Fine-tuned IndoBERT for stock market sentiment analysis: evidence from CNBC Indonesia news

Abstract

Financial sentiment analysis in Indonesian markets faces significant accuracy challenges, with existing models achieving only 78-81% accuracy. We present a fine-tuned IndoBERT-Large model for classifying sentiment in Indonesian stock market news headlines, trained on 9,819 CNBC Indonesia headlines (January 2024-March 2025). Through systematic hyperparameter optimization and stratified vocabulary-balanced splitting, our model achieved 94.20% accuracy, surpassing previous baselines by 4-16 percentage points. These results demonstrate IndoBERT's effectiveness for Indonesian financial NLP and its potential for real-time market monitoring and investment decision support systems.