<![CDATA[Financial news sentiment classification was important for understanding market narratives and investor perception, but Indonesian financial news remained challenging because it contained domain-specific terminology, numerical expressions, and imbalanced sentiment categories. This study benchmarked two transformer-based models, IndoBERT and Multilingual BERT, for classifying Indonesian financial news sentiment into negative, neutral, and positive classes. The dataset consisted of economic and financial news articles from Kontan, CNBC, and Bisnis.com during the first quarter of 2026. After preprocessing, 3,366 articles were used, consisting of 3,070 neutral, 184 negative, and 112 positive articles. The dataset was divided into training, validation, and testing sets using stratified splitting. Class weighting was applied to reduce the effect of class imbalance. The results showed that IndoBERT achieved the best overall performance, with 0.94 accuracy and 0.71 macro F1-score, while Multilingual BERT achieved 0.93 accuracy and 0.70 macro F1-score. These findings indicated that IndoBERT was more suitable for Indonesian financial news sentiment classification, although Multilingual BERT remained competitive, especially in detecting positive sentiment.]]>
Copyrights © 2026
