Hybrid Multi Input Transfer Function-Long Short Term Memory untuk Peramalan Inflasi Menggunakan Variabel Makroekonomi dan Indikator Digital

Damayanti, Riza (2026) Hybrid Multi Input Transfer Function-Long Short Term Memory untuk Peramalan Inflasi Menggunakan Variabel Makroekonomi dan Indikator Digital. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003241009-Master_Thesis.pdf Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Inflasi merupakan salah satu indikator makroekonomi yang berperan penting dalam menjaga stabilitas perekonomian suatu negara. Peramalan inflasi yang akurat diperlukan sebagai dasar dalam pengambilan kebijakan moneter dan perencanaan ekonomi. Namun, karakteristik inflasi yang dipengaruhi oleh hubungan linear maupun nonlinear menyebabkan model statistik konvensional belum mampu memberikan hasil peramalan yang optimal. Penelitian ini bertujuan untuk mengimplementasikan model Hybrid Multi Input Transfer Function Long Short-Term Memory (MITF–LSTM) dalam peramalan inflasi Indonesia menggunakan variabel makroekonomi dan narasi digital sebagai variabel eksogen, membangun model Multi Input Transfer Function (MITF), serta mengevaluasi peningkatan akurasi model hybrid dibandingkan model MITF standalone. Data yang digunakan berupa data bulanan periode Januari 2005 hingga Juli 2025 yang terdiri atas inflasi, nilai tukar, BI Rate, harga minyak dunia, Google Trends, dan Indeks Ekspektasi Harga (IEH). Tahapan penelitian meliputi pemodelan ARIMA pada variabel eksogen, pembentukan model MITF, ekstraksi residual, pemodelan residual menggunakan Baseline LSTM, Tuned LSTM, dan Bagging LSTM, kemudian menggabungkan hasil peramalan MITF dan LSTM menjadi model Hybrid MITF–LSTM. Evaluasi model dilakukan menggunakan Root Mean Square Error (RMSE), serta didukung oleh scatter plot aktual terhadap peramalan dan koefisien determinasi (R^2). Hasil penelitian menunjukkan bahwa model MITF berhasil mengidentifikasi empat variabel eksogen yang berpengaruh terhadap inflasi, yaitu nilai tukar, harga minyak dunia, Google Trends, dan Indeks Ekspektasi Harga (IEH), sedangkan BI Rate tidak dimasukkan ke dalam model karena tidak menunjukkan hubungan dinamis berdasarkan analisis Cross-Correlation Function (CCF). Model Hybrid MITF-LSTM mampu meningkatkan akurasi peramalan dibandingkan model MITF standalone. Berdasarkan indikator RMSE, model hybrid menghasilkan penurunan kesalahan peramalan hingga 24,07% pada horizon 6 bulan, 35,34% pada horizon 12 bulan, dan sekitar 12% pada horizon 24 bulan. Hasil scatter plot dan nilai koefisien determinasi (R^2) juga menunjukkan bahwa model hybrid menghasilkan peramalan yang lebih mendekati nilai aktual dibandingkan model MITF. Dengan demikian, Hybrid MITF-LSTM terbukti mampu meningkatkan akurasi peramalan inflasi Indonesia melalui kombinasi kemampuan MITF dalam memodelkan hubungan linear dan LSTM dalam menangkap pola nonlinear pada residual.
========================================================================================================================
====================
Inflation is one of the key macroeconomic indicators that plays an important role in maintaining a country's economic stability. Accurate inflation forecasting is essential for supporting monetary policy decisions and economic planning. However, the coexistence of linear and nonlinear patterns in inflation dynamics limits the forecasting performance of conventional statistical models. This study aims to implement a Hybrid Multi Input Transfer Function Long Short-Term Memory (MITF-LSTM) model for forecasting Indonesian inflation using macroeconomic and digital narrative variables as exogenous inputs, develop a Multi Input Transfer Function (MITF) model, and evaluate the forecasting performance of the hybrid model compared with the standalone MITF model. The study employed monthly data from January 2005 to July 2025, consisting of inflation, exchange rate, BI Rate, world oil prices, Google Trends, and the Inflation Expectation Indeks (IEH). The proposed methodology consisted of ARIMA modeling for exogenous variables, MITF model construction, residual extraction, residual modeling using Baseline LSTM, Tuned LSTM, and Bagging LSTM, followed by combining MITF predictions with LSTM residual predictions to form the Hybrid MITF-LSTM model. Model performance was evaluated using Root Mean Square Error (RMSE), and supported by actual-versus-predicted scatter plots and the coefficient of determination (R^2). The results indicate that the MITF model successfully identified four influential exogenous variables, namely the exchange rate, world oil prices, Google Trends, and the Inflation Expectation Indeks (IEH), while the BI Rate was excluded because it did not exhibit a significant dynamic relationship based on the Cross-Correlation Function (CCF) analysis. The Hybrid MITF–LSTM model consistently outperformed the standalone MITF model in terms of forecasting accuracy. Based on RMSE, the hybrid model reduced forecasting errors by up to 24.07% for the 6-month horizon, 35.34% for the 12-month horizon, and approximately 12% for the 24-month horizon. Furthermore, the scatter plots and the coefficient of determination (R^2) demonstrated that the hybrid model generated predictions that were closer to the actual inflation values than those produced by the MITF model. These findings indicate that the Hybrid MITF-LSTM model effectively improves Indonesian inflation forecasting by combining the strengths of MITF in modeling linear relationships and LSTM in capturing nonlinear patterns contained in the model residuals.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Inflasi, Multi Input Transfer Function, Long Short-Term Memory, Hybrid MITF-LSTM, Cross-Correlation Function.
Subjects:	Q Science
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Riza Damayanti
Date Deposited:	05 Aug 2026 02:04
Last Modified:	05 Aug 2026 02:04
URI:	http://repository.its.ac.id/id/eprint/143291

Actions (login required)

View Item