Deep Learning for Intelligent Supply Chain Optimization: Enhancing Operational Efficiency and Waste Reduction in U.S. Service Industries

Abstract

U.S. service industries face persistent inefficiencies driven by volatile demand patterns, short product life cycles, and operational fragility across logistics networks. These conditions create a structural forecasting challenge where traditional statistical models struggle to capture event-driven variability, leading to systematic overstocking, stockouts, and waste. This study addresses that challenge by developing a deep learning–driven supply chain optimization framework that integrates high-resolution calendar features, lagged demand patterns, and sequence-based neural forecasting architectures. Using the M5 dataset as a proxy for service-sector operational behavior, we benchmark classical models (Naive, Linear Regression, LightGBM) against LSTM, N-Beats, and a lightweight Transformer model designed for Colab-scale experimentation. We evaluate the forecasting outputs through a simulation of inventory policies including (s, S) and base-stock systems, capturing holding, shortage, and waste costs under multiple demand scenarios. Results show that Transformer-based models consistently outperform statistical baselines on multi-horizon forecasting, improving RMSE and MAE across varied temporal contexts. When integrated with inventory policies, deep models reduce stockouts and expired-inventory waste while lowering total operational costs compared with heuristic or statistical forecasting systems. Sensitivity and ablation analyses confirm the value of calendar encodings, lag structure, and longer receptive fields in improving predictive stability. Computational profiling confirms that optimized neural architectures can run efficiently on edge-constrained environments such as Colab GPUs. These findings demonstrate that deep learning–enabled forecasting provides a practical and high-impact path toward intelligent supply chain optimization in U.S. service industries.