Inventory Reconstruction and the Stability of Policy Comparisons under Censored Retail Demand
Retail datasets can record sales and stockouts while omitting the inventory quantities needed to train an ordering policy. We ask whether two ways of reconstructing those quantities yield the same policy comparison when uncertainty about demand on stockout days is held fixed. Using FreshRetailNet-50K, we compare a cutoff censored sample-average approximation learner with a Kaplan–Meier estimator. We refit both under the published inventory reconstruction and a hypothetical alternative that pools inventory slack. The primary analysis covers 1,133 store-product series in 96 held-out stores over 20 days. At a shortage-to-overage cost ratio of nine and a normalized loss margin of 0.02, the historical loss bounds favor one learner by at least the margin in 273 series under at least one reconstruction. Both reconstructions favor the same learner in 179 of these series. Of the remaining 94, ten switch which learner is favored. Giving each store equal weight, 8.27% of eligible series lose a common certificate (approximate store-bootstrap 95% interval, 6.66–10.03%, conditional on the evaluation dates). In all ten reversals, the cutoff learner changes its order quantity while the identified Kaplan–Meier quantity stays fixed. Inventory-balance checks limit the interpretation of either reconstruction as a physical stock history. Even with exact bounds for missing evaluation demand, the choice of training inventories can change which ordering policy the benchmark supports.
Authors
- S. M. Abdulla Al Mamun
Institutions
- University of Manchester (GB)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24374632
- Primary Topic
- Supply Chain and Inventory Management
- Type
- preprint