Operationalising Large Language Models for Financial Sentiment Analysis: The PipeFinePT Modular Pipeline

Background: Financial decision-making increasingly relies on automated analysis of large volumes of textual information.However, financial sentiment analysis resources remain scarce for low-resource linguistic contexts such as European Portuguese, limiting reproducible evaluation and adoption of advanced AI techniques within financial information systems.Objective: This study aims to design and evaluate a modular pipeline capable of operationalizing large language models (LLMs) for sentiment classification of Portuguese financial news while assessing performance, robustness, and deployment trade-offs across different model categories.Methods: A domain-specific dataset of 200 European Portuguese financial news articles was constructed and manually annotated by three independent annotators using a positive, neutral, and negative sentiment scheme.The proposed PipeFinePT pipeline integrates data acquisition, preprocessing, prompt-based annotation, and standardized evaluation.Multiple proprietary and openweights LLMs were evaluated using identical prompts and experimental conditions, with performance measured through accuracy, precision, recall, and F1-score metrics, including robustness testing via reversed input ordering.Results: Results show that LLMs can perform financial sentiment classification in European Portuguese with moderate reliability, achieving up to 0.781 accuracy, although pairwise differences among the three models with valid outputs were not statistically significant.Higher-tier proprietary models such as GPT-4.1 demonstrated stronger semantic interpretation but at substantially higher inference cost, whereas lighter proprietary variants and the open-weight LLaMA-4-Maverick-17B offered improved efficiency with reduced contextual sensitivity.Robustness experiments revealed generally stable model rankings but performance variability under altered input ordering. Conclusion:The PipeFinePT pipeline provides a reproducible framework for integrating LLM-based sentiment analysis into applied financial information systems.The findings highlight both the feasibility and limitations of deploying LLMs in low-resource financial language environments and support their use as operational analytical components in multilingual financial analytics workflows.