Machine Learning-Based Demand Forecasting and Inventory Replenishment in Smart Warehousing: A Cost-Sensitivity Analysis Using Public Retail Sales Data

Abstract

Demand uncertainty is a practical bottleneck in smart warehousing and retail inventory management. This study examines whether a machine-learning forecast can improve replenishment decisions beyond a simple moving average benchmark. Using the M5 retail sales dataset, the paper compares a 7-day moving average with a Random Forest model and then feeds both forecasts into a reorder point-based inventory simulation. The sample covers the top 100 FOODS SKUs in Walmart store CA_1 during a 28-day out-of-sample test period. Random Forest improves forecast accuracy: WAPE decreases from 39.82% to 35.03%, a relative improvement of 12.05%. Service outcomes also improve: stockout rate falls from 3.04% to 2.43%, and lost sales decrease from 724 to 537 units. The cost result is conditional. Under the baseline cost setting, the Random Forest policy carries more inventory and produces a slightly higher total cost. Sensitivity analysis shows that the model becomes economically attractive when stockout penalties are higher or holding costs are lower. These findings suggest that machine-learning forecasts can support smart warehousing by reducing service risk, but their economic value depends on inventory cost structure.