A Comparative Study of Univariate and Multivariate Time Series Forecasting for CPO Prices Using Machine Learning Techniques

Abstract

The Malaysian palm oil sector has significantly contributed to developing the domestic economy and the global palm oil market. However, the fluctuation in Crude Palm Oil (CPO) prices poses a significant risk to farmers, producers, traders, consumers, and others involved in CPO production and marketing. An accurate CPO price forecasting technique is required to aid decision-making in risky and unpredictable scenarios. Hence, this project aims to compare the performances of four-time series forecasting models, including Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and CNN-LSTM, in the context of univariate and multivariate analysis for CPO prices in Malaysia. This research methodology is based on five phases: research understanding, data understanding, data preparation, modeling, and evaluation. Monthly CPO prices, the production and export volume of CPO, selected vegetable oil prices, crude oil prices, and monthly exchange rate data from January 2009 to December 2022 were utilized. The metrics evaluation of Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) were then performed to compare and evaluate the performance of the models. Experimental analysis indicates that the CNN model trained on a multivariate dataset with carefully selected significant independent variables outperformed other models. With a configuration of 500 epochs and early stopping, it achieved remarkable results compared to models trained using a univariate approach, boasting an RMSE of 245.611 and a MAPE of 7.13.