Optimizing multi-tier supply chain ordering with LNN+XGBoost: mitigating the bullwhip effect

Purpose Supply chain ordering must balance profit, customer service and upstream order variability (the bullwhip effect). This study evaluates whether a two-stage pipeline–a Liquid Neural Network (LNN) encoder with an adaptive-leak recurrent cell followed by an XGBoost readout–offers a useful operational trade-off for multi-tier ordering, rather than assuming hybrid superiority. Design/methodology/approach An LNN encoder represents temporal demand dynamics and XGBoost regresses its frozen latent states; the encoder is explicitly distinguished from canonical Liquid Time-Constant and closed-form continuous-time networks. Under a leakage-free protocol, hyperparameters are selected on a separate development seed and frozen before testing on untouched days from ten unseen demand seeds. A matched linear-readout ablation, paired statistics with multiplicity correction, an observed-order DataCo replay, cost sensitivity analysis and an end-to-end latency benchmark are used. Findings Transformer ($1.738M), XGBoost ($1.737M) and LNN+XGBoost ($1.737M) attain nearly identical mean system profit, with no Holm-adjusted difference between the hybrid and either baseline. Relative to the exact-matched linear readout, LNN+XGBoost gains $14,477 (95% CI $4,811–$24,143; d_z = 1.07), although the family-wise Holm result (p = 0.072) does not indicate a statistically significant difference. Its manufacturer demand-referenced variance ratio is 2.61 versus 2.44 (XGBoost) and 2.48 (Transformer), with a fill rate of 0.996. The Q-network minimizes variance (1.43) but sacrifices profit and service, the DataCo replay favors standalone XGBoost and the hybrid is not latency-optimal. Originality/value The study provides the first leakage-free, ablation-controlled evaluation of an adaptive-leak-LNN–XGBoost ordering pipeline and characterizes a profit–service–bullwhip trade-off with explicit operating conditions rather than claiming universal superiority.