Data-driven analysis on inventory problem for anticipatory shipping

Abstract

Under the anticipatory shipping (AS) mode, online retailers predict inventory and ship items to various hubs before orders arrive. In fact, the inventory decision for each hub is essence for AS, since overage increases shipping costs, while underage incurs order losses, similar to a Newsvendor mode. In this study, we introduce a “Machine Learning - Newsvendor” framework and adopt big data analytics to optimize the AS inventory. Specifically, we integrate forecasting algorithm into optimization model and develop two algorithms, i.e., XGBoost-NV (eXtreme Gradient Boosting - Newsvendor) and LGBM-NV (Light Gradient Boosting Machine - Newsvendor), and then propose a two-stage hybrid algorithm C-DT-NV (Clustering-Decision Tree-Newsvendor), facilitating algorithm selection between XGBoost-NV and LGBM-NV, for homogeneous products. Through employing the algorithms over enormous amount of data in a large bakery chain, we find that XGBoost-NV (LGBM-NV) outperforms XGBoost (LGBM) in the AS inventory optimization, where the average cost decreases by 23.91% (23.57%), and C-DT-NV decreases average cost and inventory error further. Finally, we examine three influential factors (i.e., price, demand volatility and return quantity) in AS inventory within the “Machine Learning - Newsvendor” framework and find that, as product demand declines, the AS inventory decreases with price increasingly, while the AS inventory decreases with demand volatility significantly, regardless of the demand. In addition, as return quantity decreases, the AS inventory is more likely to increase when the demand stays high, but keeps constant when the demand is low enough.