Efficient Financial Sentiment Analysis via LLM Fine-Tuning with Preference Optimization

Abstract

Large Language Models (LLMs) show promise for financial text analysis, but their size can be prohibitive, and general models may lack domain-specific nuance. This paper introduces an efficient and effective framework for advanced financial sentiment analysis by fine-Tuning a LLM, Qwen2.5-7b, using Low-Rank Adaptation (LoRA). We make two primary contributions to enhance model performance: (1) a novel zero-shot Chain-of-Thought (CoT) prompting strategy tailored for financial reasoning, which guides the model to produce more accurate and interpretable sentiment classifications along with descriptive explanations; and (2) the application of Direct Preference Optimization (DPO) using responses from a more capable LLM as positive examples to further refine output quality. Our model not only classifies sentiment but also provides nuanced descriptions, offering richer insights for financial decision-making. We demonstrate the practical utility of these sentiment signals by integrating them into a long-short portfolio strategy, evaluated using standard financial metrics. Experiments show our approach achieves strong performance compared to traditional baselines and pre-fine-Tuned models, highlighting the potential of domain-Adapted LLMs for sophisticated financial applications. Ablation studies confirm the significant contributions of both CoT and DPO components.