Comparing ChatGPT-Based and KR-FinBERT-Based News Sentiment for Stock Return Predictability: Evidence from the KOSPI Market

Key Points

• This research aims to compare the sentiment indicators derived from ChatGPT and KR-FinBERT regarding their predictive power for short-term stock returns in the KOSPI market.
• Constructed sentiment scores using corporate financial news from ChatGPT and KR-FinBERT.
• Integrated sentiment indicators into various deep learning architectures, including RNN, LSTM, and iTransformer.
• Evaluated predictive performance using information coefficient (IC), mean squared error (MSE), and mean absolute error (MAE) under a strict walk-forward time series validation framework.
• ChatGPT-based sentiment consistently showed higher and more stable IC performance across the entire rolling window period, particularly excelling during volatile periods like 2020–2021.
• KR-FinBERT demonstrated superior performance with specific architectures but exhibited lower stability in predictions over time.
• Overall, the study indicates that contextual sentiment information from large language models may provide broader and more adaptable insights for short-term return predictions compared to domain-specific models.