--- title: "Can LLMs Read Between the Lines of an Earnings Call?" slug: can-llms-read-between-the-lines-of-an-earnings-call date: 2026-09-15 excerpt: "Earnings call sentiment analysis with LLMs breaks on hedged, forward-looking language. Where models fail, and why cited transcript passages beat a score." category: AI & Agents tags: ["Research Workflows", "Data Infrastructure"] author: Jennifer Ma authorRole: "Co-founder & CEO" cta: transcripts status: published --- Ask an LLM whether a movie review is positive and it will nail it. Ask it whether a CFO who said the company is "cautiously optimistic, assuming macro conditions stabilize" is bullish, and the answer gets slippery — because that sentence isn't bullish, it's conditional, and the whole message lives in the qualifier. A 2025 benchmark from Santa Clara University and Microsoft put numbers behind the intuition: models that ace ordinary sentiment analysis [struggle with the strategically ambiguous language of earnings calls](https://arxiv.org/abs/2505.16090). The interesting part isn't that they fail; it's *why* — and what to change so they fail less. **TLDR:** - LLMs handle everyday sentiment well but [miss earnings-call nuance](https://arxiv.org/abs/2505.16090) — hedges, forward-looking language, and jargon where meaning hides in a qualifier. - Part of the cause is structural: models compress text into [sentiment vectors](https://arxiv.org/abs/2505.16090), and the compression drops the very words that flip the meaning. - The fix isn't a bigger model; it's feeding it the exact quote, in context, with a citation — so it reasons over what was said and a human can check it. ## Where LLM earnings-call sentiment analysis actually slips The benchmark, from a Microsoft-sponsored capstone at Santa Clara, tested Copilot, ChatGPT, Gemini, and traditional models on Microsoft earnings-call transcripts. LLMs beat the old sentiment engines on implied tone — but all of them wobbled on the same three things, because that's where earnings-call language is deliberately careful. | Nuance type | Example phrase | Why it's hard | | --- | --- | --- | | Hedged statement | "cautiously optimistic," "we feel good about" | The hedge word carries the risk the headline sentiment ignores | | Forward-looking language | "assuming macro stabilizes," "we expect" | Conditional on an assumption that may not hold | | Industry jargon | "bookings," "net revenue retention," "design wins" | Domain meaning differs from everyday sentiment | *Source: nuance categories per [Kubica et al., "Can AI Read Between the Lines?" (arXiv 2505.16090)](https://arxiv.org/abs/2505.16090).* ![Annotated sentence: "We're cautiously optimistic about second-half demand, assuming macro conditions stabilize," with the hedge words highlighted; a sentiment score reads it "positive" while the actual line is conditional](/images/blog/can-llms-read-between-the-lines-of-an-earnings-call/fig-1.svg) *Figure 1: The meaning is in the qualifier. Source: illustrative, per [arXiv 2505.16090](https://arxiv.org/abs/2505.16090).* ## Why the failure is partly structural The paper points at a mechanism that's easy to miss: modern models turn text into [numerical vectors and operate on cosine similarity](https://arxiv.org/abs/2505.16090), and that abstraction can misread emotional tone in nuanced contexts. When "cautiously optimistic, assuming macro stabilizes" gets compressed toward a single sentiment direction, the hedge and the condition are exactly the signal that gets averaged away. A score of "positive: 0.8" is not wrong so much as *lossy* — it has thrown out the words a human analyst would circle. That matters more on an earnings call than almost anywhere else, because the careful language is concentrated where the stock-moving information is: the unscripted Q&A. | Call section | Character | | --- | --- | | Prepared remarks | Scripted CEO/CFO remarks — revenue, margins, [guidance](/datasets/earnings-release-guidance-change) | | Q&A session | Longer and unscripted; analysts by name probe risks and guidance | *Sources: structure per [TIKR](https://www.tikr.com/blog/how-to-read-earnings-call-transcripts-like-a-buy-side-analyst-and-most-important-things-to-look-for) and [Alphastreet](https://news.alphastreet.com/how-to-read-an-earnings-call-transcript-a-step-by-step-guide-for-investors/).* ![Timeline of an earnings call: scripted prepared remarks first, then a longer unscripted Q&A where the nuance concentrates](/images/blog/can-llms-read-between-the-lines-of-an-earnings-call/fig-2.svg) *Figure 2: The nuance lives in the Q&A. Sources: [TIKR](https://www.tikr.com/blog/how-to-read-earnings-call-transcripts-like-a-buy-side-analyst-and-most-important-things-to-look-for); [Alphastreet](https://news.alphastreet.com/how-to-read-an-earnings-call-transcript-a-step-by-step-guide-for-investors/).