FinBERT-FOMC: Fine-Tuned FinBERT Model with Sentiment Focus Method for Enhancing Sentiment Analysis of FOMC Minutes

Sandro Gössi, Ziwei Chen, Wonseong Kim, Bernhard Bermeitinger, Siegfried Handschuh

Abstract: In this research project, we used the financial texts published by the Federal Open Market Committee (FOMC), known as the FOMC Minutes, for sentiment analysis. The pre-trained FinBERT model, a state-of-the-art transformer-based model trained for NLP tasks in finance, was utilized for that. The focus of this research has been on improving the predictive performance of complex financial sentences, as our problem analysis has shown that such sentences pose a significant challenge to existing models. To accomplish this objective the original FinBERT model was fine-tuned for domain-specific sentiment analysis. A strategy, referred to as Sentiment Focus (SF) was utilized to reduce the complexity of sentences, making them more amenable to accurate sentiment predictions.To evaluate the efficacy of our method, we curated a manually labeled test dataset comprising 1375 entries. The results demonstrated an overall improvement of in accuracy when using SF-enhanced fine-tuned FinBERT over the original FinBERT model. In cases of complex sentences containing conjunctions like but, while, and though with contradicting sentiments, our fine-tuned model outperformed the original FinBERT by a margin of .