Data-driven adaptive demand classification: Integrating prediction performance and threshold expansion

Abstract

Demand classification is essential for improving forecasting and inventory management, particularly for products with intermittent and irregular demand patterns. However, traditional methods based on fixed thresholds, such as Average Demand Interval (ADI) and squared Coefficient of Variation (CV2), often struggle with borderline cases and typically ignore company-or industry-specific characteristics. As a result, they provide classifications that may work for average or standard environments but fail to capture the nuances of different operational contexts. This work presents an innovative, data-driven, adaptive approach that dynamically refines classification thresholds using prediction performance as a guiding metric. By optimizing the balance between forecast accuracy and material coverage, our method extends smooth demand classification to a broader range of items while keeping forecasting errors within acceptable limits. Validated on a large-scale warehouse dataset, the approach supports more responsive and resilient supply chain operations. These results demonstrate how adaptive classification mechanisms can enhance operational decision-making and advance the integration of predictive analytics into information systems for inventory optimization.