Probabilistic and Point Reconciliation in Deep Learning-Based Hierarchical Forecasting for Retail

Hierarchical retail forecasting requires predictions that are both accurate and coherent across multiple planning levels, from total demand to individual product–store series. Because these forecasts guide inventory, replenishment, storage, and distribution decisions, improving their coherence and reliability can support more efficient resource use, reduce avoidable overstock and product waste, and limit the need for emergency logistics. This study investigates how global deep-learning architectures interact with post hoc reconciliation in point and probabilistic forecasting. Using an M5-derived hierarchical and grouped structure comprising 42,840 series, we compare three MLP-oriented models, MLP, N-BEATS, and N-HiTS, with five transformer-based models, Transformer, Temporal Fusion Transformer, Informer, PatchTST, and Autoformer. All models are evaluated under a common 28-day forecasting horizon, temporal partition, Optuna-based tuning protocol, and three complete seeded runs. Coherence is imposed using Bottom-Up reconciliation and four MinTrace variants, while probabilistic forecasts are generated through residual-block bootstrap reconciliation. Point and probabilistic performance are assessed level-wise and globally using MASE and scaled CRPS, respectively. The results show that the strongest transformer-based combination outperforms the strongest MLP-based combination at every hierarchy level. PatchTST combined with MinTrace-WLS-struct is particularly effective at aggregate and intermediate levels, achieving a Total-level MASE of 0.537 and sCRPS of 0.037. At the Product–Store level, Bottom-Up reconciliation becomes preferable, with the Transformer attaining the lowest MASE of 1.367 and sCRPS of 0.912. Because granular series dominate the hierarchy-wide average, the lowest overall MASE is obtained by TFT with Bottom-Up reconciliation (1.428), whereas the lowest overall sCRPS is shared by the Transformer and Informer with Bottom-Up reconciliation (0.813). These findings demonstrate that neither the forecasting architecture nor the reconciliation method should be selected independently of hierarchy depth and forecasting objective. Strategic and tactical levels benefit primarily from PatchTST with MinTrace reconciliation, whereas highly granular operational forecasting favors Bottom-Up reconciliation with transformer-based models. From a sustainability perspective, this level-aware framework provides a basis for aligning forecasting decisions with resource efficiency, waste reduction, service reliability, and greater resilience across the retail supply chain.