Hybrid Time Series Forecasting for Semiconductor Demand: A Comparative Analysis of Prophet-Based Ensemble Models for Smooth and Erratic Demand Patterns

In this study, we compare various hybrid time series forecasting models to improve prediction accuracy for semiconductor product demand. Our investigation examines two contrasting demand patterns observed in the industry: smooth demand (regular occurrences with modest variability) and erratic demand (significant variability despite occurring at generally consistent intervals). The analysis draws from actual semiconductor product data collected between 2009 and 2022. We benchmarked conventional time series approaches—specifically ARIMA and Facebook Prophet—against our proposed hybrid frameworks that integrate Prophet with advanced machine learning techniques (Random Forest and XGBoost). The empirical results indicate that our Prophet-Random Forest hybrid architecture substantially outperforms traditional methods, delivering a 30.5% accuracy improvement for smooth demand forecasting and a 7.2% enhancement when predicting erratic demand patterns. The proposed approach offers a practical alternative to computationally intensive foundation models, achieving substantial forecasting improvements while maintaining the interpretability and operational efficiency required for semiconductor manufacturing environments.