Two earnings signals that finance AI often separates
Earnings releases first deliver numbers such as EPS and revenue surprise. Management tone, guidance, and question-and-answer credibility arrive 30 to 90 minutes later in the conference call. Financial economics has emphasized the first signal and NLP research the second, but incompatible targets and metrics made direct comparison difficult.
The authors introduce EarningsInOne, a corpus aligning earnings news, earnings-call transcripts, and intraday and next-day prices for S&P 1500 companies from 2022 through 2025.
Main pattern reported by the study
- Quantitative surprise peaked around the announcement and was largely absorbed by the next market open.
- Qualitative earnings-call sentiment peaked on the following trading day.
- The authors argue that earlier NLP evaluation, often optimized for sign-agnostic volatility error, may have hidden the directional value of language signals.
The central point is not that “AI predicts stocks.” It is a methodological claim that numbers and language arriving at different times should be compared under a common trading and evaluation framework.
Potential uses and limits
The corpus may support earnings-analysis agents, filing summarization, and event-driven financial NLP benchmarks. However, it is an arXiv preprint and has not completed peer review. Results from 2022–2025 U.S. large- and mid-cap companies may not generalize to other countries or market regimes. Trading costs, data latency, and realistic execution can also reduce backtested performance.
What readers should verify
Important follow-up checks include the released data scope, survivorship bias, timestamp alignment, transaction-cost assumptions, and independent reproduction. The findings do not guarantee returns for any security or strategy, and this article is not investment advice.