Artificial Intelligence in Supply Chain Demand Forecasting: A Structured Review of Method Evolution, Context Dependence and Research Gaps

Abstract

Demand forecasting is at the starting point of the supply chain planning link, and the forecast deviation is amplified along the chain to form a bullwhip effect, causing the backlog and shortage to occur at the same time. Machine learning and deep learning methods have entered this field in large numbers in recent years, but the literature is scattered across settings as different as retail, medicine, fresh produce and spare parts. This article is based on OpenAlex structured search to review: from 2015 to 2026, there are 2,970 records, and after qualifying the sources of journals and conferences, there are 2,227 records, of which 1,210 contain artificial intelligence method terms, and 18 representative studies are deeply synthesized. The results are threefold. The method genealogy progresses along statistical methods, tree models, deep sequence models, attention and hybrid architecture, and 91.9% of attention literature appears in 2024 and later. The statistical method has not been withdrawn. 71.9% of its literature also comes from the past three years, and it remains the main control benchmark. The accuracy advantage of artificial intelligence depends on the demand form. It is stable in continuous high-density scenarios, while the evidence in intermittent demand and truncated observation scenarios is weak. Cold start, explainability and end-to-end evaluation from forecast to inventory remain three open gaps.