CFOs Meet LLMs

Abstract: Business sentiment is a closely watched economic signal, but measuring it is slow and costly: surveys reach only a few hundred firms, arrive periodically, and take time to compile. We show that LLMs hold the potential to address these shortcomings. We prompt an LLM to role-play as the CFO of a specific company at a specific date and focus on the economic-optimism question on the Duke-Federal Reserve CFO Survey over 2002-2025. We find that the LLM reproduces individual human responses: the predicted optimism score significantly forecasts the CFO's actual answer, surviving firm and year-quarter fixed effects and a control for the most recent prior response. Predictive accuracy increases with the amount of information supplied, as both respondent history and firm characteristics improve fit, and the relationship persists under quarterly aggregation. With appropriate conditioning, LLMs may be able to serve as credible digital twins of executives, offering scalable, high-frequency expectations data for financial research and policy.

• The paper demonstrates that LLMs, when conditioned with firm attributes and historical data, accurately replicate CFO sentiment responses.
• It employs GPT-5.4 and detailed firm profiles to simulate survey answers, achieving high predictive fidelity with R² up to 0.50 when using respondent history.
• The approach enables scalable, high-frequency sentiment analysis, offering novel tools for belief formation research and policymaking.

LLMs as Digital Twins: Synthesizing CFO Sentiment at Scale

Introduction

“CFOs Meet LLMs” (2606.13812) addresses the inherent limitations of traditional surveys measuring business sentiment, specifically the Duke-Federal Reserve CFO Survey, by deploying LLMs as digital proxies for executive respondents. The paper validates LLMs’ capacity to generate synthetic yet individualized forecasts of economic sentiment, calibrated to specific firms and informed by both public information and respondent-level history. The study’s empirical focus is the longstanding CFO optimism question—“Rate your optimism about the overall U.S. economy on a scale from 0 to 100”—with human and LLM responses systematically matched across over two decades (2002–2025). The approach combines persona simulation with rigorous conditioning on firm attributes and executive response history, yielding novel insights into belief formation and the potential for scalable, high-frequency managerial expectations data.

Methodology

The methodology leverages GPT-5.4 (OpenAI, 2026), enabled with web search restricted to contemporary data, to role-play as the CFO of specific public firms at historical survey dates. The synthetic survey prompts supply detailed firm profiles (including industry, size, geography, and credit rating) and, crucially, the executive’s history of prior optimism scores and the LLM’s past forecasts for the individual, all strictly preceding the evaluation date. The canonical optimism survey question is posed verbatim to ensure comparability. For each of 6,075 matched survey responses, three independent LLM runs are generated and averaged to smooth stochasticity.

Distinguishing features of the framework include:

• Information Cutoff: Explicit context cutoffs preclude lookahead bias; the model is temporally aligned with the human respondent’s information set at the survey date.
• Respondent Conditioning: Injection of respondent-specific history (up to 12 quarters) anchors the digital twin, countering the tendency of instruction-tuned models to default to mode-seeking or bland averaging.
• Stability Controls: Multiple independent LLM replications per prompt ensure high internal consistency (ICC ≈ 0.96), with mean within-prompt SD at 2.22 points (on a 0–100 scale).

Robustness is addressed by excluding any survey data from the LLM training set due to private submission via contractual agreements, further mitigating memorization or contamination.

Empirical Findings

Individual-Level Predictive Fidelity

The LLM-generated optimism scores are significant linear predictors of the matched human CFO response at the firm-quarter level. In baseline OLS, a one-point increase in LLM optimism corresponds to a 0.57-point increase in the human score ($t = 17.98$, $R^2 = 0.27$). Importantly, this predictive relationship survives the inclusion of firm and year–quarter fixed effects—thus, the model’s fidelity is not merely attributable to persistent differences across firms or time, but also reflects within–firm, over–time variation (i.e., idiosyncratic sentiment shifts). Even after controlling for a respondent’s most recent prior answer, the LLM score remains a strong predictor (coefficients 0.29–0.53 across specifications), establishing that the model’s output involves more than mechanical extrapolation from history.

