Scenario-Oriented AI Methods for Supply Chain Demand Forecasting

Abstract

Demand forecasting is a critical function within supply chain management. Although artificial intelligence-based forecasting methods have been explored in academic research, their practical application remains limited. Rather than identifying the algorithm with the highest average accuracy, this study examines the business contexts in which each forecasting method is most appropriate. The limitations of traditional time series and regression approaches, particularly their inability to address high-dimensional, heterogeneous, or rapidly evolving demand patterns at reasonable scaling costs, are reviewed. Subsequently, machine learning, deep learning, and hybrid models are introduced. A scenario-matching framework is developed to classify these methods based on forecasting horizon, decision level, data characteristics, and management requirements, moving beyond simple enumeration. Choice rules are established for varying levels of data quality, volatility, and interpretability within the framework's operationalisation. Four persistent challenges are identified: data quality, unexpected events, generalisation, and interpretability. These issues are interconnected and collectively highlight the need for a long-term forecasting system that incorporates robust modelling, a resilient data pipeline, monitoring for distributional shifts, and an interpretable human interface. The conclusion outlines practical implications and directions for future research.