Abstract
Accurate demand forecasting is essential for informed decision-making in today's dynamic business environment, where product demand often follows diverse and shifting patterns throughout increasingly shorter life cycles driven by continuous product innovation. This study aims to develop a forecasting framework capable of accurately predicting demand across varying patterns, with particular attention to the decline phase of the product life cycle. Traditional statistical forecasting methods, such as those in the ARIMA family, generally perform well with linear trends over short horizons, whereas machine learning techniques like XGBoost are better suited for capturing complex, nonlinear patterns over longer periods. This paper introduces an adaptive, hybrid forecasting framework that integrates ARIMA-based regression models with XGBoost using a weighted ensemble strategy. Initially, the framework tests linear models; if diagnostic analysis indicates nonlinearity, it incorporates XGBoost to address these complexities. To optimize the ensemble model performance, a grid search algorithm adjusts the ensemble weights by minimizing the root mean square error (RMSE), enabling the framework to dynamically leverage the strengths of both approaches. The proposed method was validated on five datasets representing different phases of the product life cycle. Results demonstrate that the proposed framework achieved MAPE below 13% on most datasets, with up to 80% improvement over ARIMA models in cases involving high variability demand patterns. The results show that the ensemble model enhances both flexibility and accuracy, especially for demand patterns that combine linear and nonlinear components. The framework benefits from the explainability and time-series capabilities of ARIMA while utilizing XGBoost's power to model nonlinear relationships. This research underscores the practical advantages of hybrid modeling in improving demand forecasting and operational planning across various industry sectors.