The ABC storage is the most popular class-based policy for the storage location assignment in warehouses. It divides a storage area into three zones and assigns the most demanded products to the best-located zone. Despite the policy’s popularity, arbitrary zone sizes are commonly used, which can lead to major efficiency losses. We investigate how several factors, such as the warehouse layout, the demand characteristics, and the storage and routing policies, impact the solutions for the zone sizing problem. We propose a new methodology to solve it using machine learning models to predict the optimal zone sizes considering the mentioned factors. We simulate many common manual warehouse settings, such as the multi-block layout, demand distributions, and several operating policies, to observe which zone sizes lead to the best performance in each one. The data generated is used to train four regression models – ordinary least squares, regression tree, random forest, and multilayer perceptron – to predict the optimal zone sizes from the best ones observed. Computational experiments show that zone sizes provided by all models significantly improve the order picking efficiency when compared to the arbitrary zone sizes commonly used, notably for the one-zone (random policy), the twozone (20/80 rule), and the three-zone (20/30/50) systems. The proposed methodology is easily adaptable for different warehousing systems and problems when enough data is available to train the models. The resulting linear functions and decision trees are made available and can be used by practitioners for determining zone sizes for their particular warehouse.