A graph neural network-based framework for hierarchical time series forecasting in retail

Hierarchical time series forecasting in retail requires forecasts that are both accurate and aggregation coherent across multiple levels. Most existing methods follow a two-stage pipeline that first generates base forecasts and then reconciles them, which limits the exploitation of hierarchical structure in learning. We propose an end-to-end graph-based neural reconciliation framework that jointly models hierarchy-aware dependencies, temporal dynamics, and coherent forecasting within a single differentiable architecture. The framework combines graph neural networks and gated recurrent units with an explicit bottom-up reconciliation module that guarantees exact aggregation consistency. We evaluate the framework on two retail datasets, Italian and Walmart, under both without-promotion and with-promotion settings. The strongest variants, TGLP-BUN (Temporal Graph Linear Projection with Bottom-Up Nonlinear reconciliation) and TALP-BUN (Temporal Attention Linear Projection with Bottom-Up Nonlinear reconciliation), achieve the strongest overall performance against strong statistical and end-to-end deep-learning baselines. The gains are concentrated at higher levels on the sparse Italian hierarchy and are more broadly distributed across the Walmart hierarchy, especially when promotional information is included. Node-wise error distributions and Diebold–Mariano tests further show that these improvements are systematic across series rather than driven by only a small subset of nodes.