Využitelnost LLM modelů při tvorbě finančních analýz

The development of artificial intelligence, and specifically large language models, has opened new opportunities for the fields of financial analysis and planning. Given the vast volume of data these models can process, they now compete with traditional manual analysis of financial statements. The goal of this bachelor’s thesis is to evaluate the effectiveness of large language models in predicting the financial statements of automotive companies. The thesis begins with concepts of intelligence and then covers the key architectural milestones of large language models. In the practical section, an automated system was developed to extract structured data from financial statements. Using the GPT 5.4 API, the system generated predictions of financial statements (balance sheets, income statements, and cash flow statements) for the past three fiscal years of selected automotive companies. In total, 1,170 financial statements were generated, and 390 ex-post comparisons of predictions against actual values were performed. The accuracy of the system was evaluated using MAPE, normalized mean error (nME) and MASE metrics. The results demonstrate moderate to low accuracy (median MAPE ranging from 10% to 60%) – depending on the automotive company. The model achieved the best results in relatively stable balance sheet items – assets, liabilities, equity, and book value per share. The worst results were free cash flow and operating cash flow. The experiment was conducted on fully anonymized financial statements without timestamps. The thesis discusses the limitations of LLMs and their potential use in an agentic system.