Improving Supply Chain and Logistics: Predicting Demand and Optimizing Inventory with Historical Sales Data

Abstract

Abstract Supply chains increasingly operate under volatile demand, uncertain replenishment lead times, promotional shocks, and disruptions that make point-forecast-based inventory policies difficult to manage. This article develops and empirically evaluates an integrated framework that links historical point-of-sale demand forecasting with probabilistic, multi-echelon inventory optimization. The study compares six forecasting approaches Simple Moving Average, Holt-Winters Exponential Smoothing, ARIMA, XGBoost, LSTM, and a stacked ensemble using 156 weeks of weekly sales data covering 450 stock-keeping units and three distribution nodes. Models are evaluated on a 13-week holdout using MAE, RMSE, MAPE, and R², with formal Diebold–Mariano tests used to assess predictive differences. The best-performing ensemble combines XGBoost, LSTM, and Holt-Winters outputs through a ridge-regularized meta-learner. Its aggregate MAPE is 11.6%, compared with 24.1% for ARIMA, while its MAE and RMSE are also lowest. The study then converts forecast uncertainty into lead-time demand distributions through residual bootstrapping and uses these distributions to parameterize a dynamic continuous-review (s, S) policy. A discrete-event simulation with 500 Monte Carlo replications shows that the ensemble-driven policy reduces total simulated supply-chain cost from $4.286 million to $2.918 million, improves realized fill rate from 89.4% to 96.7%, and reduces warehouse-to-supplier bullwhip amplification from 3.02 to 1.78.