Paper: arXiv 2409.17933
Authors: Manish Jha, Jialin Qian, Michael Weber, Baozhong Yang
Abstract
We create a firm-level ChatGPT investment score, based on conference calls, that measures managers’ anticipated changes in capital expenditures. We validate the score with interpretable textual content and its strong correlation with CFO survey responses. The investment score predicts future capital expenditure for up to nine quarters, controlling for Tobin’s \(q\) and other determinants, implying the investment score provides incremental information about firms’ future investment opportunities. The investment score also separately forecasts future total, intangible, and R&D investments. Consistent with theoretical predictions, high-investment-score firms experience significant positive short-term returns upon disclosure, and negative long-run future abnormal returns. We demonstrate ChatGPT’s applicability to measure other policies, such as dividends and employment.
Complexity vs Empirical Score
- Math Complexity: 4.5/10
- Empirical Rigor: 8.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper’s mathematical complexity is moderate, relying on established theoretical frameworks like q-theory and standard regression models, without dense derivations. Empirical rigor is high, evidenced by extensive backtesting on real-world data (74k transcripts), validation against CFO surveys, and robustness checks.
Research Flowchart
flowchart TD A["Research Goal: Measure ChatGPT's Value in<br>Firm-Level Investment Policy Forecasting"] --> B["Methodology: Create Firm-Level<br>ChatGPT Investment Score"] B --> C["Data Source: Earnings<br>Conference Call Transcripts"] C --> D["Computation: Prompt ChatGPT to<br>Quantify Managerial Sentiment on<br>Future Capital Expenditures"] D --> E["Validation & Outcomes"] E --> F["Primary Outcome: Score predicts<br>future CapEx up to 9 quarters<br>(controlling for Tobin's q)"] E --> G["Secondary Outcomes: Predicts total,<br>intangible, & R&D investments"] E --> H["Market Reaction: Short-term positive<br>abnormal returns, Long-term negative<br>abnormal returns"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a firm‑level ChatGPT investment score extracted from earnings call transcripts to measure managers' expected changes in capital expenditures.
Source sentence
We create a firm-level ChatGPT investment score, based on conference calls, that measures managers' anticipated changes in capital expenditures.
Assumes 2
- Tobin’s q is decomposed into an expected component, a private manager signal, and a random error. model
Source sentence
We assume that the Tobin’s q consists of several components: qt+1 = qe t + qm t +ϵt+1, where qe t is the expected value of Tobin’s q, qm t is a private signal from the manager, and ϵt+1 is a random component not known to investors or the manager.
- The manager publicly discloses her private signal at the beginning of period t+1 before investment decisions. model
Source sentence
In the model, the manager publicly discloses her signal qm t at the beginning of period t +1, before investment and dividend decisions are made, which allows the market to update its expectation of qt+1 from qe t to qe t + qm t .
Data
- Earnings call transcripts text-news public
Source sentence
The primary text dataset we use encompasses earnings call transcripts from 2006 to 2020, sourced from the Seeking Alpha website.
- Duke CFO survey other public
Source sentence
We obtain the quarterly Duke CFO survey firm-level data which has been initiated and analyzed in Graham and Harvey (2001).
- Compustat fundamentals commercial
Source sentence
We utilize Compustat and CRSP to obtain corporate accounting variables and stock returns.
- CRSP market-prices commercial
Source sentence
We utilize Compustat and CRSP to obtain corporate accounting variables and stock returns.
Costs
Not applicable
Finds 5
- The ChatGPT investment score predicts future capital expenditure for up to nine quarters, even after controlling for Tobin’s q and other determinants. predictive-accuracy
Source sentence
The investment score predicts future capital expenditure for up to nine quarters, controlling for Tobin's $q$ and other determinants, implying the investment score provides incremental information about firms' future investment opportunities.
- A one‑standard‑deviation increase in the investment score is associated with a −1.83% raw return and similar negative abnormal returns over the next quarter. negative
Source sentence
A one-standard-deviation increase in the investment score corresponds to −1.83%,−1.49%, and −1.42% in raw returns, FF5-adjusted returns, and q5-adjusted returns in the quarter subsequent to the earnings call, respectively.
- The negative association between the investment score and future abnormal returns persists for up to nine quarters, with average quarterly impacts of about −1.5% (q5‑adjusted) and −1.1% (FF5‑adjusted). negative
Source sentence
On average, a one-standard-deviation increase in the investment score leads to a change of −1.54% in annualized q5-adjusted returns for each quarter n = 2 to 10, and −1.10% in annualized FF5-adjusted return for each quarter n = 2 to 6, respectively.
- The investment score is positively associated with cumulative abnormal returns in the short‑term windows [0,1], [0,3] and [0,5], yielding about a 0.6% increase per standard‑deviation. performance
Source sentence
Economically, a one-standard-deviation increase in the investment score leads to an increase of around 0.6% cumulative abnormal return.
- ChatGPT can also be applied to extract expectations about other corporate policies such as dividends and hiring. other
Source sentence
The methodology we develop, however, can be equally applied to extract firms’ expectations about other corporate policies. We consider two important discretionary policies: dividend payment and hiring.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.
8 papers in the archive : A Test of Lookahead Bias in LLM Forecasts ; A Financial Brain Scan of the LLM
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.