AI and Finance
Generative AI has emerged as a major technology that is disrupting both the finance industry and financial research methodologies. The release of ChatGPT in November 2022, followed by rapid advancements in large language models (LLMs), has led to changes in firm valuations, hiring, and profitability, as well as changes in occupational earnings. Our research shows that ChatGPT has a wide impact across firms, affecting corporate outcomes in all industries, not just the technology sector. We also show that the impact varies widely both across and within industries. In this post, which is based on our working paper, AI and Finance, we summarize research exploring how Generative AI tools influence firm value, firm decisions, and financial research. We also offer practical insights on how generative AI can enhance financial research, as well as asset management and corporate finance decisions. 
Generative AI as a Technology Shock to Firms
The launch of ChatGPT in November 2022 marks a pivotal technology shock that has significantly influenced firm values across all sectors, including finance. Generative AI is expected to alter firm valuations because, while GPT technically stands for “Generative Pre-Trained Transformer,” ChatGPT is also a GPT in the sense of a “General Purpose Technology” that enables a wide range of occupations to become more productive.
In our research in The Labor Impact of Generative AI on Firm Values, we develop a methodology for measuring firm-level exposures to this increase in productivity. For each task that is normally done in different occupations (defined by the O*NET database) we determine its potential for automation or augmentation by Generative AI. Using LinkedIn data on the composition of public firms’ employment, we then aggregate these task-level exposures into a measure of Generative AI Exposure at the firm level, which reflects what share of tasks in a firm can be made more productive using Generative AI tools. Moreover, we distinguish whether the tasks for which Generative AI can be used are the core tasks that a worker does in their job, or are supplemental tasks which are not key for that worker’s function within the firm.
An example of the resulting of our measured Generative AI exposures for different groups of white-collar workers is shown in Figure 2 below: occupations that involve more interpersonal contact or manual tasks (e.g. health care) are less exposed than occupations that are mainly based on entering texts or commands into computers (e.g. programmers).
Figure 2: Share of Tasks Exposed to Generative AI Across Selected Occupation Groups
We show in our research that financial markets seem to have revalued firms in line with our measure of exposure to the technology shock. That is, the value of firms that we would predict to have a greater productivity increase because of the technology saw a large jump in their stock prices after ChatGPT was released on November 30, 2022, compared to less-exposed firms.
The ‘Artificial Minus Human’ portfolio, which goes long the “Artificial” quintile of firms heavily exposed to Generative AI and shorts those with minimal exposure (the “Human” quintile), would have earned 44 basis points in daily returns during the two-week period following the ChatGPT release. This out-performance is shown in Figure 3: a gap in valuations opens up right after the release, and widens even further as more capabilities of the technology are revealed with the release of GPT 4 in March 2023.
Figure 3: Cumulative Abnormal Returns by Generative AI Exposure: This figure plots the cumulative abnormal returns of value-weighted quintile portfolios sorted by firms’ Generative AI exposure.
In that research, we also document that Generative AI exposure is associated with a decline in hiring by firms for the most affected roles—especially if Generative AI impacts the core tasks of the workers.
Generative AI’s Impact on Financial Research
Generative AI is also transforming financial research in several practical and impactful ways. Below, we outline key areas in which Generative AI is driving innovation and improving research methodologies, which are also summarized in Table 1 below. In the paper, we provide examples from recent research and practical advice for implementing these method:
1. Text Classification
LLMs have demonstrated powerful capabilities in processing and analyzing vast amounts of unstructured data. This capability allows researchers to efficiently analyze financial documents, such as earnings call transcripts, regulatory filings, and news articles. By using transformer-based models like GPT-4, financial researchers can extract insights about firm performance, investor sentiment, and qualitative firm characteristics that would have previously been prohibitively expensive to do at scale with human research assistants.
2. Embeddings
Embeddings generated by transformer models are increasingly being used to capture the semantic relationships within financial documents. These embeddings translate high-dimensional, complex data into lower-dimensional vectors that still retain critical meaning, enabling new forms of analysis. For example, embeddings can reveal similarities between companies based on their regulatory filings, facilitating clustering and benchmarking exercises.
3. Retrieval-Augmented Generation (RAG)
RAG models combine LLMs with an information retrieval mechanism to enable researchers to find specific information across vast financial datasets. This method is particularly useful when researchers need targeted information that is spread across multiple lengthy documents, such as regulations, corporate filings, or market reports.
4. Simulating Agent Behavior
LLMs can simulate human behaviors and responses from the perspective of different groups of respondents, such as preferences across different asset classes or responses in surveys. By simulating how agents might act under various counterfactual scenarios, researchers can obtain a preliminary measure of expected behavior in response to complex events or policy changes. These simulated responses can then inform research design, guide theory development, or serve as benchmarks for empirical studies.
5. Hypothesis Generation
Generative AI also plays a role in generating new research hypotheses and testing preliminary ideas: By leveraging the vast knowledge embedded within LLMs, researchers can rapidly generate potential hypotheses, identify potential research design flaws, and explore preliminary patterns within their data, which can then be validated using traditional econometric or experimental methods.
Generative AI holds significant potential for automating supplementary research tasks, such as writing code, drafting literature reviews, summarizing findings, and proofreading papers. These capabilities enable researchers to focus on more complex analytical work, teaching, communication of their results, and on research design tasks, while leaving repetitive, lower-level tasks to Generative AI models.
Table I: Applications of Generative AI in Finance Research
| Application | Question Types | Examples |
| Embeddings | – How to represent complex data concisely? – What are semantic relationships in data? – How to cluster similar entities? |
– Gabaix et al. (2023): Asset embeddings – Chen and Sarkar (2020): 10-K filings embeddings – Kim et al. (2024): Labor market clusters |
| Text Classification | – What is the sentiment of financial text? – How to categorize documents? – What topics are discussed? |
– Chang et al. (2024): Earnings call sentiment – Krockenberger et al. (2024): Covenant violations – Caragea et al. (2020): FinTech patent classification |
| Retrieval-Augmented Generation (RAG) | – How to find relevant information in large datasets? – How to use LLMs for classification based on many potential sources? – How to retrieve similar documents from a corpus? |
– Bartik et al. (2023): Housing regulation classification – Chen and Wang (2024): AI patent assignment to functionalities |
| Simulating Agent Behavior | – Can LLMs replicate human preference heterogeneity? – What are expected survey responses? – What would human expectations be in counterfactual scenarios? |
– Fedyk et al. (2024): Asset class preferences – Bybee (2023): Macroeconomic expectations – Hewitt et al. (2024): Experimental outcome predictions |
| Hypothesis Generation | – How to generate new research ideas or business ideas? – How to conduct qualitative research at scale? |
– Si et al. (2024): Novel research idea generation – Girotra et al. (2023): New product ideas – Ludwig and Mullainathan (2024): Feature importance in neural network predictions |
Conclusion
Generative AI represents a profound technology shock that is driving major changes in the economy and in financial research. The initial effects on firm value when ChatGPT was released likely only reflected a fraction of the economic value that will be generated by that first wave of capabilities and the even more powerful models that have followed since then. For researchers, Generative AI opens up new frontiers in terms of research productivity and large-scale data analysis by enabling workflows that will reshape the way research is done.
Legal Disclaimer:
EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
