没有合适的资源?快使用搜索试试~ 我知道了~
Can ChatGPT Forecast Stock.pdf
需积分: 5 1 下载量 189 浏览量
2023-06-30
13:31:02
上传
评论
收藏 301KB PDF 举报
温馨提示
试读
29页
ChatGPT研究资料,Can ChatGPT Forecast Stock Price Movements Return Predictability and Large Language Models
资源推荐
资源详情
资源评论
Can ChatGPT Forecast Stock Price Movements?
Return Predictability and Large Language Models
∗
Alejandro Lopez-Lira and Yuehua Tang
University of Florida
First Version: April 6, 2023
This Version May 11, 2023
Abstract
We examine the potential of ChatGPT, and other large language models, in predicting
stock market returns using sentiment analysis of news headlines. We use ChatGPT to indi-
cate whether a given headline is good, bad, or irrelevant news for firms’ stock prices. We
then compute a numerical score and document a positive correlation between these “ChatGPT
scores” and subsequent daily stock market returns. Further, ChatGPT outperforms traditional
sentiment analysis methods. We find that more basic models such as GPT-1, GPT-2, and
BERT cannot accurately forecast returns, indicating return predictability is an emerging ca-
pacity of complex models. ChatGPT-4’s implied Sharpe ratios are larger than ChatGPT-3’s;
however, the latter model has larger total returns. Our results suggest that incorporating ad-
vanced language models into the investment decision-making process can yield more accurate
predictions and enhance the performance of quantitative trading strategies. Predictability is
concentrated on smaller stocks and more prominent on firms with bad news, consistent with
limits-to-arbitrage arguments rather than market inefficiencies.
∗
We are grateful for the comments and feedback from Andrew Chen, Carter Davis, Andy Naranjo,
Nikolai Roussanov, Ben Lee, Holger K. von Jouanne-Diedrich, and ‘@jugglingnumbers.’ Emails: Alejan-
dro Lopez-Lira (corresponding author): alejandro.lopez-lira@warrington.ufl.edu, and Yuehua Tang: yue-
hua.tang@warrington.ufl.edu.
1
Electronic copy available at: https://ssrn.com/abstract=4412788
The application of generative artificial intelligence and large language models (LLMs)
such as ChatGPT in various domains has gained significant traction in recent months, with
numerous studies exploring their potential in diverse areas. In financial economics, however,
using LLMs remains relatively uncharted territory, especially concerning their ability to pre-
dict stock market returns. On the one hand, as these models are not explicitly trained for
this purpose, one may expect that they offer little value in predicting stock market move-
ments. On the other hand, to the extent that these models are more capable of understanding
natural language, one could argue that they could be a valuable tool for processing textual
information to predict stock returns. Thus, the performance of LLMs in predicting financial
market movements is an open question.
To the best of our knowledge, this paper is among the first to address this critical question
by evaluating the capabilities of ChatGPT in forecasting stock market returns. Through a
novel approach that leverages the model’s sentiment analysis capabilities, we assess the
performance of ChatGPT using news headlines data and compare it to existing sentiment
analysis methods provided by leading vendors. It is well known that stock returns are
predictable at a daily horizon using news and trained algorithms (Tetlock (2007), Tetlock,
Saar-Tsechansky, and Macskassy (2008), and Tetlock (2011) among others), possibly because
combining new information is complicated (Fedyk and Hodson (2023)). Hence, our focus is
to evaluate whether models not trained in predicting returns acquire this capability as they
become better at other tasks.
Our findings have important implications for the employment landscape in the financial
industry. The results could potentially lead to a shift in the methods used for market predic-
tion and investment decision-making. By demonstrating the value of ChatGPT in financial
economics, we aim to contribute to the understanding of LLMs’ applications in this field and
inspire further research on integrating artificial intelligence and natural language processing
in financial markets. In addition to the implications for employment in the financial industry,
our study offers several other significant contributions.
2
Electronic copy available at: https://ssrn.com/abstract=4412788
Firstly, our research can aid regulators and policymakers in understanding the potential
benefits and risks associated with the increasing adoption of LLMs in financial markets. As
these models become more prevalent, their influence on market behavior, information dissem-
ination, and price formation will become critical areas of concern. Our findings can inform
discussions on regulatory frameworks that govern the use of AI in finance and contribute to
the development of best practices for integrating LLMs into market operations.
Secondly, our study can benefit asset managers and institutional investors by providing
empirical evidence on the efficacy of LLMs in predicting stock market returns. This insight
can help these professionals make more informed decisions about incorporating LLMs into
their investment strategies, potentially leading to improved performance and reduced reliance
on traditional, more labor-intensive analysis methods.
Lastly, our research contributes to the broader academic discourse on artificial intelligence
applications in finance. By exploring the capabilities of ChatGPT in predicting stock market
returns, we advance the understanding of LLMs’ potential and limitations within the financial
economics domain. This can inspire future research on developing more sophisticated LLMs
tailored to the financial industry’s needs, paving the way for more efficient and accurate
financial decision-making.
