Adaptive demand forecasting framework with weighted ensemble of regression and machine learning models along life...

Area of Science:

• Business Analytics
• Machine Learning
• Operations Research

Background:

• Accurate demand forecasting is critical for business decision-making in dynamic markets with short product life cycles.
• Traditional methods like ARIMA excel at linear trends, while machine learning (XGBoost) handles complex, nonlinear patterns.

Purpose of the Study:

• To develop an adaptive, hybrid forecasting framework for accurate demand prediction across diverse product life cycle phases.
• To specifically address challenges in forecasting demand during the product decline phase.

Main Methods:

• An adaptive, hybrid framework integrating ARIMA and XGBoost models using a weighted ensemble strategy.
• Dynamic model selection: ARIMA for linear trends, incorporating XGBoost for nonlinearities detected via diagnostic analysis.
• Grid search optimization to determine optimal ensemble weights by minimizing root mean square error (RMSE).

Main Results:

• The hybrid framework achieved a Mean Absolute Percentage Error (MAPE) below 13% on most datasets.
• Demonstrated up to 80% improvement over traditional ARIMA models for high-variability demand patterns.
• The ensemble model showed enhanced flexibility and accuracy, particularly for demand patterns with combined linear and nonlinear components.

Conclusions:

• Hybrid modeling offers practical advantages for improving demand forecasting accuracy and operational planning.
• The framework effectively balances ARIMA's explainability and time-series capabilities with XGBoost's nonlinear modeling power.
• This approach is beneficial across various industry sectors requiring robust demand prediction.