A Residual‐Based Hybrid Forecasting System for Large‐Scale Retail Sales Prediction

Accurate sales forecasting in large‐scale retail systems is challenging due to complex temporal dynamics and latent inter‐store dependencies. While statistical models such as SARIMAX effectively capture trend, seasonality, and exogenous effects, they typically assume independence across stores and fail to exploit cross‐series relationships. To address this limitation, we propose a residual‐based hybrid forecasting framework that decouples temporal and relational modelling. SARIMAX models are first applied at the store level, and the resulting residuals are used to construct a residual‐driven inter‐store graph. A Graph Attention Network (GAT) then learns heterogeneous inter‐store dependencies to refine the forecasts. The approach is evaluated on two large‐scale datasets, Rossmann and Favorita. Results show consistent improvements over strong baselines, with up to 42% RMSE reduction on Rossmann and competitive improvements on Favorita. Diebold–Mariano tests confirm that improvements are statistically significant (). These findings demonstrate that residual‐based graph learning is an effective and interpretable approach for capturing latent inter‐store dependencies in retail forecasting.