Does Explicit Reasoning Help in Finance? A Study of Chain-of-Thought for Financial NLP

Reasoning capability is widely considered a hallmark of large language models (LLMs), and Chain-of-Thought (CoT) prompting has been extensively adopted to enhance model performance. However, its effectiveness in specialized, high-stakes domains such as finance remains underexplored. In this study, we systematically evaluate the practical utility of explicit reasoning across multiple financial tasks. We compare a diverse set of open source and commercial LLMs using two inference modes, Direct Prediction and Explicit Reasoning, on three financial tasks: Financial Named Entity Recognition (NER), Financial News Headline Classification, and Financial Sentiment Analysis (SA). Our results reveal that explicit reasoning does not consistently improve performance in financial settings. It yields limited or no gains for headline classification and sentiment analysis, and substantially degrades accuracy in NER, primarily due to misalignment with task-specific annotation conventions and domain constraints. In contrast, our zero-shot prompting with carefully crafted instructions achieves significant and consistent improvements through effective in-context task alignment. These findings indicate that, for financial applications, precise task specification and conservative inference strategies are more critical than explicit reasoning, offering practical guidance for building efficient and reliable financial AI systems.