An analytics-driven hybrid method for multi-item demand forecasting in supply chains

This study proposes a new deep learning (DL)-based approach for multi-item demand forecasting in multi-wave distribution networks. In modern merchandising supply systems, traditional forecasting techniques such as moving averages and autoregressive integrated moving average (ARIMA) models are often inadequate for capturing the dynamic nature of sales and operations planning, as they struggle with non-stationary demand, evolving market conditions, and the growing complexity of supply chain networks, resulting in forecasts that are neither sufficiently accurate nor timely. To address these limitations, this study evaluates a hybrid forecasting framework that combines poly-linear regression (PLR) with a transformer-encoder extended long short-term memory (TE-LSTM) architecture to identify latent demand patterns from large and heterogeneous datasets. An empirical analysis compares the proposed PLR-TE-LSTM model with baseline approaches such as standard LSTM across multiple products and distribution locations, demonstrating consistently superior forecasting performance in multi-product, multi-distribution center settings. The study further examines the operational impact of improved forecasting by evaluating alternative inventory review strategies under a fixed-quantity replenishment policy, showing meaningful improvements in forecast accuracy, order fulfillment performance, inventory holding costs, and service levels. The results indicate that the proposed DL framework enhances inventory-related decision-making by reducing excess inventory and stockout risks, thereby improving efficiency and responsiveness in complex distribution networks.