2028. Learning Inventory Control in the Face of Censored Demand and Shrinkage
Invited abstract in session TA-47: Retail Inventory Management, stream Retail Operations.
Tuesday, 8:30-10:00Room: Parkinson B08
Authors (first author is the speaker)
| 1. | Recep Bekci
|
| Management Sciences, University of Waterloo |
Abstract
Inventory shrinkage, driven by theft and operational inefficiencies, poses a significant financial burden on retailers, with annual losses exceeding $100 billion. A critical challenge lies in mitigating shrinkage in real time when inventory levels are unobservable and demand is uncertain. Traditional inventory control policies rely on precise stock data, which is often unavailable due to high auditing costs and latent shrinkage processes. This paper addresses this gap by formulating a dynamic inventory control problem under partial observability, where retailers can only infer stock levels through sales data and stock-out events. We propose COSIL (Cycle-based Online Shrinkage Inventory Learning), a novel algorithm that dynamically adjusts order-up-to levels by learning both demand patterns and shrinkage rates over cycles defined by stock-out occurrences. COSIL leverages gradient-based updates and a shrinkage estimation subroutine to refine inventory decisions without direct observation of stock levels. Theoretical analysis demonstrates the algorithm’s efficiency, while computational experiments validate its robustness across diverse demand and shrinkage scenarios. Our results offer retailers a practical, data-driven framework to reduce losses while adapting to evolving operational uncertainties, with implications for automated replenishment systems and loss prevention strategies.
Keywords
- Inventory
- Machine Learning
Status: accepted
Back to the list of papers