CFOs Meet LLMs
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "CFOs Meet LLMs".
Warren: Here's the gist of the paper. Business sentiment is a closely watched economic signal, but measuring it is slow and costly: surveys typically reach only a few hundred firms, arrive periodically, and take time to compile The LLM-generated optimism score predicts the individual CFO’s actual answer, even in specifications that include firm and year-quarter fixed effects as well as a control variable measuring the human CFO’s lagged response.
Ruth: The LLM framework approximates how an executive is likely to respond given their information set and prior answers, providing an individual-level benchmark for the predictable component of managerial expectations.
Warren: The authors propose and evaluate a method for generating synthetic CFO expectations with large language models (LLMs).
Ruth: They prompt the LLM to role-play as the CFO of a specific public company at a specific point in time, supplying the model with the firm’s industry, revenue, headcount, and geographic footprint.
Paper summary: Ruth: What’s interesting is that this isn't just about getting a general market feeling; it’s about linking the digital twin sentiment to the human response in a very direct way >
Warren: They found that the predictive accuracy of this LLM forecast changes depending on how much information you give it, showing a dose-response relationship >
Iris: Specifically, when you supply the LLM model with three or more prior responses from that CFO, the fit increases to nearly fifty percent because that’s when it gets really accurate >
Ruth: It’s not just about the firm profile; they found that conditioning on each CFO respondent’s own history significantly anchors the persona to the individual rather than just a group type >
Warren: And when you look at it aggregated, even after controlling for things like macroeconomic benchmarks such as the Michigan Consumer Sentiment Index and some professional forecast of real GDP growth, the mean LLM optimism remains a significant predictor of mean CFO optimism >
Iris: They also validated this individual-level prediction within a single firm by matching every synthetic score to the response of the actual CFO at that same firm in that same quarter >
Ruth: So, they’re showing that even though these individual survey responses weren't used in training the model, these predictions are out-of-corpus with respect to those specific human answers >
Warren: That individual predictive content is what they point to as notable because it goes against prior work suggesting managerial expectations are just noisy reflections of public information >
Conclusion: Ruth: So, looking at "CFOs Meet LLMs," the main thing is that this method generates synthetic CFO expectations using large language models >
Warren: The key finding is that the LLM score is a significant predictor of the actual CFO’s response at the individual level >
Iris: And accuracy scales with how much information you provide to that AI model >
Ruth: Plus, aggregate results show that mean LLM optimism stays a significant predictor of mean CFO optimism even when you control for external economic benchmarks >
Warren: The paper suggests that in the future, these large language models might be useful as synthetic executives, especially when they are conditioned on a rich firm- and respondent-level context >
Iris: This offers the prospect of a real-time, scalable individual-level sentiment measure for both practitioners and policymakers >
Duke University · Georgia State University
q-fin.CP, q-fin.GN
Submitted: 2026-06-11
Updated: 2026-11-15
Comments: 21 pages, 4 tables, 1 figure
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Business sentiment is a closely watched economic signal, but measuring it is slow and costly: surveys typically reach only a few hundred firms, arrive periodically…
Key concepts
- LLM-generated optimism score
- This is a metric produced by an LLM that simulates what a CFO would likely say. It goes beyond simple firm averages by predicting the specific answer of an individual CFO based on their company details and past responses, offering a more granular view of sentiment.
- Conditioning on Executive History
- A key technique where the LLM is fed the specific history and prior forecasts of an individual CFO. This anchors the model's prediction to that person's unique behavior rather than just a generic firm type, significantly boosting accuracy in predicting individual responses.
- Dose-Response Relationship
- This describes how increasing the amount of relevant information provided to the LLM affects its predictive accuracy. The results show that supplying more prior expectation data—specifically over three responses—monotonically increases the model's fit to actual CFO answers.
