Until very recently, there was limited evidence to suggest that data-driven machine learning (ML) methods improve sales forecasting to optimize retail operations. However, since 2015, ML-based forecasting has dominated the most recent retail sales forecasting competitions. Despite their superior performance, simple statistical forecasting techniques remain widely used in the retail industry. This is partly because, for all their advantages, these top-performing ML methods are often too complex to implement and impossible to interpret, earning the name ‘black box.’ In this dissertation, we investigate how to facilitate the deployment of these state-of-the-art forecasting methods in terms of computational requirements and implementation complexity. In addition, we investigate how the adoption of these new forecasting methods impact the day-to-day inventory management of retailers. Finally, we try to understand the key factors that drive the performance of these top-performing forecasting methods, and provide recommendations on how to implement these.