1
Our study has far-reaching implications that extend beyond the immediate context of
stock market predictions. By shedding light on the potential contributions of ChatGPT to
financial economics, we hope to encourage continued exploration and innovation in AI-driven
finance.
Related Literature
Recent papers that use ChatGPT in the context of economics include Hansen and Kazinnik
(2023), Cowen and Tabarrok (2023), Korinek (2023), and Noy and Zhang (2023). Hansen and
Kazinnik (2023) show that LLMs like ChatGPT can decode Fedspeak (i.e., the language used
1. See for example Wu et al. (2023).
3
Electronic copy available at: https://ssrn.com/abstract=4412788
by the Fed to communicate on monetary policy decisions). Cowen and Tabarrok (2023) and
Korinek (2023) demonstrate that ChatGPT is helpful in teaching economics and conducting
economic research. Noy and Zhang (2023) find that ChatGPT can enhance productivity in
professional writing jobs. Contemporaneously, Xie et al. (2023) find ChatGPT is no better
than simple methods such as linear regression when using numerical data in prediction tasks.
We attribute the difference in results to their focus on using historical numerical data to
predict, while ChatGPT excels at textual tasks. Ko and Lee (2023) finds ChatGPT may be
helpful in selecting across asset classes. Furthermore, Yang and Menczer (2023) demonstrates
that ChatGPT successfully identifies credible news outlets. Our study is among the first to
study the potential of LLMs in financial markets, particularly the investment decision-making
process.
We contribute to the recent strand of the literature that employs text analysis and ma-
chine learning to study a variety of finance research questions (e.g., Jegadeesh and Wu
(2013), Campbell et al. (2014), Hoberg and Phillips (2016), Gaulin (2017), Baker, Bloom,
and Davis (2016), Manela and Moreira (2017), Hansen, McMahon, and Prat (2018), Ke,
Kelly, and Xiu (2019), Ke, Montiel Olea, and Nesbit (2019), Bybee et al. (2019), Gu, Kelly,
and Xiu (2020), Cohen, Malloy, and Nguyen (2020), Freyberger, Neuhierl, and Weber (2020),
Lopez-Lira 2019, Binsbergen et al. (2020), Bybee et al. (2021)). Our paper makes a unique
contribution to this literature as being the first to evaluate the text processing capabilities
of recently developed LLMs such as ChatGPT in forecasting stock market movements.
Our paper also adds the literature that uses linguistic analyses of news articles to extract
sentiment and predict stock returns. One strand of this literature studies media sentiment
and aggregate stock returns (e.g., Tetlock (2007), Garcia (2013), Calomiris and Mamaysky
(2019)). Another strand of the literature uses the sentiment of firm news to predict future
individual stock returns (e.g., Tetlock, Saar-Tsechansky, and Macskassy (2008), Tetlock
(2011), Jiang, Li, and Wang (2021)). Different from prior studies, we focus on understanding
whether LLMs add value by extracting additional information that predicts stock market
4
Electronic copy available at: https://ssrn.com/abstract=4412788
reactions.
Finally, our paper also relates to the literature on employment exposures and vulnerability
to AI-related technology. Recent works by Agrawal, Gans, and Goldfarb (2019), Webb
(2019), Acemoglu et al. (2022), Acemoglu and Restrepo (2022), Babina et al. (2022), Noy
and Zhang (2023) have examined the extent of job exposure and vulnerability to AI-related
technology as well as the consequences for employment and productivity. With AI being
on a constant rise since its inception, our study focuses on understanding an urgent but
unanswered question – the capabilities of AI, and LLMs in particular, in the finance domain.
We highlight the potential of LLMs in adding value to market participants in processing
information to predict stock returns.
1 Background
ChatGPT is a large-scale language model developed by OpenAI based on the GPT (Genera-
tive Pre-trained Transformer) architecture. It is one of the most advanced natural language
processing (NLP) models developed to date and trained on a massive corpus of text data
to understand the structure and patterns of natural language. The Generative Pre-trained
Transformer (GPT) architecture is a deep learning algorithm for natural language processing
tasks. It was developed by OpenAI and is based on the Transformer architecture, which was
introduced in Vaswani et al. (2017). The GPT architecture has achieved state-of-the-art
performance in various natural language processing tasks, including language translation,
text summarization, question answering, and text completion.
The GPT architecture uses a multi-layer neural network to model the structure and
patterns of natural language. Using unsupervised learning methods, it is pre-trained on
a large corpus of text data, such as Wikipedia articles or web pages. This pre-training
process allows the model to develop a deep understanding of language syntax and semantics,
which is then fine-tuned for specific language tasks. One of the unique features of the GPT
5
Electronic copy available at: https://ssrn.com/abstract=4412788
剩余28页未读,继续阅读
资源评论
数字魔术师
- 粉丝: 462
- 资源: 320
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功