Research on Chain Supermarket Inventory Optimization based on Multi-dimensional Spatio-temporal Data

Abstract

With the increasing scale and complexity of chain supermarkets, inventory management in multi-dimensional spatio-temporal data environments becomes more challenging, as traditional methods fail to handle regional demand heterogeneity. This paper proposes an integrated inventory management approach that leverages spatio-temporal data by combining the DeepAR probabilistic forecasting model with the newsvendor model for region-specific optimization. Sales regions are divided into five areas, and four demand-influencing features are selected. The DeepAR model, with a Gaussian likelihood, estimates demand distributions for each region. Experimental results show stable training convergence and high prediction interval coverage (PICP above 0.85 for most regions), while adaptively widening intervals in data-scarce or highly seasonal regions. These forecasts are then used in the newsvendor model to determine optimal order quantities by minimizing overage and underage costs. The proposed framework enables differentiated inventory strategies, improving service levels and operational efficiency.