Sage Journals: Discover world-class research

Abstract

The introduction of ChatGPT and other tools based on artificial intelligence (AI) has the potential to revolutionize the field of education. We study how the public release of ChatGPT and the increased attention on this new large language model from OpenAI are associated with the expected returns of publicly traded firms that operate in the education sector. We also perform separate subgroup analyses for the traditional education sector and the so-called education technology sector. Using linear and threshold CAPM-GARCH models, we find that after the public release of ChatGPT, both the education sector as a whole and the education technology sector have underperformed benchmarks. Our results show that increased attention leads to lower next-day returns in the education sector as a whole and the education technology sector in particular. Additionally, during periods of higher attention, expected returns tend to decline in these two sectors. We also show that the introduction of ChatGPT or the increased interest in this AI tool in the population does not affect the traditional education sector. The introduction of ChatGPT thus has a heterogeneous effect across the various education sectors we examine, with the education technology sector receiving most of the disruption.

Keywords

ChatGPT artificial intelligence AI Twitter-based attention Edtech

Introduction

The interest in ChatGPT and artificial intelligence (AI) has recently expanded rapidly worldwide. In just a matter of months, the chatbot has become a global phenomenon with millions of users, and AI is now “on everyone’s lips.” According to OpenAI—the company behind ChatGPT—the chatbot reached 1 million users only five days after its launch in November 2022. For comparison, Instagram spent 2.5 months, Spotify 5 months, Dropbox 7 months, and Facebook 10 months to reach the same number of users (Buchholz, 2023).

ChatGPT is able to produce highly sophisticated answers to a wide range of question types and adjust the response to the academic level requested—in a matter of seconds. The new technology is particularly relevant in so-called writing-intensive courses (Herman, 2022; Lund & Wang, 2023), but tests have shown that it may also be efficiently used to solve problems within statistics, mathematics, quantitative methods, and programming, among many others (Frieder et al., 2023; Surameery & Shakor, 2023; Wardat et al., 2023). Recent video examples posted on YouTube also highlight its ability to analyze data (Freberg, 2023). In addition, it can produce excellent visualizations and interpretations of results—at the user’s request. Lo (2023) has reviewed how ChatGPT can be used in education and found that its usefulness is domain-specific. According to Kasneci et al. (2023) and Lo (2023), ChatGPT can be used by students on all levels of the education system; from the elementary level to universities. It can also be used as a tool in remote learning and empower students with disabilities. For example, ChatGPT can help students learn by correcting grammar and style, preparing reviews to aid text comprehension, and explaining key concepts. It can also be used to support assessments by generating questions and tasks from a given corpus. ChatGPT also offers innovations for instructors as well (Kasneci et al., 2023; Lo, 2023) by helping to prepare course materials (from translation to topic suggestions) or to help in the assessment process (preparing exams and questions). A recently published article even shows that the new technology can be efficiently used to assist with the production of research studies (Dowling & Lucey, 2023). The software is not without limitations, though, and numerous examples of how ChatGPT has offered strange and even false statements have been posted online and debated publicly. Na et al. (2024) have studied education-related subreddits on the social/discussion platform Reddit and performed a content analysis of posts that included the keyword ”chatgpt”. Their sample covered a sample from January 1 2022 until March 31, 2023, thus the onset of ChatGPT. They found that concerns were related to inaccuracy, lack of credibility, or depth of the answers. A study by Kasneci et al. (2023) argued that while LLMs (large language models) have the potential to improve the digital ecosystem for education, they also bring copyright issues and potential biases, and might lead to a high reliance of students and educators on model (un-authored) outcomes. Training of LLMs also tends to lag behind up-to-date knowledge (Lo, 2023). Memarian and Doleck (2023) agree with Kasneci et al. (2023) that ChatGPT might reduce the ability to learn and explore. To avoid that, good technical expertise is necessary with ChatGPT. Similar concerns were identified by Li et al. (2023), who studied discussions related to ChatGPT on the social platform X (Twitter), a tool that we also use to monitor the general public’s attention towards ChatGPT. Their analysis revealed that concerns are related to the impact on learning outcomes and skill development, but also towards policy and social issues, as the tool is not accessible to all. Interestingly, X users have also expressed concerns related to workforce replacement, which in turn might impact provided services and products in the education sector. X users, who, based on their account description, specialized in education and technology, were particularly engaged. These examples suggest that ChatGPT is an innovation that has the potential to disrupt the education sector.

As noted by Lee, Ju, and Lee (2023), the use of ChatGPT is rapidly increasing, but the research on ChatGPT in the consumer field is limited. We therefore contribute by filling a general gap in the literature as we study how the introduction of the new technology first released by OpenAI has financially affected the education sector. We use trading data from firms in the education sector that are publicly traded. We focus on publicly traded companies because the best estimate of the expected future earnings of any company is made by the people having their money on the line—investors. Saggu and Ante (2023), for example, already documented that the launch of ChatGPT had a significant positive impact on AI-related crypto assets. Apart from analyzing the education sector, we identify two distinct subsectors. One is the sector that is represented by firms that focus on operating universities and campuses and provide education-related services to these institutions. That is, the traditional education subsector. The second is represented by firms that offer digital tools and platforms to enhance learning and might be perceived either as direct competitors with the new technology offered by OpenAI or might benefit from the technology by using it in their courses. That is, the education technology (EdTech) subsector.

Investors can regard ChatGPT as both an opportunity and a threat. The introduction of this innovative language model, therefore, creates increased uncertainty over the future of the industry. Such increased uncertainty likely changes investor’s expectations and manifests into market returns. Our analysis is therefore based on both the standard linear and nonlinear (threshold) models of market return, where we examine how market returns of companies in the education sector and the two subsectors changed prior to, during, and after the introduction of ChatGPT. Our main results utilize Twitter-based¹ attention data to estimate the interest and awareness of the public regarding ChatGPT. We investigate whether an increase in attention influences the future development of stock prices and/or induces changes in the price dynamics for the education sector and the two subsectors. In that regard, we rely on the growing literature on the limited attention hypothesis (Andrei & Hasler, 2015; Barber & Odean, 2008), according to which investors gravitate toward so-called “attention-grabbing” investment opportunities. This implies that news and events deemed worthy of investors’ attention tend to be more rapidly incorporated into prices. Previous empirical studies have found that awareness of a particular topic or company can predict trading volumes (Preis et al., 2010; 1933; Bordino et al., 2012), abnormal returns (Bijl et al., 2016), and price variation (the variance of returns) (Aouadi et al., 2013; Audrino et al., 2020; Ballinari et al., 2022; Goddard et al., 2015; Guidolin & Pedio, 2021; Hamid & Heiden, 2015; Smales, 2021; Vlastakis & Markellos, 2012).

Our main contribution can be summarized in three major findings. First, we document that after the public announcement of ChatGPT, all three portfolios underperformed their benchmarks. This decline was particularly significant in the education sector and the EdTech subsector. Our second finding is that the decline in stock returns is associated with increased attention toward ChatGPT. Specifically, this heightened interest leads to lower returns for the overall education and EdTech portfolios on the following day. Third, for these two portfolios, we identify periods of increased attention that are associated with both lower returns and lower comovement with the respective benchmark. These results show that the introduction of ChatGPT to the public had a disruptive effect on the education sector, particularly affecting the EdTech sector.

The remainder of this study is organized as follows. The next section discusses ChatGPT and its potential impact on the educational sector. To explore this topic, we have chosen to include a conversation with the chatbot itself. In Section 3, we present the data and methodology used in the analyses. Section 4 presents the results, and in Section 5, we offer some concluding remarks, implications, and directions for future research.

