The number of product returns represents a considerable cost factor in e-commerce, especially in the apparel sector. The application of advanced information technologies and predictive analytics, enabling to capture and analyze massive amounts of user data, pave the way for a more efficient management of product returns and reverse logistics. However, we identify a lack of data-driven approaches in this area, especially regarding product returns prediction. In this paper, we present an ensemble selection approach for predicting product returns in the apparel sector. Computational experiments indicate that our approach produces satisfying results in terms of prediction quality. We further explore the correlation between sample sizes and computational times. Thereby, we demonstrate that the run-time increases exponentially when using more data records. To address heavy run-time overheads resulting from high processing and memory requirements of classifiers, we present a framework to embed ensemble selection processes into a highly scalable cloud environment. The framework explains the provisioning of cloud resources and parallelization of tasks according to ensemble selection processes. It further builds a basis for considering data streams, data splitting, and a dynamic adoption of changing customer behavior over time, which has not been considered in related work so far. The envisioned forecasting support system aids retailers in reducing product returns and increasing profit margins.