- The paper finds that ChatGPT, unlike DeepSeek, demonstrates significant ability to predict future stock market movements and correlate with macroeconomic variables using news data.
- The study employs ChatGPT-3.5 and DeepSeek to classify Wall Street Journal news articles (1996-2022) and uses robust statistical techniques to evaluate their predictive power.
- Results show ChatGPT predicting stock returns up to a year ahead with R
² up to 8.52%, particularly during uncertainty, highlighting its advantage over traditional methods and DeepSeek in extracting fundamental-related insights.
Analysis of "ChatGPT and DeepSeek: Can They Predict the Stock Market and Macroeconomy?"
This paper presents a rigorous exploration into the predictive capabilities of ChatGPT and DeepSeek regarding the stock market and macroeconomic indicators. The authors utilize news data from the Wall Street Journal spanning 1996 to 2022 to evaluate whether these LLMs can forecast financial trends. Their findings reveal distinct capabilities between ChatGPT and DeepSeek, particularly highlighting ChatGPT's ability to identify predictive elements in textual data, which markedly influence future stock market movements.
The authors employ a methodical approach, leveraging ChatGPT-3.5 and DeepSeek to classify news articles as either "good" or "bad" for stock market predictions. Their methodology includes zero-shot prompts with ChatGPT, fine-tuning techniques, and comparisons with other LLMs like BERT and RoBERTa. A particular strength of the paper is the use of robust statistical techniques, such as Hodrick t-statistics, which enhance the reliability of their findings.
One of the notable empirical results is ChatGPT's ability to predict stock returns based on good news, with statistically significant correlations evidenced in various forecast horizons up to one year. The paper reports an R² value of up to 8.52% when predicting market returns for up to a year ahead using the good news ratio. This predictability is primarily observed during periods of economic downturns, high economic policy uncertainty, or when the news content is particularly novel. Such findings suggest that ChatGPT can capture aspects of textual sentiment that human investors may overlook, leading to a delayed incorporation into market prices.
Contrastingly, DeepSeek, despite its competitive performance in some tasks, does not exhibit the same predictive power for future stock returns. The paper attributes this to differences in training datasets, emphasizing ChatGPT's extensive training in English, which may have enhanced its capability to discern nuanced financial information from English-language media sources.
The analysis also incorporates comparisons with traditional methods of textual sentiment analysis, such as those using predefined word lists, and finds that these methods do not capture the same predictive power as ChatGPT. The latter's emergent capacities, attributed to its large model size and advanced contextual understanding, allow it to extract and leverage information more effectively.
Additionally, this paper ventures beyond stock predictability, showing that ChatGPT-derived news indicators significantly correlate with macroeconomic variables like industrial production and VIX, indicating a broader applicability in economic forecasting contexts. However, the paper delineates a critical distinction: while ChatGPT excels in extracting and utilizing information related to macroeconomic fundamentals, DeepSeek seems more aligned with capturing investor sentiment.
This paper contributes to the growing body of literature exploring AI and LLMs in financial markets, underscoring the nuanced capabilities of models like ChatGPT. The authors recommend further research directed at other financial markets, such as bonds or commodities, to broaden our understanding of how AI tools can enhance investment strategies and economic forecasts.
In conclusion, while the findings position ChatGPT as a formidable tool for market prediction, the varying outcomes with DeepSeek underline the importance of LLM training specifics and their thematic specialization. This inquiry propels forward the discourse on how AI-driven insights can be integrated into systematic investment frameworks, potentially altering paradigms of market efficiency and investor behavior. Future developments in LLMs may further refine these applications, leading to unprecedented levels of predictability and market comprehension.