A collaborative analytics framework for artificial intelligence-based demand forecasting in supply chains

Despite substantial advances in artificial intelligence (AI), forecasting systems in practice often remain anchored in traditional, siloed approaches, leaving managers with unreliable and inconsistent signals for decision-making. Such forecasts tend to exacerbate the bullwhip effect, inflate safety stocks and generate significant inefficiencies in working capital and logistics costs. In fast-moving consumer goods (FMCG) supply chains, forecasts that fail to reach a minimum level of reliability are of limited value for operational planning, whereas sufficiently accurate forecasts provide a dependable foundation for managerial decisions. This practice-oriented study examines whether joint AI-based forecasting, combining downstream retailer data with advanced machine-learning techniques, enables firms to consistently achieve forecast reliability levels that are suitable for planning. Drawing on a unique dataset from a Central European manufacturer-retailer partnership, we compare three forecasting scenarios: a Manufacturer baseline forecast based on traditional statistical methods (ARIMA) applied to shipment data; a Manufacturer AI forecast using machine-learning models (XGBoost) trained exclusively on shipment data; and a Joint AI forecast in which machine-learning models (XGBoost) are trained on retailer warehouse outbound data. The analysis shows that Joint AI reduces total absolute error by 44.7% relative to Manufacturer AI on the test-data basis (cluster-bootstrap 95% CI: 27.8–57.3%). Under the stricter common- demand benchmark, the reduction relative to Manufacturer AI remains statistically robust at 21.2% (95% CI: 4.4–35.7%). The advantage over the traditional Manufacturer baseline is positive in point estimate but not statistically established on the common-demand benchmark. Taken together, the results indicate that, within the scope of the dyad and the two algorithmic approaches examined, collaborative data access shifts forecast errors most where their operational and financial consequences are largest, concentrating the gains of AI-based fore casting in the high-volume core of the product portfolio.