Development and Comparative Analysis of Demand Forecasting Models

Abstract

The petrochemical industry plays a vital role in the economy as a key supplier of raw materials for various industries, particularly plastic resin pellets used in manufacturing. However, economic fluctuations and the complexity of demand and supply make plastic resin pellets demand forecasting highly challenging. This study aims to develop and compare forecasting models for plastic resin pellets demand using a case study of a petrochemical manufacturer in Thailand. The product selected for this study was chosen based on the highest total sales during the study period. Data characteristics were examined using the Mann-Kendall Trend Test and the Friedman Test, while variables were selected using Mutual Information (MI). Seven forecasting models were developed and compared, including ARIMA, ARIMAX, XGBoost, and hybrid models comprising ARIMA-XGBoost, ARIMAX-XGBoost, XGBoost-ARIMA, and XGBoost-ARIMAX. Model performance was evaluated using the Mean Absolute Percentage Error (MAPE). The results indicate that, in this case study, hybrid models employing Machine Learning as the primary model tend to provide better forecasting performance than most traditional models. In particular, the XGBoost-ARIMAX model achieved the highest forecasting accuracy among all models considered.