- Synthetic Executives
- The idea that LLMs can function as realistic, simulated executives. By conditioning them on rich context like firm and respondent-level details, they can serve as scalable tools for generating synthetic data to study economic sentiment in a way traditional surveys cannot.
Terminology
Summary
Business sentiment is a closely watched economic signal, but measuring it is slow and costly: surveys typically reach only a few hundred firms, arrive periodically, and take time to compile The LLM-generated optimism score predicts the individual CFO’s actual answer, even in specifications that include firm and year-quarter fixed effects as well as a control variable measuring the human CFO’s lagged response. p. 0 Graham† Campbell R. Harvey† Manish Jha‡
†Duke University and NBER ‡Georgia State University
Abstract
Business sentiment is a closely watched economic signal, but measuring it is slow
and costly: surveys typically reach only a few hundred firms, arrive periodically,
and take time to compile. (2016)
show that CFO expectations contain extrapolative components that predict future investment. Our LLM framework approximates how an executive is likely to respond given their information
set and prior answers, providing an individual-level benchmark for the predictable component
of managerial expectations. (2016)
show that CFO expectations contain extrapolative components that predict future investment. Our LLM framework approximates how an executive is likely to respond given their information
set and prior answers, providing an individual-level benchmark for the predictable component
of managerial expectations.Where the paper says it
Where the paper says it
Where the paper says it
Methodology for Synthetic Data Generation
The authors propose and evaluate a method for generating synthetic CFO expectations with large language models (LLMs). They prompt the LLM to role-play as the CFO of a specific public company at a specific point in time, supplying the model with the firm’s industry, revenue, headcount, and geographic footprint. p. 1 We make no
claim that the model reasons in the same manner as the executive. We prompt the LLM to role-play as the CFO of a specific public company at a specific point
in time, supplying the model with the firm’s industry, revenue, headcount, and geographic
footprint, and imposing an information cutoff matching the date of the actual survey wave. In this particular quarter, our prompting
strategy successfully prevents the model from incorporating knowledge of subsequent events:
a model with access to post-cutoff information would almost certainly produce markedly more
pessimistic forecasts for this period. The system prompt directs the model to adopt a comprehensive analytical framework
that mirrors the decision-making process of a financial executive at a given point in time.Where the paper says it
Where the paper says it
Conditioning on Executive History
A key methodological contribution relative to earlier persona-based work is that we condition on each CFO respondent’s own history. Supplying the model with a person’s prior answers, and with the LLM’s own prior forecasts for that person, anchors the persona to the individual rather than to a group or a generic firm type. p. 2 (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. Supplying the model with a person’s prior
answers, and with the LLM’s own prior forecasts for that person, anchors the persona to the
individual rather than to a group or a generic firm type, and we find that it limits the collapse. A parallel pattern holds for the breadth of firm characteristics provided to the LLM: prompts
that include the full company profile produce better predictions than sparse prompts. These
dose–response relationships, where richer conditioning inputs (dose) produce systematically
stronger or more accurate model outputs (response), indicate that the model is using the
conditioning information rather than guessing, and they highlight the conditions under which
synthetic executives most faithfully reproduce human responses.Where the paper says it
Where the paper says it
Predictive Accuracy and Information Dosage
The predictive accuracy of the LLM forecast varies with the information provided, showing dose–response relationships. With respect to providing the LLM model with the CFO respondent’s own history of expectations, the fit increases monotonically with the amount of information provided. p. 3 We find that more information leads to better forecasts. With respect to
providing the LLM model with the CFO respondent’s own history of expectations, the fit
increases monotonically with the amount of information provided: from anR
2 of approximately
10% when no prior CFO responses are available, to roughly 33% with one to three historic CFO
