<![CDATA[Palm oil production forecasting in Kalimantan was important because production patterns differed across regions and changed over time. However, previous forecasting approaches often relied on limited classical models and did not sufficiently compare statistical, machine learning, and deep learning methods on regional panel time-series data. This study compared several forecasting models for palm oil production in Kalimantan using annual district and city-level data from 2010 to 2024. The dataset was transformed into panel time-series format, consisting of 56 regional series and 840 observations. Data from 2010 to 2022 were used for training, while data from 2023 to 2024 were used for testing. The evaluated models included Naive Forecast, Autoregressive Integrated Moving Average, Holt Linear Trend, Light Gradient Boosting Machine, Extreme Gradient Boosting, Long Short-Term Memory, Gated Recurrent Unit, and Neural Basis Expansion Analysis for Time Series. The results showed that Long Short-Term Memory achieved the best performance based on root mean squared error and coefficient of determination, while Naive Forecast performed best based on absolute and percentage error metrics. These findings indicated that deep learning was effective for reducing large prediction errors, but simple forecasting remained competitive for stable regional production patterns.]]>
Copyrights © 2026
