Machine Learning-Based Sentiment Classification for Detecting Financial Misinformation on Social Media

Abstract

Financial misinformation on social media is difficult to identify because misleading posts often combine emotionally charged language, incomplete evidence, promotional claims, and technically plausible financial terminology. This conceptual technical review examines how machine-learning sentiment classification can be incorporated into a broader detection pipeline for financial misinformation. The review synthesizes literature on financial-domain language models, deception detection, sentiment analysis, feature engineering, and human-in-the-loop verification. It proposes a modular architecture consisting of data acquisition, text normalization, financial sentiment representation, misinformation-risk classification, explanation, and reviewer escalation. The analysis shows that sentiment is useful as a contextual signal but is insufficient as a stand-alone indicator because legitimate financial communication may also be strongly positive or negative. Robust systems should combine contextual embeddings, source and propagation features, claim-evidence consistency, calibration, and explainable outputs. Evaluation should report precision, recall, F1-score, area under the curve, calibration error, and performance across topics and time periods. The proposed architecture provides an engineering-oriented foundation for developing transparent financial misinformation screening systems without treating automated classification as a substitute for professional verification.