responses, to nearly 50% when more than three prior responses are supplied (Table 2, Panel B). We find that more information leads to better forecasts. With respect to
providing the LLM model with the CFO respondent’s own history of expectations, the fit
increases monotonically with the amount of information provided: from anR
2 of approximately
10% when no prior CFO responses are available, to roughly 33% with one to three historic CFO
responses, to nearly 50% when more than three prior responses are supplied (Table 2, Panel B).Where the paper says it
Where the paper says it
Robustness and Macroeconomic Benchmarking
The LLM framework introduces an LLM-based measure of CFO-level economic sentiment and provides the first individual-level, within-firm validation of LLM-simulated survey responses. The relationship between human CFO sentiment and digital twin sentiment is robust to aggregation, as the mean LLM optimism remains a significant predictor of mean CFO optimism even after controlling for macroeconomic benchmarks like the Michigan Consumer Sentiment Index and the SPF professional forecast of real GDP growth. p. 9 Macro Forecasts and Consumer Sentiment
We also compare a within-quarter aggregation of the LLM’s economy-wide optimism scores
against two established measures of macroeconomic expectations. The first is the mean yearahead real GDP growth forecast (drgdp5) from the Survey of Professional Forecasters (SPF),
administered quarterly by the Federal Reserve Bank of Philadelphia and drawing on roughly 30
to 40 professional forecasters at financial institutions, consulting firms, and academic organizations.Where the paper says it
Implications for Research and Policy
The scalability opens several research designs that were previously infeasible, such as generating daily or weekly LLM sentiment scores for a full cross-section of public firms to study policy shocks. The findings suggest that in the future, LLMs may be useful as synthetic executives, particularly when conditioned on a rich firm- and respondent-level context. p. 0 We find the
same patterns hold for two other questions measuring CFO expectations: the
respondent’s optimism about their own firm and their expectation of own-firm
revenues. With appropriate conditioning, LLMs may in the future be able to serve
as digital twins of executives, offering scalable, high-frequency expectations data
for financial research and policy. Because LLM agents need not behave like humans, individual-level fidelity is not something
we can assume, and in our setting it emerges only after we condition each persona on the
respondent’s own history, without which instruction-tuned models collapse toward modal
opinions (Santurkar et al. Our framework introduces an LLM-based measure of CFO-level economic sentiment and
provides the first individual-level, within-firm validation of LLM-simulated survey responses,
approximating private forecasts using an information set of public data and a respondent’s
own past answers. Supplying the model with a person’s prior
answers, and with the LLM’s own prior forecasts for that person, anchors the persona to the
individual rather than to a group or a generic firm type, and we find that it limits the collapse. In our sample, the within-quarter dispersion of LLM scores is about 75% of the dispersion in
the human panel.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The paper presents a method for generating synthetic CFO expectations using large language models. Four main findings emerge: the LLM score is a significant predictor of the actual CFO’s response at the individual level, predictive accuracy scales with the information provided, and aggregate results show that mean LLM optimism remains a significant predictor of mean CFO optimism even after controlling for macroeconomic benchmarks. p. 0 The LLM-generated optimism score
predicts the individual CFO’s actual answer, even in specifications that include
firm and year-quarter fixed effects as well as a control variable measuring the
human CFO’s lagged response. Accuracy increases with the information provided
to the LLM, and the relation persists under quarterly aggregation. With appropriate conditioning, LLMs may in the future be able to serve
as digital twins of executives, offering scalable, high-frequency expectations data
for financial research and policy. Keywords: Large language models; CFO survey; Managerial expectations; Business
sentiment; Synthetic surveys; Digital twins; Belief formation
JEL Classification: C55, D84, E27, G30, G41
Version: September 8, 2026.Where the paper says it
Where the paper says it
References
Aher et al. (2023) Argyle et al. p. 3 Argyle et al. (2023) show that LLM-generated synthetic survey populations closely match
the distributional properties of real human samples, and Suh et al. Argyle et al. (2023) show that LLM-generated synthetic survey populations closely match
