Analisis Sentimen Berita Keuangan Berbahasa Indonesia Menggunakan IndoBERT dan LSTM untuk Klasifikasi Tren Harga Saham pada Lima Emiten Blue Chip LQ45

Abstract

Blue chip stock price movements within the LQ45 index are non-linear and stochastic, causing conventional statistical approaches to fail in generating reliable predictions. This study develops and evaluates a hybrid deep learning architecture combining IndoBERT—a BERT-based language model pre-trained on Indonesian-language corpora—with Long Short-Term Memory (LSTM) networks to classify daily stock price trend direction (up/down). Trend labels are derived from daily return values: a return greater than zero is labeled Up (1), otherwise Down (0). Sentiment scores were extracted via IndoBERT fine-tuning from 9,819 Indonesian-language financial news articles (CNBC Indonesia) and merged with historical OHLCV data and technical indicators as model features. Experiments were conducted on five LQ45 blue chip equities: BBRI.JK, BBNI.JK, BBCA.JK, BMRI.JK (Financial Sector), and GOTO.JK (Technology). The hybrid model outperformed the baseline on three of five equities, with the highest accuracy improvement on BBCA.JK (+23.21 points, from 33.93% to 57.14%) and BMRI.JK (+12.50 points, from 46.43% to 58.93%). The overall average Relative Error Reduction (RER) reached +9.81%, demonstrating that integrating IndoBERT sentiment significantly enhances LSTM-based stock trend prediction in the Indonesian capital market.