Class-Conditional Allocation of Forecasting Models for Retail Inventory Replenishment: Evidence from an ABC-Segmented Hardware Portfolio

Background: Retail inventories are heterogeneous, yet demand forecasting studies typically select one model for an entire portfolio, leaving open how forecasting effort should be allocated across inventory segments. Methods: Six models—ARIMA, SARIMA, ARIMAX, Random Forest, LightGBM and an LSTM network—were estimated on weekly demand for a Colombian hardware retailer (35,399 transactions, 1633 SKUs, January 2025–March 2026). Because 95.5% of items are intermittent or lumpy, models were fitted to three ABC-class-aggregated series of 60 observations, evaluated on a 12-week holdout and under a rolling-origin scheme, and compared against naive benchmarks using MASE and RMSSE. Results: The best model differs by class. In Class A a statistical specification wins, ARIMAX on the holdout (RMSE 843 units per week) but ARIMA under rolling origins, the two being statistically indistinguishable. Random Forest is best in Class B (255). In Class C no model improves on a last-value rule (75 against 66). Relative to that benchmark, error falls to 39.8% in Class A and 47.6% in Class B. Diebold–Mariano tests separate 16 of 45 pairs on the holdout and three of 30 under rolling origins. Conclusions: Allocating models by inventory class is operationally justified rather than statistically decisive, and the low-turnover class warrants no model at all.