💬 NLP

Fast Numbers, Slow Language: Bridging Quantitative and Qualitative Earnings Signals

This paper introduces the "EarningsInOne" corpus to bridge the quantitative and qualitative earnings analysis communities, revealing that while numeric surprises drive immediate market reactions, qualitative conference call sentiment offers a distinct, tradeable signal that peaks on the following trading day but has been previously obscured by incompatible evaluation metrics.

Original authors: Ding Yu, Zhuo Liu, Hao Zhang, Hangfeng He

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ding Yu, Zhuo Liu, Hao Zhang, Hangfeng He

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Two Different Speeds of News

Imagine a company releases its quarterly report like a magician pulling a rabbit out of a hat. This event actually happens in two distinct stages, and the paper argues that the financial world has been trying to watch both stages with the same pair of glasses, which doesn't work.

The "Fast Numbers" (The Rabbit): First, the company releases a press release with hard math: "We made $1.52 per share instead of the expected $1.44." This happens instantly. Computer algorithms (robots) read this math and trade stocks within minutes. By the time the stock market opens the next morning, the price has already adjusted to this number. The "alpha" (extra profit) from this number is gone in a flash.
The "Slow Language" (The Magic Trick): 30 to 90 minutes later, the CEO gets on a conference call to explain the numbers. They say things like, "We are cautiously optimistic, but the economy is tricky." This is qualitative language. Humans need time to listen, interpret the tone, and decide if the CEO is being honest or hiding something. This signal doesn't peak until the next trading day.

The Problem: Two Communities, Two Different Playbooks

For decades, two groups of experts studied this event, but they never talked to each other because they used completely different rulebooks:

The Financial Economists (The Math Guys): They studied the "Fast Numbers" for 50 years. They look at how stock prices move over days or weeks. They use a strategy called "Long Top, Short Bottom" (buy the best news, sell the worst news) and measure success by how much money they make relative to risk.
The NLP Researchers (The Language Guys): For the last 10 years, they studied the "Slow Language" (the conference calls). But they treated it like a school exam. They tried to predict the exact stock price movement for every single company and measured success by how close their guess was to the actual price (using a metric called MSE). They didn't have a real trading strategy; they just wanted to see if their AI could guess the right number.

The Result: The Math Guys thought the Language Guys were useless because their "exam scores" (MSE) didn't show a clear profit. The Language Guys thought the Math Guys were missing the nuance of the CEO's tone. They were speaking different languages.

The Solution: Building a Bridge

The authors built a bridge called EARNINGSINONE. It's a massive new database that lines up the press release, the conference call, and the stock prices for 1,500 US companies.

They created a Unified Rulebook:

They applied the "Long Top, Short Bottom" trading strategy to both the numbers and the language.
They used the same "scorecard" (metrics like Sharpe Ratio and Information Coefficient) to judge both.

The Discovery: Fast Numbers, Slow Language

When they ran the tests using the same rules, a clear pattern emerged:

The Numbers are Fast: If you try to trade on the "Fast Numbers" (EPS surprise) at the next morning's market open, you make almost no money. The robots already ate that profit. However, if you trade intraday (within 5–60 minutes of the news breaking), you can still make a profit. The window is tiny, but it exists.
The Language is Slow: If you try to trade on the "Slow Language" (CEO tone) immediately, it doesn't work well. But if you wait until the next trading day, the language signal becomes very powerful. The market takes time to digest the CEO's tone, and by the next day, the price moves in the direction the language predicted.

The Analogy:
Think of the earnings announcement like a sports game.

The Numbers are the final score. As soon as the referee blows the whistle, the score is on the ticker tape. Everyone knows it instantly. Betting on the score after the game starts is useless.
The Language is the post-game interview. The coach might say, "We played well, but our star player is injured." The crowd needs time to process that. By the next morning, the team's stock (or reputation) has adjusted to reflect that injury news.

Why This Matters

For the Math Guys: They missed out on short-term profits because they waited too long to trade the numbers. They need to look at the data during the day, not just the next morning.
For the Language Guys: They were using the wrong test. They were trying to predict the exact price for every stock (like guessing the exact temperature). Instead, they should have been ranking the stocks (like guessing which team will win). When they switched to the "ranking" method, their language models turned out to be very profitable.

The Bottom Line

The paper proves that numbers and language tell two different stories at two different speeds.

Numbers are for the robots: Fast, immediate, and gone by morning.
Language is for the humans (and patient algorithms): Slow, takes time to digest, and becomes valuable the next day.

By fixing the way they measure success, the authors showed that the "language" part of earnings calls is actually a real, tradeable asset that was previously hidden because researchers were using the wrong tools to find it.

Technical Summary: Fast Numbers, Slow Language: Bridging Quantitative and Qualitative Earnings Signals

Problem Statement
Financial economists and Natural Language Processing (NLP) researchers have studied earnings announcements for decades using fundamentally incompatible frameworks, preventing direct comparison of their respective signals.

