A Leakage-Free ML Framework for Cycle-Count Prioritization in Multi-Depot Distribution Networks

We study a real stock-reconciliation snapshot from
a large consumer-electronics distributor in Bangladesh covering
1,890 stock-keeping units (SKUs) across 18 depots (34,020 SKU–
depot cells). A key property of such reconciliation data is that
the recorded discrepancy is an exact arithmetic function of the
other two recorded fields; naive models that consume those
fields therefore achieve near-perfect but meaningless accuracy
through target leakage. We make this leakage explicit and instead
formulate an honest, decision-relevant task: predict, from only
contextual attributes known before a physical count, which SKU–
depot cells are most likely to exhibit a material discrepancy,
so that limited cycle-count effort can be prioritized. Among
seventeen benchmarked classifiers, a gradient-boosting model
attains a five-fold cross-validated ROC-AUC of 0.805 ± 0.006.
Translated into operations, auditing the top 20% of cells ranked
by predicted risk recovers 52.8% of all discrepancies, versus 20%
under random counting—a 2.64× efficiency gain. We further
provide permutation- and SHAP-based explainability, a budget-
constrained audit-optimization formulation, a sensitivity study
over the materiality threshold, and complementary ABC and
TOPSIS multi-criteria SKU prioritizations. We report results
candidly, including the limited predictability inherent to count
errors and the single-period nature of the data, and outline how a
second count period would enable genuine temporal forecasting.