<![CDATA[Stock price prediction remains challenging due to the nonlinear and dynamic behavior of financial markets, where price movements are influenced not only by historical market data but also by investor sentiment reflected in financial news. This study investigates the effectiveness of sentiment integration for stock price prediction by comparing FinBERT, a financial domain-specific language model, and IndoBERT, a language-specific Indonesian language model, within an LSTM-based forecasting framework. Experiments were conducted using daily stock price data and financial news collected for three Indonesian stocks (BBCA, BBRI, and BSDE) over the period 2019–2025. Sentiment scores extracted from financial news were aggregated on a daily basis and incorporated as additional features alongside historical price variables. Model performance was evaluated using MAE, RMSE, MAPE, R², and Directional Accuracy. The results indicate that sentiment-enhanced models do not consistently improve numerical forecasting accuracy compared with the baseline LSTM model. However, sentiment integration does not consistently improve directional prediction, although limited stock-specific differences are observed under certain stock-specific conditions. Comparative analysis further shows that neither FinBERT nor IndoBERT consistently outperforms the other across all datasets and metrics, suggesting that sentiment effectiveness is highly dependent on linguistic context and market characteristics. These findings highlight that sentiment information should be incorporated selectively rather than assumed to universally improve stock forecasting performance in emerging markets.]]>
