Researchers from the University of Chicago wanted to test the capabilities of GPT-4 in the fieldfinancial analysisand not just any analysis, eh, they went straight to the heart of the matter, namely predicting the future profits of companies from their financial reports!
Otherwise, it’s a professional job reserved for experienced financial analysts, because you have to know how to dissect accounting balance sheets, calculate key ratios, interpret trends… In short, it’s complex and machines really shouldn’t have competed with the flair and experience of professionals.
Except… Surprise! The results of the study are astounding.
GPT-4, without any special training, was better than the average human analyst at predicting whether profits would increase or decrease! The AI even matched specialized machine learning models, trained for hours on tons of financial data. Not a bad performance!
To achieve this, the researchers prepared the data by anonymizing the financial statements so that there was no way for GPT-4 to cheat by mining its memory. Therefore, he had to act like an adult, analyzing only the figures offered… He notices key trends, calculates the right ratios, carefully interprets the results… in short, a real analyst works!
The researchers even showed that a neural network trained solely on the feedback provided by GPT-4 was able to predict profits almost as well as artificial intelligence alone.
Obviously, GPT-4 still has problems with the twisted cases, such as small loss-making companies and human analysts maintaining an advantage when they have access to “soft” information in addition to numbers, so rest assured that AI will not steal their jobs tomorrow (but the day after ?).
Nevertheless, these results open quite exciting perspectives! First, it could democratize financial analysis by making it accessible to the general public. Else, hedge funds could use GPT-4 to discover investment opportunities.
Researchers have also tested a trading strategy based on AI predictions and it works pretty well! The study shows that a long short strategy based on GPT-4 predictions generates market-beating returns, with Sharpe ratios and alpha significant. AI appears to be particularly good at finding value in small markets, where human analysts and even traditional neural networks struggle.
But be careful, you have to be careful. Artificial intelligence playing in the stock market can be risky and we will need additional protection measures. And then let’s not forget that this remains an experimental study, not reality on the ground.
But I thought it was cool enough to tell you about it.