ChatGPT and Potential Challenges for the Educational Sector

OpenAI was formed in December 2015 and the now widely popular ChatGPT was released to the public on November 30, 2022 (Kirmani, 2022). The tool itself is based on the generative pretrained transformer (GPT) architecture, which is a type of neural network that has been trained on massive amounts of text data from the internet and other sources (Lund & Wang, 2023; Saggu & Ante, 2023). The first version of the transformer-based large language model was released in 2018 and had 117 million parameters. The second version (GPT-2) came in 2019 and had 1.5 billion parameters. GPT-3 was released in 2020 and had 175 billion parameters. In March 2023, GPT-4 was released, and this version is estimated to have more than one trillion parameters.² The chatbot has successfully completed a long list of tasks, including superior performance in a job interview for a coding position and generating movie scripts, stories, and song lyrics (Bhaimiya, 2023; Elias, 2023). How could the new tool potentially impact the educational sector? Inspired by the conversation (Lund & Wang, 2023) had with ChatGPT about a similar topic, we asked the chatbot to briefly explain how the new tool can affect the educational sector. Below, we provide a direct transcript of the conversation we had.³

Q: “Can you give a brief description of various tasks ChatGPT can solve within the educational sector?”

A: “Certainly! ChatGPT can be utilized for a wide range of tasks within the educational sector. Some of the key applications include:

• Answering Questions: ChatGPT can provide instant responses to questions asked by students, teachers, or any other users. It can handle a variety of subjects and topics, offering explanations, definitions, examples, and more.

• Homework Assistance: Students can seek help from ChatGPT with their homework assignments. It can guide them through problem-solving processes, offer suggestions, and explain concepts to improve understanding.

• Concept Clarification: If students are struggling to grasp a particular concept, ChatGPT can provide further explanations, analogies, or real-world examples to aid in comprehension.

• Language Learning: ChatGPT can assist in language learning by engaging in conversations, providing vocabulary explanations, grammar clarifications, and even offering practice exercises.

• Writing Support: Students can receive writing assistance from ChatGPT, such as generating ideas, improving sentence structures, providing feedback on drafts, or suggesting resources for research.

• Exam Preparation: ChatGPT can help students prepare for exams by answering practice questions, providing study materials, recommending relevant resources, and offering tips and strategies for effective studying.

• Personalized Learning: ChatGPT can adapt its responses based on individual learning needs and preferences. It can offer tailored recommendations, adaptive exercises, and personalized feedback to enhance the learning experience.

• Educational Content Generation: ChatGPT can assist teachers in generating educational content, such as lesson plans, quizzes, and assignments. It can also create engaging presentations or interactive learning materials.

• Research Assistance: Researchers and students can seek assistance from ChatGPT in finding relevant academic papers, summarizing research articles, or suggesting further reading material.

• Tutoring and Mentoring: ChatGPT can act as a virtual tutor or mentor, providing guidance, support, and encouragement to learners. It can offer personalized study plans, track progress, and recommend areas of improvement.

It’s important to note that while ChatGPT can provide valuable support in various educational tasks, it should not replace human educators. The model can assist in supplementing and enhancing the learning experience, but human interaction, critical thinking, and pedagogical expertise are still vital components of education.”

The educational tasks mentioned above, however, are key components of the business models of many companies operating within the educational sector, particularly within the EdTech domain. The release of ChatGPT to the public has the potential to disrupt the sector, creating considerable uncertainty about its future. It is, therefore, of particular interest to examine how the introduction of the new technology affects the market’s perception of such companies’ future prospects.

Material and Methodology

Education Sector

Our analysis is centered around three portfolios (sets of companies) that are created from U.S. based companies that belong to the “education services sector” (as indicated via the SIC Code 82).⁴ In our analysis, we included only companies that have had a market capitalization of at least 50 million USD over our sample period and a volume of at least 0.1 million USD also over the sample period from January 1, 2022, until June 16, 2023.⁵ In total, 18 companies satisfied these requirements and were used to form our first portfolio that we refer to as the education sector. The average market capitalization of the portfolio over the sample period is 27.469 billion USD. Define P_i,t as the closing price on day t = 0, 1, 2, .., T and R_i,t = ln(P_i,t) − ln(P_i,t−1) as the continuous returns of stock i = 1, 2, …, N = 18, that is part of the Education sector. If we let MC_i,t denote the corresponding market capitalization, the daily market capitalization weighted portfolio return of the education sector is given as ⁶

R_{E, t} = \sum_{i = 1}^{N} \frac{R_{i, t} M C_{i, t}}{\sum_{i = 1}^{N} M C_{i, t}} .

(1)

Including an additional company, with a smaller market capitalization than 50 million USD over our sample period, would give that one company a weight of 0.21% or less in the portfolio.

The introduction of the ChatGPT will likely have a larger impact on education technology companies, where products and services are facilitated through information technologies, by offering online learning tools, platforms, courses, learning management systems. In contrast, traditional education institutions mostly focus on owning and operating colleges, universities, and campuses. This led us to examine two distinct subsectors—the former is referred to as the EdTech sector, and the latter as the Traditional Education sector. Based on the products and services the companies offer, we assigned each company into one of the two subsectors, resulting in nine companies for each subsector. This classification is somewhat subjective, which is why we explore alternative subsector compositions in Section 4.3. The EdTech portfolio has an average market capitalization over the sample period of 16.600 billion USD, that is, the EdTech portfolio is about 50% larger than the portfolio of the Traditional Education subsector. The final list of all 18 companies, along with brief descriptions, can be found in Appendix A1. The resulting portfolio returns for traditional education (R_TE,t) and EdTech (R_ET,t) are:

\begin{aligned} R_{TE, t} & = \sum_{i \in TE} \frac{R_{i, t} M C_{i, t}}{\sum_{i \in TE} M C_{i, t}} \\ R_{ET, t} & = \sum_{i \in ET} \frac{R_{i, t} M C_{i, t}}{\sum_{i \in ET} M C_{i, t}} \end{aligned}

(2)

To estimate the potential impact of the public release of ChatGPT on the stock prices of the education sector, we use two methodologies that build upon the standard Capital Asset Pricing Model: i) the linear CAPM-GARCH model (Generalized AutoRegressive Conditional Heteroskedasticity) and ii) the threshold (in-mean) CAPM-GARCH model.

The CAPM-GARCH Model Framework

To examine whether the dynamics of the returns changed after the public release of ChatGPT, we estimate three specifications of the CAPM model for each of the three portfolios. In the first specification, we model portfolio returns only via market returns. This approach captures the sensitivity of the portfolio to market returns, R_m,t, that is the association between price changes on the market and the respective portfolio. Given the technological nature of ChatGPT and the widespread use of new technologies in education, we use the NASDAQ-100 as our representative market index, which we refer to as the technology market index. In the robustness section, we also include results that utilize the S&P 500 stock market index. To examine whether the period after the public release of ChatGPT has led to systematic changes to the returns of the various education portfolios, we introduce a second specification that includes the event dummy, E_t. The event dummy takes value one from November 30, 2022, to the end of the sample period and zero otherwise. In the third specification, instead of using the event dummy, we include the Twitter-based attention measure, A_t. The attention measure corresponds to the daily sum of geolocated tweets in the U.S. that contained the word “ChatGPT.” If the public release of ChatGPT affected investors’ expectations, we would expect that increased public interest in ChatGPT would lead to noticeable changes in the returns of education-related portfolios. This assumption is in line with the attention-based literature. However, it’s also possible that the relationship could be reversed—meaning that fluctuations in stock prices could actually drive public attention. To account for this possibility, we also incorporate lagged attention in our asset pricing equation (3). The advantage of using the attention to capture a specific event is that it i) does not need a specific event date and is thus useful if the effect is spread over a potentially long period of unknown length, ii) it shows variability, in contrast to an event-based indicator variable (i.e., E_t). A general specification of the CAPM-GARCH model takes the following form:

\begin{aligned} R_{t} & = γ_{0} + γ_{1} R_{m, t} + γ_{2} E_{t} + γ_{3} \ln A_{t - 1}^{D} + ϵ_{t} \\ ϵ_{t} & = σ_{t} η_{t}, η_{t} \overset{i i d}{\sim} f (η_{t}, λ, ξ) \\ σ_{t}^{2} & = ω + α ϵ_{t - 1}^{2} + β σ_{t - 1}^{2} \\ f (x, λ, ξ) & = {(2 π)}^{- 1 / 2} ξ^{- 1} {(x^{2} + 1)}^{1 / 2} e^{- 2^{- 1} {(ξ^{- 1} (\sinh^{- 1} (x) - λ))}^{2}} \end{aligned}

