Modelling and Forecasting Fresh Agro-Food Commodity Consumption Per Capita in Malaysia Using Machine Learning
Rayner Alfred,Christylyn Leikson,Bonaventure Boniface,Geoffrey Harvey Tanakinjal,Assis Kamu,Mori Kogid,Stephen L. Sondoh,Nolila Mohd Nawi,Nalini Arumugam,Ryan Macdonell Andrias
DOI: https://doi.org/10.1155/2022/6106557
2022-11-13
Mobile Information Systems
Abstract:This study focuses on identifying and analyzing spending trend profiles and developing the per capita consumption models to forecast the fresh agro-food per capita consumption in Malaysia. Previous published works have looked at statistical and machine learning methods to forecast the demand of agro-food such as ARIMA and SVM methods. However, ordinary least squares (OLS) and neural network (NN) methods have shown better results in modelling time series data. For that reason, the primary objective of this study is to model and forecast the consumption per capita (PCC) of several selected fresh agro-food commodities in Malaysia using the OLS and NN methods. The secondary objectives of the paper include investigating the performance of OLS against NNs with three different topologies, discussing the correlation between Malaysia GDP per capita and the agro-food commodity PCC, and finally assessing whether the PCC data are increasing over time or decreasing over time and whether the trend in either direction is statistically significant by using the Mann–Kendall statistical test. Based on the results of the agro-food consumption per capita (PCC) forecasting, several critical agro-food commodities are also identified in this work. The material of the study consists of the per capita consumption of thirty-three (33) agro-food items that can be categorized into rice, livestock, vegetables, fisheries, and fruits, total gross domestic product (GDP) per capita, and the total population of Malaysia between 2010 and 2017. Based on the results obtained, the neural network () model was found to produce the lowest total MSE of 17.95, for all 33 fresh agro-food investigated in this study. Several agro-food commodities have been identified as having significant positive (e.g., rice, spinach, cabbage, celery cabbage, eggplant, cucumber, poultry, lamb, squid, tuna, star fruit, jackfruit, durian, sweet corn, and coconut) or negative (e.g., pork, mackerel, papaya, guava, mangosteen, pineapple, banana, rambutan, and watermelon) trends using the Mann–Kendall trend test. This study also demonstrated that the production of critical agro-food commodities (e.g., rice, chili, cabbage, celery cabbage, poultry (chicken/duck), beef, lamb, crab, mango, and coconut) should be improved to ensure self-sufficiency ratios (SSRs) of more than 100% to accommodate the increased projected consumption in Malaysia by the year 2025. This paper concludes that neural network methods produce better prediction, and future works include forecasting agro-food demand based on other independent variables such as weather conditions, disease outbreak, and stock market trends. There is a need to explore further the capability of ensemble models or hybrid models based on deep learning methods using multi-source data, as these have been shown to improve the performance of the base model. With these ensemble models combined with multi-source data, a more comprehensive analysis of the PCC can be obtained.
computer science, information systems,telecommunications