Retail store-SKU level replenishment planning with attribute-space graph recurrent neural networks

Abstract

Large retail chains routinely make replenishment decisions for inventory at the store Stock-Keeping-Unit (SKU) level, relying heavily on accurate demand forecasting to meet target service levels. We critically examine the current state-of-the-art in demand forecasting, revealing a notable gap in existing models, specifically the neglect of intricate cross-SKU promotional effects. To address this gap, we introduce an innovative solution: the Attributed Space Graph Recurrent Neural Network (ASG-RNN), which encodes demand history and drivers' history using a recurrent neural network, and captures cross-SKU promotional interactions through a graph convolution network. The graph convolutions operate on an attribute space graph constructed using K-nearest neighbor algorithm based on SKU-level static attributes. Initial empirical tests were conducted on a dataset comprising 200 SKUs in the Coffee category from a retail chain with 93 stores. The analysis demonstrates that ASG-RNN, on average, reduces loss by 4.1% compared to the best benchmark, resulting in average inventory cost savings of 4.94%. Notably, for promoted SKUs, ASG-RNN reduces loss by 13% on average, and it can achieve a 30% reduction in loss when the target service rate is high. Further experiments conducted on four additional product categories, including Beer, Laundry Detergent, Milk, and Mayo, yield similar results, showcasing the robustness of ASG-RNN's performance. These findings underscore the practical significance and cost-effectiveness of our innovative approach in enhancing demand forecasting and optimizing replenishment decisions for large retail chains.