Advanced Prompting Guide for Investors: Analyzing Earnings Calls
The press release tells you what happened last quarter. The call tells you how management feels about the next one. Most of us read the first and skim the second, because a transcript runs for pages and the interesting bits are buried in the Q&A.
That's where AI earns its keep. The right ChatGPT prompts for stock analysis can pull out who sounded confident, who dodged a question, and which risks only surfaced because an analyst pushed. The wrong ones give you a cheerful summary that tells you nothing you couldn't get from the headline.
This guide gives you eight prompt templates you can copy today, plus a way to chain them into a one-page brief. They work in ChatGPT and in any other major chatbot. If you're still choosing a tool, start with our list of top AI stock analysis tools. And if you're wondering how far to trust the numbers a chatbot hands back, read our ChatGPT vs. Perplexity accuracy test first.
Why the Way Executives Talk Is Worth Measuring
This isn't just a hunch. Researchers have been studying earnings call language for years, and a few findings are worth knowing before you start prompting:
- Tone moves prices. A study built on the Loughran & McDonald financial word lists found that the tone of a transcript has real predictive power over how the stock reacts after the call (RAE study).
- Dodging shows up in the numbers. A 2026 paper in the Journal of Behavioral Finance used language-model embeddings to measure how closely executives' answers match the analyst's question. Executives who answered more directly were associated with higher returns, and humans rated low-similarity answers as evasive far more often (University of Limerick).
- Incoherent answers hurt. Researchers who studied 6,802 S&P 500 calls found that incoherent answers were linked to next-day stock moves and to next-quarter earnings (UT McCombs).
- Non-answers cluster in predictable places. A Frankfurt team found them more often after follow-up questions, negatively toned questions, and questions about the future (summary here).
One honest caveat. These are academic results, built with purpose-made models over thousands of calls. A prompt in a chat window won't reproduce them, and it isn't a trading system. What it can do is help you read one call more carefully than you would on your own, in a fraction of the time.
Before You Paste Anything: Four Ground Rules
Most bad AI output comes from the setup, not the model. Get these right and every template below works better.
- Use the full transcript. Not a summary, not a news article about the call. If the transcript is long, paste it in two or three chunks and say so ("This is part 1 of 3, wait for the rest").
- Fence the model in. Tell it to use only the text you provided, and give it permission to answer "not stated." Without that permission, a chatbot will happily fill gaps with plausible guesses.
- Ask for quotes. Every claim should come with the speaker's name and the words they used. You can't judge tone from a paraphrase, and quotes are easy to check.
- Keep facts and opinions apart. Make the model label what it observed and what it thinks that means. You want to disagree with the second part without doubting the first.
The Eight Prompt Templates
Swap the square-bracket placeholders for your own details. Keep the rules lines intact, since they do most of the work.
1. The two-part read: prepared remarks vs. Q&A
Prepared remarks are what management wants you to hear. The Q&A is what analysts wanted to know. Comparing them is the fastest way to find the gap.
You are helping me read an earnings call transcript.
Use ONLY the text I paste below. If something isn't in the text, write "not stated."
Company: [COMPANY] | Quarter: [e.g. Q3 2026]
Task:
1. Split the call into prepared remarks and Q&A.
2. For each part, give the three points management pushed hardest.
3. Tell me one topic that got airtime in the Q&A but almost none in the prepared remarks.
4. Give two direct quotes that sum up the overall message, with speaker names.
Format: two short sections, plain English.
Transcript:
[PASTE TRANSCRIPT]2. The tone and confidence audit
Skip the "is it bullish or bearish?" question. It invites a vague answer. Ask for the actual hedging words instead.
Use ONLY the transcript below. For each executive who speaks, find:
- Confident language (firm commitments, specific numbers, plain statements)
- Hedged language (e.g. "we hope," "we believe," "subject to," "depending on")
Output a table with these columns:
Speaker | Confident language (quote) | Hedged language (quote) | Overall read (one sentence)
Then answer: do the CEO and CFO sound aligned? Label each point either
"Observed" (backed by a quote) or "My read" (your interpretation).
Do not give a sentiment score.
Transcript:
[PASTE TRANSCRIPT]3. The evasion check
This one goes straight at the Q&A, which is where the research says the signal lives.
Go through every analyst question in the Q&A below. For each one:
1. Summarize the question in 12 words or fewer.
2. Classify the answer: Direct / Partial / Deflected / Not answered.
3. Quote the key sentence of the answer.
4. Give a one-line reason for your label.
Rules: if you're torn between two labels, choose the weaker one and say so.
Then list the three questions with the weakest answers and tell me
what follow-up question I'd want to ask each executive.
Transcript:
[PASTE Q&A SECTION]4. The risk extractor
Risk factors in a 10-K are written by lawyers. Risks on a call slip out in conversation, often as an aside. Catch the asides.
List every risk, headwind, or uncertainty mentioned in this call, including
offhand remarks. Use ONLY the transcript.
Table columns:
Risk | Who said it | Prepared remarks or Q&A | Quote | Quantified? (yes/no) | Severity (low/med/high) and why
Then split the list in two:
(a) risks management raised on their own
(b) risks that only came up because an analyst asked
Tell me which list is longer and what that might suggest.
