verifying the critical contributions of the lightweight attention mechanism and shallow MLP structure to fusion performance. Moreover, the attention weights in the fusion module serve as sample-level interpretability indicators, clearly illustrating the contribution of each base model to the final prediction, thereby enhancing model transparency and business interpretability. In summary, the proposed fusion model not only improves prediction performance but also balances interpretability and deployment efficiency, demonstrating potential for practical business applications.