Adaptive Inventory Control under Non-Stationary Retail Demand: a Deep Reinforcement Learning Study on the M5 Dataset
Academic Conferences Series · 2026-04 · 21 (1)
DOI: 10.62381/acs.cess2026.08 ↗Kurz gefasst
Fachartikel in der Zeitschrift Academic Conferences Series, 2026-04. Verfasst von Yongjia Xie.
Abstract (Originalsprache)
Retail demand is not stable over time. Its average level and its variability change because of trend, seasonality, promotions, and shifts in consumer behavior. A fixed inventory rule that is tuned on past demand can become wrong when the demand level moves to a new range. This paper studies that problem with real data. We take one high volume product from the public M5 Walmart dataset and build a daily, single product, lost sales inventory environment that is driven by the real demand of that product. We train a deep reinforcement learning agent with Proximal Policy Optimization to set the daily order quantity. We compare the learned policy against two classic policies, a base stock policy and an (s, S) policy, both tuned on the training period. The held out test period for this product has a much lower mean demand than the training period, which makes it a clear test of adaptation. The deep reinforcement learning policy lowers the total cost on the test period by 13.5 percent, from 19705 to 17052 in the same cost units. Almost all of this gain comes from holding cost, which the agent cuts by about 48 percent because it stops over stocking once demand falls. The cost gain comes with a lower service level. The fill rate of the learned policy is 88.2 percent against 99.1 percent for the fixed rule, which reflects the chosen cost weights. The result shows that an adaptive learned policy can track a new demand level and reduce cost, while a fixed policy keeps acting on an old level. We report the full setup, the data statistics, and the results so the study can be reproduced.
Zitieren
APA
Xie, Y. (2026). Adaptive Inventory Control under Non-Stationary Retail Demand: a Deep Reinforcement Learning Study on the M5 Dataset. Academic Conferences Series, 21(1). https://doi.org/10.62381/acs.cess2026.08
Vancouver
Xie Y. Adaptive Inventory Control under Non-Stationary Retail Demand: a Deep Reinforcement Learning Study on the M5 Dataset. Academic Conferences Series. 2026;21(1). doi:10.62381/acs.cess2026.08
BibTeX
@article{Xie2026,
author = {Xie, Yongjia},
title = {Adaptive Inventory Control under Non-Stationary Retail Demand: a Deep Reinforcement Learning Study on the M5 Dataset},
journal = {Academic Conferences Series},
year = {2026},
volume = {21},
number = {1},
doi = {10.62381/acs.cess2026.08}
}
EUVIMED