Inventory policy decisions can depend heteroge neously on product and operational characteristics, making a single global feature-to-policy relationship potentially restrictive. This study develops an adaptive partitioning framework for feature-driven inventory policy estimation, referred to as DAPT. The method partitions the feature space using depth-limited threshold-based splits and estimates a regularized Ridge policy within each terminal region. A single global Ridge policy using the same features and historical information serves as the benchmark. The empirical study uses an HKU inventory dataset and evaluates both policy classes using chronological expanding window validation. Under the prespecified configuration λ = 1, Monika Budania CentraleSup´ elec (Universit´ e Paris-Saclay) 3 rue Joliot-Curie Gif-sur-Yvette, France [16]–[19]. A natural response to such heterogeneity is to allow different regions of the feature space to use different decision relationships. Tree-based methods provide one mechanism for constructing such regions, while local linear models provide a relatively interpretable way of estimating region-specific relationships. Tree-based approaches have been considered in inventory and replenishment applications [12], [15]. However, the combination of partitioning and local regularized linear estimation should not itself be regarded as a new modeling principle. K¨unzel et al. explicitly study tree-based estimators with linear aggregation functions and consider Ridge regres sion within tree leaves [8], [11]. nmin = 1000, Dmax = 2, and K = 12, DAPT achieves a lower empirical policy-cost proxy than the global Ridge benchmark in all four chronological evaluation windows. The mean relative improvement is 4.51%, with a median of 4.60% and observed improvements ranging from 1.85% to 6.99%. The results pro vide empirical evidence that adaptive feature-space partitioning with region-specific regularized linear policies can reduce the specified empirical inventory policy-cost measure in the studied setting. The study does not establish global optimality, statistical consistency, universal superiority, or external validity beyond the evaluated inventory environment. Index Terms—Adaptive partitioning; inventory policy learning; Ridge regression; feature-driven inventory; contextual inventory management; empirical inventory cost; data-driven inventory optimization.

—Adaptive partitioning; inventory policy learning; Ridge regression; feature-driven inventory; contextual inventory management; empirical inventory cost; data-driven inventory optimization.