PERBANDINGAN MODEL ARMA-GARCH DAN LSTM MULTILAYER UNTUK PERAMALAN HARGA SAHAM ICBP.JK

Abstract

Peramalan harga saham merupakan aspek penting dalam pengambilan keputusan investasi karena pergerakan harga saham bersifat fluktuatif dan volatil. Penelitian ini bertujuan membandingkan kinerja model ARMA–GARCH dan model deep learning LSTM multilayer dalam meramalkan harga saham PT Indofood CBP Sukses Makmur Tbk (ICBP.JK). Tahapan pemodelan ARMA–GARCH dilakukan melalui proses uji stasioneritas, identifikasi ordo melalui ACF/PACF, serta estimasi parameter. Sementara itu, LSTM multilayer dibangun melalui normalisasi data dan optimasi hyperparameter. Evaluasi performa model dilakukan dengan kriteria Mean Absolute Percentage Error (MAPE). Hasil penelitian menunjukkan bahwa model ARMA(1,1)–GARCH(1,1) menghasilkan MAPE sebesar 4.49%, sedangkan LSTM multilayer menghasilkan MAPE sebesar 1.22%. Meskipun kedua model dikategorikan memiliki kinerja sangat baik, LSTM multilayer terbukti lebih unggul dalam menangkap tren jangka panjang. Hal ini menunjukkan bahwa pendekatan deep learning lebih efektif untuk memodelkan data keuangan yang memiliki volatilitas yang tinggi.

Stock price forecasting is a crucial element in investment decision-making due to the volatile and fluctuating nature of stock price movements. This study aims to compare the performance of the classical ARMA–GARCH model and the deep learning multilayer LSTM model in forecasting stock prices of PT Indofood CBP Sukses Makmur Tbk (ICBP.JK). The ARMA–GARCH modeling process involves stationarity testing, order identification through ACF/PACF plots, and parameter estimation. Meanwhile, the multilayer LSTM model is constructed through data normalization and hyperparameter optimization using the grid search method. Model performance was evaluated using the Mean Absolute Percentage Error (MAPE) criterion. The results indicate that the best-fitting ARMA(1,1)–GARCH(1,1) model produced a MAPE of 4.49%, whereas the multilayer LSTM model achieved a lower MAPE of 1.22%. Although both models are categorized as having "very good" forecasting performance, the multilayer LSTM proved superior in capturing long-term trends. These findings suggest that the deep learning approach is more effective for modeling financial data characterized by high volatility.