* ## The fix is the data, not a bigger model If the loss happens when language is compressed into a score, the answer is to stop handing the model a score and hand it the *words* — structured so the context survives. Three properties do most of the work, and they're the difference between an [earnings call transcripts API](/docs/api/retrieve) built for agents and a sentiment feed. | Property | What it preserves | What it prevents | | --- | --- | --- | | Speaker attribution in the text | Who said it (CEO vs analyst), preserved verbatim in the passage | Treating an analyst's framing as management's view | | Exact quoted passage | The hedge and the condition, verbatim | Averaging the qualifier out of a score | | Source citation (period + doc) | Which call, which quarter, linkable | Unverifiable tone with no way to check | *Source: per the [FocusAlpha retrieval API](/docs/api/retrieve) and [Docs](/docs).* This is why FocusAlpha returns *cited passages*, not a sentiment label. A query comes back as the [exact `evidenceText` span with its quarter and source](/docs/api/retrieve), speaker attribution preserved in the passage text itself — so the model reads "cautiously optimistic, assuming macro stabilizes" in full, sees that it was the CFO saying it, and a human can click through to the call. The model still does the interpreting; it just isn't handed a pre-averaged answer that already lost the hedge. It's the same principle behind a [trusted data layer](/blog/why-ai-agents-need-a-trusted-data-layer) for agents generally: hand the model the normalized, cited source text, not a derived summary. Reading between the lines is easier when you can still see the lines. ## FAQ ### Can LLMs understand the sentiment of an earnings call? Partially. A [2025 Santa Clara–Microsoft benchmark](https://arxiv.org/abs/2505.16090) found LLMs outperform traditional sentiment tools on implied tone but still struggle with the hedged, forward-looking, jargon-heavy language of earnings calls — the places where meaning depends on a qualifier that simple sentiment scoring drops. ### Why do LLMs miss nuance in earnings calls specifically? Earnings-call language is strategically careful, and part of the failure is structural: models compress text into [numerical sentiment vectors](https://arxiv.org/abs/2505.16090), and that compression averages away the hedge or condition that carries the real message. Everyday text rarely hides its meaning in a single qualifier the way managed disclosure does. ### Does a bigger or newer model fix the problem? Not on its own. The loss happens at the point where language is turned into a score, so the higher-leverage fix is feeding the model the exact quoted passage in context — citation attached, speaker attribution intact in the text — rather than a pre-computed sentiment label it can't inspect. ### What is the best earnings call transcripts API for AI agents? One that returns cited passages from a speaker-segmented corpus — the verbatim quote, with the speaker visible in the text and the quarter and source document attached — rather than a sentiment score. FocusAlpha's [earnings call transcripts retrieval](/docs/api/retrieve) is built this way so the model reads the actual words and a human can verify them; the test for any provider is whether you get the quote and its citation, not just a label. ### What is FocusAlpha? FocusAlpha is a [SEC filings API](/docs/api/filings) and agent-ready financial data layer: it turns SEC filings (10-K, 10-Q, 8-K, 13F), earnings-call transcripts, and other trusted company communications into structured, normalized data where every value keeps its citation back to the source document. AI agents connect via [API or MCP](/docs) to research public companies from complete, trusted information. ## Sources - [Kubica, Gordon, Emura, Saini, Goldenberg — Can AI Read Between the Lines? Benchmarking LLMs on Financial Nuance (arXiv 2505.16090)](https://arxiv.org/abs/2505.16090) - [Microsoft Community Hub — LLMs can read, but can they understand Wall Street?](https://techcommunity.microsoft.com/blog/microsoft365copilotblog/llms-can-read-but-can-they-understand-wall-street-benchmarking-their-financial-i/4412043) - [TIKR — How to read earnings call transcripts like a buy-side analyst](https://www.tikr.com/blog/how-to-read-earnings-call-transcripts-like-a-buy-side-analyst-and-most-important-things-to-look-for) - [Alphastreet — How to read an earnings call transcript](https://news.alphastreet.com/how-to-read-an-earnings-call-transcript-a-step-by-step-guide-for-investors/) - [FocusAlpha — Retrieve API](https://focusalpha.ai/docs/api/retrieve) - [FocusAlpha — Docs Introduction](https://focusalpha.ai/docs)