Duke-Fed CFO Survey Analysis

• The Duke-Fed CFO Survey is a quarterly assessment capturing CFOs' forward-looking views on both macroeconomic and firm-level conditions.
• It combines stable panel data with rotating questionnaire content to facilitate robust econometric analysis and benchmark corporate sentiment.
• Recent studies leverage the survey to validate LLM-derived proxies and digital twins, demonstrating significant alignment with traditional CFO responses.

The Duke-Federal Reserve CFO Survey is a quarterly survey of corporate chief financial officers and senior financial executives that elicits forward-looking assessments about business conditions and firm decisions. Commonly known as the Duke CFO Survey or Graham–Harvey CFO Survey, it is typically run by Duke University’s Fuqua School of Business and, in many waves, in collaboration with the Federal Reserve Banks of Richmond and Atlanta. In recent arXiv research, the survey functions both as a timely barometer of corporate sentiment and as a benchmark for evaluating LLM-derived expectation measures and executive “digital twins” (Bybee, 2023, Graham et al., 11 Jun 2026).

1. Institutional setting and chronology

The survey is referred to as the “Duke or Graham and Harvey survey of chief financial officers (CFOs), started in 1998 by John Graham and Campbell Harvey,” and is noted as running from 2000 to the present. It is typically conducted quarterly and surveys corporate CFOs; in the more recent panel-based work, the target population is described as CFOs and senior financial executives. Duke University administered it until mid-2020; since then it has been jointly conducted with the Federal Reserve Banks of Richmond and Atlanta (Graham et al., 11 Jun 2026).

This institutional configuration matters because the survey combines academic continuity with policy relevance. Its quarterly cadence gives it a recurring observational structure suitable for both cross-sectional and panel econometric work, while the participation of Federal Reserve Banks situates it within a broader infrastructure for monitoring business expectations. A plausible implication is that the survey’s value derives not only from the level of reported sentiment, but also from its stable timing and repeated elicitation of beliefs from decision-makers directly involved in financing, investment, and employment planning.

2. Questionnaire architecture and principal measures

The survey asks CFOs about a broad set of macro and firm-specific quantities. A focal question in one line of research is expected U.S. stock market returns over the next 12 months. Other common topics include expectations for GDP growth, inflation, capital spending, employment, and firm-level plans and constraints. In this form, the survey provides forward-looking measures of corporate decision-makers’ beliefs about both aggregate and firm-level conditions (Bybee, 2023).

A distinct long-panel application uses the economic-optimism item about the U.S. economy. Its wording and scale are reported as: “Rate your optimism about the overall U.S. economy on a scale from 0-100, with 0 being the least optimistic and 100 being the most optimistic.” Although the wording is stable, not every quarterly wave includes this item; restricting to waves that do yields 93 year-quarters during 2002–2025. The item is singled out because it is the single question with the necessary longitudinal consistency to support a long, matched panel at the person-firm level, which is critical for within-person validation (Graham et al., 11 Jun 2026).

The coexistence of broad rotating content and a small set of stable questions is methodologically important. It permits the survey to operate simultaneously as a general-purpose barometer of corporate sentiment and as a source of repeated measures suitable for fixed-effects designs, persistence controls, and quarter-level aggregation tests.

3. Position in the study of expectations and nonrational beliefs

The survey is widely used as a barometer of corporate sentiment and macro-financial expectations. Researchers and policymakers use it to track corporate sentiment, assess nonrational belief components, and monitor expected economic and market conditions. In the return-expectations setting, the survey is treated as a direct measure of subjective beliefs rather than as a revealed-preference proxy (Bybee, 2023).

One empirical regularity emphasized in this literature is extrapolation in expected returns. The return-expectations relation is written as

$$E_t(r_{t+H}) = \alpha + \gamma \sum_{i=1}^{k} r_{t-i} + \delta X_t + \varepsilon_t, \quad \gamma>0,$$

where $X_t$ can include controls such as dividend yield or term spread. In the reported evidence, CFO expected returns are positively related to recent realized returns, consistent with $\gamma>0$. A second regularity is that subjective expected returns are negatively related to future realized returns:

$$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$

The same work also describes CFO expected returns as disconnected from objective expected-return predictors, with weak or opposite-signed correlations relative to variables such as $\log(D/P)$ and composite expected-return indices. These features place the survey within a research program on deviations from full-information rational expectations, even though the paper’s formal underreaction regressions are run against the Survey of Professional Forecasters rather than on CFO macro expectations themselves.