Mark that last part as "My read."
Transcript:
[PASTE TRANSCRIPT]5. The quarter-over-quarter shift
One call is a snapshot. Two calls show direction. This is often where the best insight is, because small changes in wording say more than big claims.
Below are two transcripts from the same company.
A = previous quarter, B = current quarter. Use ONLY these texts.
Compare them and tell me:
1. Topics that appear in B but not in A (new)
2. Topics that appear in A but not in B (dropped)
3. Changes in how management talks about [demand / margins / pricing / headcount]
4. New qualifiers or softer verbs in B compared with A
For every point, quote both versions side by side.
Transcript A:
[PASTE]
Transcript B:
[PASTE]6. The guidance language audit
"We expect," "we aim to," and "we hope to" sound similar. They aren't. This template sorts them out.
Extract every forward-looking statement from this transcript.
Classify each as:
- Numeric guidance (a specific figure or range)
- Directional ("we expect growth")
- Aspirational ("we aim to," "we hope to")
Flag any condition attached to it (words like "assuming," "barring," "if").
Quote each one with the speaker's name.
At the end, tell me what share of the guidance is numeric versus vague.
Transcript:
[PASTE TRANSCRIPT]7. The chained brief (the one that ties it together)
Once you've run the individual templates, or if you're short on time, use this single prompt to run them in sequence and produce a one-page brief.
Work through this transcript in three stages. Use ONLY the text provided.
Quote speakers by name. Write "not stated" where the text is silent.
Stage A (extract): the five most important claims management made,
and every risk mentioned.
Stage B (test): for each analyst question, label the answer
Direct / Partial / Deflected / Not answered. Note any hedging language
from the CEO and the CFO.
Stage C (synthesize): write a one-page brief with these headings:
1. What management wants me to believe
2. What the Q&A suggests instead
3. Three things I should verify in the 10-Q or 8-K before acting
4. Two questions I'd ask on the next call
Keep each section under 120 words.
Transcript:
[PASTE TRANSCRIPT]8. The skeptic's pass
Language models tend to agree with whatever framing you've built up. Break that habit with a deliberate counter-read.
Now argue against the brief you just wrote. Using only evidence from the
transcript, make the strongest case that your reading was too generous
(or too harsh). Quote the lines a skeptic would point to.
End with one sentence on which reading the text supports better, and why.How to Run Them in Practice
A routine that works well: run template 1 for orientation, then 3 and 4 for the Q&A and risks, and finish with 7 and 8 for the brief and the counter-read. If you follow a company over several quarters, add template 5 every time a new call lands. It takes about as long as making coffee, and over a year it builds a record of how management language changes before the numbers do.
Keep everything inside one chat session so the model remembers the transcript, but start a fresh session for each company. Mixing two companies in the same conversation is a good way to get quotes attributed to the wrong CEO.
Mistakes That Quietly Ruin the Output
- Asking for a verdict. "Should I buy?" gets you a hedged paragraph. Ask for evidence, then decide yourself.
- Trusting the quote without checking. Copy any line you plan to rely on and search for it in the transcript. Models occasionally tidy up wording or merge two sentences.
- Pasting a truncated transcript. If the Q&A got cut off, the model won't tell you. Check that the last analyst name in your paste matches the end of the call.
- Skipping the numbers. Tone tells you how management feels. The 8-K and 10-Q tell you what happened. Verify figures on SEC EDGAR.
- Treating one call as a trend. Some executives are simply cautious talkers. Compare them with their own past calls, not with a more upbeat CEO.
A Quick Verification Checklist
- Search every quote you plan to use in the original transcript.
- Confirm any number against the press release or filing.
- Re-run the evasion check once with different wording. If the labels flip, treat them as weak signals.
- Write down your own view before you read the AI's, then compare.
Frequently Asked Questions
What are the best ChatGPT prompts for stock analysis?
The best ones set clear boundaries (use only this text), demand quotes, and ask for observations separately from interpretation. Open-ended prompts like "analyze this call" produce generic summaries.
Can ChatGPT tell me if an executive is lying?
No, and you shouldn't ask it to. It can point out hedging, deflection, and shifts in wording. Those are clues worth investigating, not proof of anything.
Where do I find earnings call transcripts?
Many companies post a webcast or transcript on their investor relations page. Data platforms and some finance sites also publish transcripts. Check that the one you use is complete before pasting it.
Does this work in tools other than ChatGPT?
Yes. The templates are plain text and work in any major chatbot. Results vary by model, so test the same prompt on the tool you actually use.
What if the transcript is too long?
Split it into parts, tell the model how many parts are coming, and send them in order. Run the Q&A through templates 3 and 4 separately if needed.
Disclaimer: This article is for general educational and informational purposes only and does not constitute personalized financial, investment, tax, or legal advice. AI tools can misquote or misread documents. Investing carries risk, including the possible loss of principal. Always perform independent research and consult a qualified professional before making financial decisions.
Comments
Post a Comment