Target eliminates unknown out-of-stocks across ~2,000 stores using ensemble ML Inventory Ledger

In brick-and-mortar retail, on-shelf availability is typically measured by whether the system records zero inventory — an approach that assumes on-hand records are accurate. When Target conducted physical audits across its nearly 2,000 U.S. stores, it found that half of all out-of-stock events were invisible to its systems: the recorded quantity showed available inventory while shelves were actually empty. Across an assortment of more than 100,000 SKUs, shrinkage sources — shipping and receiving errors, theft, misplaced items, and system glitches — continuously eroded record accuracy. Because replenishment triggers only when the system records zero inventory, unknown out-of-stocks stall reorder entirely, leaving products unavailable to guests and suppressing sales with no automated correction possible.

Target engineered the Inventory Ledger, an event-driven, stateless system that journals every inventory change — sales, replenishment, back-room put-away, and order fulfillment — across every item at every location in real time. Built on RocksDB and MongoDB with a custom sharding scheme to handle Target's transaction volume, the Ledger became the integration point for an ensemble of Machine Learning models (gradient boosted trees and neural networks) that infer unknown out-of-stocks from patterns in product, sales, replenishment, and inventory signals. Thousands of specialized models, each restricted to a specific product category or process-defect signature, feed an arbitration engine that selects one winning correction per item per store each day. Those corrections flow automatically to inventory accounting systems — triggering replenishment without requiring store team-member intervention. Prior to full automation, human verification served as an early-stage gate to validate model accuracy before scaling.

The Inventory Ledger and ML ensemble were deployed across nearly every product category in all ~2,000 Target stores. Key outcomes include:

• 360,000 inventory transactions per second processed at peak load, with up to 16,000 inventory position requests served per second
• Physical audits confirmed the original problem scale: 50% of out-of-stocks had been unknown to systems prior to the program
• Inventory corrections produced substantial sales lift for products that would otherwise have remained unavailable to guests
• Hardware-based detection (shelf-edge cameras, weight sensors, light sensors) was eliminated entirely — the data-driven ensemble proved more cost-effective and maintainable at nationwide scale

Models continue to learn and improve as labeled training data accumulates.

• Assign measurement to a team independent of solution developers — this removes bias and enables clean comparisons when multiple competing correction signals are under evaluation.
• Build an arbitration engine that selects one correction per item per store; manual team-member verification does not scale as the volume of correction suggestions grows.
• Specialized, domain-restricted models outperform a single general model across heterogeneous retail assortments — category-specific patterns require category-specific training data and features.
• Purely data-driven inference can be more cost-effective than hardware at large physical footprints, where camera and sensor costs compound across thousands of stores.