4. News-based LLM proxies and the survey’s stock-market expectation series

One arXiv contribution constructs an expectations proxy by querying GPT-3.5 on historical Wall Street Journal articles from 1984–2021. The underlying archive is the Dow Jones Historical News Archive. The procedure excludes non-economic content by subject tags and headline patterns, keeps core sections and weekdays, requires minimum article and headline lengths, and randomly samples 300 articles per month to control API costs. Each article is passed through a standardized zero-shot prompt asking whether the news will increase or decrease a target variable such as “the S&P 500 index,” “the consumer price index,” or “real GDP,” and requesting four outputs: a categorical answer in $$\{\text{increase},\text{decrease},\text{uncertain}\}$$, confidence, magnitude, and a brief explanation (Bybee, 2023).

Article-level outputs are aggregated into a balance statistic,

$$F^{gpt}_{t}(X^k_{t+h}) = \frac{\sum_{i\in A_t}\mathbb{I}(\text{Increase})^k_i - \sum_{i\in A_t}\mathbb{I}(\text{Decrease})^k_i}{\sum_{i\in A_t}\mathbb{I}(\text{Increase})^k_i + \sum_{i\in A_t}\mathbb{I}(\text{Decrease})^k_i},$$

where $A_t$ indexes articles in period $t$, $k$ labels the variable, and $X_t$0 is the forecast horizon, which is unknown for the LLM. For the CFO comparison, the LLM series is formed at monthly and multi-month frequencies and then aligned to the survey’s quarterly release cadence. The CFO survey explicitly asks about 12-month-ahead stock market returns; the LLM horizon is not explicitly known, so the comparison relies on frequency alignment and smoothing rather than exact horizon matching. The three-month average is the default aggregate in subsequent analysis.

Using 76 quarterly observations in the overlapping sample, the reported correlations between the LLM-based return-expectation proxy and the CFO survey are as follows:

LLM aggregation window	Correlation	t-stat
One-month average	0.49	4.84
Two-month average	0.56	5.88
Three-month average	0.59	6.21
Optimal EWMA	0.59	6.24

The t-statistics are formed using Newey–West standard errors with a 12-month lag. The three-month average is the default specification, but the one-month, two-month, and optimally smoothed EWMA variants also yield statistically significant correlations. The same study reports that the LLM and CFO measures exhibit closely aligned “belief moments”: correlation with past 12-month returns is $X_t$1 for the CFO-aligned GPT series and $X_t$2 for the CFO survey; correlation with mutual fund flows into equities is $X_t$3 for the GPT series and $X_t$4 for the CFO survey; and the signs on correlations with $X_t$5, changes in $X_t$6, CAY, changes in CAY, and the GS and GW composite expected-return indices broadly match across the two series. This suggests that the LLM-derived measure is not merely contemporaneously correlated with the survey, but also reproduces the survey’s established departures from objective-return benchmarks.

5. Person-firm panels and LLM “digital twins”

A later arXiv study uses the survey to evaluate whether an LLM can reproduce individual CFO responses by role-playing as the CFO of a specific company at a specific date. The analysis focuses on public-company respondents so that responses can be linked to externally verifiable firm characteristics and to public information available to the model. The resulting matched panel contains 6,075 firm-quarter observations with a non-missing optimism response over 2002–2025. For each observation, the study generates three independent LLM forecasts and averages them, yielding 18,225 raw LLM runs and 6,075 mean forecasts (Graham et al., 11 Jun 2026).

The LLM setup uses GPT-5.4 via Duke’s institutional Responses API. Submitted survey data are contractually excluded from training, and the model is given a hard information cutoff equal to the respondent’s completion date. Prompts include executive name and title, company name and ticker, industry, prior-year sales revenue bracket, total employees bracket, percent foreign revenue, headquarters region or state, ownership, and credit rating. They also include respondent history: up to 12 most recent pre-cutoff optimism scores reported by the same person within the preceding 12 calendar quarters, the LLM’s own prior forecasts for that person, and the person’s pre-cutoff lifetime mean optimism. The model is instructed to output only a single number on the 0–100 optimism scale.

The individual-level statistical framework defines $X_t$7 as the human CFO’s optimism score and $X_t$8 as the matched three-run mean LLM score. Baseline and fixed-effects specifications are

$X_t$9

$\gamma>0$0

with extensions that add the respondent’s immediate prior answer $\gamma>0$1 or most recent prior answer $\gamma>0$2. Standard errors are two-way clustered by firm and year-quarter in individual-level regressions and Newey–West with lag 4 at the quarterly level.

The reported estimates show that the predicted optimism score significantly forecasts the CFO’s actual answer. In the baseline regression without fixed effects, $\gamma>0$3, $\gamma>0$4, $\gamma>0$5, and $\gamma>0$6. With firm fixed effects only, $\gamma>0$7 and $\gamma>0$8; with quarter fixed effects only, $\gamma>0$9 and $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$0; with firm and quarter fixed effects jointly, $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$1, $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$2, $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$3, and $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$4. In the subsample with the respondent’s immediate prior answer available, the coefficient remains positive after controlling for persistence and joint fixed effects: $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$5, $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$6, $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$7, with $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$8 on the lagged human response. Under the “any-prior” robustness design, the corresponding coefficient is $$\rho = \operatorname{corr}\big(E_t(r_{t+H}),\, r_{t+H}\big) < 0.$$9, $\log(D/P)$0, $\log(D/P)$1.

