Skip to content

Knowledge base

How to anticipate analyst questions before an earnings call

By Artificial Societies
Published
Updated

To anticipate analyst questions before an earnings call, model each analyst who covers the stock from their public record and ask what they most want management to answer. Artificial Societies, which models audiences as networks of AI personas, simulated 14 sell-side analysts for a leading hyperscaler, and they pressed on demand durability and financing commitments rather than the margin reset the company expected, according to its case study. On the live call, the questions ran the same way. Your team gets each analyst’s likely question, in that analyst’s own words, before the call.

Artificial Societies has also tested the method against public transcripts, preregistering its simulated analysts’ questions before the NVIDIA, Marvell and Credo calls in 2026. For Marvell and Credo, detailed personas came closer than basic profiles (name, title, employer and coverage only) to what analysts actually asked in substance and phrasing, according to the September 2026 study. Analysts’ own past questions remained a strong baseline for wording and identity, and the briefing given to the personas strongly steered what came back.

How did a hyperscaler use Artificial Societies to prepare its earnings call?

Artificial Societies supported the investor relations team of a leading hyperscaler before a quarterly earnings call, according to its hyperscaler earnings-call case study. Revenue was about to double year on year while rising memory costs pulled the margin guide down. Regulation around selective disclosure meant the team could not test its prepared remarks, its guidance range or its view of what the market would fixate on with the small, regulated audience that mattered.

Artificial Societies built four societies around the company’s stock, matching each persona one-to-one to a real individual from public data and social activity:

SocietyWho it modelsPersonas
Sell-Side AnalystsEvery analyst who had asked a question on the company’s last eight earnings calls14
Buy-Side InvestorsPortfolio managers, analysts and chief investment officers at the institutions that hold the stock250
Corporate EcosystemCustomers, suppliers, partners and policymakers around the company, including a subgroup of its own employees570
Financial MediaBusiness and finance journalists at major outlets and specialist trade titles1,600+

Source: Artificial Societies, hyperscaler earnings-call case study, engagement undated.

The work ran in three stages: each analyst’s most likely question, investor reaction to the guide and to versions of the prepared remarks, then how the story would travel through the ecosystem and the press. The societies returned nearly 125,000 responses within 48 hours, start to finish, the case study records.

The company expected the margin reset to dominate the call. The simulated analysts largely looked past it and pressed instead on the durability of demand and the scale of the company’s financing commitments. On the live call, the questions ran the same way, although the case study reports that pattern rather than a count of matched questions. Script testing showed which version of the remarks landed best with investors, and the simulated press coverage matched the real next-morning framing: growth in the headline, the margin reset as a cost story, financing as the emerging risk.

Because every persona was simulated, the hyperscaler’s team could test unreleased numbers and draft scripts that could never be shown to a real investor, and no draft, number or script left Artificial Societies’ private, secured environment. Our page on investor communications testing covers script and guidance testing and summarises the US rule on selective disclosure.

What did Artificial Societies’ preregistered earnings call study find?

Artificial Societies ran three preregistered simulations and compared each with the call transcript, as Dr Linda Wei and Edoardo Chidichimo report in Testing Simulations Against Reality (September 2026). NVIDIA was descriptive; Marvell and Credo were controlled comparisons with hypotheses fixed in advance. Each set a detailed persona against a basic profile holding only the analyst’s name, title, employer and coverage, with the same model, instructions and briefing, so the comparison isolated what the extra detail contributed.

The samples were small, the authors note. Marvell had 21 analysts on the roster, of whom 9 could be evaluated, and 525 generated questions; Credo had 18 on the roster, 13 of whom asked questions on the call, and 840 generated questions. The post highlights three findings:

  1. 1.

    Detailed personas came closer to the real questions. For Marvell and Credo, their questions resembled the real ones more closely in substance and phrasing than the basic profiles did. For Credo, they did not improve the prediction of broad topics; the gain was in what analysts asked about those topics and how they put it.

  2. 2.

    Detailed personas produced questions distinctive to the analyst. With a facts-only Credo briefing, and with names, firms and introductions removed, an evaluator picked the right analyst from a line-up of five more often with detailed personas than with basic profiles, and more often than the 20% expected by chance.

  3. 3.

    The briefing strongly influenced what came back. For Marvell, almost every briefed response focused on gross margins, a controversy the briefing had named.

The post’s Credo example shows what a close match looks like. Before Credo’s first-quarter fiscal 2027 call on 1 September 2026, one analyst’s simulated question and his real one both opened with congratulations on the record quarter and asked the same executive by name to go a level deeper. Both hung the question on the $600 million optical target management had set and asked about the mix of components inside it. The match was not exact: the simulated question asked about gross margins and the quarterly ramp, while the real analyst picked up a topic management had raised on the call and extended the question into fiscal 2028.

Where does simulation fit in earnings call Q&A preparation?

