Machine Learning Applications for Fraud Detection and Financial Sentiment

This review examines machine learning applications in two major financial tasks: fraud detection and financial sentiment analysis. For fraud detection, research demonstrates that ensemble models, particularly gradient boosting methods such as CatBoost, XGBoost, and LightGBM, achieve high performance in identifying fraudulent transactions under extreme class imbalance, with reported F1 scores above 0.90. Deep learning approaches, including recurrent and convolutional architectures, further enhance detection by modeling sequential transaction behavior, while graph neural networks capture relational structures within transaction networks. In financial sentiment analysis, studies show that transformer-based models such as FinBERT and fine-tuned GPT variants substantially improve classification accuracy compared with traditional methods. Domain-specific corpora, including FiQA and Financial PhraseBank, provide benchmarks for polarity classification and market impact analysis. Emerging multimodal and federated learning frameworks integrate text, speech, and distributed training to address data privacy and modality heterogeneity. The review highlights observed methodological trends, comparative performance metrics, and dataset characteristics, providing an evidence-based synthesis of academic findings up to 2025.