the distributional properties of real human samples, and Suh et al. Models tuned on human feedback are known to
compress the diversity of human opinion: Santurkar et al. (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. Models tuned on human feedback are known to
compress the diversity of human opinion: Santurkar et al. (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. The first simulates a population of forecasters to recover the dispersion of beliefs rather
than any one person’s answer. Bali et al. The first simulates a population of forecasters to recover the dispersion of beliefs rather
than any one person’s answer. Bali et al. For each of 6,075 survey responses from public companies
1Fried et al. (2026) demonstrate that the Duke-Fed CFO survey tracks future inflation; and Baslandze et al. For each of 6,075 survey responses from public companies
1Fried et al. (2026) demonstrate that the Duke-Fed CFO survey tracks future inflation; and Baslandze et al. With appropriate conditioning, LLMs may in the future be able to serve
as digital twins of executives, offering scalable, high-frequency expectations data
for financial research and policy. Keywords: Large language models; CFO survey; Managerial expectations; Business
sentiment; Synthetic surveys; Digital twins; Belief formation
JEL Classification: C55, D84, E27, G30, G41
Version: September 8, 2026. With appropriate conditioning, LLMs may in the future be able to serve
as digital twins of executives, offering scalable, high-frequency expectations data
for financial research and policy. Keywords: Large language models; CFO survey; Managerial expectations; Business
sentiment; Synthetic surveys; Digital twins; Belief formation
JEL Classification: C55, D84, E27, G30, G41
Version: September 8, 2026. First, we impose information cutoffs that
instruct the model not to access post-survey data, and any respondent history we supply is itself
dated strictly before the cutoff. We find supportive evidence that this instruction was followed
in that the first-quarter 2020 LLM forecast does not anticipate the COVID-era reduction in
optimism that occurred ex post. First, we impose information cutoffs that
instruct the model not to access post-survey data, and any respondent history we supply is itself
dated strictly before the cutoff. We find supportive evidence that this instruction was followed
in that the first-quarter 2020 LLM forecast does not anticipate the COVID-era reduction in
optimism that occurred ex post. , 2024), inflation expectations (Faria-e Castro and Leibovici, 2024;
Wu et al. , 2025), market research (Brand et al. , 2024), inflation expectations (Faria-e Castro and Leibovici, 2024;
Wu et al. , 2025), market research (Brand et al. Models tuned on human feedback are known to
compress the diversity of human opinion: Santurkar et al. (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. Models tuned on human feedback are known to
compress the diversity of human opinion: Santurkar et al. (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. (2025), Sarkar and Vafa (2025), and Gao et al. Memorization
refers broadly to the model’s retention of facts from its training corpus (Carlini et al. (2025), Sarkar and Vafa (2025), and Gao et al. Memorization
refers broadly to the model’s retention of facts from its training corpus (Carlini et al. 1 Graham (2022) further documents the central role that executive expectations play in corporate investment and planning decisions. Coibion et al. 1 Graham (2022) further documents the central role that executive expectations play in corporate investment and planning decisions. Coibion et al. Coibion et al. (2018)
use a large survey of corporate managers and show that their macroeconomic expectations
are widely dispersed, depart systematically from professional forecasts, and feed back into
company decisions. Coibion et al. (2018)
use a large survey of corporate managers and show that their macroeconomic expectations
are widely dispersed, depart systematically from professional forecasts, and feed back into
company decisions. Models tuned on human feedback are known to
compress the diversity of human opinion: Santurkar et al. (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. Models tuned on human feedback are known to
compress the diversity of human opinion: Santurkar et al. (2023) find that such models place
nearly all of their probability mass on a single answer to opinion-survey questions, matching
the modal view of a group rather than the spread of views within it, and that prompting the
model to adopt a group’s persona does not repair this. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. , 2024), inflation expectations (Faria-e Castro and Leibovici, 2024;
Wu et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. , 2024), inflation expectations (Faria-e Castro and Leibovici, 2024;
