Peramalan Nilai Kurs USD Menggunakan Model Artificial Neural Network Berbasis Fitur ARIMA-EGARCH dengan Mempertimbangkan Variabel Eksogen

Siregar, Pangeran Fajar Halomoan (2026) Peramalan Nilai Kurs USD Menggunakan Model Artificial Neural Network Berbasis Fitur ARIMA-EGARCH dengan Mempertimbangkan Variabel Eksogen. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221101-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Nilai tukar Rupiah terhadap Dolar AS (Kurs USD) memiliki tingkat volatilitas tinggi yang sulit direpresentasikan secara utuh oleh pemodelan linier parametrik konvensional. Penelitian ini bertujuan untuk memodelkan dinamika Kurs USD menggunakan metode Fungsi Transfer dengan harga minyak mentah Brent sebagai variabel eksogen, serta mengatasi asimetri volatilitas melalui pendekatan hibrida ARIMA EGARCH-Neural Network (NN). Analisis awal pada data harian mengindikasikan perlunya penanganan observasi yang hilang menggunakan algoritma Kalman Filter, serta penerapan transformasi Box-Cox dan diferensiasi orde pertama untuk mencapai kestabilan data. Hasil estimasi Fungsi Transfer menunjukkan bahwa parameter eksogen harga Brent tidak memberikan pengaruh linier yang signifikan secara statistik, sehingga arah penelitian dialihkan sepenuhnya pada pendekatan univariat. Model rata-rata ARIMA([1,6],1,[6]) menyisakan efek heteroskedastisitas yang kemudian berhasil diatasi oleh model varians EGARCH(2,2) dengan spesifikasi distribusi Skewed Student-t. Guna mengakomodasi sisa galat (noise) non-linier, arsitektur hibrida dibangun dengan memanfaatkan nilai aktual Kurs USD, estimasi (fitted) ARIMA, dan varians kondisional EGARCH sebagai fitur masukan pada jaringan NN. Melalui optimasi hyperparameter (window size 20, 64 unit neuron, tingkat dropout 0,1, fungsi aktivasi RELU, dan optimizer ADAM), evaluasi peramalan out-of-sample menggunakan mekanisme direct recursive forecasting menghasilkan kinerja yang sangat optimal. Model ini menghasilkan tingkat kesalahan Root Mean Square Error (RMSE) sebesar Rp 59,01 dan Mean Absolute Percentage Error (MAPE) 0,265%. Tingkat akurasi model terbukti semakin presisi pada horizon peramalan jangka pendek (2 minggu) dengan RMSE mencapai Rp 17,64 dan MAPE 0,09%. Arsitektur hibrida ini terbukti secara empiris mampu mengestimasi volatilitas asimetris dengan sangat akurat dan mengungguli kinerja pemodelan parametrik tunggal.
=======================================================================================================================================
The exchange rate of the Indonesian Rupiah against the US Dollar (USD) exhibits high volatility that is difficult to be fully represented by conventional parametric linear modeling. This study aims to model the USD exchange rate dynamics using the Transfer Function method with Brent crude oil price as an exogenous variable, and to overcome volatility asymmetry through a hybrid ARIMA EGARCH-Neural Network (NN) approach. Initial analysis of the daily data indicated the need for missing observation handling using the Kalman Filter algorithm, along with Box-Cox transformation and first-order differencing to achieve data stability. The Transfer Function estimation results showed that the exogenous parameters of the Brent price provided no statistically significant linear effect, thus the research direction shifted entirely to a univariate approach. The ARIMA([1,6],1,[6]) mean model left heteroscedasticity effects, which were successfully addressed by the EGARCH(2,2) variance model with a Skewed Student-t distribution specification. To accommodate the remaining non-linear noise, a hybrid architecture was constructed utilizing the actual USD values, ARIMA fitted estimations, and EGARCH conditional variance as input features for the NN. Through hyperparameter optimization (window size 20, 64 neuron units, dropout rate 0.1, RELU activation, and ADAM optimizer), the out-of-sample forecasting evaluation using a direct recursive forecasting mechanism yielded highly optimal performance. The model produced a Root Mean Square Error (RMSE) of Rp 59.01 and a Mean Absolute Percentage Error (MAPE) of 0.265%. The accuracy of the model proved increasingly precise over a short-term forecasting horizon (2 weeks), achieving an RMSE of Rp 17.64 and a MAPE of 0.09%. This hybrid architecture is empirically proven to estimate asymmetric volatility with high accuracy and outperforms individual parametric modeling.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	ARIMA-EGARCH, Fungsi Transfer, Harga Minyak Brent, Kurs USD, Neural Network, Peramalan, ARIMA-EGARCH, Brent Oil Price, Forecasting, Neural Network, Transfer Function, USD Exchange Rate
Subjects:	H Social Sciences > HA Statistics > HA30.3 Time-series analysis Q Science > QA Mathematics > QA280 Box-Jenkins forecasting Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Pangeran Fajar Halomoan Siregar
Date Deposited:	05 Aug 2026 03:33
Last Modified:	05 Aug 2026 03:33
URI:	http://repository.its.ac.id/id/eprint/143761

Actions (login required)

View Item