Explainable Machine Learning Models for Demand Forecasting and Inventory Management in Smart Commerce

Smart commerce requires demand forecasting that are highly accurate and transparent. However, the deep learning models operate as black boxes and struggles to justify inventory decisions. To achieve both these aims, this paper presents a hybrid model that combines Temporal Fusion Transformer (TFT) with stabilized SHAP and counterfactual explanations. TFT is trained on an IoT-augmented Walmart data (5 stores, 100 items, 1,431 days) and adds a temporal regularization loss that decreases SHAP explanation variance by 42.3 percent to 18.1 percent. An innovative closed-form derivative conditions SHAP values on sensitivity of inventory cost, which facilitates gradient-based policy optimization. The framework has 8.3% MAPE and 91.2% directional accuracy, which is 1.2 percentage points higher than N-BEATS. XAI-optimized inventory policies lowered safety stock by 16.8% and still provide 98.5% service in a 180-day randomized trial, which resulted in a 12.4k cost savings per product. The system achieves SLAs of 142 ms at 142 ms of latency deployed using FastAPI and MLflow. This presents the earliest, auditable, deployable, explainable automated replenishment, EU AI Act transparency requirements alongside providing quantifiable financial results in smart business settings.