Financial Economics: Focuses on quantitative surprise (numeric EPS/revenue deviations from analyst estimates). Historically, this community evaluates signals at the "next market open" horizon using long/short trading strategies (top decile vs. bottom decile) and metrics like Information Coefficient (IC) and Sharpe ratio. Recent literature suggests that for non-microcap stocks, the exploitable alpha from quantitative surprise has compressed to the announcement moment itself, disappearing by the next open.
NLP: Focuses on qualitative signals from Earnings Conference Call Transcripts (ECTs), such as management tone, guidance, and Q&A credibility. These studies typically evaluate signals using pointwise metrics like Mean Squared Error (MSE) over volatility or sign-agnostic targets, often without a practical trading framework (e.g., treating all stocks as equally tradeable rather than ranking them).

This methodological divergence leaves two gaps: financial economists miss the short-horizon alpha of quantitative signals by waiting for the next open, while NLP researchers fail to demonstrate the real economic value of qualitative signals because their evaluation metrics (MSE) do not align with directional trading performance.

Methodology
The authors introduce EARNINGSINONE, a unified corpus and evaluation framework designed to bridge these communities.

Dataset (EARNINGSINONE):
- Covers the S&P 1500 universe (2022–2025), comprising 5,428 earnings events across 701 tickers.
- Aligns three data modalities for the same events: earnings news (with precise timestamps), ECT transcripts, and intraday/next-day price data.
- Uses LLMs (GPT-5-mini) to extract numeric actuals from news articles and generate sentiment scores from ECTs and news.
Unified Evaluation Framework:
- Trading Mechanism: Applies a canonical "long top-decile, short bottom-decile" strategy to both quantitative and qualitative signals, moving away from the NLP community's "trade-all" approach.
- Entry Anchors & Horizons:
  - Fast (Numbers): Enters at FirstActual (the first news article confirming EPS/revenue, typically 1–5 minutes post-release) and holds for intraday horizons (5–60 minutes).
  - Slow (Language): Enters at NextOpen (the following day's market open) and holds for next-day horizons (up to 390 minutes).
- Metrics: Evaluates performance using IC (Spearman rank correlation), Q5−Q1 (quintile spread), and Sharpe ratio, rather than MSE.
Signal Construction:
- Quantitative: EPS and revenue surprises derived from news (LLM-extracted) and vendor data.
- Qualitative: Scalar sentiment scores ([-1, 1]) derived from full ECT transcripts, with ablation studies on masking numbers or grounding with numeric headers.

Key Results
The unified framework reveals a "clean speed separation" between the two signal types:

Fast Numbers: Quantitative surprise peaks immediately at the announcement (FirstActual).
- At FirstActual, the strategy achieves an IC of 0.03 and a Sharpe ratio of 1.81.
- By NextOpen, this alpha is largely arbitraged away (IC drops to -0.02, Sharpe to -0.31), confirming that numeric information is processed instantly by algorithms.
- Including revenue surprises (Joint signals) and using high-quality vendor data improves performance, but the decay by next open remains consistent.
Slow Language: Qualitative ECT sentiment peaks at NextOpen.
- Sentiment(ECT) evaluated at NextOpen yields an IC of 0.11 and a Sharpe ratio of 2.28, outperforming all quantitative variants at that horizon.
- This confirms that qualitative signals require overnight human interpretation and are tradeable the following day.
- Ablation studies show that masking numbers in the ECT reduces performance (Sharpe 0.40), while grounding the sentiment with known numeric beats/misses improves tail concentration (Sharpe 2.65).
Orthogonality: The two signals are largely orthogonal (Spearman $\rho$ = 0.21). ECT sentiment captures management framing and tone beyond the headline beat/miss, while news sentiment correlates more strongly with the numeric outcome ( $\rho$ = 0.38).
Baseline Performance: Naïve baselines (Random L/S, Long-only, Ticker-Mean) and prior NLP baselines (Embedding-based MLP) significantly underperform the unified signal-based strategies.

Significance and Claims
The paper claims to bridge the gap between financial economics and NLP by demonstrating that:

Speed Separation is Real: Quantitative surprise is a "fast" signal exhausted within minutes, while qualitative sentiment is a "slow" signal that matures overnight.
Evaluation Alignment Matters: NLP signals possess real economic value but were previously obscured by misaligned metrics (MSE) and lack of a trading framework. When evaluated with financial-standard metrics (IC, Sharpe) and a long/short framework, ECT sentiment emerges as a robust, tradeable alpha source.
Complementarity: The two communities were studying complementary phenomena. Financial economists can recover short-horizon alpha by evaluating at announcement timestamps, while NLP researchers can validate the tradeability of language signals by adopting quantitative asset management evaluation tools.

The authors conclude that language signals may show better value when evaluated at their "natural speed," a finding applicable to other data-rich domains mixing structured and textual signals. The work is presented as a research contribution; the authors explicitly state that the trading signals are for research purposes and do not constitute investment advice.