Klasifikasi Jual/Beli Saham Berdasarkan Data Historis dan Analisis Sentimen Berita Menggunakan Deep Learning

Hasanat, Ichlasul (2024) Klasifikasi Jual/Beli Saham Berdasarkan Data Historis dan Analisis Sentimen Berita Menggunakan Deep Learning. Diploma thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201091-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (5MB)

Abstract

Pasar saham merupakan tempat yang menarik bagi para investor dan peneliti. Pergerakannya dikendalikan oleh banyak faktor di seluruh dunia, baik dari informasi publik, kondisi sosial, kebijakan makroekonomi, ataupun kondisi politik dari suatu negara. Dengan banyaknya faktor tersebut, tentu berita-berita mengenai saham dapat dijadikan suatu fitur yang kuat untuk mempertimbangkan keputusan dalam proses jual ataupun beli saham. Fitur tersebut didapatkan dengan mengidentifikasi sentimen dari berita tersebut. Baik itu positif, ataupun negative. Dalam hal ini, digunakan metode analisis sentimen untuk menentukan sentimen berita saham pada website CNBC Indonesia, Bisnis, dan Kontan.
Pada penelitian ini, digunakan model transformer yang telah di-pretrained seperti FinBERT, FinancialBERT, dan FinBERT-tone untuk menilai sentimen berita saham. Untuk mendukung model tersebut, digunakan fitur indikator teknikal saham seperti Moving Average, Moving Average Convergence Divergence (MACD), Bollinger Bands, dan Relative Strength Index (RSI). Kedua dataset tersebut kemudian digabungkan dengan menyesuaikan index dan dilakukan proses training menggunakan metode deep learning dan ensemble learning berupa Voting Classifier. Kedua metode ini mengacu pada empat jenis model Recurrent Neural Network (RNN), yaitu Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), dan Bidirectional Gated Recurrent Unit (Bi-GRU).
Model Voting Classifier menggunakan semua metode deep learning dengan memanfaatkan semua fitur indikator teknikal dan label sentimen meghasilkan nilai akurasi terbaik saat diuji menggunakan strategi emiten BBRI, yaitu dengan strategi target 50 hari, 4 label klasifikasi, dan sliding window 30 hari. Akurasi yang didapat menggunakan strategi ini mencapai 0.9422. Model ini juga mampu menghasilkan rata-rata akurasi sebesar 0.8549 jika diuji satu persatu pada semua emiten saham. Jika model ini diuji dengan sektor yang sama dan sektor yang berbeda, yaitu dengan sektor finansial dan sektor bahan dasar, rata-rata akurasi yang didapatkan adalah sebesar 0.8734 dan 0.8365.
============================================================
The stock market is a fascinating place for investors and researchers. Its movements are controlled by many factors around the world, be it public information, social conditions, macroeconomic policies, or political conditions of a country. With so many factors, news about stocks can be used as a powerful feature to consider decisions in the process of buying or selling stocks. The feature is obtained by identifying the sentiment of the news. Be it positive, or negative. In this case, the sentiment analysis method is used to determine the sentiment of stock news on the CNBC Indonesia, Bisnis, and Kontan websites.
In this research, the sentiment analysis method will use pretrained transformer models such as FinBERT, FinancialBERT, and FinBERT-tone, to assess the sentiment of stock news. Then to support the model, stock technical indicator features such as Moving Average, Moving Average Convergence Divergence (MACD), Bollinger Bands, and Relative Strength Index (RSI) are used. The two datasets are then combined by adjusting the index and the training process is carried out using deep learning and ensemble learning methods in the form of Voting Classifier. Both methods refer to four types of Recurrent Neural Network (RNN) models, namely Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), and Bidirectional Gated Recurrent Unit (Bi-GRU).
The Voting Classifier model using all deep learning methods by utilizing all technical indicator features and sentiment labels produces the best accuracy value when tested using the BBRI issuer strategy, namely with a 50-day target strategy, 4 classification labels, and a 30-day sliding window. The accuracy obtained using this strategy reached 0.9422. This model is also able to produce an average accuracy of 0.8549 when tested one by one on all data. If the model is tested with the same sector and different sectors, namely the financial sector and the basic materials sector, the average accuracy obtained is 0.8734 and 0.8365.

Item Type:	Thesis (Diploma)
Uncontrolled Keywords:	Analisis Sentimen, Berita Saham, Deep Learning, Ensemble Learning, Indikator Teknikal, Transformer ============================================================ Deep Learning, Ensemble Learning, Sentiment Analysis, Stock News, Technical Indicators, Transformer
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.9.D343 Data mining. Querying (Computer science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Ichlasul Hasanat
Date Deposited:	01 Aug 2024 06:26
Last Modified:	01 Aug 2024 06:26
URI:	http://repository.its.ac.id/id/eprint/110625

Actions (login required)

View Item