Large language models can turn earnings call transcripts into structured research: changes in guidance, explanations for margin pressure, unanswered analyst questions and comparisons with previous quarters. Their most useful role is not to declare a stock a buy or sell. It is to make management’s statements easier to inspect and challenge.
A useful earnings call analysis should answer three questions: what changed, what supports that interpretation, and what still needs checking? A polished summary without traceable evidence does not meet that standard, however confident it sounds.
What LLMs Can Extract From Earnings Calls
Start with bounded tasks rather than an open request to “analyze the company.” Ask the model to identify revenue drivers, capital spending plans, hiring intentions, pricing changes and revisions to management’s outlook. Keep reported results separate from forecasts and conditional statements.
There is research support for extracting managerial expectations from transcripts. The NBER working paper on ChatGPT and corporate policies constructs an investment score from conference calls and finds that it predicts subsequent capital expenditure after controlling for other investment determinants. That finding concerns corporate spending expectations; it does not establish a ready-made stock trading strategy.
Separate Prepared Remarks From Analyst Questions
Review management’s presentation and the question-and-answer session separately. For the presentation, extract claims and supporting figures. For analyst questions, identify the issue raised, the executive’s response and whether that response addresses the question.
Consider a hypothetical retailer reporting stronger revenue while declining to quantify promotional spending. A useful output would flag the missing margin detail. It should not jump from “not quantified” to “management is hiding deteriorating profitability.” The former is an observation; the latter requires evidence.
Build a Repeatable Analysis Workflow
Begin with the correct company, fiscal quarter and call date. Preserve speaker labels, transcript locations and document versions. Check doubtful passages against the recording where available, especially numbers, negatives and statements about guidance.
Supply the earnings release and previous guidance alongside the transcript. Broader reconciliation belongs in a workflow for combining news, price data and company filings with AI; the call analysis itself should remain focused on management’s statements and changes.
Use a stable output structure so that reviewers can compare quarters without translating a new format each time.
| Field | Required content |
|---|---|
| Claim | A reported result, forecast or management explanation |
| Evidence | Supporting passage, speaker and transcript location |
| Change | Difference from the supplied previous disclosure |
| Review status | Verified, ambiguous, conflicting or not disclosed |
For longer transcripts, test a section-by-section approach, followed by a reconciliation pass across the extracted findings. Do not assume that fitting an entire document into a model’s input means every passage receives equal attention. The 2024 Lost in the Middle study found that the models tested often performed worse when relevant information appeared in the middle of long inputs. Treat this as a reason to test retrieval coverage, not a performance estimate for every newer model.
Use a Prompt That Requires Evidence
Using only the supplied documents, identify changes in revenue expectations, margins, capital expenditure and demand. For each finding, return a supporting quotation, speaker, document location, relevant period and comparison with prior guidance. Separate management statements from your interpretation. Mark missing evidence as “not disclosed.” Do not infer figures or issue a trade recommendation.
This prompt creates a reviewable assignment. Add company metrics where needed, such as bookings, occupancy or subscriber retention. Keep their definitions consistent across quarters, and flag any change in how management calculates them.
Check Numbers Outside the Language Model
Use the model to locate figures and identify the intended calculation, then verify the arithmetic with a spreadsheet or another deterministic calculation tool. Preserve currency, units, reporting period and whether a measure uses GAAP or an adjusted basis.
Financial numerical reasoning deserves separate evaluation. The FinQA financial question-answering benchmark was designed around questions requiring reasoning across financial text and tables. Its 2021 results are not a scorecard for current models, but the task provides a useful testing pattern: verify both the answer and the calculation that produced it.
Suppose hypothetical operating margin rises from 12% to 14%. That is an increase of two percentage points, or 200 basis points, rather than a 2% relative increase. Also check whether guidance refers to the next quarter or full year. Correct arithmetic applied to the wrong period still gives the wrong answer.
Treat Sentiment as an Interpretation, Not a Fact
Prefer topic-level assessments to a single positive or negative score. A company can sound confident about demand while expressing concern about collections. Require supporting passages for each assessment and allow mixed or unclear results.
Do not infer vocal stress from a text transcript. Assess observable features instead: whether management quantifies an answer, repeats earlier wording, changes its forecast or leaves a question unresolved. Avoid presenting these features as proof of deception.
Fabricated quotations and plausible but unsupported explanations remain review risks. NIST’s generative AI risk profile identifies confidently false output, information integrity and privacy risks. For earnings analysis, require quotation matching, document checks and approved handling of confidential research notes.
Evaluate the Research Output Before Using It
Build a manually checked sample of calls containing guidance changes, ambiguous answers and conflicting disclosures. Measure extraction accuracy, missed findings, quotation accuracy and numerical errors separately. A concise summary can still omit the quarter’s most consequential statement.
Keep the model version, prompt and input documents with each result. When changing the workflow, rerun the same evaluation sample before comparing outputs. If historical findings feed a strategy test, address timestamp integrity and information leakage through backtesting controls for AI trading systems.
The final deliverable should be a research brief with evidence, unresolved questions and checked calculations. Let the model organize the call; keep the investment judgment accountable to the underlying documents.