Artificial intelligence may soon have a continued presence in the financial services industry. According to forthcoming research from the University of Chicago Booth School of Business, Large Language Models (LLMs), a type of AI trained in understanding and generating content, can outperform certain financial analysts in predicting future earnings. The researchers shared their preliminary findings in an unrevised draft. By using chain-of-thought prompting, which breaks down complex reasoning tasks into smaller steps, these models, particularly generative pre-trained transformers (GPTs), have shown an accuracy of 60. 4%. This is 7 percentage points higher than the average prediction of analysts, according to the study. What makes this remarkable is that the language models were only provided with the balance sheet and income statement, without any additional narratives or context. The study discovered that with simple instructions, the models could analyze financial statements and make predictions on future earnings comparable to the consensus forecasts of analysts in their first month. The researchers wrote that these results indicate the potential for GPT to perform financial statement analysis better than human analysts, even without specific narrative contexts. They emphasized the importance of human-like step-by-step analysis, which helps the model follow the typical steps performed by analysts. The report also found that the language model's forecasts were more valuable when human biases or inefficiencies, such as disagreements, were present. However, like humans, GPT's predictions were not perfect. Accuracy was lower for smaller firms, firms with higher leverage ratios, firms recording losses, or firms with volatile earnings.

In such cases, the context tends to matter more, and predictions can be less accurate. While both GPT and analysts struggle with predictions for smaller or loss-making firms, analysts are generally better equipped to handle complex financial circumstances. Analysts incorporate soft information and external context beyond financial statements, which contributes to their expertise. The report concluded that the potential of LLMs, such as GPT, to democratize financial information processing is noteworthy and should interest investors and regulators. Language models can become more than just a tool for investors; they could actively participate in decision-making. However, the authors cautioned that AI performance may differ in real-world scenarios. It remains to be seen whether AI can significantly enhance human decision-making in financial markets. The authors stated that GPT and human analysts are complementary rather than substitutes.