Where Pooling Pays Off : A Machine Learning Approach to Estimating Sales Uplift Through Inventory Pooling in Omnichannel Fashion Retail

Inventory pooling has long been recognized as an effective strategy to reduce inventory-related costs by aggregating stock over multiple locations. However, most of the previous research conducted on the subject has focused on the cost-side and less attention has been given to the revenue-side effects, particularly the sales uplift achieved by improved product availability. Estimating this uplift is often challenging because of the inherent problem that historical sales data are often censored by stockouts, meaning that the observed demand does not fully reflect true customer demand. This study uses a two-phase methodology to investigate how machine learning models can be used to improve estimates of sales uplift through inventory pooling in an omnichannel fashion retail environment. In the first phase the models were developed and evaluated using uncensored weekly sales and inventory data from a fashion retail company. Two tree-based machine learning (ML) models, Random forest and XGBoost, were compared with the company’s existing Fair Share of Business (FSOB) model and a naive K-Nearest Neighbors baseline. In phase two the models were applied on the full dataset to reconstruct demand which was later used to carry out inventory constrained simulations under different substitution rates (0%, 25%, 50%, and 75%) and pooling scenarios (full pooling, store-to-online and online-to-store). The results show that the tree-based ML models outperform the benchmark approaches when evaluated on uncensored demand observations, with Random Forest achieving the strongest overall predictive performance, followed by XGBoost. However, because the true demand associated with censored data is inherently unobservable, superior predictive accuracy on uncensored data cannot be assumed to directly translate into more accurate sales uplift estimates. Across all simulation scenarios, the benchmark models generated considerably larger unconstrained sales uplift estimates, however, much of this uplift could not be realized once inventory constraints were applied. Random Forest consistently estimated the lowest sales uplift while XGBoost’s estimates ended up in between. The simulations further revealed that online-to-store pooling generated sales uplift levels close to those achieved under full pooling, whereas store-to-online pooling provided only limited benefits. The findings demonstrate the importance of accurately reconstructing censored demand when evaluating inventory pooling initiatives and highlight the potential of machine learning methods to support data-driven inventory decisions. The study contributes to the literature by linking demand reconstruction and inventory pooling analysis in a fashion retail context and provides practical insights for retailers seeking to improve inventory utilization and product availability.