CFOs Meet LLMs

Battle Count: 75

The paper provides a scalable, high-frequency measure of executive economic sentiment that could serve as an alternative or complement to traditional survey-based sentiment indicators. The Duke-Fed CFO Optimism Index has historically led hiring and GDP growth, and CFO expectations predict corporate investment (Gennaioli et al., 2016). An LLM-generated version available at higher frequency and broader coverage could inform macro-aware trading strategies, sector rotation, and risk management. However, the paper focuses on validation rather than direct trading applications, and the sentiment measure is a single optimism score rather than a multi-dimensional signal. The R² of ~0.27-0.49 at the individual level and ~0.72 at the quarterly aggregate level suggests meaningful but imperfect predictive power. The method could enable daily/weekly sentiment scores for the full cross-section of public firms, which would be valuable for systematic strategies.

The core methodology involves prompt engineering with GPT-5.4 via OpenAI's Responses API, which is relatively straightforward. However, the full pipeline requires: (1) constructing firm-executive profiles from survey data with name standardization and missing-data imputation, (2) building respondent history panels with chronological processing, (3) implementing information cutoffs and web-search constraints, (4) running 3 independent API calls per observation (18,225 total calls), (5) matching synthetic scores to actual survey responses, and (6) running OLS regressions with two-way clustered standard errors and fixed effects. The prompt design is sophisticated with multiple conditioning layers. Access to the specific LLM version and institutional API agreements adds practical barriers. The econometric analysis is standard but the data construction and matching pipeline is non-trivial.

The paper provides detailed system prompts (Appendix B), sample user prompts (Appendix C), variable definitions (Table A1), and data construction procedures (Appendix A). However, the Duke-Fed CFO Survey individual responses are confidential and not publicly available. The LLM used is GPT-5.4 accessed via Duke's institutional agreement, which may not be replicable by other researchers. The exact API parameters and model version details are partially specified. The methodology is well-documented but full replication requires access to proprietary survey data and the specific LLM version.