Wu et al. (2023), Lopez-Lira et al. (2025), Sarkar and Vafa (2025), and Gao et al. (2023), Lopez-Lira et al. (2025), Sarkar and Vafa (2025), and Gao et al. Operating in the
intersection of corporate strategy and the macroeconomic environment, they aggregate information about customers, suppliers, competitors, and capital markets into forward-looking
assessments of business activity. Research spanning a quarter century has shown that measured CFO expectations are economically meaningful: Gennaioli et al. Operating in the
intersection of corporate strategy and the macroeconomic environment, they aggregate information about customers, suppliers, competitors, and capital markets into forward-looking
assessments of business activity. Research spanning a quarter century has shown that measured CFO expectations are economically meaningful: Gennaioli et al. Our framework introduces an LLM-based measure of CFO-level economic sentiment and
provides the first individual-level, within-firm validation of LLM-simulated survey responses,
approximating private forecasts using an information set of public data and a respondent’s
own past answers. 2 We also address the memorization concerns raised by Glasserman and Lin
2For a survey of machine learning in economics and finance, see Korinek (2023). Our framework introduces an LLM-based measure of CFO-level economic sentiment and
provides the first individual-level, within-firm validation of LLM-simulated survey responses,
approximating private forecasts using an information set of public data and a respondent’s
own past answers. 2 We also address the memorization concerns raised by Glasserman and Lin
2For a survey of machine learning in economics and finance, see Korinek (2023). CFOs Meet LLMs
John R. Graham† Campbell R. CFOs Meet LLMs
John R. Graham† Campbell R. Graham† Campbell R. Harvey† Manish Jha‡
†Duke University and NBER ‡Georgia State University
Abstract
Business sentiment is a closely watched economic signal, but measuring it is slow
and costly: surveys typically reach only a few hundred firms, arrive periodically,
and take time to compile. Graham† Campbell R. Harvey† Manish Jha‡
†Duke University and NBER ‡Georgia State University
Abstract
Business sentiment is a closely watched economic signal, but measuring it is slow
and costly: surveys typically reach only a few hundred firms, arrive periodically,
and take time to compile. First, we impose information cutoffs that
instruct the model not to access post-survey data, and any respondent history we supply is itself
dated strictly before the cutoff. We find supportive evidence that this instruction was followed
in that the first-quarter 2020 LLM forecast does not anticipate the COVID-era reduction in
optimism that occurred ex post. First, we impose information cutoffs that
instruct the model not to access post-survey data, and any respondent history we supply is itself
dated strictly before the cutoff. We find supportive evidence that this instruction was followed
in that the first-quarter 2020 LLM forecast does not anticipate the COVID-era reduction in
optimism that occurred ex post. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. A second strand is cautionary. Henning et al. A second strand is cautionary. Henning et al. (2024) find that chatbot behavior in standard economic games is broadly similar to that of human subjects, though not identical to it. Manning and Horton (2025) further demonstrate that
grounding LLM personas in economic theory and calibrating them against human data substantially improves predictive fidelity in novel settings, providing a methodological foundation
3 (2024) find that chatbot behavior in standard economic games is broadly similar to that of human subjects, though not identical to it. Manning and Horton (2025) further demonstrate that
grounding LLM personas in economic theory and calibrating them against human data substantially improves predictive fidelity in novel settings, providing a methodological foundation
3 Graham† Campbell R. Harvey† Manish Jha‡
†Duke University and NBER ‡Georgia State University
Abstract
Business sentiment is a closely watched economic signal, but measuring it is slow
and costly: surveys typically reach only a few hundred firms, arrive periodically,
and take time to compile. Graham† Campbell R. Harvey† Manish Jha‡
†Duke University and NBER ‡Georgia State University
Abstract
Business sentiment is a closely watched economic signal, but measuring it is slow
and costly: surveys typically reach only a few hundred firms, arrive periodically,
and take time to compile. 3 Our setting satisfies the no-leakage condition formalized by Ludwig
et al. (2025), since the target variable was contractually excluded from the training corpus, and
it avoids the most severe category in the leakage taxonomy of Kapoor and Narayanan (2023),
