Domain-Aware Preprocessing Improves Financial Headline Sentiment Classification Across TF-IDF and Transformer Architectures

Abstract

This paper investigates the interaction between text preprocessing strategies and feature representation in financial headline sentiment classification. Using the Financial News Headline dataset comprising 10,686 entity-annotated financial news headlines, we evaluate three preprocessing pipelines of increasing domain specificity across six traditional machine learning configurations and a fine-tuned FinBERT transformer model. Our central finding is that domain-aware preprocessing has a representation-dependent effect: it improves TF-IDF-based classifiers by 2.63 percentage points in macro F1 but degrades FinBERT by 2.38 percentage points. This occurs because TF-IDF representations permanently lose sentiment signal when stopwords are removed, whereas FinBERT's contextual attention mechanism already captures the semantic relationships that domain-aware preprocessing manually preserves. Standard NLP stopword removal degrades TF-IDF classification by 2.18 percentage points because the default NLTK English stopword list removes directional words (up, down), negation words (not, no), and magnitude words (more, less) that carry critical sentiment signal in financial text. A stacking ensemble of four diverse classifier families achieves the best traditional ML macro F1 of 0.7727 (95% CI: 0.754-0.790), while fine-tuned FinBERT on raw headlines achieves 0.8328 (95% CI: 0.816-0.849). SMOTE-ENN synthetic oversampling proves counterproductive on sparse TF-IDF features, reducing macro F1 by 5.69 percentage points compared to class weighting alternatives. These results provide actionable guidance for practitioners selecting preprocessing strategies, model architectures, and imbalance handling techniques for financial sentiment classification systems.