ARCA: Forecasting demand for device accessories at Amazon

Multi-horizon and multi-lead time forecasting is a well-established area in machine learning, particularly in demand forecasting, which plays a critical role in a product’s lifecycle. Accurate forecasts support key operational functions, including inventory management, financial planning, promotion planning, and supply chain optimization. Traditional demand forecasting methods typically rely on learning sales patterns directly from historical data of a given product. However, forecasting demand for accessories—products that are purchased in conjunction with main devices (e.g., covers or headphones for tablets)—introduces additional complexities. In this paper, we propose a novel forecasting technique for accessories that leverages their inherent attach rate patterns to main devices. Additionally, we introduce a correction module to mitigate biases in the forecasts of the main products, thereby improving the accuracy of accessory predictions. While our primary focus is forecasting accessories for Amazon devices, the proposed methodology is broadly applicable to any product that exhibits an attach rate dependency. The proposed model was deployed in production within Amazon in July 2024 and has since been generating daily accessory forecasts across 17 countries and two channels: Online (Amazon website) and Offline (third-party retailers).