Inventory Out of Stock Inference Using Machine Learning With Ensemble Strategies

Abstract

Inventory Record Inaccuracy (IRI) poses a significant risk to business continuity, particularly for Maintenance, Repair, and Overhaul (MRO) enterprises. While rigorous record-keeping is intended to ensure data integrity, it paradoxically often leads to IRI, manifesting as Inventory Freezing or Phantom Inventory. This study aims to demonstrate a comprehensive data analytics pipeline—from data acquisition and transformation to analysis and decision-making—using a machine learning approach. The research employed Random Forest Classifier (RFC) and Extreme Gradient Boosting (XGB) models on a dataset of 10,369 observations from a Jakarta-based warehouse, enhanced with 16 supply chain features and 12 Boolean indicators for missing data. The results indicate a clear precision-recall trade-off: RFC achieved 85% precision with 50% recall, while XGB attained 93% recall with 29% precision after hyperparameter tuning. To overcome this, two ensemble strategies are proposed: Consensus (voting) and Cascading (pipeline) ensembles, which offer management practical options to optimize stock-checking efforts. The key implication is that no single model is superior, but a strategic combination of models can effectively predict Out-of-Stock (OOS) events, reducing reliance on heuristic inventory checks