HYBRID MODELS FOR DEMAND FORECASTING IN INVENTORY MANAGEMENT: DESIGN, EXPERIMENTAL EVALUATION, AND INTEGRATION WITH ERP/WMS

Abstract

Under conditions of high demand volatility, traditional inventory management methods demonstrate insufficient accuracy due to nonlinear dynamics and complex external factors. The aim of this study is to develop, test, and empirically validate a hybrid demand forecasting system that integrates applied informatics tools (ARIMA, LSTM, ensemble methods) to solve complex forecasting and inventory optimization problems. Solving complex demand forecasting and inventory management tasks requires the application of modern applied informatics tools. The methodology includes a rolling backtest on data from a major logistics company (1,000 stock-keeping units, 36 months), optimization of model weights on a validation set, and a shift from point forecasts to a quantile-based inventory replenishment policy. The results show an 18–19% reduction in symmetric mean absolute percentage error (sMAPE), a 12% decrease in excess inventory, a 40% reduction in stockouts, and a 19,5% increase in inventory turnover. The scientific novelty lies in a formalized weighted aggregation scheme and the integration of uncertainty estimates into managerial decision-making. The practical value consists of the solution’s readiness for integration into ERP/WMS systems and its operational sustainability through MLOps practices.