Truace
Subscribe
57/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

AI reading of 125,830 earnings calls predicts GDP and jobs quarters ahead, paper finds

Study title: Generative AI, Managerial Expectations, and Economic Activity

Evidence type
Observational studyPreprint, not yet peer reviewed
Sample
n = 125,830
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2496cite or check it

Study title: Generative AI, Managerial Expectations, and Economic Activity

Economists at Georgia State, Rochester Institute of Technology and Purdue used OpenAI's GPT-5 mini to read 125,830 corporate earnings-call transcripts from 2006 to 2025 and score how optimistic managers were about the US economy on a seven-point scale. The resulting AI Economy Score predicted real GDP growth, industrial production, employment and wages beyond standard predictors.

Read the study abstract

Using generative AI, we extract managers’ economic expectations from over 120,000 corporate conference call transcripts. The resulting AI Economy Score predicts GDP growth, production, employment, and wages five or more quarters ahead, beyond survey forecasts and other existing measures. Industry- and firm-level scores capture sector- and firm-specific activity, and a composite of firm, industry, and macroeconomic expectations further improves forecasting power and horizon. Predictability concentrates among hist…

Generative AI, Managerial Expectations, and Economic Activity
"Frankfurt Stock Exchange (Ank Kumar) 03" by Ank Kumar, CC BY-SA 4.0.

The gain

In a working paper analysing 125,830 earnings calls, an AI-derived score of manager optimism raised the share of next-quarter GDP growth explained from 43.8% to 48.6% and kept predicting GDP for five quarters.

In brief

Adding the score to a model that already included the GZ credit spread raised the share of next-quarter GDP growth explained (R2) from 0.438 to 0.486. For employment, the score alone gave an R2 of 31.1%. It predicted GDP for five quarters, employment for up to five and wages for roughly ten.

The paper says the forecasting power sits mainly with managers whose past forecasts were accurate and unbiased, and with firms that are less diversified. That pattern fits managers holding real information rather than noise. Industry-level and firm-level versions of the score also forecast sector GDP and firm sales, employment and wages, for up to four years in some cases.

Main points

  1. The data covers 125,830 earnings calls from 5,513 companies, 2006 to 2025, scored in two passes: first extracting statements on the US economy, then rating optimism from 1 to 7.
  2. A one-standard-deviation rise in the score is associated with 2.67%, 1.23% and 0.21% higher next-quarter growth in industrial production, employment and wages.
  3. In a head-to-head test against the Survey of Professional Forecasters, the score kept its predictive power and was comparable in strength to the survey forecast, and both stayed significant together.
  4. Rerunning the analysis on anonymised transcripts with firm and date information removed still produced a significant predictor, addressing the risk that the model had seen the outcomes in training.
  5. An out-of-sample test using only calls from 2024Q2 to 2025Q4, after GPT-5 mini's training cutoff, still showed the firm-level score predicting firm sales and the industry score predicting industry GDP.
  6. Replicating the method with Google's open-source Gemma model gave similar results, so the finding does not depend on one AI model.
  7. The average score over 80 quarters was 4.252, just above the neutral midpoint of 4. The score did not anticipate the early-2020 pandemic collapse but did track the recovery.

The full story

The authors built the score in two steps because earnings calls run about 7,500 words and mostly concern the firm. GPT-5 mini first pulled out sentences where management spoke about US economic conditions, then a second request scored the whole call's economic content. Manual review of about one thousand calls found the outputs similar to human assessments, according to the authors.

On timing, the score turned down ahead of the 2008 financial crisis, before the trough in realized GDP. It fell alongside, not ahead of, output in the first quarter of 2020. A vector autoregression found that a positive shock to the score predicted higher consumption, investment and output for 8 or more quarters, followed by tighter monetary policy.

At the industry level, a one-standard-deviation rise in the adjusted industry score corresponded to 1.16 percentage points of annual employment growth and 1.65 points of annual wage growth. At the firm level, a one-standard-deviation rise in the firm component corresponded to 3.41 points of annual employment growth and 2.82 points of annual wage growth.

The authors say the score is built from public transcripts and gives researchers and policymakers a cheaper alternative to costly manager surveys. Earlier versions of the paper used GPT-3.5 Turbo and obtained comparable results.

Sources

1 on file
  1. NBER Working Paperdoi:10.3386/w35949
    Peer-reviewed2026-10-01

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

AI reading of 125,830 earnings calls predicts GDP and jobs quarters ahead, paper findshttps://truvace.com/item/generative-ai-managerial-expectations-and-economic-activity

  1. Who gets this benefit, and who was left out of the study or story?

    The gain In a working paper analysing 125,830 earnings calls, an AI-derived score of manager optimism raised the share of next-quarter GDP growth explained from 43.8% to 48.6% and kept predicting GDP for five quarters.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on an observational study with a sample of 125,830, from NBER Working Paper. It is a preprint, not yet peer reviewed.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The authors say that telling the model the call date does not eliminate look-ahead bias, since its training data extend past that point, and the score failed to anticipate the early-2020 pandemic collapse.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Business