Financial Sentiment Classification Using Traditional Machine Learning, BERT, FinBERT, and Zero-Shot FLAN-T5

Financial sentiment classification is a challenging natural language processing task because financial text often combines specialized terminology, implicit sentiment, numerical information, and context that cannot always be captured by simple lexical patterns. This study presents a controlled comparison of five approaches for three-class financial sentiment classification: TF-IDF with Logistic Regression, TF-IDF with a Linear Support Vector Machine (SVM), BERT, FinBERT, and zero-shot FLAN-T5. The experiments use the Financial PhraseBank dataset, which contains 4,846 financial statements labeled as negative, neutral, or positive. The principal experiment uses a stratified 70/15/15 train-validation-test split with random seed 42, and all five models are evaluated on the same test partition. Logistic Regression obtains 73.31% accuracy and a 63.35% macro F1 score, while Linear SVM obtains 74.69% and 68.15%, respectively. BERT reaches 85.01% accuracy and 83.18% macro F1, whereas FinBERT achieves 88.72% accuracy and 87.43% macro F1. Zero-shot FLAN-T5 records 37.69% accuracy and 36.15% macro F1. To provide additional statistical context, bootstrap confidence intervals and exact McNemar tests are applied to the fixed test predictions. A separate three-seed robustness experiment is also conducted for FinBERT using seeds 42, 123, and 2024. Across these runs, FinBERT obtains a mean accuracy of 85.50% with a standard deviation of 1.53 percentage points and a mean macro F1 of 83.26% with a standard deviation of 2.33 percentage points. This robustness analysis is intentionally reported separately from the principal five-model comparison because equivalent multi-seed experiments were not completed for the other models. Overall, the results indicate that financial-domain language modeling can provide a meaningful advantage for this financial sentiment classification setting under the stated experimental protocol.