Explainable LSTM-Based Demand Forecasting for Multi-Channel Retail Supply Chain Performance Analysis

Modern retail supply chains face unprecedented complexity due to multi-channel distribution, demand volatility, and increasing customer expectations. This study presents a comprehensive evaluation framework integrating ten state-of-the-art machine learning algorithms with explainable AI methodologies to address both predictive accuracy and model transparency in supply chain demand forecasting. We evaluate diverse algorithmic families including ensemble methods, gradient boosting variants, kernel machines, and recurrent neural networks on the DataCo Global multi-channel retail supply chain dataset (2016–2018, 180,519 transactions). Experimental results reveal LSTM’s exceptional performance achieving RMSE of 42.20, R2 of 98.79%, and MAPE of 0.41%, demonstrating 39-fold improvement over traditional approaches. The integrated explainable AI framework using LIME unveils temporal lag features and delivery scheduling as critical demand determinants, transforming black-box predictions into actionable business intelligence for supply chain practitioners. Research purpose: To develop an evaluation framework bridging predictive performance and model interpretability in retail supply chain demand forecasting through systematic comparison of advanced machine learning approaches. Research motivation: Contemporary supply chain forecasting research lacks comprehensive frameworks jointly optimizing prediction accuracy and model transparency. Multi-channel retail operations require forecasting systems delivering both superior performance and explainable insights for strategic inventory management, distribution planning, and operational decision-making. Research design, approach, and method: Multi-algorithmic evaluation of ten algorithms on real-world supply chain transaction data spanning two years. Models evaluated using RMSE, R2, and MAPE metrics, with LIME-based interpretability analysis to identify key demand drivers and validate model reasoning. Main findings: LSTM achieves superior temporal dependency modeling with 98.79% R2 accuracy. Gradient boosting methods show competitive performance while ensemble approaches reveal metric inconsistencies suggesting overfitting. Explainability analysis identifies temporal lag features, moving averages, and seasonal patterns as primary predictive drivers, with delivery timing significantly influencing demand patterns. Practical/managerial implications: The framework enables supply chains to transition from reactive to predictive management paradigms. LSTM-driven forecasting optimizes inventory positioning, reduces stockouts, and minimizes holding costs. Explainable components build stakeholder trust and provide actionable insights for operational adjustments.