Peramalan Harga Minyak Sawit Dunia Menggunakan Panel Data Spasial, Dekomposisi, dan Model Generatif

Ghufron, Muhammad Zakky (2025) Peramalan Harga Minyak Sawit Dunia Menggunakan Panel Data Spasial, Dekomposisi, dan Model Generatif. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025231089-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Minyak sawit memiliki peran penting dalam perekonomian Indonesia sebagai komoditas ekspor utama. Ketidakstabilan produksi dan stok minyak sawit yang dipengaruhi oleh faktor iklim, serta permintaan minyak sawit yang terus meningkat berdampak pada fluktuasi harga dunia. Penelitian ini bertujuan untuk mengembangkan model peramalan harga minyak sawit berdasarkan indikator ekonomi dan perubahan iklim dengan pendekatan berbasis panel data spasial, dekomposisi deret waktu dan arsitektur Transformer generative artificial intelligence. Penelitian ini mencakup beberapa tahap utama. Pertama, dilakukan imputasi nilai hilang pada data deret waktu dengan pengembangan metode GAIN menggunakan Gaussian Random Generator. Kedua, dilakukan analisis data panel untuk mengetahui pengaruh indikator ekonomi dan iklim terhadap fluktuasi harga. Ketiga, dikembangkan metode Seasonal-Adjusted Empirical Mode Decomposition (SAEMD) untuk ekstraksi pola tren dan musiman agar meningkatkan akurasi peramalan harga. Keempat, model FEDFormer dan metode peramalan lain seperti Random Forest (FR), Linear Regression (LR), dan Long-Short Term Memory (LSTM) digunakan untuk melakukan peramalan harga minyak sawit dalam jangka waktu 1, 3, 5, dan 7 tahun mendatang. Evaluasi model pada penelitian ini menggunakan beberapa metrik diantaranya Root Mean Square Error (RMSE), R-square (R2) Sliced Wasserstein Distance (SWD), Euclidean Distance (ED), Akaike Information Criterion (AIC), dan Bayesian Information Criterion (BIC). Hasil dari tahap pertama penelitian menunjukkan bahwa Gaussian-GAIN mampu memberikan hasil peramalan serta distribusi yang lebih baik dibanding GAIN biasa, yaitu dengan nilai SWD dan RMSE LSTM pada angka 1129,67 dan 818,45, sementara tanpa metode Gaussian didapatkan angka 1228,28 dan 953,60. Tahap kedua berhasil dilakukan regresi data panel menggunakan model FEM, yang mana diperoleh beberapa variabel yang berkorelasi positif seperti suhu maksimum, serta jumlah stok yang berkorelasi negatif terhadap harga sawit. Lalu pada tahap ketiga, penggunaan SAEMD dan LSTM dapat secara signifikan meningkatkan hasil peramalan sebesar 341,71 pada nilai RMSE, sementara dengan metode STL sebesar 437,95, dan 557,29 tanpa dekomposisi. Pada tahap terakhir, hasil peramalan menggunakan kombinasi SAEMD dan FEDFormer berhasil dilakukan, dan menghasilkan nilai RMSE paling baik daripada metode lain, yaitu sebesar 219,36.
=================================================================================================================================
Palm oil plays an important role in the Indonesian economy as a major export commodity. The instability of palm oil production and stocks that caused by climate factors, combined with the increasing demand for palm oil have an impact on world price fluctuations. This study aims to develop a palm oil price forecasting model based on economic indicators and climate change with a spatial data panel-based approach, time series decomposition and Transformer generative artificial intelligence architecture. This study includes several main stages. First, missing value imputation is carried out on time series data by developing the GAIN method using Gaussian Random Generator. Second, panel data analysis is carried out to determine the effect of economic and climate indicators on price fluctuations. Third, the Seasonal-Adjusted Empirical Mode Decomposition (SAEMD) method is developed to extract trend and seasonal patterns to improve price forecasting accuracy. Fourth, the FEDFormer model is used to forecast palm oil prices for the next 1, 3, 5, and 7 years. Model evaluation in this study uses several metrics including Root Mean Square Error (RMSE), R-square (R2) Sliced Wasserstein Distance (SWD), Euclidean Distance (ED), Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). The results from the first stage demonstrate that the Gaussian-GAIN method yields superior imputation and forecasting performance compared to standard GAIN, achieving SWD and LSTM RMSE values of 1129.67 and 818.45, respectively, versus 1228.28 and 953.60 without the Gaussian approach. In the second stage, panel data regression using the Fixed Effects Model (FEM) reveals several variables with significant correlations, such as maximum temperature showing a positive relationship, while stock levels exhibit a negative relationship with palm oil prices. The third stage shows that SAEMD with LSTM significantly improves forecasting results, reducing the RMSE to 341.71, compared to 437.95 using STL and 557.29 without decomposition. In the final stage, the combination of SAEMD and FEDFormer produces the best forecasting performance, achieving an RMSE of 219.36, outperforming other method

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Data Panel, Dekomposisi, Ekonomi, Minyak Sawit, Perubahan Iklim, Transformer
Subjects:	T Technology > T Technology (General) > T174 Technological forecasting T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Muhammad Zakky Ghufron
Date Deposited:	05 Aug 2025 02:31
Last Modified:	05 Aug 2025 02:31
URI:	http://repository.its.ac.id/id/eprint/126597

Actions (login required)

View Item