Optimizing AI-Driven Replenishment Policies Using Response Surface Methodology: A Multi-Echelon Simulation Study
AI doesn't always translate to better inventory economics, even if the demand forecasts it generates are better, if the rules engineered behind replenishment models are still "legacy" and based on legacy planning conditions. This paper proposes a sequential response surface methodology (RSM) protocol to provide an approach for policy tuning after forecasting in a four-echelon Saudi fast-moving consumer goods (FMCG) simulation. FMCG Co. is a generic placeholder rather than a registered company, and all case data are synthetic, generated using realistic assumptions for the Saudi FMCG sector. The analysis is based on 1,248 weekly demand observations for the three regions and the four SKU families, a static LightGBM forecasting engine, and a periodic-review replenishment policy. These five controllable factors were screened using a Resolution V fractional factorial design, steepest descent, a three-factor central composite design, quadratic response modelling, constrained optimization, Derringer-Suich desirability, and held-back confirmation. For the cost-led response-surface search, safety-stock multiplier, review period and demand-signal smoothing were kept. Total cost showed a significant center-point curvature (F = 80.49, p = .0029) and the fitted quadratic models accounted for over 99.9% of the variance in total cost, fill rate, and bullwhip with no significant lack-of-fit tests. The strict model-based constrained solution (k = 1.775, R = 5.095 days, alpha = 0.343) predicted an annual cost of SAR 46.146 million, a 96.000% fill rate, and a bullwhip ratio of 1.595. The desirability solution with a balance of the three factors (k = 2.253, R = 3.520 days, alpha = 0.3635) predicted better service and stability at higher cost. A previous candidate, based on cost, was rejected as infeasible, while the only directly confirmed feasible candidate was able to deliver 97.605% fill and 1.546 bullwhip at SAR 49.706 million. The results indicate that forecast modernization and replenishment-policy modernization are two different interventions and that the policy re-tuning is important to translate the forecasting capability into system-level performance while maintaining interpretability and awareness of constraints.