Uncertainty-Aware Inventory Forecasting Using Quantile Gradient Boosting and a Diagnostic Risk Index

Accurate demand forecasting plays a central role in inventory planning for retail supply chains, yet most operational forecasting systems rely on point estimates that overlook uncertainty and SKU-level risk. This paper presents an uncertainty-aware forecasting framework based on Quantile Gradient Boosting Decision Trees and introduces a novel Diagnostic Risk Index (DRI) to identify volatile and high-risk inventory items. The proposed method trains multiple quantile models to construct predictive intervals and combines interval width and forecast deviation into an interpretable risk score. Three extended variants of the DRI-price-weighted DRI, feature-importance-weighted DRI, and relative DRI-are examined through an ablation study. Experiments on the full M5 retail dataset demonstrate competitive forecasting performance while uncovering SKU-level volatility patterns that are not reflected in conventional accuracy metrics. The findings show that uncertainty-based risk diagnostics can support practical inventory management decisions, such as targeted replenishment and exception-based monitoring. The proposed framework is fully reproducible and scalable to large-scale retail forecasting environments.