in which the outcome to be predicted is itself present in the training data. 3 Our setting satisfies the no-leakage condition formalized by Ludwig
et al. (2025), since the target variable was contractually excluded from the training corpus, and
it avoids the most severe category in the leakage taxonomy of Kapoor and Narayanan (2023),
in which the outcome to be predicted is itself present in the training data. , 2020; Lopez-Lira and Tang, 2023; Uddin et al. , 2024; Wang et al. , 2020; Lopez-Lira and Tang, 2023; Uddin et al. , 2024; Wang et al. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. We go further by testing whether the LLM
persona predicts the response of the specific human at the same firm and quarter, and by
showing how that fidelity depends on the information the model is given. In a non-economic
setting, Park et al. We go further by testing whether the LLM
persona predicts the response of the specific human at the same firm and quarter, and by
showing how that fidelity depends on the information the model is given. In a non-economic
setting, Park et al. (2023), Lopez-Lira et al. (2025), Sarkar and Vafa (2025), and Gao et al. (2023), Lopez-Lira et al. (2025), Sarkar and Vafa (2025), and Gao et al. The final section offers
some conclusions. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. The final section offers
some conclusions. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. Construction of the LLM Optimism Scores
We construct LLM-generated forecasts to simulate the expectations of individual CFOs participating in The CFO Survey, conducted quarterly by Duke University and the Federal Reserve
sentiment classification (Tetlock, 2007; Tetlock et al. , 2008; Loughran and McDonald, 2011; Hansen and
McMahon, 2016), financial analysis (Cao et al. , 2020; Lopez-Lira and Tang, 2023; Uddin et al. , 2024; Wang et al. , 2020; Lopez-Lira and Tang, 2023; Uddin et al. , 2024; Wang et al. With appropriate conditioning, LLMs may in the future be able to serve
as digital twins of executives, offering scalable, high-frequency expectations data
for financial research and policy. Keywords: Large language models; CFO survey; Managerial expectations; Business
sentiment; Synthetic surveys; Digital twins; Belief formation
JEL Classification: C55, D84, E27, G30, G41
Version: September 8, 2026. With appropriate conditioning, LLMs may in the future be able to serve
as digital twins of executives, offering scalable, high-frequency expectations data
for financial research and policy. Keywords: Large language models; CFO survey; Managerial expectations; Business
sentiment; Synthetic surveys; Digital twins; Belief formation
JEL Classification: C55, D84, E27, G30, G41
Version: September 8, 2026. We show that large language models hold the potential
to address these shortcomings. We prompt an LLM to role-play as the CFO of a
specific company on a specific date, for every public-company CFO who responded
to the Duke–Federal Reserve CFO Survey between 2002 and 2025, and answer a
question about economy-wide optimism. However, despite their informational value, managerial expectations are
expensive to collect (Bloom and Van Reenen, 2007), available only at a fixed frequency, and are
limited to the several hundred companies whose executives choose to participate in a given
survey wave. In this paper, we explore whether it is possible to generate credible CFO-style
expectations at scale, without the cost, the small sample, and the delay of a human survey. However, despite their informational value, managerial expectations are
expensive to collect (Bloom and Van Reenen, 2007), available only at a fixed frequency, and are
limited to the several hundred companies whose executives choose to participate in a given
survey wave. In this paper, we explore whether it is possible to generate credible CFO-style
expectations at scale, without the cost, the small sample, and the delay of a human survey. We show that large language models hold the potential
to address these shortcomings. We prompt an LLM to role-play as the CFO of a
specific company on a specific date, for every public-company CFO who responded
to the Duke–Federal Reserve CFO Survey between 2002 and 2025, and answer a
question about economy-wide optimism.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Business sentiment is a closely watched economic signal, but measuring it is slow and costly: surveys typically reach only a few hundred firms, arrive periodically, and take time to compile The LLM-generated optimism score predicts the individual CFO’s actual answer, even in specifications that include firm and year-quarter fixed effects as well as a control variable measuring the human CFO’s lagged response. The LLM framework approximates how an executive is likely to respond given their information set and prior answers, providing an individual-level benchmark for the predictable component of managerial expectations.
