AI-Driven Predictive Analytics for Intermittent Demand Forecasting in Wholesale Distribution: A Cost-Based Benchmarking Study
DOI:
https://doi.org/10.63125/kncce104Keywords:
Intermittent demand, Croston method, Machine learning, Neural networks, Inventory cost, Forecast evaluation, Wholesale distribution, Systematic reviewAbstract
Wholesale distributors carry tens of thousands of stock-keeping units whose demand is intermittent: most periods record no sales and the remainder record quantities that vary widely. Forecasting such demand decides stock investment and service, yet the literature judges forecasting methods mostly by accuracy statistics that are known to misrank methods for inventory purposes. This systematic review synthesises the literature published before 2020 on classical, artificial-intelligence (AI) and machine-learning (ML) methods for intermittent demand as a literature-based benchmark organised by the type of evaluation used. A reproducible search of the Crossref index with 28 strings returned 5,500 records (2,946 unique); 2,288 titles and 379 abstracts were screened against pre-registered criteria and 232 studies were included and coded with a 40-field form covering context, data, methods, comparative design, accuracy metrics, inventory and cost measures, demand classification and ten reporting items; an independent 10% re-screening estimated consistency. The literature is recent, spread over many venues and concentrated in spare-parts settings; Croston-type estimators remain the reference methods, neural networks are the dominant AI family, and tree ensembles and deep networks appear only at the end of the period. 15% of the studies evaluate forecasts with an inventory or cost measure, and among the studies that compare AI/ML with classical methods, the family reported best differs between accuracy-based and cost-based evaluation in 1 of the 3 cases that report both. Wholesale distribution is explicitly studied in 5 studies. A cost-based benchmarking protocol for a wholesale distributor is derived from the evidence: SKU classification, a classical reference set, rolling-origin evaluation, inventory simulation under a stated policy with holding and shortage costs, and a reporting checklist. The review is limited by single-reviewer, abstract-based screening and by one bibliographic index.


