Stock Price Prediction Using ChatGPT

Attempts to predict future stock market prices have developed into a core research area within Asset Pricing Theory. Recently, research has been actively conducted using Transformer architectures to capture the market's nonlinear movements, or approaches like TabGPT that predict future prices based on past trading records.

In particular, the study by Lopez-Lira and Tang(2023) is drawing attention as the first research to verify the effectiveness of using ChatGPT for news analysis to predict stock prices. According to the study's results, classifying news headlines into positive, negative, and neutral categories using ChatGPT was found to have a significant effect on the stock returns of the following day.

While existing basic language models such as GPT-1 and BERT showed limitations in predicting returns, the latest LLM demonstrated the ability to understand the subtle context of news through extensive pre-training and to reason about its impact on future stock prices.

The analysis used daily returns and news headlines from the U.S. CRSP database, and to control for ChatGPT's randomness, the experiment was conducted with the Temperature value set to 0. Additionally, the 'GPT Score' derived by ChatGPT showed low correlation with sentiment scores from existing news providers, suggesting that it contains new information not captured by existing methods.