Order picking is the most labour-intensive warehouse activity, and travel rather than retrieval consumes most of the picking cycle. Storage slotting is therefore a high-leverage operational lever, yet slotting decisions remain driven by backward-looking demand summaries such as periodic ABC classification. This paper develops and validates an integrated framework in which a machine learning model forecasts near-future product-level picking demand, a genetic algorithm (GA) consumes that forecast to assign products to physical locations, and the resulting layout is evaluated by replaying real, unmodified picking waves through an aisle-aware routing simulation. The framework is applied to a public order-picking dataset from a footwear warehouse in Sherbrooke, Quebec (de Assis et al. 2025). On a temporal holdout, LightGBM with a Tweedie objective forecasts 20-business-day horizon-total picks at R² = 0.849, outperforming XGBoost, naive persistence and an identically-informed Ridge baseline. A replicated-stock GA then assigns 2,456 Reference–Size combinations across 2,292 locations under the warehouse’s real 18-unit capacity. A central methodological finding is that naive layout comparison on this dataset is confounded: benchmark layouts replicate stock far more heavily than the GA, and replica volume mechanically shortens routes independently of placement quality. Under an equalized per-product stock budget, the forecast-driven layouts lead clearly, reducing travel by 22.5% (GA-Ridge) and 20.0% (GA-ML) against random storage, while class-based storage achieves 6.5% and dedicated and hybrid layouts fall below random. Because the warehouse picks with manual trolleys, benefit is accounted in labour hours: 1,700–1,912 hours per year, roughly 0.9–1.0 full-time equivalents.

Keywords

Warehouse slotting, storage location assignment, demand forecasting, genetic algorithm and order picking.