Social media platforms have become key sources of real-time investor sentiment, influencing market movements in Indonesia's stock market. However, accurate sentiment extraction in Bahasa Indonesia remains challenging due to informal language and evolving financial terminology. This paper presents a comparative study of sentiment analysis models applied to Indonesian stock market tweets. We evaluate a traditional SVM model using TF-IDF features against two variants of the transformer-based IndoBERT model (base and large) on the ID-SMSA dataset. IndoBERT-base achieves the highest accuracy (97.72%), significantly outperforming SVM (90.22 %). A novel Out-of-Vocabulary (OOV) evaluation systematically quantifies vocabulary coverage impacts on model performance, revealing that IndoBERT's sub word tokenization offers superior robustness to domain-specific vocabulary compared to TF-IDF-based approaches. These results demonstrate the effectiveness of con-textual language models in financial sentiment classification tasks. © 2025 IEEE.