Forecasting demand for new products is a significant challenge in supply chain management due to the paucity of historical data, which hinders the practical application of classic stock adjustment to order volatility. The paper proposes a machine learning-based dynamic forecasting approach that integrates internal and external feature variables to construct a preliminary forecasting model supported by a rolling compensation system to adjust for inaccuracies. The approach involves the following steps. First, a new product is categorized using a classification model based on existing product properties to facilitate the application of the most relevant preliminary fore-casting pattern. Then, during the post-launch time, residual discrepancies between actual orders and forecasted orders are dynamically adjusted using an ARMA-based compensation model. Residual patterns are detected using a run chart to facilitate the continued adaptation of models to changing patterns. An approach is applied to tray data in a cooperative effort between academia and industry. The experimental results demonstrate that the approach reduces prediction error by approximately 18.7% compared to classic models and achieves better adaptation to market patterns in a shorter period. The work enhances new product demand forecasting accuracy and facilitates real-time monitoring of multiple product groups, offering the manufacturing industry a forward-looking decision support system.

Authors:
Chien-Chih Wang, Ming Chi University of Technology, Taiwan
Che-Yu Hung, Ming Chi University of Technology, Taiwan

About the Presenter(s)
Dr. Chien-Chih Wang, Professor of Industrial Engineering and Management at MCUT, specializes in data analysis, Lean Six Sigma, and machine vision. His current project focuses on AI-driven quality inspection for smart manufacturing.

See this presentation on the full schedule – Friday Schedule

Bookmark