The study also reports a marked information dose-response. With no respondent history, $\log(D/P)$2 and $\log(D/P)$3; with 1–3 prior answers, $\log(D/P)$4 and $\log(D/P)$5; with more than 3 prior answers, $\log(D/P)$6 and $\log(D/P)$7. When fewer than 10 firm-profile fields are available, $\log(D/P)$8 and $\log(D/P)$9; with 10–12 fields, $$\{\text{increase},\text{decrease},\text{uncertain}\}$$0 and $$\{\text{increase},\text{decrease},\text{uncertain}\}$$1. The same paper reports high within-prompt stability across stochastic replications: pairwise correlations of approximately $$\{\text{increase},\text{decrease},\text{uncertain}\}$$2–$$\{\text{increase},\text{decrease},\text{uncertain}\}$$3, mean within-prompt standard deviation of $$\{\text{increase},\text{decrease},\text{uncertain}\}$$4, intraclass correlation of $$\{\text{increase},\text{decrease},\text{uncertain}\}$$5 with 95% confidence interval $$\{\text{increase},\text{decrease},\text{uncertain}\}$$6, and reliability of the three-run mean of approximately $$\{\text{increase},\text{decrease},\text{uncertain}\}$$7.

At the quarterly level, LLM mean optimism remains significant when matched to quarter-level CFO averages and compared with standard macro gauges. Across 93 quarters, the coefficient on quarterly LLM mean optimism is $$\{\text{increase},\text{decrease},\text{uncertain}\}$$8 when Michigan is included, $$\{\text{increase},\text{decrease},\text{uncertain}\}$$9 when the SPF is included, and $$F^{gpt}_{t}(X^k_{t+h}) = \frac{\sum_{i\in A_t}\mathbb{I}(\text{Increase})^k_i - \sum_{i\in A_t}\mathbb{I}(\text{Decrease})^k_i}{\sum_{i\in A_t}\mathbb{I}(\text{Increase})^k_i + \sum_{i\in A_t}\mathbb{I}(\text{Decrease})^k_i},$$0 in a horse race with both Michigan and SPF; in the horse race, Michigan and SPF are not significant once the LLM measure is included. This indicates that, under the paper’s conditioning scheme, the synthetic series retains explanatory power after aggregation and relative to conventional sentiment indicators.

6. Interpretation, limitations, and research uses

The survey’s role in recent LLM work is twofold. First, it serves as a benchmark for whether a text-based expectations proxy can recover salient features of corporate beliefs from journalistic narratives. Second, it serves as a validation target for synthetic respondent-level forecasts conditioned on firm information and respondent history. In both applications, the survey is treated as a direct measure of subjective expectations rather than as an objective truth criterion (Bybee, 2023, Graham et al., 11 Jun 2026).

Several limitations are explicit. In the news-based comparison, variable coverage is incomplete: the survey spans a broad set of macro and firm-level expectations, but the empirical comparison focuses on stock market return expectations at a 12-month horizon. The macro underreaction tests are conducted against the SPF, not against CFO macro questions. The LLM expectations are driven by Wall Street Journal articles and the prompt, and aggregation uses a balance statistic rather than numeric point forecasts. Differences in content and framing between journalistic narratives and CFO decision frameworks may therefore introduce biases. Horizon alignment is also imperfect because the survey asks explicitly about 12-month-ahead returns while the LLM horizon is implicit.

In the digital-twin application, residual variation remains material because human CFOs have private information that an LLM cannot access. Predictions are weakest for first-time respondents, where there is no own-history to anchor the model. Structural breaks can challenge any model trained on historical regularities. The paper further notes that mean absolute error and root mean squared error are not reported. Calibration is imperfect but structured: relative to human responses, LLM scores are slightly lower in level at the individual level, slightly compressed in cross-sectional dispersion within a quarter, somewhat more volatile at the quarterly aggregate, and prone to “modal opinion” collapse if not anchored with respondent history. The inclusion of respondent history is therefore not a cosmetic design choice; omitting it “collapses dispersion toward modal answers,” whereas history anchoring restores cross-sectional heterogeneity and improves fit.

These findings motivate a specific interpretation of the survey’s current research role. The strong CFO–LLM correlations suggest that LLMs can recover salient features of corporate beliefs from textual and firm-specific information, but the resulting objects remain belief proxies. They do not constitute ground truth and should be used alongside traditional surveys. A plausible implication is that the Duke-Federal Reserve CFO Survey will remain foundational even in an environment of scalable LLM measurement, because it provides the low-frequency, high-credibility anchor against which higher-frequency synthetic belief series can be calibrated, stress-tested, and interpreted.