COMPARATIVE ANALYSIS OF ARTIFICIAL INTELLIGENCE METHODS AND ALGORITHMS FOR INVENTORY DEMAND FORECASTING

Abstract

Accurate demand forecasting is essential for inventory availability, working-capital efficiency and service continuity, yet retail demand is affected by promotions, seasonality, intermittent purchases and changing market conditions. This article presents a critical, narrative comparison of statistical, machine-learning and deep-learning approaches to inventory demand forecasting. ARIMA and intermittent-demand methods are assessed alongside Random Forest, XGBoost, LightGBM, long short-term memory networks, DeepAR and Temporal Fusion Transformers. The review examines modelling assumptions, information requirements, computational burden, interpretability, forecast uncertainty and relevance to replenishment decisions. Evidence from the M5 forecasting competition and methodological research indicates that no algorithm is uniformly superior: performance depends on the structure of the series, the availability of explanatory variables and the evaluation design. Tree-based ensembles offer a practical option for feature-rich retail data, while probabilistic neural models can support distribution-aware inventory decisions when sufficient related series are available. A conceptual hybrid forecasting framework and an evaluation protocol connecting forecast accuracy to inventory costs are proposed. The paper reports a literature-based synthesis rather than newly conducted experiments.