Lumpy Demand Forecasting Based on Prophet-ML Hybrid Model

This study proposes a machine learning-based hybrid forecasting approach for lumpy demand, which is characterized by intermittency and high variability. The proposed approach adopts Prophet as a baseline forecasting model and improves its predictions through a residual correction mechanism that explicitly accounts for the sign of forecast errors. Residuals are classified by their direction, and separate regression models are trained to correct over- and under-forecasting errors. Experimental results on two datasets show that it achieves substantial improvements under cost structures where over-forecasting penalties are dominant, particularly in comparison with baseline forecasting models. These findings indicate that the proposed hybrid model corrects redisuals in a direction that effectively mitigates the increase in inventory holding costs associated with over-forecasting in lumpy demand environments. Sensitivity analysis further demonstrates that the proposed method provides stable performance across different correction coefficient settings, highlighting its robustness in cost-sensitive inventory environment.