A financial agent for fundamental analysis: an empirical investigation in the Brazilian stock market

Large Language Models (LLMs) have recently enabled agentic systems for complex decision-making, including financial trading. However, existing Finance-AI research has primarily emphasized short-term signals such as sentiment and price prediction, leaving the ability of LLMs to perform Fundamental Analysis, a core, structured reasoning task, largely unexplored. In this work, we study whether LLMs can reliably compute fundamental financial indicators from real-world financial reports. We formulate Fundamental Analysis as a deterministic numerical reasoning task grounded in long-form financial documents, requiring the computation of 32 predefined indicators spanning accounting, valuation, profitability, leverage, and efficiency. We evaluate reasoning-enabled and non-reasoning LLMs at two model scales under workflow-based and agentic tool-augmented architectures, and analyze the impact of reflection mechanisms on numerical accuracy and robustness. Models are assessed using Normalized Mean Absolute Error (NMAE) and token-efficiency metrics. We further examine downstream decision quality by integrating model outputs into an LLM-based investment agent. Experiments on companies from the Brazilian stock market provide insights into LLM reasoning, grounding, and generalization beyond U.S.-centric settings.