BNY’s practice note on the earnings conference call (opens in a new tab) (2025) recommends a preparation timeline. Key themes are set four weeks before the call, the press release, script and Q&A are drafted at two weeks and finalised at one, and the team rehearses with a mock Q&A the day before. The note advises teams to “run a simulated Q&A exercise” that asks the difficult questions, and to re-review the key questions from competitors’ calls in the final days. David Calusdian of Sharon Merrill Advisors recommends Q&A messaging organised around major topic areas, in bullets rather than prose, and rehearsing the hard questions out loud (July 2026 (opens in a new tab)).

A topic-based Q&A book tells management what will come up, and says less about who will ask and what answer each analyst wants. The harder work, Artificial Societies’ earnings calls post argues, is anticipating which part of the story needs explaining, which assumptions might be challenged, and where two analysts may want very different answers. The topic list can also be wrong: in Artificial Societies’ hyperscaler case study, the company expected its margin reset to dominate a call that turned on demand and financing. A mock Q&A run by colleagues draws its questions from the same room that wrote the list. We would add simulation to the book in four steps:

  1. 1.

    Build a persona for every analyst likely to ask a question; in the hyperscaler case study, that was everyone who had asked one on the company’s last eight calls.

  2. 2.

    Ask each persona for the question it most wants answered, with several candidates per analyst; the Marvell and Credo studies used five.

  3. 3.

    File the candidates under each topic in the Q&A book, tagged with the analyst, the angle and any number the question anchors on.

  4. 4.

    Rehearse the executive against the analyst’s version of the question, not only the topic heading.

How should you brief simulated analysts without steering them?

Brief simulated analysts with the company’s facts and leave out the debates you expect, because in Artificial Societies’ Marvell simulation a briefing that named gross margins pulled almost every briefed response towards them. For Credo, Artificial Societies ran every briefed arm, or version, of the study twice: once with a briefing that named the live debates and once with the same facts only. Removing named debates cut the share of the personas’ primary questions about gross margins from 97% to 20%, but did not produce an improvement in topic overlap that met the preregistered threshold, Artificial Societies reports (September 2026). The authors add that this sensitivity to framing is not unlike human behaviour.

The facts-only Credo briefing is also where the study’s identity result appeared: an evaluator matched real questions to the right analyst more often from detailed personas than from basic profiles. From that design we would recommend four practices to a preparation team; they are our reading of the study, not results it reports:

  • Brief the personas with the results, the guidance and the public record, and keep your list of expected issues out of the main briefing.

  • Run a named-debates version beside it as a diagnostic, and treat the difference between the two as a measure of how much the briefing is doing.

  • Keep an unbriefed run, which the study included as a separate arm, as a check on what the personas raise unprompted.

  • If the simulated analysts agree with the topic you named, count that as your assumption echoed back, not as confirmation.

How do you check the predictions after the call?

Check predictions against the transcript only if you wrote them down before the call, together with the rules for scoring them. The Center for Open Science defines preregistration as specifying a research plan in advance and submitting it to a registry (Center for Open Science (opens in a new tab)). Artificial Societies registered its predicted questions and evaluation plans on the Center’s Open Science Framework (OSF), whose registrations create “time stamped, read-only versions of a project” (OSF (opens in a new tab)). Without that record, it is easy to pick an impressive match from a large set of simulated questions after the call, which says little about how the method performed overall, Artificial Societies’ earnings calls post notes.

Artificial Societies’ study scored each analyst’s five candidates against their real questions on four measures, each answering a different question a preparation team would ask:

MeasureThe question it answersHow the study scored it
TopicDid we predict what they asked about?Overlap between the topics a blind coder, who did not know which version produced each question, assigned to the candidates and to the real questions (Jaccard overlap: shared topics divided by all topics used)
MeaningDid we predict the substance?Similarity between text embeddings, numerical representations of meaning, of each candidate and its closest real question
Wording and voiceDid we predict how they asked it?How much the candidates helped a separate small language model predict the real first question, token by token
IdentityCould you tell which analyst asked?A five-way line-up, names and firms removed, judged by a model from a different family; chance is 20%

Source: Dr Linda Wei and Edoardo Chidichimo, Artificial Societies, Testing Simulations Against Reality, September 2026.

A team without that apparatus can still run the version the study used for NVIDIA: mark each predicted question as a close match, a partial overlap or no match once the transcript is out. BNY’s note already suggests evaluating the quality and depth of the questions posed after each call. Scoring them against a list written two weeks earlier turns that review into a test of the preparation.

What can simulated analysts not tell you?

Simulated analysts can miss what an analyst learns during the call itself. On Credo’s 1 September 2026 call, one analyst built his real question on a topic management raised during the call, which no simulation run beforehand could hear, Artificial Societies’ earnings calls post shows. The September 2026 post reports results only for the questions asked in the room. It does not establish how accurately a simulation measures second- and third-order effects, or how information spreads afterwards, and its 9 and 13 evaluable analysts are what the authors call a small number.