(3)

The first line is the ’mean-equation’, which is a linear regression model of interest, while the remaining equations describe the behavior of errors that are not assumed to be constant or to be drawn from a normal distribution, but instead a more flexible framework of error is employed, the GARCH, common in the finance literature. Thus, we allow the error term ϵ_t to be a mixture of time-varying volatility σ_t that is modeled via a GARCH model of Bollerslev (1986) and an independent and identically distributed η_t. We assume that η_t follows the Johnson-SU distribution (Johnson, 1949a, 1949b), which is an unbounded transformation of the normal density, where λ is the skewness and ξ the kurtosis parameter. The distribution is flexible enough to capture potential (left) skewness and heavy-tailed returns of financial assets. The distribution is frequently used in empirical research (e.g., Lyócsa, Molnár, & Výrost, 2021) and recommended by Choi and Nam (2008) when modeling financial asset returns within a GARCH model framework. We use the reparameterized version of the distribution as implemented by Ghalanos (2020).

Including the event dummy variable and the attention variables in the same model results in collinearity⁷ that inflates the estimated standard errors. Such specifications are therefore not considered. The specification described above is just a comprehensive description of the most general specification. In the robustness section, we consider an alternative specification.

Threshold CAPM-GARCH Model

The public release of ChatGPT might have induced a structural break in i) the expected education portfolio returns and ii) the comovement the education portfolio returns have with the market returns. Hence, if market expectations about the future prospects of the firms operating in the educational sector start to change, we might observe a decoupling from the dynamics of the general market and increased exposure to the attention. A break in comovement is not captured in the linear CAPM-GARCH model. In such cases, the threshold regression offers an estimate of such a “kink point.”⁸ Motivated by the threshold linear regression (see, e.g., Hansen, 2022), we allow state-variable-induced changes in the mean equation of equation (3). Specifically, the model implies that the returns behave according to one of the two models (regimes), depending on the state in which a given state-variable Q_t is:

\begin{aligned} R_{t} & = \{\begin{cases} γ_{0,1} + γ_{1,1} R_{m, t} + ϵ_{t, 1}, Q_{t} < ν \\ γ_{0,2} + γ_{1,2} R_{m, t} + ϵ_{t, 2}, Q_{t} \geq ν \end{cases} \\ ϵ_{t} & = σ_{t} η_{t}, η_{t} \overset{i i d}{\sim} f (η_{t}, λ, ξ) \\ σ_{t}^{2} & = ω + α ϵ_{t - 1}^{2} + β σ_{t - 1}^{2} \\ f (x, λ, ξ) & = {(2 π)}^{- 1 / 2} ξ^{- 1} {(x^{2} + 1)}^{1 / 2} e^{- 2^{- 1} {(ξ^{- 1} (\sinh^{- 1} (x) - λ))}^{2}} \end{aligned}

(4)

where the indices 1 and 2 indicate that separate parameters are estimated for regimes 1 and 2, respectively. Q_t is the state or threshold variable, and ν is the estimated threshold. We set the state variable to be either i) Twitter-based attention, A_t−1, or ii) the linear time trend t = 1, 2, …, T. In the former case, we can estimate the implicit effect of attention on the dynamics of education sector portfolio returns. In the latter case, the linear time trend allows us to endogenously date the unconditional switching point in the relationship between the education sector portfolios and market index returns.

The model is linear in the coefficients, except for the threshold parameter ν. The model is estimated in two steps. In the first step, we estimate the threshold in the mean equation by minimizing the residual sum of squares. In the second step, we estimate the resulting CAPM-GARCH specification under different regimes. To assess the validity of our specifications, we check the serial correlation of the standardized and squared standardized residuals (e.g., ϵ_t/σ_t) and test for the asymmetric volatility effects via the Engle and Ng (1993) test. We estimate robust standard errors following the approach of White (1982). Further robustness checks are provided in the corresponding section.

Results

Preliminary Data Analysis

In Table 1, we report key descriptive statistics over the entire sample period. The average daily return was −0.02% (−4.89% annually (p.a.)) for the educational sector, 0.021% (5.51% p.a.) for the traditional education subsector and −0.051% (−12.11% p.a.) for the EdTech subsector. The corresponding return for the benchmark technology market index was −0.036% (−8.6% p.a.). After November 30, 2022, the average daily returns were −0.078% (−17.77% p.a.), −0.034% (−8.1% p.a.) and −0.112% (24.66% p.a.) for the whole education sector and for the two subsectors, respectively. During the same sample period, the benchmark technology index had an average daily return of +0.13% (38.73% p.a.). We observe similar heterogeneity for the daily average number of tweets containing the word “ChatGPT.” Over the entire sample period, the daily average amounts to 5.8 but increases to 107.49 for the subperiod from November 30, 2022, to June 16, 2023. Table 1 also shows considerable variation in returns. While the minimum daily return of the whole education sector is −7.34%, it jumps to −9.029% for the EdTech subsector, which also proved to be the portfolio with the highest variance. A daily standard deviation of 2.315 for the EdTech sector corresponds to an annualized volatility of 36.75%. The traditional education sector is considerably less volatile, with a daily standard deviation of 1.411% (22.40% annualized volatility).⁹ We also observe that attention is quite persistent over the entire sample period, where even the 22^nd order of auto-correlation is at 0.82. Such persistence is typical for attention measures (see, e.g., Lyócsa et al., 2023; Lyócsa & Plíhal, 2022).

Table 1.

Data Characteristics.

Variables	Mean	SD	Min	Q1	Median	Q3	Max	Skewness	Kurtosis	ρ(1)	ρ(22)	EL
Panel A: Market returns
Education sector	−0.020	1.803	−7.344	−1.023	−0.068	1.055	5.243	−0.140	4.106	−0.027	−0.044
Traditional	0.021	1.411	−5.116	−0.866	−0.017	0.895	5.298	0.110	3.810	−0.508	−0.038
EdTech	−0.051	2.315	−9.029	−1.460	−0.185	1.391	6.348	−0.159	3.969	−0.020	−0.036
Technology market	−0.036	1.810	−5.297	−1.223	−0.094	1.269	7.093	−0.048	3.363	−0.009	−0.082
Panel B: Attention
Tweets [log]	1.758	2.271	0.000	0.000	0.000	4.564	5.210	0.541	1.327	0.993	0.820	***

Notes: All values are calculated from daily percentage returns using a sample from January 1, 2022 until June 16, 2023. SD denotes standard deviation, Q1 and Q3 are lower and upper quartiles and ρ(1), ρ(22) is the auto-correlation of the given order.* **, ** denote the statistical significance of the auto-correlation of the Escanciano & Lobato (2009) test (EL columns).

In Figure 1, we visualize the key variables in the study. Before the event date, the education sector outperformed the benchmark market index (see the upper-left panel of Figure 1). However, following the public release of ChatGPT, the performance of the education portfolios seems to decline in comparison to the market portfolio. This can be more formally examined by the cumulative average abnormal returns (CAARs). CAARs measure a portfolio’s return relative to expectations; a positive CAAR indicates that the asset performed better than anticipated. These are visualized for each portfolio in the upper-right panel of Figure 1.¹⁰ The underperformance over the following eight months that coincides with the period after the public release of ChatGPT is considerable, with the EdTech subsector being the most affected.

