Decoupling Expiration Risk Classification and Discount Rules in a Web Inventory Monitoring Prototype

Abstract

Recorded expiration status and future inventory loss are different decision targets. This study audits a date-based monitoring architecture, a historical K-nearest neighbor (KNN) experiment, and independent markdown rules. Source-document recovery established that the 1,000 simulated labels were deterministic date bands: High at seven days or fewer, Medium at 8–30 days, and Low beyond 30 days. The revised architecture assigns date status directly and retains KNN only as an offline diagnostic. The historical K = 9 confusion matrix gives accuracy 0.735, macro F1 0.502, balanced accuracy 0.488, High recall 0.441, and Medium recall 0.063; 32 of 59 High records were classified Low. These results are not an independently selected-model estimate. A new sensitivity grid evaluated 135 hypothetical combinations of demand response, sell-through, cost, and discount. Discounting improved accounting profit in 60 of 108 discounted scenarios and worsened it in the remainder; these counts describe a designed grid, not real-world success probabilities. Forty black-box cases passed for a new Python reference decision core, without validating the original PHP application. The findings support direct treatment of known dates and conditional economic review of markdowns. Original-data ablation, matched classifier comparisons, renewed validation, and application-level testing remain dependent on the original CSV and source code.