E-commerce fraudulent returns present a significant challenge for retailers, leading to financial losses and operational inefficiencies. This paper introduces a hybrid machine learning solution to detect fraud in return orders at both pre- and post-product receipt stages. We propose a novel Customer Return Score (CRS) system that evaluates customer behavior to flag potentially fraudulent activities. High-scoring customers may receive automatic refunds for low-value returns, while higher-risk returns undergo a more rigorous process using Random Forest, Support Vector Machines (SVM), and Logistic Regression models. Among these, Random Forest emerged as the best-performing model, particularly in handling imbalanced datasets. This approach reduces financial losses, minimizes manual intervention, and improves the overall efficiency of fraud detection in e-commerce returns. As part of future work, we aim to integrate Reinforcement Learning (RL) for adaptive fraud detection strategies and blockchain technology to secure transaction data with tamper-proof records.