Figure 1.

The upper-left panel shows a comparison of the education sector (blue line), the traditional education subsector (purple line), the EdTech subsector (black line), and the technology market index (NASDAQ-100). The upper-right panel shows cumulative average abnormal returns (CAAR). The CAARs are calculated for the period starting 5 days prior to the release date of ChatGPT (November 30, 2022). The CAARs are found as the cumulative market cap-weighted sums of abnormal returns of individual companies making up the education sectors. The abnormal returns were calculated using a single market model with the technology index representing the market returns. The estimation window goes from early 2022 until November 22, 2022. The results using an expanding estimation window are similar. The vertical dashed line corresponds to the introduction of ChatGPT to the public on November 30, 2022. The lower panel shows daily attention to “ChatGPT” on Twitter in the United States.

The “buzz” about ChatGPT started with its release on November 30, 2022. Before that, there were only a few tweets posted in the U.S. containing the word “ChatGPT” (see the lower panel of Figure 1). The attention to “ChatGPT” seems to be related to the observed decrease in the value of the EdTech sector. These preliminary results suggest an underperformance of the education sector and its subsectors. We will formally test whether we can attribute these changes in the pricing dynamics to the period around the event date and to the attention toward ‘ChatGPT’.

Empirical Results

CAPM Model Framework

Education Sector

In Table 2, we present results for the whole education sector portfolio of the CAPM-GARCH model under three specifications. Model 1 in Panel A shows that there is no indication of an abnormal return, as the constant term is not significantly different from zero when examining the whole sample period. The coefficient for the technology market index at 0.715 is below 1, which suggests that the education sector is less volatile than the technology market in general. The results in Panels B and C show that, overall, the volatility is highly persistent. Interestingly, the results from the GARCH part of the model show no asymmetry, but both are heavy-tailed, as indicated by the significant kurtosis parameter. It thus appears that the returns of the education sector portfolio suffered extreme price shocks over the sample period, which is well captured by the flexible Johnson-SU distribution, which empirically supports our choice for not employing the normal distribution for residuals.

Table 2.

Market model - Education sector.

Models	1	2	3
Panel A: Mean equation
Constant	0.013	0.101	0.105
	[0.062]	[0.077]	[0.078]
Technology market index	0.715***	0.720***	0.720***
	[0.034]	[0.035]	[0.035]
After event (0 = ’no’, 1 = ’yes’)		−0.233**
		[0.119]
Daily attention			−0.052**
			[0.026]
Panel B: Variance equation
ω	0.023	0.028*	0.028*
	[0.019]	[0.017]	[0.017]
α	0	0	0
	[0.007]	[0.010]	[0.010]
β	0.984***	0.980***	0.980***
	[0.008]	[0.002]	[0.002]
Panel C: Distribution parameters
λ - skewness	−0.145	−0.153	−0.149
	[0.158]	[0.163]	[0.164]
ξ - kurtosis	1.541***	1.568***	1.571***
	[0.244]	[0.257]	[0.257]
Panel D: Model diagnostics
Persistence	0.984	0.980	0.980
ρ_x(1)	−0.016	−0.029	−0.029
$ρ_{x^{2}} (1)$	−0.039	−0.037	−0.037
Sign bias test (p-value)	0.739	0.836	0.839
Log-Likelihood	−571.2	−569.2	−569.1
R²	53.4%	54.0%	54.0%
AIC	3.177	3.171	3.171
BIC	3.252	3.257	3.256
HQ	3.207	3.205	3.205

Notes: The* ** and *** denote statistical significance at the 10%, 5% and 1% significance levels. Standard errors are heteroscedasticity consistent estimates of White (1982) ρ_x(1) $(ρ_{x^{2}} (1))$ denotes first-order auto-correlation coefficients of (squared) standardized residuals. The Sign bias test corresponds to the p-value of the joint test of Engle & Ng (1993). R² is the squared correlation coefficient between observed and fitted observations.

In Model 2, we add an event dummy variable, taking value one from November 30, 2022, to the end of the sample period. The results show a negative, significant, and substantial long-term effect of −0.233. The education sector underperformed the benchmark index after the public release of ChatGPT, at least until the end of our sample period in June 2023. In Model 3, we include a one-day lagged attention measure. The results show that the lagged Twitter-based attention measure has a negative and significant association with next-day returns of the education sector with a coefficient of −0.052. A doubling of the current day’s attention (a 100% increase) induces a −0.052% decrease in returns. The results suggest that at least part of the decline in the performance of the education sector can be attributed to the increased attention to “ChatGPT.” For both Models 2 and 3, we have not observed any significant changes in the dynamics of the error terms. Volatility persistence and heavy-tail behavior of shocks only slightly improved the fit.

The Traditional Education and EdTech Subsectors

The introduction of ’ChatGPT’ might have a different impact on the two non-overlapping subset of companies, the traditional education and technology-driven education sectors (EdTech). The results in Tables 3 and 4 show that this might indeed be the case. Several results stand out and show that for the two portfolios, which represent two different subsectors, the pricing dynamics are distinctive. The traditional education sector shows substantially lower sensitivity to technology market index returns, ranging from 0.420 to 0.423 across the three specifications, while the corresponding sensitivity of the EdTech subsector is 0.907–0.915. Moreover, neither the after-event dummy variable nor the lagged attention variable have significant effects in the traditional education sector models. In the EdTech sector models, on the other hand, these variables have negative and statistically significant impacts on portfolio returns. The event dummy variable shows a sizable long-term negative effect of −0.278, and the attention variable effect is −0.063, meaning that a doubling of attention leads to a decline in the next day’s returns of −0.063%, on average. These results imply that companies operating in the traditional education sector are much more stable and do not seem to be disturbed by the introduction of ChatGPT.

Table 3.

Market model - Traditional Education Sector.

Models	1	2	3
Panel A: Mean equation
Constant	0.039	0.092	0.088
	[0.062]	[0.080]	[0.080]
Market index	0.420***	0.423***	0.423***
	[0.035]	[0.036]	[0.036]
After event (0 = ’no’, 1 = ’yes’)		−0.142
		[0.119]
Daily attention			−0.029
			[0.025]
Panel B: Variance equ ation
ω	0.002	0.001	0.001
	[0.013]	[0.013]	[0.013]
α	0.000	0.000	0.000
	[0.010]	[0.010]	[0.010]
β	0.999***	0.999***	0.999***
	[0]	[0]	[0]
Panel C: Distribution parameters
λ - skewness	0.304	0.306	0.304
	[0.300]	[0.314]	[0.313]
ξ - kurtosis	2.172***	2.235***	2.232***
	[0.350]	[0.380]	[0.380]
Panel D: Model diagnostics
Persistence	0.999	0.999	0.999
ρ_x(1)	−0.044	−0.050	−0.049
$ρ_{x^{2}} (1)$	0.120	0.133	0.131
Sign bias test (p-value)	0.072	0.045	0.048
Log-Likelihood	−570.3	−569.7	−569.7
R²	30.1%	30.5%	30.5%
AIC	3.172	3.174	3.174
BIC	3.247	3.260	3.260
HQ	3.202	3.208	3.208

Table 4.

Market model - EdTech Sector.

