Researchers from the University of Chicago decided to test GPT-4’s capabilities in financial analysis, and they didn’t just pick any ordinary analysis. They went straight to the heart of the matter: predicting companies’ future earnings from their financial statements!
Typically, this is expert-level work reserved for seasoned financial analysts who can dissect balance sheets, calculate key ratios, and interpret trends. In short, it’s complex, and machines weren’t really expected to compete with the intuition and experience of the pros.
Except… surprise! The study results are astonishing.
GPT-4, without any specific training, proved better than the average human analysts at predicting whether earnings would increase or decrease! The AI even matched the performance of specialized machine learning models that had been trained for hours on tons of financial data. Not bad for a performance!
To achieve this, the researchers prepared the data by anonymizing the financial statements so GPT-4 couldn’t cheat by referencing its memory. It had to manage on its own, analyzing only the provided figures… identifying key trends, calculating the right ratios, finely interpreting the results… in short, performing real analyst work!
The researchers even showed that a neural network trained solely on the comments provided by GPT-4 could predict earnings almost as well as the AI itself.
Of course, GPT-4 still struggles with tricky cases, like small companies that are losing money, and human analysts retain an advantage when they have access to “soft” information in addition to the numbers. So rest assured, the AI isn’t taking their jobs tomorrow (but the day after tomorrow?).
Nevertheless, these results open up rather exciting prospects! Firstly, it could democratize financial analysis, making it accessible to the general public. Secondly, hedge funds could use GPT-4 to detect investment opportunities.
The researchers even tested a trading strategy based on the AI’s predictions, and it works pretty well! The study shows that the “long short” strategy based on GPT-4’s predictions generates returns higher than the market, with significant Sharpe ratios and alphas. The AI seems particularly adept at finding value in small-cap stocks, where human analysts and even traditional neural networks struggle more.
But caution is needed. An AI playing the stock market can be risky, and more safeguards will be necessary. And let’s not forget that this remains an experimental study, not real-world application.
But I thought it was cool enough to share with you.
