Calibrated Quantile Transformers for Cost-Aware Inventory Control in Simulated Multi-Echelon Supply Networks

Accurate demand quantiles do not automatically yield low-cost network inventory decisions. We evaluate a series-aware quantile Transformer linked to a rolling residual correction that shrinks item-specific errors toward pooled calibration residuals. The corrected 0.90 demand quantile drives retailer order-up-to decisions under shared distribution-center supply, with explicit replenishment lead times. To distinguish the value of calibration from forecasting architecture, a LightGBM quantile model and fixed-buffer controls are subjected to the same simulation. 10 independently generated synthetic networks each contain 48 product-store demand series over 176 weeks, and all policies use the same chronological 31-week holdout. Across the networks, the calibrated Transformer lowers average simulated cost by 13.1% relative to static stock and 3.6% relative to its uncalibrated quantiles. The network-paired mean incremental cost reduction is 3.54% with a 95% bootstrap interval [2.33%, 4.78%]. The study isolates decision postprocessing in a controlled simulation rather than making a real-retailer deployment claim.