Models	1	2	3
Panel A: Mean equation
Constant	−0.006	0.102	0.108
	[0.075]	[0.089]	[0.089]
Technology market index	0.907***	0.914***	0.915***
	[0.045]	[0.045]	[0.045]
After event (0 = ’no’, 1 = ’yes’)		−0.278**
		[0.135]
Daily attention			−0.063**
			[0.029]
Panel B: Variance equa tion
ω	0.001	0.002	0.002
	[0.017]	[0.016]	[0.016]
α	0.000	0.000	0.000
	[0.007]	[0.007]	[0.007]
β	0.999***	0.999***	0.999***
	[0.000]	[0.000]	[0.000]
Panel C: Distribution parameters
λ - skewness	−0.251	−0.278	−0.272
	[0.186]	[0.207]	[0.206]
ξ - kurtosis	1.618***	1.636***	1.636***
	[0.271]	[0.298]	[0.298]
Panel D: Model diagnostics
Persistence	0.999	0.999	0.999
ρ_x(1)	−0.012	−0.023	−0.024
ρ_x²(1)	−0.036	−0.036	−0.036
Sign bias test (p-value)	0.147	0.133	0.152
Log-Likelihood	−669.3	−667.6	−667.4
R²	51.1%	51.6%	51.6%
AIC	3.716	3.712	3.711
BIC	3.791	3.798	3.797
HQ	3.746	3.746	3.745

We also observe that while volatility is highly persistent for both portfolios, the EdTech sector has lower kurtosis. The fit of the model for the EdTech sector is also much higher. These results imply that the two subsectors have distinctive price dynamics that are captured for the whole sample period and the after-event period. In the next section, we will explore the differences between the pricing dynamics in greater detail.

Threshold Regression Framework

Education Sector

Table 5 presents the results from the threshold regression models for the education sector. We apply two models, with the first model using the lagged Twitter-based attention as the state variable, meaning that the coefficients of the linear regression (part) will change depending on the level of interest in the population. The threshold is estimated to be 2 tweets. Hence, the dynamics between the educational sector returns and the broader market index change when attention goes beyond this level. This closely corresponds to the time when ChatGPT was released to the public. In the low-attention regime, the sensitivity to the market is 0.769, and the constant, although positive, does not show abnormal positive returns at 0.115. This indicates that the returns in the low-attention regime comove strongly with those of the benchmark technology market index. In the high-attention regime, however, the coefficient loaded on the market index drops significantly by −0.227 to 0.542. The change is of substantial magnitude. Moreover, the constant in the second regime is −0.109. Although the coefficient in the model is insignificant, the change from 0.115 to −0.109 (a change of −0.224) is significant. We thus observe not only a decoupling of the education sector from the technology market index but also a decline in the expected returns, where both effects are associated with increased attention.

Table 5.

Threshold Market model - Education sector.

Models	1			2
Panel A: Threshold of the state-variable
Daily attention - tweets	2
Time				21.10.22
Panel B: Mean equatio n
Regimes	Low	High	Δ	Before	After	Δ
Constant	0.115	−0.109	−0.224**	0.102	−0.057	−0.159
	[0.074]	[0.090]	[0.113]	[0.070]	[0.094]	[0.111]
Technology market index	0.769***	0.542***	−0.227***	0.795***	0.535***	−0.260***
	[−0.224]	[0.051]	[0.065]	[0.038]	[0.053]	[0.068]
Panel C: Variance equation
ω	0.022			0.019
	[0.017]			[0.011]
α	0.000			0.000
	[0.006]			[0.011]
β	0.984			0.987
	[0.008]			[0.004]
Panel D: Distribution parameters
λ - skewness	−0.164			−0.199
	[0.149]			[0.140]
ξ - kurtosis	1.529			1.572
	[0.244]			[0.267]
Panel E: Model diagnostics
Persistence	0.984			0.987
ρ_x(1)	−0.022			−0.025
ρ_x²(1)	−0.035			−0.035
Sign bias test (p-value)	0.771			0.754
Log-Likelihood	−564.4			−562.7
R²	54.8%			55.3%
AIC	3.151			3.141
BIC	3.247			3.237
HQ	3.189			3.179

Notes: The* ** and *** denote statistical significance at the 10%, 5% and 1% significance levels. Δ denotes changes from the low (before) to the high (after) regime. Standard errors are heteroscedasticity consistent estimates of White (1982) ρ_x(1) $(ρ_{x^{2}} (1))$ denotes first-order auto-correlation coefficients of (squared) standardized residuals. The Sign bias test corresponds to the p-value of the joint test of Engle & Ng (1993). R² is the squared correlation coefficient between observed and fitted observations.

In Model 2, we estimate the same model specification but now use the linear time trend rather than the number of tweets as the state variable. The threshold date is estimated to be October 21, 2022. During the ’Before’ period, which corresponds to the observations before October 21, 2022, we observe higher comovement with the market, while after that period, not only the comovement but also the expected returns have declined. These results are comparable to the attention-based results of the first specification.

Concerning model diagnostics, both models show high persistence of the latent volatility of the shocks, with significant kurtosis of the shocks. Overall, the results allow us to associate structural changes in the pricing dynamics to both attention and the period around the release of ChatGPT to the public.

The Traditional Education and EdTech Subsectors

Finally, we associate changes concerning the traditional education and EdTech subsectors in Tables 6 and 7. As in the linear case, the two subsectors behave differently. The traditional education subsector (see Panel A of Table 6) shows that the optimal threshold for the attention to be 2 with a decline in the expected return and in the sensitivity to the technology market index for higher attention, although these changes in the coefficients were insignificant. On the other hand, with the EdTech subsector (see Panel A of Table 7), the estimated optimal threshold was much higher, at 102 tweets, and the model shows a considerable decline in the expected returns from 0.123 in the low-attention period to −0.401 in the high-attention period. The change is statistically significant. The EdTech sector closely followed the technology market index in the low-attention period, with the sensitivity coefficient being close to one at 0.942. In the high-attention period, the parameter measuring comovement declined to 0.754, showing signs of decoupling. However, the decline of −0.188 is not significant. The increased attention is thus associated with substantially more negative expectations toward the EdTech subsector than the traditional education sector.

Table 6.

Threshold Market model - Traditional Education Sector.

Models	1			2
Panel A: Threshold of the state-variable
Daily attention - tweets	2
Time				3.2.23
Panel B: Mean e quation
Regimes	Low	High	Δ	Before	After	Δ
Constant	0.098	−0.050	−0.148	0.100	−0.141	−0.241*
	[0.080]	[0.090]	[0.120]	[0.074]	[0.119]	[0.146]
Technology market index	0.437***	0.374***	−0.063	0.430***	0.373***	−0.057
	[−0.148]	[0.061]	[0.074]	[0.038]	[0.113]	[0.120]
Panel C: Variance equation
ω	0.001			0.024
	[0.013]			[0.02]
α	0.000			0.000
	[0.010]			[0.014]
β	0.999***			0.982***
	[0.000]			[0.003]
Panel D: Distribution parameters
λ - skewness	0.295			0.329
	[0.297]			[0.332]
ξ - kurtosis	2.200***			2.302***
	[0.391]			[0.571]
Panel E: Model diagnostics
Persistence	0.999			0.982
ρ_x(1)	−0.053			−0.059
$ρ_{x^{2}} (1)$	0.134			0.143
Sign bias test (p-value)	0.050			0.024
Log-Likelihood	−569.2			−568.6
R²	30.6%			31.0%
AIC	3.177			3.173
BIC	3.273			3.270
HQ	3.215			3.212

Table 7.

Threhold Market model - EdTech Sector.

