Recent research has uncovered that large language models like ChatGPT and BERT can be helpful in researching stock prices and public opinion. In a paper called “Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models,” University of Florida professors Alejandro Lopez-Lira and Yuehua Tang tested OpenAI’s ChatGPT to see if it could evaluate and assess the sentiment of news headlines.
Sentiment analysis, which is used for determining if a news headline expresses positive, negative, or neutral sentiments about a company or subject, has been found to help improve accuracy of market observer’s predictions. By comparing ChatGPT’s results with a sample of company share performance post evaluation, the researchers discovered that the model provided predictions that were statistically significant, which is better than other simpler language models.
ChatGPT was favored because of its more advanced language understanding capabilities that allow it to pick up on the subtlety of news headlines. The implications of this result implies that ChatGPT could help predict stock market movements based on headline sentiments.
However, Alejandro Lopez-Lira, assistant professor of finance at University of Florida and one of the paper’s co-authors, believes that ChatGPT alone may not be enough to provide effective sentences analysis on current events. He suggests that GPT should be fine-tuned or even further contextualized in order to produce more reliable forecasts.
In a similar paper called, “Language Models Trained on Media Diets Can Predict Public Opinion,” researchers from MIT and Harvard suggest that large language models fine-tuned with media diets can predict opinions of groups exposed to the particular media. Google scientist and MIT doctoral candidate Eric Chu elaborates that the media diet models seek to predict the opinion of media consumers through training a language model on their media consumption.
The authors concluded that these models could be used to improve accuracy of public opinion polling, but also open up discussion regarding how media affects the public. They posited that these models can be used to detect any kinds of systemic bias, echo chambers, and filter bubbles that might be occurring in media content.
Alejandro Lopez-Lira is an assistant professor of finance at the University of Florida and co-author of the paper “Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models”. He has a strong background in financial engineering, specifically on the topics of portfolio theory, asset pricing, and quantitative finance.
Eric Chu is a research scientist at Google, who was a MIT Doctoral Candidate at the time of his research project. He worked on the paper “Language Models Trained on Media Diets Can Predict Public Opinion”. In the paper, he discussed the potential of media diet models to accurately predict the opinion of media consumers by training a language model on media they consume.
