Online shopping is popular among citizens due to convenience, competitive pricing, and flexible return policies. However, flexible return policies can cause E-commerce retailers who sell seasonable items to suffer losses from excess inventories and outdated products. No specific machine learning-based prediction models are available to predict the return of orders from E-commerce retailers for seasonal sale items. This research proposed an improved machine learning-based prediction model for the return of orders from seasonal sale items of E-commerce retailers in China using past order data. By adopting the CRISP-DM method, a comparison of two machine learning models, namely the K-nearest neighbor (KNN) and random forest (RF), found that RF outperformed KNN with a higher recall value when the SMOTE method was adopted to reduce the impact of unbalanced data on the model performance. This research also identified the important predictors that caused a high return rate. The findings will enable E-commerce merchants to identify possible highly returnable items and make necessary adjustments in production to reduce unnecessary expenses and improve operational efficiency.