Models	1			2
Panel A: Threshold of the state-variable
Daily attention - tweeets	102
Time				28.10.22
Panel B: Mean equation
Regimes	Low	High	Δ	Before	After	Δ
Constant	0.123	−0.401**	−0.523***	0.109	−0.095	−0.204
	[0.077]	[0.167]	[0.185]	[0.088]	[0.109]	[0.134]
Technology market index	0.942*	0.754***	−0.188	1.012***	0.679***	−0.333***
	[−0.523]	[0.129]	[0.144]	[0.050]	[0.080]	[0.100]
Panel C: Variance equation
ω	0.007			0.001
	[0.021]			[0.012]
α	0.000			0.000
	[0.007]			[0.005]
β	0.997***			0.999***
	[0.001]			[0.000]
Panel C: Distribution parameters
λ - skewness	−0.451			−0.398*
	[0.289]			[0.229]
ξ - kurtosis	1.680***			1.744***
	[0.245]			[0.301]
Panel D: Model diagnostics
Persistence	0.997			0.999
ρ_x(1)	−0.027			−0.011
ρ_x²(1)	−0.034			−0.034
Sign bias test (p-value)	0.299			0.081
Log-Likelihood	−668.9			−661.8
R²	50.7%			53.0%
AIC	3.942			3.686
BIC	4.042			3.782
HQ	3.982			3.724

After switching the state variable to a linear time trend, the optimal threshold was identified to be on February 3, 2023, for the traditional education sector. In Figure 1, we observe that around February–March 2023, the traditional education portfolio started to decouple from the technology market index. Furthermore, we note a sharp decline in the cumulative abnormal returns depicted in the upper-right panel of Figure 1. After the threshold period, the returns of the traditional education sector underperformed and became slightly less sensitive to the development of the technology market index. The optimal threshold for the EdTech sector is estimated to be October 28, 2022, and is associated with a shift in expected returns from 0.109 to −0.095, albeit statistically insignificant. However, the sensitivity to the technology market index decreased from 1.012 to 0.679, a significant −0.333 decline. Moreover, we also observe that the shocks (see Panel C of Table 7) show high kurtosis and are also skewed to the left, that is, extreme negative shocks are more likely.

To summarize our main empirical results, the linear CAPM-GARCH models show that upon the public release of ChatGPT, the expected returns declined in the education sector. Moreover, the increased attention of the general public toward ChatGPT was followed by lower returns the next day, on average. The nonlinear threshold CAPM-GARCH model associated the changing pricing dynamics of the education portfolios with increased attention. The threshold CAPM-GARCH models also reveal that ChatGPT has a much stronger association with the EdTech sector than with the traditional education sector. Overall, the evidence shows that investors’ perception of the education sector, particularly the EdTech subsector, changed considerably after the release of ChatGPT and with increased attention toward ChatGPT.

Discussion and Robustness Analysis

In this section, we discuss the sensitivity of the results reported in the previous section concerning i) alternative model specifications, ii) use of equally weighted portfolios, iii) use of a different market index, and iv) assuming different variance equations and error distribution assumptions. All the results from these analyses are available upon request.

Within-Portfolio Variation

Part of the variation in portfolio returns might be driven by a few stocks in the portfolio. For example, returns of a few larger companies in the education sector might be declining, which impacts our conclusions for the whole (sub) sector. The heterogeneity of cross-sectional returns can be captured by the cross-sectional absolute deviation:

C A S D_{t} = N^{- 1} \sum_{i = 1}^{N} | R_{i, t} - {\bar{R}}_{t} |

(5)

The larger the CASD_t, the more distinct are returns in the given (sub)sector, and a positive (negative) coefficient would imply that when portfolio returns increase (decrease), the change is driven by only a few companies. A non-significant coefficient implies herding behavior instead, that is returns in the portfolio behave similarly (lower heterogeneity/variance of returns). We estimate the CASD_t for all three portfolios and use both current and lagged cross-sectional deviations in the mean equation of the CAPM-GARCH model:

R_{t} = γ_{0} + γ_{1} R_{m, t} + γ_{2} \ln A_{t - 1}^{D} + γ_{3} C A S D_{t} + γ_{4} C A S D_{t - 1} + ϵ_{t}

(6)

Regardless of the education portfolio employed, the γ₄ coefficient for the lagged CASD_t−1 was not significant. However, the current CASD_t was negative and significant for the whole education sector, and an even stronger effect was found for the traditional education sector. Hence, lower returns were associated with increased cross-sectional variation, which suggests that such portfolio returns are driven by a few companies (i.e., no herding behavior within the portfolio). However, the γ₂ coefficient remained negative and significant for the overall education sector and the EdTech sector at −0.051 and −0.064, while it remained insignificant for the traditional education sector.

Equally Weighted Portfolios

Our main results are based on the market capitalization weighted portfolios (see Equations (1) and (2)), and the results might be driven by large companies only. We reestimate the linear CAPM-GARCH models with equally weighted portfolio returns. Our results hold and are very close to our initial estimates reported in Tables 2 –4.

Under the nonlinear threshold CAPM-GARCH model, with attention being the state variable, we obtain results that do not change our main conclusions. For the education sector, the optimal threshold moved from 2 to 101, and we find higher decoupling for the high-attention period and a larger decline in expected returns. The results for the traditional education sector show changes in a similar direction but are not significant, similar to the main results. For the EdTech sector, we find the same optimal threshold at 102 daily tweets, but the sensitivity of returns does not change between the low- and high-attention regimes. On the other hand, we find a larger decline in expected returns. Hence, the results still show that the EdTech sector was affected negatively and the most among the portfolios.

The Benchmark Market Index: S&P 500

Our main results are based on the use of the NASDAQ-100 market index, which is overrepresented by growth-oriented, technology-driven companies. As an alternative, we reestimate all our models with the S&P 500 as the benchmark market index. Although the numerical results are most affected by this change in the market index, in summary, our main conclusions hold. The EdTech subsector shows a decline in expected returns associated with the event and increased attention in these analyses.

Specifically, for the linear CAPM-GARCH model, we find higher comovement across all education sectors, that is, the education sector is more sensitive to changes in the S&P 500 than to changes in the NASDAQ-100. We find a significant decline in expected returns after the public announcement of ChatGPT for the education and EdTech sectors and also a negative and significant coefficient for the lagged attention to the ChatGPT variable with similar magnitudes as reported in the main results.

Under the threshold model with the attention measure as the state variable, we find that for the education and EdTech portfolios, the optimal threshold is 102 daily tweets. In the high-attention regime, we do not find significant changes in comovement but a larger decline in the expected returns. The price dynamics of the traditional education subsector have not changed significantly. This shows that the choice of the benchmark index has not changed our main conclusions.

Alternative Error Distributions and Variance Equations

In the linear CAPM-GARCH models, we observed a high level of volatility persistence driven by the persistence of latent volatility. For such empirical scenarios, a suitable alternative might be the integrated GARCH model of Engle and Bollerslev (1986). Such models led to similar results, with the EdTech subsector having slightly larger effects. We also considered the asymmetric model (Glosten et al., 1993), but we found only occasionally significant negative asymmetric volatility effects, and the results again remained unchanged.

Finally, we considered two additional error distributions, the skewed Student-t and the skewed generalized error distribution (see Fernández & Steel, 1998). Our results remained robust also when using these alternative error distributions.

Alternative Assignment to EdTech and Traditional Education Subsectors

The assignment of companies into EdTech and Traditional education subsectors is to some extent subjective as not all products and services of a company are necessarily part of the EdTech or traditional education sector. However, not all companies have the same impact on the subsector’s portfolio returns. Therefore, we decided to study how our results change after we exclude the two largest (in terms of market capitalization) companies from each subsector (one at a time), as these companies have the highest impact on the returns of the respective subsector. This accounts for company choice uncertainty but also for assignment uncertainty of the largest companies.

For the EdTech subsector the two companies are Grand Canyon Education (LOPE) and PowerSchool Holdings (PWSC). Removing LOPE or PWSC led only to minor changes in the results, and not of qualitative importance. For example, the threshold date moved to October 26, 2022, after excluding LOPE from EdTech, as opposed to October 28, 2022. For the traditional education subsector, removing Graham Holdings Company (GHC) was also inconsequential with the threshold date moving to November 11, 2022. Otherwise, the results were qualitatively unchanged. The removal of Strategic Education did not lead to any major changes in the single equation model for the traditional education sector. Neither the ChatGPT event dummy nor the attention measure has a significant effect in this model. However, the threshold model now suggests that the most likely regime shift occurred on May 26, 2022. Yet there are only minor changes between the before and after regimes, which is consistent with our previous findings.

