Thailand’s Maize Prices Forecasting Using Ensemble Technique

Abstract

Crop price forecasting is crucial for farmers, policymakers, and investors. This paper aims to propose suitable machine learning models for forecasting Thailand’s maize prices by implementing and comparing various machine learning algorithms, including regression trees (RT), support vector regression (SVR), ensemble bagging with RT and SVR as the base learner (Bag-RT and Bag-SVR), and random forest (RF). The dataset used in this study is collected from two main sources: the Office of Agricultural Economics in Thailand (OAE) and the investing.com website from January 2002 to August 2023, consisting of 260 records and 53 features. Given the dataset's numerous independent variables, we applied the recursive feature elimination combined with the Pearson correlation feature selection method to reduce feature dimensions by focusing on the most relevant features. The prediction models were trained using 10-fold cross-validation and evaluated using three metrics: R-squared (R2), mean absolute error (MAE), and root mean square error (RMSE). The top-performing model, Bag-SVR, achieved the best R2 value of 0.961, MAE of 0.234, and RMSE of 0.315, followed by the SVR model with R2 value of 0.959, MAE of 0.251, and RMSE of 0.333. In contrast, the RT model demonstrated the lowest performance scores with an R2 value of 0.846, MAE of 0.44, and RMSE of 0.617. In conclusion, our study emphasizes the influence of feature selection on model performance and showcases the potential of machine learning models for accurate maize price forecasting in Thailand.