Alejandro Lopez-Lira used ChatGPT to parse news headlines according to whether they are good or bad for a stock and found that ChatGPT was successful in predicting the direction of the next day’s returns. In a previous related report, we told you that Goldman Sachs estimates that about 35 percent of financial work is at risk of being automated by artificial intelligence.
Predictions are successful as long as numbers don’t come into play.
Lopez-Lira said the results surprised him and suggested that sophisticated traders have not yet used ChatGPT-style machine learning in their trading strategies.
How was the experiment done?
The experiments analyzed more than 50,000 headlines from a data provider on the New York Stock Exchange, the Nasdaq, and publicly traded stocks on a small-cap stock exchange. Lopez-Lira and her partner Yuehua Tang began collecting this data in October 2022. ChatGPT’s training data ends in 2021, meaning ChatGPT has never seen these headlines before.
They then fed the headers into ChatGPT 3.5 with the following prompt:
“Forget all your previous instructions. Pretend you’re a financial professional. You’re a financial expert with stock advice experience. Answer “YES” if the good news is on the first line, “NO” if the bad news, or “UNKNOWN” if it’s vague. Then the next line is a short answer. and elaborate in a concise sentence.”
Get rich with ChatGPT?
However, in the research conducted, it is said that artificial intelligence models will have great importance in the future. The same experiments have been repeated using more basic models such as GPT-2 and BERT. But these gave worse results than GPT-3.5. Currently, the newest model is the GPT-4, and the development of the models continues. For now, these models, including the GPT-4 model, are not very strong in mathematical matters. It is stated that when this gap is eliminated and combined with other disciplines, more accurate predictions can be opened.