Financial Statement Analysis with Large Language Models

A recent study has found that large language models (LLMs) like GPT-4 can outperform professional financial analysts in predicting the direction of future earnings

The paper investigates the capabilities of large language models (LLMs), specifically GPT-4, to perform financial statement analysis traditionally done by professional analysts. It discusses whether LLMs can predict the future earnings of companies purely based on numerical financial statements, without narrative context. The study finds that LLMs outperform human analysts and other machine learning models in predicting earnings changes, potentially revolutionizing the role of LLMs in financial decision-making.

Key Findings:

Performance of LLMs: When provided with anonymized and standardized financial statements, LLMs were able to predict future earnings directions more accurately than human financial analysts and a comparable state-of-the-art machine learning model.

Chain of Thought (CoT) Prompting: LLMs performed better when using CoT prompting, which mimics human analytical processes, suggesting that the step-by-step reasoning enhances prediction accuracy.

Comparative Advantage: LLMs have a relative advantage in situations where human analysts typically struggle, such as when dealing with complex, ambiguous data devoid of contextual clues.

Economic Insights: Despite not being trained specifically on financial tasks, LLMs generate useful economic insights from the data, indicating a broader potential applicability across various domains of financial analysis.

Implications for Trading: Trading strategies based on predictions from LLMs showed higher returns, demonstrating the practical financial implications of integrating LLMs into investment decision processes.

Methodology:

The study involved feeding GPT-4 standardized forms of balance sheets and income statements and instructing it to analyze and predict the direction of future earnings without any contextual information.

Both simple and CoT prompts were tested to evaluate the model's performance under different instructional conditions.

The performance was benchmarked against human analysts' predictions and other machine learning models using historical data from the Compustat database covering multiple decades.

Conclusion:

The paper concludes that LLMs like GPT-4 can effectively analyze financial statements and predict earnings directions with high accuracy. This capability surpasses that of traditional human analysts in many cases, suggesting a transformative potential for LLMs in financial analysis and decision-making. The findings advocate for the integration of LLMs into financial practices, where they can complement and enhance human expertise, particularly in interpreting large volumes of financial data efficiently.

Source: Financial Statement Analysis with Large Language Models

Alex G. Kim1 Maximilian Muhn2 Valeri V. Nikolaev

May 20, 2024