Peramalan Harga Saham Menggunakan Metode Hybrid Bidirectional Long Short Term Memory Dan Multi Head Attention

Rinaufaldi, Ahmad Fauzan (2026) Peramalan Harga Saham Menggunakan Metode Hybrid Bidirectional Long Short Term Memory Dan Multi Head Attention. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026221143-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Pasar modal memiliki peran vital bagi pembangunan ekonomi suatu negara karena kemampuannya yang menyediakan sumber pembiayaan bagi pembangunan nasional. Aset penting pasar modal yang kian diminati ialah saham, karena kemampuannya dalam memberikan keuntungan melalui capital gain dan dividen bagi para investor. Salah satu saham blue chip di Indonesia adalah PT Bank Central Asia TBK (BBCA), karena memiliki reputasi baik dan kapitalisasi pasar terbesar untuk bank swasta di Indonesia, sehingga menjadi pilihan investor sebagai instrumen investasi yang relatif stabil. Akan tetapi, harga saham sangat dipengaruhi oleh volatilitas pasar dan pengaruh ekonomi makro seperti pergerakan Indeks Harga Saham Gabungan (IHSG) pada pasar saham Indonesia, akibatnya menjadi sulit untuk diprediksi secara akurat menggunakan metode konvensional. Oleh karena itu, diperlukan model peramalan harga saham yang mampu menangkap pola kompleks harga saham agar investor dapat mengambil keputusan yang akurat. Tugas akhir ini mengembangkan model peramalan harga saham BBCA menggunakan pendekatan hybrid Bidirectional Long Short-Term (BiLSTM) dan Multi-Head Attention (MHA), yang menggabungkan kemampuan BiLSTM dalam memahami pola temporal dua arah serta mekanisme Multi-Head Attention dalam menyoroti fitur penting pada data time series, dengan pendekatan multivariat dua variabel input menggunakan data IHSG sebagai variabel eksternal. Model dengan hasil terbaik pada penelitian ini diperoleh dari BiLSTM dengan proporsi data 70:30, menggunakan konfigurasi LSTM Units sebesar 128, learning rate 0,005, dan batch size 32, yang menghasilkan nilai MAPE sebesar 1,07267%. Namun, berdasarkan rata-rata keseluruhan eksperimen, model BiLSTM-MHA menunjukkan performa yang lebih konsisten dengan nilai mean MAPE sebesar 1,19070%, lebih baik dibandingkan BiLSTM sebesar 1,24851%. Hal ini menunjukkan bahwa penambahan mekanisme Multi-Head Attention tidak secara signifikan meningkatkan performa terbaik, namun mampu meningkatkan stabilitas prediksi. Meskipun demikian, peningkatan konsistensi tersebut diikuti oleh konsekuensi berupa waktu komputasi yang lebih tinggi. Oleh karena itu, pemilihan model terbaik bergantung pada kebutuhan pengguna, apakah lebih mengutamakan efisiensi komputasi atau stabilitas performa prediksi.
================================================================================================================================
The capital market plays a vital role in a country's economic development due to its ability to provide financing for national development. Stocks are an increasingly popular capital market asset due to their ability to provide profits through capital gains and dividends for investors. One of the blue chip stocks in Indonesia is PT Bank Central Asia TBK (BBCA), due to its good reputation and the largest market capitalization for a private bank in Indonesia, making it a relatively stable investment instrument for investors. However, stock prices are heavily influenced by market volatility and macroeconomic influences such as the movement of the Jakarta Stock Exchange Composite (JKSE) on the Indonesian stock market, making it difficult to accurately predict using conventional methods. Therefore, a stock price forecasting model capable of capturing complex stock price patterns is needed so that investors can make accurate decisions. This final project develops a BBCA stock price forecasting model using a hybrid Bidirectional Long Short-Term (BiLSTM) and Multi-Head Attention (MHA) approach, which combines the capabilities of BiLSTM in understanding two-way temporal patterns and the Multi-Head Attention mechanism in highlighting important features in time series data, with a multivariate approach with two input variables using JKSE data as an external variable. The model with the best results in this study was obtained from BiLSTM with a data proportion of 70:30, using a configuration of LSTM Units of 128, a learning rate of 0,005, and a batch size of 32, which produced a MAPE value of 1,07267%. However, based on the overall average of the experiment, the BiLSTM-MHA model showed a more consistent performance with a mean MAPE value of 1,19070%, better than BiLSTM at 1,24851%. This shows that the addition of the Multi-Head Attention mechanism does not significantly improve the best performance, but is able to improve the stability of predictions. However, this increase in consistency is followed by the consequence of higher computation time. Therefore, the selection of the best model depends on the user's needs, whether prioritizing computational efficiency or stability of prediction performance.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Bidirectional Long Short-Term Memory, Multi-Head Attention, Multivariat, Peramalan, Saham, Forecasting, Multivariate, Stock.
Subjects:	H Social Sciences > HB Economic Theory > Economic forecasting--Mathematical models. T Technology > T Technology (General) > T174 Technological forecasting
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Ahmad Fauzan Rinaufaldi
Date Deposited:	21 Jul 2026 04:08
Last Modified:	21 Jul 2026 04:08
URI:	http://repository.its.ac.id/id/eprint/135743

Actions (login required)

View Item