Pengembangan Model Prediksi Harga Pasar Pupuk dengan Metode Statistik dan Machine Learning

Rachmadi, Hilmi Zharfan (2024) Pengembangan Model Prediksi Harga Pasar Pupuk dengan Metode Statistik dan Machine Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201268-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (6MB) | Request a copy

Abstract

Industri pupuk global telah menghadapi volatilitas harga yang signifikan selama lebih dari satu dekade terakhir. Fluktuasi harga bahan baku, dinamika geopolitik, kebijakan ekspor, dan gangguan rantai pasokan menjadi faktor-faktor eksternal yang memengaruhi harga pupuk. Rusia, sebagai eksportir pupuk terbesar di dunia, berperan penting sebagai pemasok lima jenis pupuk: potassium chloride (KCL), rock phosphate (RP), diammonium phosphate (DAP), urea, dan triple superphosphate (TSP). Konflik geopolitik yang melibatkan Rusia telah menghambat pasokan pupuk-pupuk tersebut, sehingga meningkatkan fluktuasi harga. Dalam kondisi ini, PT X, salah satu produsen pupuk terbesar di Asia Tenggara, menghadapi tantangan dalam memprediksi harga pasar pupuk dengan akurat sebagai bahan pertimbangan dalam menetapkan harga jual yang kompetitif setiap bulan. Penelitian ini mengimplementasikan model prediksi harga pasar pupuk DAP, KCL, RP, TSP, dan urea menggunakan model statistik AutoARIMA dan model machine learning XGBoost. Kedua model digunakan untuk memprediksi harga kelima jenis pupuk selama 3, 6, dan 12 bulan mendatang. Hasil prediksi dievaluasi menggunakan metrik kesalahan Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), dan Mean Absolute Percentage Error (MAPE). Untuk meminimalkan metrik kesalahan, dilakukan eksplorasi terkait pemrosesan dan transformasi data yang meliputi penanganan data pencilan dengan imputasi median, power transform, differencing, scaling, dan pembuatan variabel lag. Setelah uji coba pada kelima dataset harga pupuk, model XGBoost menghasilkan metrik kesalahan yang lebih rendah dibandingkan AutoARIMA dalam sebagian besar skenario, baik untuk rentang prediksi pendek, maupun panjang. Kedua model menunjukkan peningkatan metrik kesalahan seiring bertambahnya rentang waktu prediksi dengan AutoARIMA mengalami peningkatan yang lebih signifikan. Berdasarkan kriteria MAPE, model XGBoost konsisten menghasilkan prediksi yang tergolong "Sangat Baik" untuk semua jenis pupuk dan rentang waktu, sedangkan performa AutoARIMA bervariasi tergantung jenis pupuk dan rentang waktu prediksi. Keunggulan XGBoost mungkin disebabkan oleh kemampuannya dalam menangani hubungan dan pola yang kompleks serta nonlinear dalam data harga pupuk. Meskipun demikian, AutoARIMA masih menunjukkan potensi untuk prediksi jangka pendek, terutama untuk pupuk RP yang cenderung lebih mudah diprediksi. Parameter terkait imputasi data pencilan dan orde differencing menjadi dua parameter terpenting untuk model XGBoost, sedangkan parameter imputasi data dan power transform menjadi yang terpenting untuk AutoARIMA. Menariknya, penggunaan data pelatihan tanpa imputasi data pencilan lebih sering menghasilkan metrik kesalahan terendah. Hal ini mengindikasikan bahwa data harga pupuk mengandung kenaikan eksponensial yang membuat penerapan imputasi data pencilan cenderung berdampak negatif terhadap akurasi prediksi.
======================================================================================================================================
The global fertilizer industry has experienced significant price volatility over the past decade. Factors such as raw material price fluctuations, geopolitical dynamics, export policies, and supply chain disruptions influence fertilizer prices. Russia, the world's largest fertilizer exporter, plays a crucial role in supplying five types of fertilizers: potassium chloride (KCL), rock phosphate (RP), diammonium phosphate (DAP), urea, and triple superphosphate (TSP). Geopolitical conflicts involving Russia have disrupted the supply of these fertilizers, leading to increased price volatility. In this context, PT X, one of the largest fertilizer producers in Southeast Asia, faces the challenge of accurately predicting fertilizer market prices to help set competitive selling prices each month. This research implements a market price prediction model for DAP, KCL, RP, TSP, and urea fertilizers using AutoARIMA and XGBoost, representing a statistical model and a machine learning model. Both models predict the prices of these fertilizers for the next 3, 6, and 12 months. The prediction results were evaluated using MAE, RMSE, and MAPE error metrics. To minimize these errors, various data processing and transformation techniques were explored, including handling outliers with median imputation, power transform, differencing, scaling, and generating lag variables. After testing on all five fertilizer price datasets, the XGBoost model produced lower error metrics than AutoARIMA in most scenarios for both short and long prediction ranges. Both models showed increased error metrics as the prediction time span increased, with AutoARIMA experiencing a more significant rise. Based on the MAPE criterion, the XGBoost model consistently produced predictions classified as "Excellent" for all fertilizer types and timescales, while AutoARIMA's performance varied depending on the fertilizer type and prediction timescale. The superiority of XGBoost may be due to its ability to handle complex and nonlinear relationships and patterns in fertilizer price data. Nonetheless, AutoARIMA still shows potential for short-term predictions, especially for RP fertilizers, which tend to be more predictable. Parameters related to outlier data imputation and differencing order are the most important for the XGBoost model, while data imputation and power transform parameters are the most crucial for AutoARIMA. Interestingly, using training data without outlier data imputation often resulted in the lowest error metric. This might suggest that the fertilizer price data contains an exponential increase, making the application of outlier data imputation likely to negatively impact prediction accuracy

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Multi-step time series forecasting, ARIMA, XGBoost, prediksi harga pasar pupuk, Multi-step time series forecasting, ARIMA, XGBoost, fertilizer market price forecasting
Subjects:	S Agriculture > S Agriculture (General) > S633.5 Fertilizers 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 > 55201-(S1) Undergraduate Thesis
Depositing User:	Hilmi Zharfan Rachmadi
Date Deposited:	01 Aug 2024 03:17
Last Modified:	17 Sep 2024 03:57
URI:	http://repository.its.ac.id/id/eprint/111468

Actions (login required)

View Item