Conclusion and Implications

In this study, we argue that ChatGPT has the potential to disrupt the education sector. We use data from the U.S. stock markets to test the impact of the public release of ChatGPT in late 2022 on the price behavior of the education sector and two subsectors that we created: i) the traditional education sector and ii) the EdTech sector. The former is represented by companies that operate campuses and universities and offer traditional services in the education sector. The latter includes firms that operate in the intersection between education and technology.

Using a sample from January 2022 until June 2023, we first find that all three portfolios (overall, traditional, and EdTech) underperformed the benchmark market index after the public release of ChatGPT. However, using a linear CAPM-GARCH model with flexible error distribution, we were unable to confirm that the underperformance of the traditional education subsector is associated with the event or with the attention of the general public toward ChatGPT. Our second result is that contrary to the traditional education sector alone, the overall education and EdTech subsector show that i) after the event, expected returns significantly declined and that ii) increased attention toward ChatGPT is associated with a decrease in next day’s expected returns. These results are much stronger for the EdTech sector, which shows that ChatGPT will likely have a larger impact on technology-driven education companies (e.g., Udemy, Coursera, Chegg, PowerSchool).

We complement our analysis using a nonlinear threshold CAPM-GARCH, where the mean equation coefficients are subject to a structural change as a function of the attention toward ChatGPT (or linear time trend). The analysis reveals our third main result that for the overall education and particularly EdTech subsector, the expected returns decline in the high-attention regime. We also find that returns of the education sector and its subsectors decouple with the benchmark market return, possibly indicating disruption of a long-run relationship. This result was not confirmed when using the S&P 500 instead of the NASDAQ-100 as the benchmark market index. Although the results under the S&P 500 indicated a larger decline in the expected returns in high-attention regimes and after the announcement of ChatGPT. Overall, our main results remained robust when considering different mean and variance equation specifications and changes from the market capitalization weighted education portfolios to equally weighted portfolios.

Our results suggest that, at least from a short-term perspective, investors expect that the introduction of ChatGPT and the resulting revolution of LLMs poses challenges to the EdTech subsector.¹¹ Such challenges can be attributed to the threat that some of the services and products will be replaced. For example, language polishing and translation services are being routinely replaced by LLMs already. It might also be the case that new opportunities emerge from ChatGPT and related AI technologies (such as competing LLMs, like Microsoft copilot, for example). Companies may, for example, integrate new tools into their existing toolbox to increase student learning beyond what the stand-alone chatbot applications can offer. Innovations thus create threats and opportunities as well. Our analysis is short-term in nature and the market may well have overreacted though (see, e.g., Howe, 1986). The situation could thus turn around when investors have more information about the potential future impact of the new technology. Yet there are good reasons to believe that the entry of ChatGPT and similar LLMs have the potential to disrupt the educational sector completely. Future studies should, therefore, focus on replicating and extending the present study when more data becomes available.

Disclosure

Statement: During the preparation of this work, the authors used ChatGPT in an interview to obtain insights into how the new tool could affect the education sector. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Supplemental Material

Supplemental Material - Has ChatGPT Disrupted the Education Sector in the U.S.?

Supplemental Material for Has ChatGPT Disrupted the Education Sector in the U.S.? by Erik Haugom, Štefan Lyócsa and Martina Halousková in Social Science Computer Review

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by the Grantová Agentura České Republiky, nr. 22-27075S.

ORCID iD

Martina Halousková

Supplemental Material

Supplemental material for this article is available online.

Notes

Author Biographies

Erik Haugom was born in Lillehammer, Norway, on February 7, 1982. He holds a PhD in Managerial Economics, Finance and Operations from the Norwegian University of Science and Technology and a master’s degree in Economic Marketing from Copenhagen Business School. He is employed as a Professor at Inland Norway University of Applied Sciences. His special fields of interest include pricing analytics, energy price and volatility modeling and forecasting, risk analysis, risk management, and econometric modeling and forecasting in general.

Štefan Lyócsa was born in Prešov, Slovakia, on February 16, 1982. He holds a PhD in Business Economics from the University of Economics in Bratislava, Slovakia. He is a Professor of Finance at the Masaryk University, Brno, Czech Republic and a researcher at the Slovak Academy of Sciences in Bratislava. His research interests are divided into i) market risks (volatility, value-at-risk, expected shortfall), specifically the role that sentiment and attention is playing in modelling market risks, ii) credit risks, designing more accurate credit-risk models. He is also interested in inter-disciplinary research and data-science (e.g. labor market, regional disparities, behavior aspects of decision making, conformity to social norms).

Martina Halousková was born in Žilina, Slovakia on December 2, 1996. She is a PhD candidate in Finance at Masaryk University in Brno, Czech Republic. She holds a master's degree in Corporate Finance from University of Economics, Bratislava. Her dissertation topic is on volatility forecasting, particularly on the stock market, with an emphasis on (i) the role of investor sentiment and attention and (ii) on the role impactful market events. She is also interested in techniques of natural language processing and text analysis and their applications in various subfields of finance research (e.g. asset pricing, market risk predictions and labour market).

References

Ali

(2023, April). GPT-1 to GPT-4: Each of OpenAI’s GPT Models Explained and Compared. Retrieved 7 August 2023, from https://www.makeuseof.com/gpt-models-explained-and-compared/

Andrei

Hasler

(2015). Investor attention and stock market volatility. Review of Financial Studies, 28(1), 33–72. https://doi.org/10.1093/rfs/hhu059

Aouadi

Arouri

Teulon

(2013). Investor attention and stock market activity: Evidence from France. Economic Modelling, 35(1), 674–681. https://doi.org/10.1016/j.econmod.2013.08.034

Audrino

Sigrist

Ballinari

(2020). The impact of sentiment and attention measures on stock market volatility. International Journal of Forecasting, 36(2), 334–357. https://doi.org/10.1016/j.ijforecast.2019.05.010

Ballinari

Audrino

Sigrist

(2022). When does attention matter? The effect of investor attention on stock market volatility around news releases. International Review of Financial Analysis, 82. https://doi.org/10.1016/j.irfa.2022.102185

Barber

B. M.

Odean

(2008). All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors. Review of Financial Studies, 21(2), 785–818. https://doi.org/10.1093/rfs/hhm079

Bhaimiya

(2023). Chatgpt could hypothetically get hired as an entry level coder if it interviewed at google, internal document reportedly says. https://www.businessinsider.com/chatgpt-google-could-get-hired-as-coder-internal-document-report-2023-2?r=US&IR=T

Bijl

Kringhaug

Molnár

Sandvik

(2016). Google searches and stock returns. International Review of Financial Analysis, 45(1), 150–156. https://doi.org/10.1016/j.irfa.2016.03.015

Bollerslev

(1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics, 31(3), 307–327. https://doi.org/10.1016/0304-4076(86)90063-1

10.

Bordino

Battiston

Caldarelli

Cristelli

Ukkonen

Weber

(2012). Web search queries can predict stock market volumes. PLoS One, 7(7), Article e40014. https://doi.org/10.1371/journal.pone.0040014

11.

Buchholz

(2023, July). Threads Shoots Past One Million User Mark at Lightning Speed. Retrieved 4 August 2023, from https://www.statista.com/chart/29174/time-to-one-million-users/

12.

Choi

Nam

(2008). Asymmetric and leptokurtic distribution for heteroscedastic asset returns: The su-normal distribution. Journal of Empirical Finance, 15(1), 41–63. https://doi.org/10.1016/j.jempfin.2006.06.009

13.

