Synergizing transformer-based models and financial sentiment analysis: A framework for generative AI in economic decision-making

Abstract

This study introduces a new way to analyze financial sentiment by combining advanced transformer-based models with generative artificial intelligence (AI) to better understand the language and context of financial discussions. The objective is to enhance the predictive accuracy of market behavior through improved understanding of investor sentiment. The proposed sentiment analysis framework leverages six domain-specific datasets: Social Sentiment Indices (X-Scores), Fin-SoMe, SemEval-2017 Task 5, Fin-Lin, Sanders, and Taborda. These datasets, primarily sourced from social media, reflect diverse investor perspectives. Generative AI models, like GPT-3.5 and GPT-4, are used to create more data, and the meaning of words is enhanced using techniques like BERT and Word2Vec. The model is trained with a cross-entropy loss function and fine-tuned using Few-shot Learning, Chain-of-Thought reasoning, and ReAct strategies, ensuring computational efficiency. Experimental results show consistent improvements across all datasets in accuracy, precision, recall, specificity, and F1 score. The use of generative AI and transformer architectures makes the model stronger and better at understanding how investors feel in real financial situations. This research contributes to the field of explicable AI in finance by demonstrating the impact of domain-adapted models and generative techniques in advancing sentiment analysis. The findings offer practical value for investors and analysts seeking data-driven insights into market dynamics and decision-making processes.