Simpler inputs remained competitive in places. In Artificial Societies’ study, analysts’ past questions remained a strong baseline for wording and identity, and for Credo the detailed personas did not improve the prediction of broad topics. Adding analysts’ published research notes did not establish an additional benefit in the small group for which notes were available.

How do you start with Artificial Societies before your next call?

Artificial Societies’ one published earnings-call engagement, for a leading hyperscaler, ran as follows. The investor relations team brought what it already had: prepared remarks, a guidance range and a view of what the market would fixate on. Artificial Societies built societies of analysts, investors, the corporate ecosystem and the press around the stock, asked for each analyst’s likely question, tested the guide and the script versions, and traced how the story would travel. A pre-call briefing arrived within 48 hours, with full access to Radiant, our platform, to re-run scenarios and re-test the script as the numbers firmed up. The societies are standing rather than one-off, so they were ready for the next quarter. We do not publish pricing; it is usage based and scoped with our team. To plan a simulation for your next earnings call, book a demo.

Frequently asked questions

What tools help identify analyst questions for an earnings call?

BNY’s 2025 practice note describes AI tools that review past earnings call information and current market dynamics to “generate a list of anticipated questions from stakeholders”. Artificial Societies simulates each analyst instead, from that person’s public record. For a leading hyperscaler, its 14 simulated sell-side analysts pressed on demand and financing rather than the margin reset the company expected, and the live call’s questions ran the same way, according to its case study.

How do companies respond to analyst questions during earnings calls?

Well-prepared management teams answer from agreed key messages per topic rather than memorised answers, David Calusdian of Sharon Merrill Advisors advises (July 2026). They agree who is best placed to answer which topics, practise handoffs between the CEO and CFO, and rehearse bridging from a narrow or negative question to a more strategic answer. Analyst simulation adds each analyst’s own version of the question to rehearse against.

Can we test unreleased numbers with simulated analysts?

Yes. In Artificial Societies’ hyperscaler engagement, unreleased numbers and draft scripts were tested on simulated audiences, and none of them left Artificial Societies’ private, secured environment. Whether a test fits your disclosure obligations is a question for counsel. BNY’s 2025 practice note advises companies to involve their technology and legal teams when adopting AI tools, because earnings content includes material non-public information that should be shared on a need-to-know basis.

How far ahead of the earnings call should we run the simulation?

Run it where the Q&A book is drafted, which BNY’s 2025 practice note places two weeks before the call. Artificial Societies’ hyperscaler engagement ran 48 hours from start to finish, according to the case study. A first run can therefore feed the draft book, and a second, on the same standing societies, can follow the final numbers before the rehearsal.

How many analysts can a simulation include?

Include every analyst likely to ask a question, and give investors, the corporate ecosystem and the press societies of their own. The rosters in Artificial Societies’ September 2026 study ran to 21 analysts for Marvell and 18 for Credo, and the hyperscaler’s largest society, Financial Media, held more than 1,600 personas. Artificial Societies builds societies of 12 to 3,500 personas, as its method and evaluation page states.

Does analyst simulation work for a first earnings call after an IPO?

Not on published evidence: Artificial Societies’ earnings-call work so far does not include a first call. The method needs a public record per analyst, and the Credo personas drew on analysts’ previous questions and other public material. A newly listed company also has no call history from which to draw the list of analysts, which for the hyperscaler meant everyone who had asked a question on its last eight calls. A first-call simulation would therefore be untested.

What data are simulated analysts built from?

Public data only, in the hyperscaler case study and the September 2026 study. The hyperscaler’s personas were each matched one-to-one to a real individual using public data and social activity. The Credo analysts were built from public records only, including their previous questions. The study also tried adding analysts’ published research notes, which did not establish an additional benefit in the small group for which notes were available.

Is a detailed persona better than asking a language model for likely questions?

Better than a basic profile, on published evidence, but not yet shown against every way of prompting a model. In Artificial Societies’ September 2026 study, the comparison held the model, the instructions and the briefing constant and changed only what the persona knew. The study also ran a one-line prompt and a web-search persona, and the post reports no separate result for either.

How accurate are AI-predicted analyst questions?

No single accuracy rate exists for this task yet: Artificial Societies’ September 2026 study reports the direction of its results on 9 Marvell and 13 Credo analysts, not match rates. That is a different task from the Survey Evaluation Report (January 2026), which records 86% distribution accuracy across 1,000 surveys against a 91% human self-replication ceiling. Our page on AI persona accuracy explains the survey measures.

How much does earnings call simulation cost?

There is no published price. Artificial Societies’ pricing is usage based and scoped with our team for each engagement, with no rate card. To get a figure for your own earnings call, talk to our team; in the same conversation we can take you through the earnings-call study’s time-stamped predictions and results, as its authors offer to.

Sources