Dowling

Lucey

(2023). Chatgpt for (finance) research: The bananarama conjecture. Finance Research Letters, 53(1), 103662. https://doi.org/10.1016/j.frl.2023.103662

14.

Elias

(2023, January 31). Google is asking employees to test potential ChatGPT competitors, including a chatbot called ‘Apprentice Bard’. Retrieved 8 August 2023, from https://www.cnbc.com/2023/01/31/google-testing-chatgpt-like-chatbot-apprentice-bard-with-employees.html

15.

Engle

R. F.

Bollerslev

(1986). Modelling the persistence of conditional variances. Econometric Reviews, 5(1), 1–50. https://doi.org/10.1080/07474938608800095

16.

Engle

R. F.

V. K.

(1993). Measuring and testing the impact of news on volatility. The Journal of Finance, 48(5), 1749–1778. https://doi.org/10.2307/2329066

17.

Escanciano

J. C.

Lobato

I. N.

(2009). An automatic portmanteau test for serial correlation. Journal of Econometrics, 151(2), 140–149. https://doi.org/10.1016/j.jeconom.2009.03.001

18.

Fernández

Steel

M. F.

(1998). On bayesian modeling of fat tails and skewness. Journal of the American Statistical Association, 93(441), 359–371. https://doi.org/10.2307/2669632

19.

Freberg

(2023). 7 ways to use chatgpt for analytics. https://www.youtube.com/watch?v=XXXXXX

20.

Frieder

Pinchetti

Griffiths

R.-R.

Salvatori

Lukasiewicz

Petersen

P. C.

Chevalier

Berner

(2023). Mathematical capabilities of chatgpt. arXiv preprint arXiv:2301.13867.

21.

Ghalanos

(2020) Introduction to the Rugarch Package. Version 1.3-1, Technical Report V. http://cran.r-project.org/web/packages

22.

Glosten

L. R.

Jagannathan

Runkle

D. E.

(1993). On the relation between the expected value and the volatility of the nominal excess return on stocks. The Journal of Finance, 48(5), 1779–1801. https://doi.org/10.1111/j.1540-6261.1993.tb05128.x

23.

Goddard

Kita

Wang

(2015). Investor attention and fx market volatility. Journal of International Financial Markets, Institutions and Money, 38(3), 79–96. https://doi.org/10.1016/j.intfin.2015.05.001

24.

Guidolin

Pedio

(2021). Media attention vs. sentiment as drivers of conditional volatility predictions: An application to brexit. Finance Research Letters, 42(5), 101943. https://doi.org/10.1016/j.frl.2021.101943

25.

Hamid

Heiden

(2015). Forecasting volatility with empirical similarity and google trends. Journal of Economic Behavior & Organization, 117(5), 62–81. https://doi.org/10.1016/j.jebo.2015.06.005

26.

Hansen

(2022). Econometrics. Princeton University Press.

27.

Herman

(2022). The end of high-school English. The Atlantic (09 Dec 2022).

28.

Howe

J. S.

(1986). Evidence on stock market overreaction. Financial Analysts Journal, 42(4), 74–77. https://doi.org/10.2469/faj.v42.n4.74

29.

Johnson

N. L.

(1949a). Bivariate distributions based on simple translation systems. Biometrika, 36(3-4), 297–304.

30.

Johnson

N. L.

(1949b). Systems of frequency curves generated by methods of translation. Biometrika, 36(Pt. 1-2), 149–176.

31.

Kasneci

Seßler

Küchemann

Bannert

Dementieva

Fischer

Gasser

Groh

Günnemann

Hüllermeier

Krusche

Kutyniok

Michaeli

Nerdel

Pfeffer

Poquet

Sailer

Schmidt

Seidel

Kasneci

(2023). Chatgpt for good? On opportunities and challenges of large language models for education. Learning and Individual Differences, 103(3), 102274. https://doi.org/10.1016/j.lindif.2023.102274

32.

Kirmani

A. R.

(2022). Artificial intelligence-enabled science poetry. ACS Energy Letters, 8(1), 574–576. https://doi.org/10.1021/acsenergylett.2c02758

33.

Lee

S. E.

Lee

K.-H.

(2023). Service chatbot: Co-Citation and big data analysis toward a review and research agenda. Technological Forecasting and Social Change, 194(4), 122722. https://doi.org/10.1016/j.techfore.2023.122722

34.

Fan

Lee

Hemphill

(2023). Chatgpt in education: A discourse analysis of worries and concerns on social media. Education and Information Technologies, 29(9), 10729–10762. https://doi.org/10.1007/s10639-023-12256-9

35.

C. K.

(2023). What is the impact of chatgpt on education? A rapid review of the literature. Education Sciences, 13(4), 410. https://doi.org/10.3390/educsci13040410

36.

Lund

B. D.

Wang

(2023). Chatting about chatgpt: How may AI and gpt impact academia and libraries? Library Hi Tech News.

37.

Lyócsa

Š.

Halousková

Haugom

(2023). The US banking crisis in 2023: Intraday attention and price variation of banks at risk. Finance Research Letters, 57(2), 104209. https://doi.org/10.1016/j.frl.2023.104209

38.

Lyócsa

Š.

Molnár

Výrost

(2021). Stock market volatility forecasting: Do we need high-frequency data? International Journal of Forecasting, 37(3), 1092–1110. https://doi.org/10.1016/j.ijforecast.2020.12.001

39.

Lyócsa

Š.

Plíhal

(2022). Russia’s ruble during the onset of the Russian invasion of Ukraine in early 2022: The role of implied volatility and attention. Finance Research Letters, 48(1), 102995. https://doi.org/10.1016/j.frl.2022.102995

40.

Memarian

Doleck

(2023). Chatgpt in education: Methods, potentials and limitations. Computers in Human Behavior: Artificial Humans, 1(2), 100022. https://doi.org/10.1016/j.chbah.2023.100022

41.

Staudt Willet

K. B.

Shi

Hur

Kim

(2024). Initial discussions of chatgpt in education-related subreddits. Journal of Research on Technology in Education, 1(2), 1–18. https://doi.org/10.1080/15391523.2024.2338091

42.

Preis

Reith

Stanley

H. E.

(2010). Complex dynamics of our economic life on different scales: Insights from search engine query data. Philosophical Transactions of the Royal Society A: Mathematical, Physical & Engineering Sciences, 368(1933), 5707–5719, https://doi.org/10.1098/rsta.2010.0284

43.

Saggu

Ante

(2023). The influence of chatgpt on artificial intelligence related crypto assets: Evidence from a synthetic control analysis. Finance Research Letters, 55(2), 103993. https://doi.org/10.1016/j.frl.2023.103993

44.

Smales

L. A.

(2021). Investor attention and global market returns during the COVID-19 crisis. International Review of Financial Analysis, 73(2), 101616. https://doi.org/10.1016/j.irfa.2020.101616

45.

Surameery

N. M. S.

Shakor

M. Y.

(2023). Use chat gpt to solve programming bugs. International Journal of Information Technology & Computer Engineering (IJITC) ISSN: 2455-5290, 3(31), 17–22. https://doi.org/10.55529/ijitc.31.17.22

46.

Vlastakis

Markellos

R. N.

(2012). Information demand and stock market volatility. Journal of Banking & Finance, 36(6), 1808–1821. https://doi.org/10.1016/j.jbankfin.2012.02.007

47.

Wardat

Tashtoush

M. A.

AlAli

Jarrah

A. M.

(2023). Chatgpt: A revolutionary tool for teaching and learning mathematics. Eurasia Journal of Mathematics, Science and Technology Education, 19(7), em2286. https://doi.org/10.29333/ejmste/13272

48.

White

(1982). Maximum likelihood estimation of misspecified models. Econometrica: Journal of the Econometric Society, 50(9), 1–25. https://doi.org/10.2307/1912526

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.40 MB