Can managers’ earnings-call remarks help forecast the economy?

An earnings call is mostly about one company. But between financial results and analysts’ questions, managers also discuss consumers, inflation, borrowing costs and demand. Can those remarks collectively reveal where the economy is heading?

Manish Jha, Jialin Qian, Michael Weber and Baozhong Yang investigate this in Generative AI, Managerial Expectations, and Economic Activity. The supplied November 2026 version is an arXiv preprint; its record does not establish peer review. This is an empirical study combining computational text analysis with historical economic data, not an experiment demonstrating that expectations cause growth.

Turning conversations into a score

The researchers analyze 125,830 earnings-call transcripts from 5,513 publicly listed companies over 2006–2025. Their main tool, GPT-5 mini, processes each call in two stages.

First, it extracts management statements relevant to US economic conditions, excluding material such as company-specific results and routine disclaimers. Second, it reads the extracted statements together and assesses management’s outlook for the coming quarter. Scores range from 1, severe expected deterioration, to 7, strong expected expansion, with 4 representing mixed or neutral signals. Calls without usable economic content are excluded rather than treated as neutral.

The national AI Economy Score is the equally weighted average across firms with valid scores each quarter. The AI therefore interprets managers’ expressed views; it is not simply asked to invent an economic forecast. Nevertheless, scoring can involve inference from indirect comments about customers, costs or hiring.

What the historical results show

The researchers test whether the score adds information beyond interest-rate measures, corporate credit spreads and past economic growth. For real GDP, adding it raises the model’s R-squared from 0.438 to 0.486. In ordinary language, the model explains 48.6% rather than 43.8% of the variation in its historical growth target.

That is an improvement in historical fit, not a demonstrated reduction in live forecasting errors. There is also an important timing detail: the baseline target compares GDP in the following quarter with GDP in the preceding quarter. It is not simply growth from the current quarter to the next.

In the baseline regressions, the score remains statistically informative about GDP through five quarters ahead. Results also link it to subsequent production, employment and wages, although strength varies by outcome and specification. For example, its next-quarter employment association is not statistically significant when the credit-spread measure is included alongside it.

Compared with the Survey of Professional Forecasters, the score delivers similar historical fit. Including both produces a modest further improvement, with both measures significant only at the weaker 10% threshold. The evidence supports a possible complement to surveys, not a clear replacement.

Different levels, different signals

The study separates national expectations from sector-specific and firm-specific differences. The industry component measures how a sector’s outlook differs from the national average; the firm component measures how one company’s outlook differs from its sector’s average.

Crucially, that firm component still concerns the manager’s view of the US economy, not a direct prediction of the company’s own sales. Nevertheless, it predicts subsequent company outcomes. Some industry and firm associations persist at horizons reaching four years, although firm-level wage results do not persist that far.

A separate composite combines 14 expectation measures, including revenue, employment, investment and prices. Its weights are estimated using earlier firm observations. This extends the reported GDP association through eight quarters, with the final horizon significant at the 10% level. The trade-off matters: the composite fits less well than the simpler economy score at shorter horizons.

Useful information—or hindsight?

Forecasts from historically accurate managers retain predictive information, while those from historically inaccurate managers contribute little additional signal. This is consistent with informed observations rather than indiscriminate optimism. It does not establish the underlying causal mechanism.

A major concern is AI hindsight: a model reading an old transcript may already know what happened afterward. Removing identifying information leaves a predictive association, but anonymization cannot prove that all historical clues disappeared. The researchers also use a smaller model to test whether identities remain recoverable, limiting what failed identification establishes.

A stronger check examines calls from 2024’s second quarter through 2025’s fourth quarter, which the authors identify as beyond GPT-5 mini’s training cutoff. Firm and industry components retain predictive associations in their respective outcome tests. The national component does not, and the period is too short for a reliable national forecasting regression.

The central promise is therefore narrower than an AI economic crystal ball: corporate conversations contain measurable forward-looking information. Whether that information reliably improves national forecasts in real time remains less firmly established than the historical results.