An Explainable Optimization Framework for Demand-Driven Inventory Decisions Using SHAP and Mixed Integer Programming

Abstract

Demand uncertainty complicates inventory decision-making and requires decision-support systems that are both accurate and transparent. However, existing studies primarily emphasize either demand forecasting or inventory optimization, with limited attention to integrating explainability into a unified decision-making framework. This study develops and evaluates an Explainable Optimization Framework that combines Extreme Gradient Boosting (XGBoost) for demand forecasting, SHapley Additive exPlanations (SHAP) for model interpretability, and Mixed Integer Programming (MIP) for inventory optimization. The framework was developed following the Design Science Research methodology, encompassing problem identification, artifact development, demonstration, evaluation, and communication. Model performance was assessed using rolling-origin backtesting to provide a robust evaluation under dynamic and uncertain demand conditions. Forecasting accuracy was measured using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Experimental results demonstrate that XGBoost achieved the highest forecasting accuracy, with an RMSE of 43.87, MAE of 33.92, and MAPE of 7.61%. The forecasted demand was subsequently incorporated into the MIP optimization model, resulting in a 22.3% reduction in total inventory cost, an increase in service level from 91.2% to 96.8%, an improvement in fill rate from 89.7% to 95.4%, a 57.1% reduction in stockout frequency, and an increase in inventory turnover from 5.8 to 7.2 compared with the baseline approach. SHAP analysis identified historical demand, promotional activities, and product price as the most influential variables affecting demand predictions, providing transparent explanations that enhance managerial trust and support informed inventory decisions. Overall, the proposed framework demonstrates that integrating forecasting, explainability, and mathematical optimization into a unified decision pipeline significantly improves operational efficiency, inventory performance, decision transparency, and resilient data-driven supply chain management across diverse industrial sectors