Financial news sentiment classification can support the extraction of market-relevant information from unstructured financial text. This study compares a general-purpose lexicon-based method, Valence Aware Dictionary and Sentiment Reasoner (VADER), with three supervised classifiers: Random Forest, Multinomial Naive Bayes, and linear Support Vector Machine (SVM). The experiments use 4,846 labelled statements from the Financial PhraseBank dataset to classify sentiment into positive, neutral, and negative categories. The supervised classifiers use term frequency–inverse document frequency (TF-IDF) features and an 80:20 train-test split, while VADER serves as a rule-based baseline. Based on the reported evaluation, SVM achieves the highest accuracy (0.76) and macro F1-score (0.69), followed by Random Forest, with an accuracy of 0.75 and a macro F1-score of 0.66. VADER yields the lowest reported performance, with an accuracy of 0.55 and a macro F1-score of 0.48. The confusion matrix indicates that positive and negative statements are frequently misclassified as neutral. These findings suggest that supervised classifiers provide stronger sentiment-classification performance than the general-purpose lexicon under the evaluated conditions. However, the results do not establish direct improvements in financial-risk forecasting, and further validation is required to assess generalisability across datasets and market conditions
Copyrights © 2026