Information Dose-Response

The predictive fit scales monotonically with the amount of respondent history and firm-specific features provided in the prompt. When no respondent history is included (i.e., first-time responses), the model achieves $R^2 \approx 0.10$; with more than three historical data points, fit approaches $R^2 \approx 0.50$. Broader firm profiles (≥10 fields) also yield improved performance ($R^2$ up to 0.32), evidencing genuine data utilization rather than prior-based guessing.

Heterogeneity and Aggregation

The predictive quality varies over time, with higher $R^2$ in later quarters where richer response history is available. When data are aggregated at the quarterly level, mean LLM optimism remains a significant and independent predictor of aggregate CFO sentiment, even after controlling for macro benchmarks such as the Michigan Consumer Sentiment Index and the SPF professional GDP forecast. In combined regressions, the LLM score absorbs the significant explanatory power, while neither Michigan nor SPF remain significant after adjusting for LLM-driven variance.

Disagreement and Cross-Sectional Heterogeneity

Conditioning on respondent history restores cross-sectional dispersion in LLM outputs, with within-quarter SDs approaching those observed in human responses (disagreement ratio ≈ 1.4). This is a marked improvement over persona-prompted LLMs without own-history conditioning, which typically yield unduly compressed forecasts centered near the mode.

Theoretical and Practical Implications

Expanding the Study of Belief Formation

By demonstrating that individual-level LLM digital twins can faithfully reproduce much of the human-specific signal in sentiment surveys, the paper opens new pathways for belief formation research. The technology enables:

• Counterfactual and High-Frequency Experimentation: Synthetic sentiment can be sampled for any public firm at arbitrary frequencies, enabling event studies and policy shock analysis not feasible with low-frequency, low-coverage human surveys.
• Synthetic Survey Experiments: The model can be conditioned on hypothetical firm profiles or peer histories to study attribution effects and information propagation.
• Decomposition of Public vs. Private Signals: The residual variance—where LLMs cannot match human responses—can serve as a proxy for the incremental value of private information.

Policy Relevance

Real-time, high-frequency synthetic sentiment could supplement, or in some cases substitute for, lagged and costly survey data in monetary policy, financial stability monitoring, or business cycle research. For example, LLM-simulated sentiment could provide continuous updates analogous to the CFO Optimism Index, unconstrained by survey periodicity or sample selection. The approach also promises a scalable “synthetic Beige Book” for central banking.

Limitations and Future Directions

Several limitations are identified:

• Private Information and First-Time Respondents: LLM predictions cannot incorporate private or forward-looking data unavailable in public records, and predictive accuracy is diminished in the absence of respondent history.
• Model Training and Contamination Risk: Although contractual data exclusion and information cutoffs minimize direct lookahead risk, indirect memorization—via correlated non-private signals in training data—cannot be entirely excluded.
• Dynamic Weighting and Peer Borrowing: For aggregation, weighting twins by historical accuracy, or borrowing peer-response history where direct executive data are unavailable, represent promising avenues to enhance robustness and coverage.

Prospective research should pursue live, prospective, out-of-sample validation as new human survey waves are collected, and more granular time-stamping of model information states.

Conclusion

“CFOs Meet LLMs” (2606.13812) demonstrates that LLMs, suitably conditioned, can serve as accurate digital twins for C-suite executives in traditional sentiment surveys. The models capture both firm-level heterogeneity and within-person dynamics, especially when supplied with respondent-specific history. The findings challenge the current reliance on costly and infrequent human surveys for managerial expectations and provide a scalable, programmatic alternative that preserves much of the individual-specific informational content. This methodology introduces new research and policy tools for analyzing firm belief formation, information aggregation, and economic forecasting, with implications for both the academic finance literature and the practical operation of policy-oriented business sentiment indices.