AI-powered inventory optimization models: a strategic framework for improving stock management in US supply chains

Retail inventory management depends on dependable demand forecasts plus inventory rules that balance holding cost, ordering cost, and ‎stockout risk under uncertainty. Advanced machine learning models now appear frequently in demand forecasting research. Their real value ‎emerges only when forecast accuracy, uncertainty representation, and interpretability connect clearly to operational inventory outcomes. This ‎study investigates how forecasting approaches relate to inventory performance within one coherent, explainable evaluation framework. This ‎work develops an end-to-end inventory optimization framework using publicly available US retail demand, pricing, and calendar data. The ‎framework integrates feature-engineered demand forecasting with baseline statistical methods, machine learning models, and probabilistic ‎forecasting through LightGBM quantile regression. Forecast outputs feed directly into an (s, S) inventory policy optimized through simulation. Evaluation relies on rolling-origin back testing, inventory cost measures, fill rate, stockout counts, robustness experiments, and SHAP-‎based explainability. For the selected high-volume SKU, exponential smoothing produced the lowest point forecast error, exceeding naive ‎benchmarks plus a LightGBM point forecasting model. LightGBM quantile regression showed higher point error than exponential smoothing, while offering useful demand uncertainty ranges. Inventory simulations revealed policy parameters plus cost assumptions exerted greater influence on service levels plus stockouts than small gains in forecast accuracy. Back testing showed that conservative inventory policies ‎maintained high fill rates even when driven by simple forecasts. Explainability results showed recent demand features plus seasonal signals ‎dominated machine learning predictions, while a linear surrogate model reproduced most model behavior. The findings show inventory ‎outcomes depend primarily on policy design, cost calibration, and uncertainty treatment rather than forecasting model sophistication. Accurate point forecasts alone fail to guarantee effective inventory control. The proposed framework emphasizes integrated evaluation, simulation, ‎and explainability as essential components when applying AI-based forecasting to retail inventory decisions‎.