Aher et al. (2023) Argyle et al. (2023) Bai et al. (2023) Bali et al. (2026) Balke and Petersen (2002) Barrero (2025) Baslandze et al. (2026) Ben-David et al. (2013) Bordalo et al. (2020) Boutros et al. (2025) Brand et al. (2023) Brown (1910) Cameron et al. (2011) Cao et al. (2024) Carlini et al. (2023) Coibion et al. (2018) Cook et al. (2023) Correia (2015) Dillion et al. (2023) Faria-e Castro and Leibovici (2024) Fried et al. (2026) Gao et al. (2025) Gennaioli et al. (2016) Glasserman and Lin (2023) Gormsen and Huber (2024) Goulding et al. (2026) Graham et al. (2019) Graham and Harvey (2001) Gu et al. (2020) Hansen et al. (2025) Hansen and McMahon (2016) Heaton et al. (2017) Henning et al. (2025) Horton et al. (2023) Jha et al. (2024) Kapoor and Narayanan (2023) Korinek (2023) Li et al. (2024) Lopez-Lira et al. (2025) Loughran and McDonald (11) Ludwigson (2004) Malmendier and Tate (2015) Manning and Horton (2025) Mei et al. p. 3 (2024) find that chatbot behavior in standard economic games is broadly similar to that of human subjects, though not identical to it. Manning and Horton (2025) further demonstrate that
grounding LLM personas in economic theory and calibrating them against human data substantially improves predictive fidelity in novel settings, providing a methodological foundation
3Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Scale synthetic survey generation for high-frequency policy shocks by using LLMs to generate daily or weekly sentiment scores for a full crosssection of public firms, circumventing traditional quarterly constraints in event-study designs.
-
Enable synthetic survey experiments by prompting LLMs with counterfactual company characteristics to study how expectations respond to specific policy shocks, which has no natural analog in human data.
-
Implement history conditioning as the primary mechanism for improving individual-level prediction accuracy, allowing the model to approximate private forecasts by using
the respondent’s own past answers
andthe LLM’s own prior forecasts for that person.
-
Improve synthetic executive fidelity by including a rich information set, specifically ensuring prompts include both
the full company profile
and conditioning on therespondent’s own history,
which increases the fit to nearly 50% in the best-case scenarios. -
Develop a scalable, individual-level sentiment measure for practitioners and policymakers that can offer a
real-time, scalable, individual-level sentiment measure
by approximating private forecasts using an information set of public data and a respondent’s own past answers.
Abstract
Business sentiment is closely watched, but measurement is slow and costly: surveys reach a limited number of firms, arrive periodically, and take time to field and compile. We show that large language models hold the potential to address some of these shortcomings. We prompt an LLM to role-play as the CFO of a specific company on a specific date, for every public-company CFO who responded to the Duke--Federal Reserve CFO Survey from 2002- 2025. We find that LLM-generated forecasts reasonably match the cross-section of CFO expectations, with accuracy increasing with the amount of information provided to the LLM.
Sources
Related papers
- AFNO: Universal Certifiably-Arbitrage-Free Operator Learning Framework for European Call-Price Surfaces
- Diffusion models for dynamic volatility surface generation and data-driven hedging
- Rethinking Synthetic Scenario Realism: Compatibility, Not Fidelity, Drives Hedging Performance
- CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining
- Diversification Across Predictors: Heterogeneous Ensemble Learning for Cross-Sectional Return Prediction
- DSINC beyond Heston: the 3/2 and other non-affine stochastic-volatility models